COPYRIGHT AND CITATION CONSIDERATIONS FOR THIS THESIS/ DISSERTATION

Transcription

1 COPYRIGHT AND CITATION CONSIDERATIONS FOR THIS THESIS/ DISSERTATION o Attribution You must give appropriate credit, provide a link to the license, and indicate if changes were made. You may do so in any reasonable manner, but not in any way that suggests the licensor endorses you or your use. o NonCommercial You may not use the material for commercial purposes. o ShareAlike If you remix, transform, or build upon the material, you must distribute your contributions under the same license as the original. How to cite this thesis Surname, Initial(s). (2012) Title of the thesis or dissertation. PhD. (Chemistry)/ M.Sc. (Physics)/ M.A. (Philosophy)/M.Com. (Finance) etc. [Unpublished]: University of Johannesburg. Retrieved from: (Accessed: Date).

2 MARKOV MODELLING AND BIT ERROR RATE ANALYSIS OF IN-VEHICLE POWER LINE COMMUNICATION by Mark D Wilson A dissertation submitted for the fulfilment of the requirements for the degree MASTER S in ELECTRICAL AND ELECTRONIC ENGINEERING SCIENCE at the UNIVERSITY OF JOHANNESBURG Study Leader: Mrs. Reolyn Heymann Co-Study Leader: Prof. Hendrik Ferreira December 2013

3 Abstract In-vehicle Power Line Communication (PLC) is an emerging technology that can easily benefit the automotive industry by reducing the amount of wires (and hence cost, weight and complexity) for vehicle wire harnesses. The reduction in weight would also lead to less fuel consumption. This dissertation aims at taking the research of this technology a step towards fully understanding the vehicle s power line as a communication medium. We investigate the bit error characteristics of a readily available transceiver on the vehicle s power bus. To do so, we develop and perform bit error recording over the medium to get experimental results with the battery line under different operating conditions. Using the first set of these results, we parametrise different kinds of Markov models to see which one simulates the channel best. Using the preferred model, we then model the rest of the sets of results so that we can simulate the channel s bit error characteristics under these different conditions. Using these models, we demonstrate how these simulations can be used to evaluate the performance of different error detection and correction techniques. In particular, we evaluate the error detection mechanisms used in the popular in-vehicle Local Interconnect Network (LIN) protocol, in addition to some simple error correction techniques. ii

4 Declaration I, Mark Wilson, hereby declare that all designs and solutions contained within this document are as a result of my own investigation. Some concepts may have been obtained from similar or relevant work, books, articles or consultations with other individuals and are appropriately referenced. Some standard design equations or methods have been used and have been suitably referenced. I also declare that all data used from other projects, books, and publications are properly referenced according to the IEEE standard referencing technique. The final design and solution are as a result of my own efforts and all relevant sources have been cited. M.D. WILSON DATE iii

5 Acknowledgements Reolyn Heymann Firstly, I would like to thank my supervisor, Reolyn Heymann, for her constant guidance throughout my Master s Degree. I greatly appreciate all the advice she has given me throughout these two years to make this venture less daunting than it could have been. Also, I am very grateful her for all the opportunities she has provided for me to learn about, take part in and experience within the academic world. In addition to my dissertation topic and research, I have learned a lot and met interesting people; which is mainly a result of Reolyn s efforts to include me in many different academic activities. Professor Hendrik Ferreira I would like to thank my co-supervisor, Professor Hendrik Ferreira. Prof. Ferreira has built up mountains of knowledge throughout his career and is a fantastic resource to tap into. His dedication to holding research group meetings every week to discuss everyone s work and dispense advice has really helped. His advice has been a valuable contribution to the shaping of this dissertation. Wendy Smith The telecommunications research group s secretary, Wendy, works extremely hard to make sure everything is in order with regards to anything administrative within the research group. She really goes out of her way to accommodate our administrative needs. Thanks, Wendy. iv

6 Telecommunications Research Group To everyone else in the telecommunications research group for our weekly discussions, and to those who work hard to organise the labs and lab equipment; thank you. Yair Maryanka Yair works at Yamar Electronics, who provided the in-vehicle power line communication transceivers. His support is greatly appreciated. He was never reluctant to answer my many questions. I trust Yamar Electronics is going to do well in the near future, when in-vehicle PLC becomes more demanding. Altech and Altech Netstar Altech sponsored my 3rd and 4th undergraduate years and were supportive enough to allow and sponsor me for my two year Master s degree as well. For this, I am grateful. They had taken a large financial strain off of me and my family and I will do my best to honour the bursary contract by working hard at their company. Under this sponsorship from Altech, I did a fair amount of vocational work at Altech Netstar. I had learned a lot while doing work there, and it has helped ease the transition from the academic world into industry. More importantly, the engineers there were always encouraging and found interest in my research, which boosted the confidence I had in my Master s work and motivated me to work harder. My Family and Close Friends I would like to acknowledge my family and close friends for their constant support and encouragement. It is much easier to work hard at something when everyone around you is encouraging and supportive of what you re working towards. My parents deserve a special mention for supporting my decision to continue my studies with this Master s degree. v

7 Contents Abstract Declaration Acknowledgements Table of Contents ii iii iv xi List of Figures xii List of Tables xvii Abbreviations and Acronyms xviii 1. Introduction Introduction Advantages of In-Vehicle PLC Dissertation Objective and Scope Document Overview Contributions Literature Study Introduction In-Vehicle Communication CAN Protocol vi

8 2.3.1 Message Format Coding LIN Protocol Message Structure Integration of LIN and CAN DC Power Line Communication Feasibility of In-Vehicle Power Line Communication Legal Considerations and Regulations In-Vehicle DC Power Line Communication Channel Characterisation Issues and Constraints Technical Requirements: Arduino Picoscope Markov Modelling Gilbert Markov Model Fritchman s Markov Model Error Control Coding Information Rate Modulo Arithmetic Parity Bits Checksum Conclusion vii

9 3. Bit Error Recording Introduction Experimental Platform Signal Waveform Error Vector and Gap Recording Sequence of Bits Matlab Send/Receive Arduino Sampling Picoscope Sampling Conclusion Experimental Results Introduction Experimental Procedure Presentation of Results Error Free Run (Gap) Distribution Graph Burst Length Distribution Burst-Interval Length Distribution Cluster Length Distribution Experimental Results No Devices On Cell Phone Charger Plugged In viii

10 4.4.3 Windscreen Wipers On Results Summary and Discussion Conclusion Markov Modelling Introduction Gilbert Markov Model Fritchman Markov Model Finite State Spreading Chain Error Spreading Chain Model Comparisons Expected BER Simulated BER Average Length of Error Bursts R-Squared Value Number of States (N) Block Error Probabilities Choosing the Best Model Error Spreading Chain Model Gilbert Model Fritchman and Finite State Spreading Chain Models The Best Model Models for the Channel Under Different Conditions ix

11 5.8.1 Cell Phone Charger Plugged In - Engine Off Windscreen Wipers On - Engine Off No Devices On - Engine On Cell Phone Charger Plugged In - Engine On Windscreen Wipers On - Engine On Modelling Summary Conclusion Channel Error Coding Introduction LIN Error Detection Single Bit Parity Check Two Bit Parity Check Checksum Error Correction Repetition Coding Rectangular Coding Hamming Coding Error Detection/Correction Comparison Conclusion Conclusion Thesis Summary Literature Study x

12 7.1.2 Bit Error Recording Experimental Results Markov Modelling Channel Error Coding Contributions Further Research References Appendices A. The Yamar SIG60 Integrated Circuit B. Bit Sequence Generator Algorithm C. Goodness of Fit D. Op-Amp Signal Booster xi

13 List of Figures 2.1 Relative Network Speed Versus Cost Graph An Example Network Structure of In-Vehicle Communication Impulsive Noise in a Vehicle DC Bus Amplitude Response of Vehicle Battery Power Line Binary Symmetric Channel Gilbert s Markov Model Fritchman s Transition Diagram Transmitted and Received Signals in Time Domain Transmitted and Received Signals in Frequency Domain Basic Setup for Matlab Send/Receive Method Matlab Send/Receive Results - Error Vector Pattern Plot Arduino Experimental Setup Arduino Setup - Timing Diagram The Long Stop Bit Anomaly Picoscope Experimental Setup Basic Flow Diagram of Picoscope Application xii

14 3.10 Final Results - Typical Error Vector Pattern Plot Final Results - Zoomed in Section of Error Vector Pattern Plot Sample Error Vector Illustrating a Burst, Burst-Interval and Cluster Error Distribution Statistics for the In-Vehicle Power Line Experiments - No Devices On - Engine Off Typical Error Vector Pattern for Transmission with No Devices on - Engine On Typical Error Vector Pattern for Transmission with No Devices on - Engine On - Boosted Transmission Error Distribution Statistics for the In-Vehicle Power Line Experiments - No Devices On - Engine On Typical Error Vector Pattern for Transmission with Cell Phone Charger Plugged In - Engine Off Typical Error Vector Pattern for Transmission with Cell Phone Charger Plugged In - Engine Off - Boosted Transmission Error Distribution Statistics for the In-Vehicle Power Line Experiments - Cell Phone Charger Plugged In - Engine Off Typical Error Vector Pattern for Transmission with Cell Phone Charger Plugged In - Engine On Typical Error Vector Pattern for Transmission with Cell Phone Charger Plugged In - Engine On - Boosted Transmission Error Distribution Statistics for the In-Vehicle Power Line Experiments - Cell Phone Charger Plugged In - Engine On Typical Error Vector Pattern for Transmission with Windscreen Wipers On - Engine Off xiii

15 4.13 Typical Error Vector Pattern for Transmission with Windscreen Wipers On - Engine Off - Boosted Transmission Error Distribution Statistics for the In-Vehicle Power Line Experiments - Windscreen Wipers On - Engine Off Typical Error Vector Pattern for Transmission with Windscreen Wipers On - Engine On Typical Error Vector Pattern for Transmission with Windscreen Wipers On - Engine On - Boosted Transmission Error Distribution Statistics for the In-Vehicle Power Line Experiments - Windscreen Wipers On - Engine On Typical Simulation run using Gilbert s Model Experimental Error Free Run Distribution with the Fitted Curve Transition Diagram for the Finite State Spreading Chain Model Experimental Data for P (0 m 1 1) and the Fitted Curve Expected BERs of the Different Size Transition Matrices The Mirrored Spreading Chain Transition Diagram Fitted Curve for the Error Run Probability Distribution Expected BER for Different Sized Transition Matrices Experimental Error Vector Pattern and Typical Simulation Run Error Vector Patterns Error Free Run Distribution Graph of the Different Models and the Experiment Error Run Distribution Graph of the Different Models and the Experiment 82 xiv

16 5.12 Distributions of the Different Models and the Experiment Block Error Probabilities for Blocks of Length Block Error Probabilities for Blocks of Length Block Error Probabilities for Blocks of Length Distribution Graphs for In-Vehicle Experiment (Solid Line) and Corresponding Simulation (Dashed Line) for Cell Phone Charger Plugged In with Engine Off Distribution Graphs for In-Vehicle Experiment (Solid Line) and Corresponding Simulation (Dashed Line) for Windscreen Wipers On with Engine Off Distribution Graphs for In-Vehicle Experiment (Solid Line) and Corresponding Simulation (Dashed Line) for No Devices On with Engine On Distribution Graphs for In-Vehicle Experiment (Solid Line) and Corresponding Simulation (Dashed Line) for Cell Phone Charger Plugged In with Engine On Distribution Graphs for In-Vehicle Experiment (Solid Line) and Corresponding Simulation (Dashed Line) for Windscreen Wipers On with Engine On Coding Scheme Overview Diagram Rectangular Code Illustration A.1 Function of the SIG60 in a Power Line Network A.2 SIG60 Signal on a Power Line A.3 SIG60 Signal on a Power Line Versus the Output of the Receiving Device 132 B.1 Bit Sequence Algorithm xv

17 D.1 Booster Circuit Schematic xvi

18 List of Tables 4.1 Summary of BERs for Experimental Results Comparison Table for the Four Different Models Summary of Models Based on In-Vehicle Experimental Results Bit Error Rate Decrease for 1-bit Parity Check Bit Error Rate Decrease for 2-bit Parity Check Bit Error Rate Decrease for Checksum with 2 Data Bytes Bit Error Rate Decrease for Checksum with 4 Data Bytes Bit Error Rate Decrease for Checksum with 8 Data Bytes Bit Error Rate Decrease for Repetition Coding on Top of Checksum Bit Error Rate Decrease for Rectangular Coding on Top of Checksum Bit Error Rate Decrease for Hamming Coding on Top of Checksum Bit Error Rate Decrease Performance for Bit Error Correction Techniques Applied on Top of Checksum Error Detection xvii

19 Abbreviations and Acronyms PLC CAN DC AC EMI UART ECU IC LIN PSD BER Power Line Communication Control Area Network Direct Current Alternating Current Electromagnetic Interference Universal Asynchronous Receiver/Transmitter Electronic Controller Unit Integrated Circuit Local Interconnect Network Power Spectral Density Bit Error Rate xviii

20 CHAPTER 1 Introduction All progress is precarious, and the solution of one problem brings us face to face with another problem. - Martin Luther King, Jr. 1.1 Introduction This dissertation is on in-vehicle DC power line communication; a topic that, if mastered, can have many benefits in various fields of applications (as described within the chapter). A brief description of power line communication is as follows: communicating data through the power line of a system as to integrate both data and power transmission over the same line. There are two main types of power lines: AC and DC. This dissertation deals with DC power lines and how to incorporate power line communication into DC power lines. The reason for this is so that the methods can be applied to vehicles (which normally have DC power). With increasing costs of copper [1], large wire harnesses used in vehicles are getting more and more expensive and they are also becoming very big and difficult to fit into the latest and more slimmer chassis designs. The weight of these wires have also become comparable to the chassis weight itself, as chassis designs become lighter, but with more electronics. This increase in wire harness weight would also increase fuel consumption. A major problem of vehicle manufacturing is that the complexity, weight and price of in-vehicle wiring is all too high and is also increasing as new vehicles are designed with even more electronics to ensure safer and more comfortable journeys. Producing cable harnessing is not an automated process due to the steps involved [2] and they are thus manually manufactured. Decreasing the size of the harness will make it easier and 1

21 Section 1.2: Advantages of In-Vehicle PLC Chapter 1 cheaper to build, as well as decrease the vehicle s weight. The root cause of the complexity of the wiring is all the different communication wires to and from the different devices in the vehicle. If all these wires can be removed, it would make the wiring in a vehicle a lot more simple. There are two feasible ways to do this: ˆ Use wireless communication throughout the vehicle, such as bluetooth, between all the devices, eliminating the need for communication wires. ˆ Use DC power line communication (PLC) so that all communication is done through the power line, also eliminating the need for dedicated communication wires. The major issue with using bluetooth or some other wireless protocol is that now every device has to be fitted with bluetooth or some other wireless module/capabilities, which could land up making the devices substantially more expensive. Another problem is that it is a lot easier to interfere with wireless than wired communication, making it unreliable for in-vehicle communication (especially for the critical devices/services). Implementing PLC would also require implementing some kind of module that will enable the devices to communicate over the power line; though it would be cheaper than implementing wireless modules. DC power lines also have predictable noise and impedance characteristics [3] that would allow for higher bandwidth. Further motivation for developing DC PLC is that it can be applied to a number of fields, not only in vehicles, but in robotics and medical devices as well. One could even imagine the future consisting of systems with only one power chord from the main power supply plugging into a central device. This central device then only needs one pair of conductors to provide both power and communication to all other devices of system. The implication would be that no more complicated wiring would be required to set up one s, otherwise wire-intensive, system. 2

22 Section 1.3: Dissertation Objective and Scope Chapter Advantages of In-Vehicle PLC There are several reasons for developing an in-vehicle power line communication solution: ˆ Power line communication will imply less wiring and cables throughout vehicle chassis, making it easier to construct, lighter (hence more fuel efficient) and cheaper. ˆ Delivering power to devices is the most important aspect of a system (since the device would not even work without power). There would therefore already be the infrastructure of power lines, and a robust one at that to ensure devices never go off. Thus using the same line to communicate would ensure the communication link is not broken either. ˆ Gaining access to a power line communication network would provide the convenience of just having to tap into the power line for both power and communication. This would imply that one could use the already installed power access points (e.g. cigarette lighters) as communication access points as well. ˆ Power line communication has even been investigated as a redundant communication channel for CAN (Controller Area Network) communication in vehicles [4]. This would be advantageous in that if a vehicle required guaranteed communication, then it could have two communication systems but not double the amount of wiring. ˆ As mentioned above, in-vehicle power line communication can also be applied to other systems such as those in robots and factories. 1.3 Dissertation Objective and Scope The objectives of this dissertation are to: ˆ Find an appropriate means to implement DC Power Line Communication. ˆ Using this mean to design an experimental setup for bit error rate detection. 3

23 Section 1.4: Document Overview Chapter 1 ˆ Use experimental data to create Markov models and compare these models with each other and the experimental data. ˆ These Markov models can then be used to simulate the bit error characteristics and help evaluate different channel coding schemes. The scope of the dissertation includes (but is not limited to) the investigation of power line communication, communication protocols, Markov modelling and bit error rate analysis. The dissertation also includes detailed design, testing, fault finding and analysis of experimental setups. 1.4 Document Overview This dissertation document covers the entirety of the project: the literature study, all the designs, implementations, implementation issues, experiments undertaken, results, conclusions and all references used. Chapter 1 is an introduction to the project and provides motivation for why this research is to be pursued. It also outlines the overall objectives of the research. Chapter 2 contains a summary of all the literature that was reviewed in order to proceed with the project. In particular it reviews the most used in-vehicle communication protocols, the feasibility of in-vehicle DC PLC, the vehicle power line as a communication channel, the PLC transceivers that will be used and other literature that is appropriately required for the later chapters. Chapter 3 presents the methods used to attempt and finally achieve bit error recording. There are three methods used, two of which provide insufficient results, and the third is the one used to obtain final results. The results are presented in multiple ways, all of which are used to help compare the models created in the following chapter. Chapter 4 presents the results of further in-vehicle experimentation, where we perform bit error recording on the channel under different conditions. 4

24 Section 1.5: Contributions Chapter 1 Chapter 5 presents the creation of four different Markov models to be used to simulate the bit error characterisation of the channel. All these models are compared to each other and the actual results obtained in Chapter 3 to determine which model is the better one. Using the conclusion of the better model, we create several models, each based on a set of results from Chapter 4. Chapter 6 uses the models created in Chapter 5 to evaluate the simple error detection mechanisms used in the Local Interconnect Network (LIN) protocol as well as simple error correction. Chapter 7 completes this dissertation with concluding remarks. 1.5 Contributions Parts of this dissertation (mainly from Chapters 3 and 5) were published in a conference paper at the IEEE Africon 2013 conference [5]. The title of the paper was: Markov Modelling of In-Vehicle Power Line Communication. The corresponding author was Mark Wilson (author of this dissertation) with co-authors Reolyn Heymann (supervisor of this dissertation) and Hendrik Ferreira (co-supervisor of this dissertation). Preliminary work done for this Africon publication was also presented under the Recent Results section of ISPLC 2013, where the work is not published anywhere, but is presented for peer review. A paper based on Chapter 4 s results was submitted for ISPLC 2014, but the acceptance status was not known at the time of submission of this dissertation. 5

25 CHAPTER 2 Literature Study Study the past, if you would define the future. - Confucius 2.1 Introduction This chapter provides a detailed literature survey of the key aspects of the project. It is important to perform as much research as possible before beginning a project as to gain as much knowledge on the subject as possible and to make sure one does not spend hours re-inventing the wheel. It is also important to be sure that no legal limitations are violated and that typical standards are adhered to as much as possible to ensure easier usability. The key standards in in-vehicle communication are investigated as well as how they are used practically in industry. This chapter also presents current literature research on the topic of in-vehicle power line communication as well as different Markov models, which is essential to analysis provided later in the dissertation. 2.2 In-Vehicle Communication There are a number of different protocols used for in-vehicle communication [6]: - Byteflight - Flexray. - Media Oriented Systems Transport (MOST). 6

26 Section 2.3: CAN Protocol Chapter 2 - Controller Area Network (CAN). - Local Interconnect Network (LIN). Byteflight was developed by BMW and was designed for high safety features. It is thus a high speed protocol but is quite expensive to implement. Flexray is a high speed, high reliable communication protocol that was designed for x- by-wire purposes to replace mechanical connections such as those between the brake/throttle pedals and the actual brakes/fuel injection actuators. MOST, as implied by the name, was designed for in-vehicle media such as the music system, headrest televisions and other media devices. CAN and LIN are the most used protocols for in-vehicle communication. They are used for the most general communication purposes such as switching indicators, head lights, back lights, controlling the window motor drivers, seat controllers etc. These are the protocols that in-vehicle PLC would most likely be replacing or be integrated with and are described in more detail below. 2.3 CAN Protocol Controller Area Network (CAN) is the most widely used protocol for in-vehicle communication [7] and it is thus important to understand how it works if a custom communication network is to be integrated. It was developed by Bosch [8] for the specific purpose of relatively high speed reliable in-vehicle communication at low cost (compared to Flexray and Byteflight). Some of the key features of CAN [9]: ˆ CAN uses a single terminated twisted pair cable. ˆ CAN is a multi-master protocol. ˆ A speed of up to 1 Mbit/s can be achieved with CAN (at up to 40 m). 7

27 Section 2.3: CAN Protocol Chapter 2 ˆ The typical maximum data rate is 40 kbytes/s. ˆ The nodes in a CAN network do not have addresses, but rather the messages have message identifiers. Each node looks at the identifier to see if the message is useful to it. The identifier also contains the message priority. ˆ CAN uses a similar scheme to carrier sense multiple access with collision avoidance (CSMA/CA), however, when two nodes attempt to send messages at the same time, only the lower priority message backs off to send at a later time, instead of both Message Format The CAN protocol has two logical values (as do many networking protocols); high and low. The high state is referred to as recessive and the low state is referred to as dominant. The reason for this is explained below under the message ID explanation. Below is the general format of a CAN message [10]: SOF Message ID RTR Control Data CRC ACK EOF A brief description of each field is provided: Start of Frame (SOF): This indicates the start of a message and consists of a single dominant (low) bit. Message Identifier: The message identifier field is 11 bits long. Messages with a lower numerical value are more prioritised, and thus, if two messages are trying to be sent at the same time, the more dominant message will hold the line at zero, indicating the less prioritised message to back off until a later time and allow the more dominant message to send. 8

28 Section 2.4: LIN Protocol Chapter 2 RTR: Control: Data: CRC: ACK: End of Frame (EOF): The Remote Transmission Request (RTR) field is a single bit that indicates if the node is requesting a transmission or is making a transmission. This bit is dominant if the message is a transmission and is recessive if the message is a request. The control field consists of six bits, four of which indicate the length of the data, followed by two dominant bits which are reserved for later expansion. The data frame can contain up to 8 bytes of data (each byte is 8 bits). This is the cyclic redundancy code derived from BCH coding (used for error coding). The acknowledgement field consists of two bits. They are both sent by the transmitter as recessive and when a receiver receives a successful transmission (after checking the CRC sequence) it makes the first bit dominant. The second bit is always recessive such that the acknowledging bit is always surrounded by at least one recessive bit on either side. Each message is ended by seven recessive bits (regardless if its a request message or a data message) Coding Message frames are coded by the method of bit stuffing, where, if the transmitter detects five consecutive bits of the same logic values, it inserts a complementary bit. Thus, if there are six consecutive bits of the same value then an error has occurred. CAN also makes use of CRC (as mentioned above) for error coding. 9

29 Section 2.4: LIN Protocol Chapter LIN Protocol Local Interconnect Network (LIN) is a slow speed UART based protocol. It was developed to be a low cost alternative to CAN for the lower priority devices such as window drivers, windscreen wiper drivers, sun roof controller, etc. LIN is a single-master multislave network protocol that is usually used as a sub-network of CAN. Some of the key features of LIN are [11]: ˆ The LIN protocol allows for speeds up to 20 kbps. ˆ It is a single master protocol with multiple slaves (up to 15). ˆ The LIN master device is usually also CAN compatible, to communicate with the rest of the vehicle. ˆ LIN is UART based. ˆ LIN is a one-wire protocol Message Structure The LIN message structure is as follows [12]: SYNCH. BREAK SYNCH. BYTE IDENT. BYTE DATA CHECK BYTE > , 32 or 64 8 The first three frames (SYNCH. BREAK, SYNCH BYTE and IDENT. BYTE) form the message header, which is always sent by the master. The rest of the frames contains the data and check bytes which could be sent by either the master or one of the slaves. Below is a small description of each field: Synchronization Break: The master holds the line low for a longer time length than a data byte. This way the slave detects a synchronization break. 10

30 Section 2.5: Integration of LIN and CAN Chapter 2 Synchronization Byte: Identifier Byte: Data Bytes: Checksum Byte: This byte is used for slave self-synchronization and is sent by the master. It consists of one low start bit, followed by eight alternating bits and one high stop bit. The duration of the eight bits defines the baudrate to be used by the slave. This byte defines the content and length of the data in the message. One must refer to the protocol specification for how exactly these parameters are represented within the identifier, but it consists of four message identifying bits, two data length defining bits followed by two parity bits. This byte is also always sent by the master. The length of the data is defined by the identifier byte and can be 2, 4 or 8 bytes long. Data is either sent by the master or the slave, depending on what the identifier byte specified. The checksum is computed on the data bytes. Its value is equal to the inverted sum of all the data bytes in modulo 256. This byte is sent after the data by the same node that sent it. NOTE: All bytes start with a low start bit and end with a high end bit, as per the norm for UART based protocols (the synchronization break is not a byte and thus does not contain these). The mechanisms for error detection in LIN are the checksum bytes (performed over the data bytes) and the parity bits (in the identifier byte). 2.5 Integration of LIN and CAN CAN was first introduced as a way to standardise most in-vehicle communication busses. It was designed to be high speed and reliable - two things that come at a cost. As 11

31 Section 2.5: Integration of LIN and CAN Chapter 2 vehicles were designed to have more and more sensor networks, electronic controller units (ECUs) and other electronic devices, the need for a more cheaper network protocol for the less important in-vehicle services came about. LIN was then invented. LIN is designed to be a sub-network of CAN. It usually handles the smaller networks of sensors or motor controllers so that each single node does not require a CAN interface. Instead, just the master node of the LIN requires CAN bus compatibility. Figure 2.1 shows how LIN and CAN compare in terms of cost and speed (as well as some of the other in-vehicle protocols). Figure 2.1: Graph Showing the Relative Cost and Network Speed of the Different Protocols (from [13]) Figure 2.2 shows how LIN and CAN are integrated into an in-vehicle communications network. Note that the LIN networks are small sub-networks of the CAN network. This is the most general representation of most in-vehicle networks, where the main nodes communicate via CAN, and the sub-networks, that consist of sensors and small motor drivers for the electric windows and seats, communicate via LIN. 12

32 Section 2.6: DC Power Line Communication Chapter 2 Figure 2.2: (from [12]) An Example Network Structure of In-Vehicle Communication 2.6 DC Power Line Communication The concept of communicating over power lines has actually been around since the 1890 s. In the 1950 s it was used at 10 Hz to control relays and town lighting [14]. Broadband networking over power lines was only achieved at the end of the 20th century. Most power line communication research and development was focused on achieving communication over the main 50/60 Hz household AC supply (commonly referred as the low voltage network). Recently, however, there has been a noticeable increase in DC power line communication research output (mainly for vehicles and smart homes ). This section investigates the research that has been performed in this area Feasibility of In-Vehicle Power Line Communication There has already been a fair amount of research on in-vehicle power line communication, each with its own motivations. Most of them being the reduction in cost, size and complexity of the wiring harness. However, we must also take a look at what the cost of implementing power line communication is, and how reliable it can be, specifically 13

33 Section 2.6: DC Power Line Communication Chapter 2 when compared to LIN and CAN. The advantages PLC offers in the automotive industry as well as the issues it faces are given in [15]. The most important issues being the reliability and speed of the network. CAN networks can operate up to speeds of 1 Mbps while LIN only usually operates at 20 kbps. Both LIN and CAN are considered reliable network protocols, and any other network that aims at replacing them or running alongside them must be just as reliable. Another important issue is cost. The cost of implementing the PLC network (the extra transceivers/integrated circuits and other components) must be less than the copper/wiring it replaces. This, however, is quite likely since the cost of semiconductors and ICs is quite low compared to copper, and gets cheaper everyday (whereas copper is getting more expensive). One more important issue, as suggested by [15], is that different subsystems usually have very different requirements, and thus very careful insulation or isolation from one another is an important design aspect to keep in mind if one is to implement PLC (since PLC will implement all communication on one line). One possible way of doing this is communicating on different frequencies over the same power line, but this may eventually cause the different communications to interfere with each other and thus cause unreliable communication. This may be a very important limiting factor to keep in mind when implementing an in-vehicle PLC. With the recent research and development on in-vehicle PLC, it is possible to communicate at CAN speed as presented in [4]. It is also easily possible to achieve LIN communication over the power line as discussed later in this chapter. In-vehicle PLC can therefore be considered reasonably feasible (at least for further research and development). 14

34 Section 2.6: DC Power Line Communication Chapter Legal Considerations and Regulations There are regulations that make a car roadworthy; such as the steering wheel must be big enough to enable full view of the dials, the car must be high enough such that the exhaust does not scrape the road at any point, vehicles may only have certain colour lights on the underside of themselves and many more. The official roadworthy test sheet is freely available online and was also published in the Government Gazette - 23 November 2005, (issue no ). This test sheet is basically a list of checks that are ticked and signed off by a registered test station employee. At the time of writing, there appears to be no regulations anywhere in the world that deals with In-Vehicle Power Line Communication directly. According to an e- mail response from the United Nation s Economic Commission for Europe (UNECE) in Geneva: there are not UN Regulations dealing directly with in-vehicle Power Line Communication (PLC). The full response suggests that the only regulations that may be of relevance is UN Regulations number 10 (which deals with electromagnetic compatibility (EMC) in vehicles) and UN Regulations number 100 (which deals with electric power trained vehicles). These regulations are freely available online [16]. Other important bodies regulating EMC is the CISPR (French acronym translating to International Special Committee on Radio Interference ) and the American Federal Communications Commission (FCC). Both of these bodies regulate EMC in their respective regions (most countries all over the world adopt regulations based on their regulations). It is important that any radiation that may be caused by in-vehicle Power Line Communication does not infringe the regulations set by these regulatory bodies In-Vehicle DC Power Line Communication Channel Characterisation There has been a fair amount of research on the type of noise and how the channel looks on a vehicle power bus. Rouissi et al. [17] took a look at the impulsive noise present on a switched on, idling vehicle s power line. Figure 2.3 shows their results which is a comparison between 15

35 Section 2.6: DC Power Line Communication Chapter 2 their experimental values versus their own model. From their experimental results, it is shown that the low amplitude impulse noise decreases as the frequency increases. Comparison Between Middleton s Model and Actual Measure- Figure 2.3: ments [17] The amplitude response of signals over the battery line in a vehicle as a function of frequency is looked at by van Rensburg et al. in [18]. Figure 2.4 shows their results, which indicates that frequencies between 10 and 30 MHz shows a promising band to communicate on due to the higher amplitude response. The band between 5 and 7 MHz is also promising due to the relatively flat response within that band. Similar experiments are performed with slightly different parameters in [19], [20], [21], [22], [23], [24] and [25]. Some of them measure at different points in different vehicles with different in-vehicle services turned on/off, etc. However, they all provide similar results to Figure 2.4 (where the better bands are between 5 and 7 MHz or between 10 and 30 MHz), which implies that for better communication results it is important to communicate within these bands. It is interesting to note that [26] performed the same experiment on a full electric vehicle and provided results that showed a similar shape except with a lot more dips in the transfer function. This is thought to be due to the motor drives. 16

36 Section 2.6: DC Power Line Communication Chapter 2 Figure 2.4: Amplitude Response of Vehicle Battery Power Line [18] Issues and Constraints The issues and constraints identified with in-vehicle PLC are: ˆ Interfering with the DC power supply line to inject communication onto it may affect some devices that are powered by the power line. It is important to keep this in mind and check the limitations of the power line after communication has been implemented on it. ˆ Some devices (especially as you plug them in or turn them on) may draw a spike of current which could cause noise in the power line. This may lead to impulsive noise that causes bit errors. ˆ A major issue of having a single communication line with multiple devices is wanting them to communicate simultaneously. Managing this would be an issue that has to be solved through protocols or special modulation techniques. ˆ Mechanical vibrations of a vehicle may cause circuitry to behave abnormally and may create extra noise. It is important to test the communication nodes under the vibrational conditions of a vehicle. ˆ Creating a new type of network in an already network-populated arena would mean making the new network as flexible as possible as to be able to easily in- 17

37 Section 2.6: DC Power Line Communication Chapter 2 terface with current networks within the in-vehicle communication arena. This means that the new network would have to be able to communicate with standard in-vehicle devices, most of which are already CAN and LIN compatible [27]. It would thus be very advantageous to create an interface that will translate LIN/- CAN messages onto the PLC network. This will make the PLC concept more appealing due to its compatibility with current devices. ˆ The power line will run throughout the whole vehicle, and implementing communication on it may cause the line to emit electromagnetic interference (EMI) which could interfere with any device along the power line, or any nearby wireless devices operating close to the bandwidth of the PLC network Technical Requirements: The technical requirements of an in-vehicle PLC network would include: ˆ There shall not be any dedicated communication lines between PLC nodes. ˆ The system must be able to run on the standard Lead Acid automotive battery (usually 12 V but can vary up to 24 V). ˆ The introduction of communication onto the power line cannot interfere with the transmission of power to devices, i.e. in-vehicle devices must not operate improperly because of lack of power due to the effects of the PLC. ˆ The communication network must support at least 20 kbps (in accordance with the CAN and LIN protocols [28][12]). ˆ The nodes should be able to convert CAN and LIN protocol messages into the PLC network, and then back into CAN and LIN protocol messages, so as to ensure compatibility with the majority of devices. ˆ The bit error rate must be low enough for reliable communication. This will depend strongly on the application of the communication network. ˆ The PLC network must physically be able withstand the environment of a vehicle. This includes withstanding temperatures of between -10 C and 62 C (these temperature values are based on how hot and cold vehicles can get [29]). 18

38 Section 2.9: Markov Modelling Chapter 2 ˆ Other physical requirements would include PLC nodes needing to be relatively small enough to fit into vehicles as communication nodes and contain required interface ports for ease of use and/or installation. 2.7 Arduino This section is a short introduction to the microprocessor development platform called Arduino that was used in certain parts of the project. According to the Arduino website [30]: Arduino is an open-source electronics prototyping platform based on flexible, easy-to-use hardware and software. It s intended for artists, designers, hobbyists, and anyone interested in creating interactive objects or environments. As implied by this description, Arduino was mainly developed for easy micro-controller interaction. They make use of ATMega microprocessors, and developed their own easyto-use software platform so that end users can easily implement their own programs on the controller. This makes it an ideal solution for anyone who has not much experience in embedded programming and/or micro-controller implementation. 2.8 Picoscope Picoscope is a range of PC-based oscilloscopes developed by Pico Technology. The particular model used in this work was the Picoscope It has a bandwidth of 10 MHz, a sampling rate of 100 MS/s, buffer memory of 8 ks and 8 bit resolution. For full specifications and further details refer to Pico Technology s website [31]. 19

39 Section 2.9: Markov Modelling Chapter Markov Modelling Although plenty of characterising has been done in terms frequency response and noise on the vehicle s power line, there has been less research in terms of statistical bit error distribution and bit error rate. This is very important for practical implementation as one needs to know if the applied protocols have good enough error detection and correction for reliable communication. This dissertation investigates the statistical bit error distribution and bit error rate, and creates Markov models based on experimental results. Here we describe two different kinds of models based on literature that will be used and how we go about getting their parameters. Two more models, that to our knowledge are not presented in any current literature, are described in Chapter Gilbert Markov Model The first model is the Gilbert model [32]. Other than the binary symmetric channel model, this is the most simple Markov model. Even with it s simplicity, it has been used to accurately model a number of channels. See [32], [33] for examples. The Gilbert model consists of a two state Markov chain, as illustrated in Figure 2.6. Each state, however, represents a binary symmetric channel (illustrated in Figure 2.5, where there is an equal probability that a 0 will error into a 1, and 1 will error into a 0). The first state, referred to as state G ( good state), has an error probability of zero. The second state, referred to as state B ( bad state), has error probability 1 h (where h is the probability of no error occurring). Each time step presents a probability to switch between each state as well as a probability of an error occurring within that state (if the channel is in the bad state). This implies two random calculations per time step when simulating this model. The probability of going from state G to state B is P, and the probability of going from 20

40 Section 2.9: Markov Modelling Chapter 2 Figure 2.5: Binary Symmetric Channel Figure 2.6: Gilbert s Simple Markov Model [32] state B to state G is p. The three parameters P, p, and h are required to fully define the Gilbert model. Unfortunately, these parameters are not directly observable from experimental data. However, Gilbert provides the following equations in his paper [32]: 1 p = q = ac b 2 2ac b(a + c), h = 1 b q, P = ap 1 h a, (2.1) 21

41 Section 2.9: Markov Modelling Chapter 2 where a, b and c are parameters that are directly observable from measured data and can be given by: a = P (1), b = P (1 1), c = P (111) P (101) + P (111). (2.2) P (1) denotes the average probability of an error (given by the number of measured errors divided by the number of transmitted bits). P (1 1) is the probability that an error has occurred given the condition that the last bit was an error (given by the amount of errors that occurred after an error divided by the amount of errors measured). The parameter c is the probability of a bit being an error given the condition that the previous bit as well as the following bit are both errors and is given by: c = (1 h)q2 q 2 + pp. (2.3) Creating a Gilbert model based on results is now simply a case of calculating a, b and c from experimental results and substituting their values into the equations for the parameters given by Fritchman s Markov Model Fritchman s Markov model is a more general model that has more than two states. Most commonly it consists of multiple error-free states and one error state. Unlike the Gilbert model, each state does not represent a binary symmetric channel, but the state which the channel is going into determines if the next bit is an error or not. Figure 2.7 illustrates this concept a little more clearly. In this figure, going from state E 1 to state 22

42 Section 2.9: Markov Modelling Chapter 2 E 4 will give an error. And going from E 4 to any other state will give no error. This implies one random calculation per time step when simulating. From here we will refer to any state that produces an error (i.e. E 4 in the figure) as an error-state and the other states as error-free states. Figure 2.7: Transition Diagram for a Fritchman Model with Three Zero- Error States [33] In this Fritchman s model, once in the bad state it is possible to either stay in the error-state or go to any of the good states but it is not possible to go from a good state to any of the other good states. There are a number of different methods for obtaining the parameters of this Fritchman s model. A specific gradient method for curve fitting is used in [33] while reference [34] uses more of a linear algebraic approach. In this dissertation we will be using the curve fitting method (described below), which was also used by Fritchman himself [35]. We first define some parameters: Let N be the number of states, k the number of good states and thus N = k + 1 for the type of model we are considering (since there is only one bad state). The probability of there being m error-free bits in a row given that the channel has come from a bad state (i.e. the probability of a error free run of length m) would be given by: 23

43 Section 2.9: Markov Modelling Chapter 2 P (0 m 1) = = k P N,i (P i,i ) m P i,i k α i βi m, (2.4) i=1 i=1 where P a,b is the probability of the channel going from state a to state b in a time step. Note that α and β are there to simplify the equation and state N represents the bad state. The experimental results will consist of a list of gaps (error free runs). To calculate P (0 m 1) we count the number of m error-free bits in a row and divide it by the total number of errors. We can then plot P (0 m 1) with respect to m. This will provide a set of data points to which we can fit a curve of a form given by equation 2.4. The equation of the fitted curve will then contain the parameters for the model. The transition matrix (defined below) for the Fritchman model will then be of the following form: β β 1 0 β β β N 1 1 β N 1 α 1 β 1 α 2 β 2... α N 1 β N 1 N 1 1 α i β i i=1 (2.5) Transition Matrix The transition matrix is a general way of storing the state transition parameters of a Markov Chain, making simulation algorithms more general as well as provides a bit of extra analysis. The transition matrix is often notated by P. Each state in the Gilbert model is a binary symmetric channel and thus one cannot represent the model using a transition matrix. But for many other models, such as the 24

44 Section 2.10: Error Control Coding Chapter 2 Fritchman model, this is possible. The transition matrix is always a square matrix. Each row of the matrix represents a state and each column (or component) in that row gives the probability of going into the other respective states. For example, if the channel is in state 1, then we look at row 1. The probability of it staying in state 1 is the component in row 1, column 1. The probability of it going into state 2 is the component in row 1, column 2, etc. With this matrix it is possible to calculate the stationary probability vector. The components of this vector essentially gives us the percentage of time that the channel will spend in each state. The percentage of time spent in the bad state(s) is then the expected BER of the channel. Below are the equations to compute the vector [36]: L = lim k D k, P = QLQ 1, (2.6) where D is the diagonal matrix containing the eigenvalues of P and the columns of Q are their respective eigenvectors. L is referred to as the limit matrix. The matrix P has identical columns and represents the stationary probability vector. This is useful because we can use it to predict the effectiveness of a model simulating experimental results without having to even perform the simulation Error Control Coding Later in the dissertation, we will be analysing the effectiveness of simple error detection (used in the relevant LIN protocol) and correction techniques over the power line channel. Here, we provide explanations for the basic concepts required to understand these techniques, as well as how the techniques work. 25

45 Section 2.10: Error Control Coding Chapter 2 Error coding is not the core of this dissertation, and thus we do not provide a very detailed discussion for these concepts, but provide a basic foundation of the fundamental concepts Information Rate A very important concept to keep in mind when we analyse these methods in Chapter 6 is the information rate which they provide. The term information rate refers to the ratio of actual information bits to the total amount of bits being sent over the channel and is very often used to evaluate the efficiency of error coding schemes [37] [38] [39]. This is important because error detection and correction coding very often makes use of some kind of redundancy in order to achieve their purpose of error detection/correction. The simple equation that gives us the information rate, ρ, is: ρ = k n, (2.7) where k is the number of information bits and n is the number of the bits that are going to be transmitted. Note that this equation has a more general form for more complex coding schemes, but for the purposes of this dissertation, this form is sufficient Modulo Arithmetic Modulo m refers to the remainder of a number when it is divided by m (where m is referred to as the modulus). E.g. 15 modulo 11 is 4, because 15 divided by 11 is 1 with a remainder of 4. In this dissertation we are mainly concerned with modulo 2 and the addition of 1 s and 0 s in modulo 2. We therefore do not delve too deeply in this subject and only show the necessary features and properties of modulo 2 arithmetic. The notation used in this dissertation for modulo 2 addition will be the symbol. 26

46 Section 2.10: Error Control Coding Chapter 2 Another important symbol used is the symbol, which will denote the inverse of a bit (and we will see shortly how this relates to modulo 2 addition). Examples are shown below: 0 1 = = = = 1 1 = 0 = = 1 = 0 1 (2.8) An important feature demonstrated by the examples in 2.8 of the modulo 2 addition of 1 s and 0 s is that the answer is always 0 if there are an even number of 1 s in the sum and it is always 1 if there is an odd number of 1 s in the sum. Another important feature is that the inverse of a bit is equivalent to the addition of a 1 in modulo Parity Bits Using parity bits is the simplest mechanism for error detection. A parity bit is a bit that you add to your sequence of information bits, where the value of this extra bit depends on the number of 1 s in your sequence. The simplest example is inserting a single parity bit to your sequence which would be equal to a 1 if you have an odd number of 1 s or a 0 if you have an even number of 1 s (thus the parity bit is equal to the modulo 2 sum of the information bits). This is known as an even parity. An odd parity is when you insert a 0 parity bit if there are an odd number of 1 s or a 1 parity bit if there are an even number of 1 s (which makes the parity bit equal to the inverse of the modulo 2 sum of the information bits). Below is how to calculate the 27

47 Section 2.10: Error Control Coding Chapter 2 parity bit using modulo 2 addition: Even Parity: P = D 0 D 1 D 2... D n, Odd Parity: P = (D 0 D 1 D 2... D n ), (2.9) where D k denotes the kth bit in the information sequence that is n bits long. It is easy to deduce that a single parity check bit will be able to detect an odd number of errors in the sequence, but not an even number of errors [37], because the parity of a sequence will change when a bit inverts, but will change back when a second inversion occurs, and so on. A single even parity check bit is often used in serial protocols where information is sent in bytes that consists of 7 information bits and the 1 parity bit. In Chapter 6, we evaluate the use of a single parity bit for communication over the vehicle s power line, and we use this same scheme with 7 information bits and 1 parity bit. It is also possible to have 2 or more bit parity check bits on the same set of information bits; which enables the parity detection mechanism to detect more combinations of errors than the single parity bit (which is equal to the number of combinations that have an odd number of errors) - this is investigated further in Chapter 6. Here, we explain how the LIN protocol discussed earlier in the chapter uses two parity bits in an attempt to protect its identifier byte. The identifier byte is constructed as follows: D 0 D 1 D 2 D 3 D 4 D 5 P 0 P 1 Where D 0 - D 5 are the six identifying bits that determine the intended address and length of data. P 0 and P 1 are the parity bits and are calculated by: P 0 = D 0 D 1 D 2 D 4, P 1 = (D 1 D 3 D 4 D 5 ). (2.10) 28

48 Section 2.11: Conclusion Chapter 2 We can see that P 0 is an even parity and P 1 is an odd parity. Further analysis of this error detection scheme is provided in Chapter Checksum Another error detection technique we will analyse in Chapter 6 is the checksum byte (also used in the LIN protocol). This is a simple yet effective technique in error detection. There are different ways of implementing a checksum. We will be using the method employed by the LIN protocol. A checksum byte is an extra byte that is added to the end of the data bytes, and, in the case of LIN, is equal to the inverted sum of the numerical values of the data bytes in modulo 256. To understand it more clearly, the steps involved in calculating the checksum are: ˆ Add the numerical values of the 8-bit data bytes together. ˆ Get the value in modulo 256. ˆ Convert this value into a binary 8-bit byte. ˆ Invert the bits of this byte (which is the equivalent of subtracting the value from 255). The checksum byte in the LIN protocol is calculated over all the data bytes that are sent (which could be 2, 4 or 8 bytes) Conclusion This chapter describes the two most widely used in-vehicle communication protocols (that being CAN and LIN); what they are mainly used for, how each network protocol is implemented as well as the integration between these two networks. The feasibility of in-vehicle DC PLC is also discussed. This included a critical look at the major issues that this new venture may have to overcome to become a practical 29

49 Section 2.11: Conclusion Chapter 2 network standard. The discussion ended with the conclusion that in-vehicle DC PLC is feasible enough to encourage further research efforts in its direction. A discussion on the vehicle power line as a communication channel was also presented and literature in this regard was reviewed. Also, two commonly used Markov models were presented. We then provide an introduction to some of the hardware used for the experimental method which includes the Arduino and the Picoscope. All their basic features and functionalities were presented. Lastly, we provide explanations for some core concepts for basic error detection coding. 30

50 CHAPTER 3 Bit Error Recording People love chopping wood. In this activity one immediately sees results. - Albert Einstein 3.1 Introduction This chapter is dedicated to the development of the bit error recording methods used and how bit errors were detected - which was in itself a large challenge. First, the signals on the power line at the transmitting transceiver and the receiving transceiver are shown to demonstrate the attenuation of the transmitted signal, which is the most probable cause of the errors in transmission in this chapter (together with noise). Sections 3.4 and 3.5 describe important concepts that are needed to understand the results and how the information was transmitted. Sections 3.6 and 3.7 present the first two methods used to attempt bit error recording but were unable to provide full results. Section 3.8 presents the final method used to successfully perform bit error recording as well as the results of the bit error recording exercise. 3.2 Experimental Platform All experiments in this dissertation were performed on a 1998 Toyota Tazz using SIG60 DC PLC transceivers (all technical information and motivation of these transceivers 31

51 Section 3.3: Signal Waveform Chapter 3 are presented in Appendix A). One of the transceivers was connected directly to the vehicle s battery with minimal wire length. The second transceiver was connected to the cigarette lighter of the vehicle through a 1 meter cable. The experiments in this chapter were performed with the car key in the accessories position (i.e. the engine was not running). The experiments were performed at baud rate modulated onto a 6.5 MHz carrier frequency (6.5 MHz being the carrier frequency recommended by Yamar). Of course, performing the experiments at different carrier frequencies and baud rates would possibly produce different results, but we use the above parameters as a basis to develop our experimental method as well as our methods for creating models. 3.3 Signal Waveform Before performing the actual bit error recording, the waveforms of the transmitted signal and the received signal is presented in Figure 3.1. These waveforms in the frequency domain are presented in Figure 3.2. It is clear that there is a large attenuation of the signal through the vehicle s power line, most probably due to the vehicle s various loads. The RMS of the transmitted signal is mv while the received signal has an RMS of mv. This is an attenuation of db. The noise level at the carrier frequency (6.5 MHz) appeared to be constant throughout the vehicle with an RMS value of mv. This gives us a Signal-to-Noise Ratio (SNR) of db at the receiving end. One will also notice two other peaks in Figure 3.2. The one at 5.5 MHz is believed to be a subsequent result of the internal workings of the SIG60 IC. When the SIG60 transmits, it transmits at two carrier frequencies, both of which are user defined through configuration of the IC. The IC can then easily switch between the two carrier frequencies by internally switching between two externally connected bandpass filters (each passing the corresponding carrier frequency). This feature was designed so that the 32

52 Section 3.3: Signal Waveform Chapter 3 application (whatever it may be) has the ability to quickly switch between transmission frequencies if necessary or desirable. This feature is mainly used for two reasons. One is to avoid noise - if the one frequency band is getting too noisy, the transceivers can switch to another, less noisy frequency. The other reason is to possibly have two communication channels on the same line. In our experiments, we use the default configuration of the IC (which is recommended by the manufacturer) which makes the two carrier frequencies 5.5 MHz and 6.5 MHz. The small peak observed in Figure 3.2 is thus the 5.5 MHz frequency that is not fully suppressed by the 6.5 MHz bandpass filter. The other small peak observed in the figure is at around 1 MHz. Investigation of the circuit setup indicates that this peak is a result of the regulator that provides the SIG60 IC with 3.3 V. Figure 3.1: Transmitted and Received Signals in Time Domain 33

53 Section 3.4: Error Vector and Gap Recording Chapter 3 Figure 3.2: Transmitted and Received Signals in Frequency Domain 3.4 Error Vector and Gap Recording In this chapter, results are presented through error vectors. The error vector is essentially a string of 1 s and 0 s, where the 1 s show where errors have occurred and 0 s represent the error-free bits. It is not to be confused with the actual strings of 1 s and 0 s that were transmitted. Graphically illustrating the error vector versus bit number nicely shows the basic behaviour of bit errors on the channel, and we will often refer to this kind of illustration as the error vector pattern. However, storing an error vector can be a large waste of memory on a computer, since millions of bits are sent and for each bit a 1 or 0 value has to be assigned. An easier way of storing how many, and in what positions, errors have occurred is to just record how many bits were error-free before each error occurred. This is known as gap recording (because it records the length of the error-free gaps ). Below is an example of 10 bits that is transmitted over a noisy medium that causes errors as shown: 34

54 Section 3.6: Matlab Send/Receive Chapter 3 Transmitted Sequence: Received Sequence: Error Vector: Gap Recording: [2; 3; 0; 2] (3.1) The number of values that need to be stored for gap recording is equal to one plus the number of errors occurred which, for practical channels with BERs less than 10 2, is a small fraction of the number of bits sent. This makes storing the results and handling the data easier. 3.5 Sequence of Bits There was one similarity across all experimental setups and that is the sequence of bits used to send/receive over the line. The problem here was that one had to start with a low start bit to get the transmission started and have low bits regularly enough to ensure continuous transmission (due to the UART nature of the SIG60 transceivers). To keep transmission simple, it was chosen to simply send ASCII characters one after the other. We start with ASCII character 33 and end with 126 and repeat the sequence. This sequence of bits gives us 48.3% of the bits being high bits and 51.7% being low bits which is very close to an even distribution of 1 s and 0 s. This sequence also allows one to create a simple algorithm to be able determine what the next bit should be without having to store the entire sequence on the receiving side. This algorithm is presented in Appendix B. 3.6 Matlab Send/Receive The first method which was used to obtain preliminary results was a simple case of preloading the same string onto two separate Matlab interfaces on two different laptops 35

55 Section 3.6: Matlab Send/Receive Chapter 3 and then sending the string over the powerline and comparing the received string with the pre-loaded string. The basic setup is illustrated in Figure 3.3. The reason why this was used only to obtain preliminary results was because of a major flaw; that is it cannot detect if an error has occurred in the start or stop bits. APPARATUS: ˆ Two laptops with MATLAB R2012a software. ˆ Two SIG60 Evaluation Boards (Schematics available online [40]). ˆ Two SIG60 Evaluation Board USB/Serial Converter. Figure 3.3: Basic Setup for Matlab Send/Receive Method Due to the nature of the UART in the serial/usb converter, Matlab will only receive the bytes that start with a low start bit and end with a high stop bit. If one of these bits is an error then the whole byte will be an error and this makes it sometimes impossible to tell exactly which bits were errors and which were not errors. It also makes it very difficult to resync and tell what errors may have occurred later in the string. Figure 3.4 shows an example of one of the longer runs of these preliminary results. The figure plots the error vector pattern as described earlier. Due to the scale of the image, 36

56 Section 3.7: Arduino Sampling Chapter 3 one stem often represents a small burst of errors (which is evident when one zooms in on a stem). Figure 3.4: Matlab Send/Receive Results - Error Vector Pattern Plot The BER (bit error rate) for these results was an average of We will compare these results with results shown below to demonstrate that it is not very accurate. 3.7 Arduino Sampling The next step was to sample each bit without a UART so that we can check the start and stop bits as well. The first attempt at this was using an Arduino microcontroller platform to check each bit as it comes and count how many error free bits were in between each error and store these counts (thus recording the error-free gaps as described earlier in the chapter). Unfortunately, the timing on most micro-controllers is not accurate enough to sample in exact enough intervals to provide reliable bit detection. Thus an external 555 timer was used, together with interrupt routines on the microcontroller, to provide the necessary timing to correctly sample each bit. Figure 3.5 shows the basic setup. APPARATUS: ˆ Two laptops. 37

57 Section 3.7: Arduino Sampling Chapter 3 ˆ One Arduino Micro-controller connected to 1 Laptop (to view/save results). ˆ One 555 Timer circuit. ˆ Two SIG60 Evaluation Boards (Schematics available online [40]). ˆ One SIG60 Evaluation Board USB/Serial Converter. Figure 3.5: Arduino Experimental Setup The micro-controller has two interrupts, where one would activate on a change in the serial line (rising or falling edge) and reset the timer. This makes sure the timer does not drift due to slight timing differences in the timer and the bit timing on the serial line. The second interrupt would detect a rising edge on the timer, and this tells the controller to read the digital value of the serial line. Figure 3.6 shows how the timing works on this setup, where the controller reads the digital state of the serial line at the rising edges of the timer and the timer is reset at each change of the serial line to reset the offset of the timer. The initial offset between the timer and the serial line would be a few microseconds due to inherent delays in the micro-controller s switching operations. This is ideal because one does not want to sample right at the edge of each bit. After each bit of the same value (no timer reset) the offset between the serial timing and clock timing becomes a little bit longer (due to the slight inaccuracy of the clock) until the serial line changes 38

58 Section 3.7: Arduino Sampling Chapter 3 Figure 3.6: Arduino Setup - Timing Diagram state at which point the clock is reset again. This means that there is a limit to how many bits of the same value can be in a row before the clock loses synchronisation with the serial line. However, the amount of bits of the same logic value in a row would be quite high (at least 50) to cause this loss in synchronisation. Due to the nature of the bit stream being sent (with a maximum of seven bits of the same value in a row), this is extremely unlikely. This method of bit recording works perfectly when used directly on an ordinary serial/usb converter from the laptop. However, the SIG60 transceivers display an odd characteristic that made this method futile; and that is that every few bytes there is a stop bit that is about 1.5 times the length of every other bit. This is due to the transceiver communicating a little bit faster than the specified baud rate to compensate for any timing infractions of the powerline signals. Thus, every now and then, the stop bit is a bit longer to resync (Figure 3.7 demonstrates this behaviour, where the stop bit is about 84µs instead of the usual 52µs). In normal serial communication operations this poses no problem but for bit error recording it makes for a challenge, since the Arduino would often count this stop bit as two bits. 39

59 Section 3.8: Picoscope Sampling Chapter 3 Figure 3.7: The Long Stop Bit Anomaly The only way to overcome this was to sample alot quicker than once per bit. It was possible to sample twice per bit using this Arduino method, but it was not possible to sample any quicker (due to the speed limitations of the switching and reading operations of the micro-controller). 3.8 Picoscope Sampling The Picoscope 2204 was introduced in Chapter 2. The specifications of the scope allow for a much higher sampling rate than needed. Another nice feature of this PC-based oscilloscope is that drivers and programming libraries are provided so that one can easily create an application to interface with the oscilloscope. Figure 3.8 shows the basic setup of this method. APPARATUS: ˆ 2 x Laptops. ˆ 1 x Picoscope 2204 connected to 1 Laptop (to view/save results). ˆ 2 x SIG60 Evaluation Boards (Schematics available online [40]). ˆ 1 x SIG60 Evaluation Board USB/Serial Converter. 40

60 Section 3.8: Picoscope Sampling Chapter 3 Figure 3.8: Picoscope Experimental Setup For the purpose of this experiment the oscilloscope was made to stream all readings directly to the application (which was written in C). The application then used a simple algorithm to determine if each bit in the stream was correct or not and performed gap recording as described previously. Figure 3.9 shows the basic flow diagram of the application where the oscilloscope sends its measured samples to the application in blocks. The oscilloscope waits until its one memory buffer is full, then sends that block of samples through a USB but at the same continues to sample the transceiver s output serial line, storing the samples in the next memory block. The application then goes through each block of data as it is passed and checks the samples. By the time the next block is ready to be sent from the oscilloscope to the application, the application would have finished analysing the previous block and be waiting for it. The sampling rate of the scope was 2 µs. This gives us about 26 samples per bit (since each bit would be 52 µs at a baud rate of ). Thus the application checks if there are 24 samples of the same logic level in a row (to allow for bits that may be a bit shorter than usual) and if there is then a bit has been detected. This solves the issue of the longer stop bits being detected twice because half a bit would measure to be about 41

61 Section 3.8: Picoscope Sampling Chapter 3 Figure 3.9: Basic Flow Diagram of Picoscope Application only 13 or so samples and would be discarded. Figure 3.10 shows typical results from this experiment (over the span of 2 million bits). The experiment was performed 10 times to gain good averaged results. This gives us a total of bits from the experiments. The average bit error rate was found to be

62 Section 3.8: Picoscope Sampling Chapter 3 If one were to zoom in on the results to have the same amount of bits as for the preliminary results then it can be shown that the preliminary results obtained by the Matlab Send/Receive method are not incorrect, but do not contain enough information for accurate analysis. As with the Matlab send/receive results, one stem presented here often represents a small burst of errors due to the scale of the image. Figure 3.11 illustrates this by showing the zoomed in section of the very first stem. It is very rare that a single error occurred by itself with large gaps on both sides. Figure 3.10: Final Results - Typical Error Vector Pattern Plot The results are fully presented in the next chapter. 43

63 Section 3.9: Conclusion Chapter 3 Figure 3.11: Final Results - Zoomed in Section of Error Vector Pattern Plot 3.9 Conclusion Three experimental methods were used before the final results were obtained. Each method and corresponding setup was explained in this chapter. The conclusion of each method is provided: The Matlab Send/Receive was the simplest and cheapest (provided two laptops are available), but did not provide the full picture of the results due to its inability to record start and stop bits. The Arduino method was a lot more complicated, but still relatively cheap. This method worked perfectly with an ordinary serial device, but was futile when using the SIG60 transceiver due the device s tendency to have a longer stop bit to compensate for the timing of the actual signal. The final method to obtain full results was using the Picoscope 2204 PC based oscilloscope. This method was relatively expensive because of the scope itself, and relatively complicated (due to the program code). 44

64 Section 3.9: Conclusion Chapter 3 The BER of the transmission over the channel was In addition to the error vector of a typical experimental run, the results were presented through a few statistical distributions: gap distribution graph, error distribution graph, burst distribution, burst-interval and cluster distributions. All of these forms (among others) are used later to help model the channel and compare model simulations with the experimental results. 45

65 CHAPTER 4 Experimental Results Insanity: doing the same thing over and over again and expecting different results. - Albert Einstein 4.1 Introduction In the previous chapter, we developed a test bed and experimental procedure that we can now use to perform more bit error recording under different conditions. In this chapter we do just that; use the developed experimental platform and procedure to obtain more results of interest and present them in a fashion that is able to display all of the interesting characteristics of the results. 4.2 Experimental Procedure As mentioned, we will use the experimental method that gave us results in the last chapter (that is the Picoscope Sampling method). Due to the vehicle being a fairly old model, the power line was not very accessible at different points in the vehicle. Thus, we leave the transmitting transceiver at the battery and the receiving transceiver connected through the cigarette lighter onto the back seat. The variance of these experiments lies in the operating condition of the vehicle (i.e. with different devices in the vehicle turned on/off). Throughout the experimentation it was found that some of the experiments produced 46

66 Section 4.3: Presentation of Results Chapter 4 very error-rich results. For the possibility that these results may hide some important statistical characteristics that would be difficult to observe under such error-rich results, we perform (and present the results of) the same experiments but with a boosted signal from the transmitter. This helps confirm some of the characteristics of some of the operating conditions that the experiments are performed under. To boost the signal we use an op-amp circuit as presented in Appendix D, and the booster was set to boost the signal voltage 1.5 times. Any graphs that have a dashed line represent the results for the boosted transmission signal. Each experiment was performed 10 times at bits each, to get a total of transmitted bits per experiment. 4.3 Presentation of Results For each set of results the following will be presented: ˆ Typical Error Vector Pattern ˆ The Average Bit Error Rate ˆ Error-Free Run Distribution Graph ˆ Burst Length Distribution Graph ˆ Burst-Interval Length Distribution Graph ˆ Cluster Length Distribution Graph The concepts of typical error vector patterns and bit error rates were presented in the last chapter, and thus do not require explanation, but do take note that some of the different error vector patterns are plotted over different spans of bits. This was done to try represent the main characteristic of the errors and bursts as clearly as possible. The other concepts listed above are explained below. Take note that it is common practice to present the error-free run, burst, burst-interval and cluster distribution graphs onto one set of axes as practised by the authors of [41], 47

67 Section 4.3: Presentation of Results Chapter 4 [42] and [43]. We will present the graphs in the same manner. It is important, however, that when viewing the graphs one must remember that the error-free run distribution graph is not a relative cumulative frequency like the other three distributions, but represents a probability (as described below) Error Free Run (Gap) Distribution Graph The first important aspect to the results is the probability distribution of error free runs. This is essentially P (0 m 1) plotted with respect to m, i.e. the probability of m error free bits occurring after an error bit. This graph is also important for parametrising the Fritchman model. Note that the error free run distribution graph will also be referred to as the gap distribution graph Burst Length Distribution A burst of errors can be defined by a group of bits in the transmission sequence that starts with an error bit and ends in an error bit and maintains a specified bit error density (denoted by 0 ) between (and including) these two errors [41]. To calculate the length of a burst, one looks for the first error, then for each error afterwards, one calculates the bit error density of the interval between the first error and the current error. If the density falls below 0, then the length of the burst is equal to the number of bits between (and including) the two error bits when the density was last above (or equal to) 0. The burst length distribution (which we will refer to as burst distribution from here onwards) is then calculated as the cumulative relative frequency of the lengths of the bursts. Cumulative relative frequency is calculated by adding the relative frequency of all the previous length bursts up to and including the current length burst and is given 48

68 Section 4.3: Presentation of Results Chapter 4 by equation 4.1: k i=0 n i N (4.1) Where, in this case, k is the burst length that we are calculating the cumulative relative frequency for, n i is the number of bursts that occurred that have length i and N is the total number of bursts that have occurred. By the nature of probability, this sum will always tend to 1 as k tends to infinity. Figure 4.1 shows what a typical burst may look like in an error vector (as well as bust-interval and clusters which are explained below). Figure 4.1: Cluster Sample Error Vector Illustrating a Burst, Burst-Interval and For our results, we let 0 = Burst-Interval Length Distribution As shown in Figure 4.1, a burst-interval is the number of bits between the bursts defined above. The burst-interval length distribution (which we will refer to as burst-interval distribution from here onwards) is the cumulative relative frequency of the lengths of these intervals Cluster Length Distribution The cluster length distribution (which we will refer to as cluster distribution from here onwards) is the cumulative relative frequency of the length of clusters (which are 49

69 Section 4.4: Experimental Results Chapter 4 illustrated in Figure 4.1). 4.4 Experimental Results We present the results in pairs - each operating condition of the vehicle without the booster and with the booster (with the exception of the first results - see below). Take note that when the engine was on, it was idling No Devices On Engine Off This is the same results we achieved in the previous chapter, where the bit error recording method was developed. As mentioned, the BER of these results was Refer to Figure 3.10 for a typical error vector pattern. When boosted with the booster, there were zero errors (even when measured for longer than transmitted bits). We thus have no graphs to present for that exercise, but it is important to take note of. Figure 4.2 shows the distribution graphs for the non-boosted transmission Engine On The next results we present is when the engine was running, but still no devices are on. With no boosted transmission the bit error rate was This is about double of that when the engine was not running. Figure 4.3 shows a typical error vector pattern for this transmission over bits. When the transmission is boosted, we observe a very low bit error rate of A typical error vector pattern over bits is presented in Figure

70 Section 4.4: Experimental Results Chapter 4 Figure 4.2: Error Distribution Statistics for the In-Vehicle Power Line Experiments - No Devices On - Engine Off Figure 4.5 shows the distribution graphs. Remember that the solid lines represents the non-boosted transmission results while the dashed lines represents the boosted transmission results. From this figure we can observe a sharp increase in the burst lengths between 8-9 bits. This implies that the bursts of noise on the channel under the condition of no devices connected and the engine running are about µs long. 51

71 Section 4.4: Experimental Results Chapter 4 Figure 4.3: Typical Error Vector Pattern for Transmission with No Devices on - Engine On Figure 4.4: Typical Error Vector Pattern for Transmission with No Devices on - Engine On - Boosted Transmission 52

72 Section 4.4: Experimental Results Chapter 4 Figure 4.5: Error Distribution Statistics for the In-Vehicle Power Line Experiments - No Devices On - Engine On Cell Phone Charger Plugged In Engine Off In this recording exercise, the car key was still in the accessories position, and the output of the cigarette lighter was split between the receiver and a cheap, no-name brand cell phone cigarette lighter charger with a Samsung Galaxy S3 cell phone being charged. The bit error rate of the transmission without the transmitting booster was This already shows that the presence of an extra device heavily affects the transmission over the battery bus. Figure 4.6 shows a typical error vector pattern that has been scaled down to 1000 bits. It is interesting to note that the bursts do appear to occur at a particular frequency. This is discussed further below. The bit error rate of the transmission with the transmitting booster was This provides much clearer results, as shown in a typical error vector pattern shown in Figure 4.7; which was over a span of bits. 53

73 Section 4.4: Experimental Results Chapter 4 The distribution graphs are presented in Figure 4.8. There are some interesting characteristics shown in these graphs. Firstly, both burst length distributions show a sharp increase at a length of about 8-9 bits, similar as for when no devices are plugged in with the engine running. This may indicate that either the burst noise caused by the cell phone charger lasts for the same amount of time; 8-9 bits ( µs), or it may be an indication of the characteristics of the transceivers themselves, or a general characteristic of the battery bus. The burst-interval distribution of the non-boosted transmission results has a sharp increase also around 8-9 bits, which indicates that the burst-intervals are of about the same length. This implies that the charger is giving off some noise burst every 442 µs on average. This kind of periodic noise is known in the field of power electronics and it has been studied [44] [45]. It is also interesting to note that the maximum cluster length for both transmissions was 7; which correlates to both transmissions having similar burst lengths. The error-free run graphs shows the large improvement in transmission when the transmission is boosted. 54

74 Section 4.4: Experimental Results Chapter 4 Figure 4.6: Typical Error Vector Pattern for Transmission with Cell Phone Charger Plugged In - Engine Off Figure 4.7: Typical Error Vector Pattern for Transmission with Cell Phone Charger Plugged In - Engine Off - Boosted Transmission 55

75 Section 4.4: Experimental Results Chapter 4 Figure 4.8: Error Distribution Statistics for the In-Vehicle Power Line Experiments - Cell Phone Charger Plugged In - Engine Off Engine On Performing the bit error recording exercise while the cell phone charger is plugged on and the engine was running produced at least one interesting point - the bit error rate with no transmission boost was , which is lower than when the engine was off. It is possible that it was just a good sample that we got for this experiment, and a bad sample for the previous experiment, but this is doubtful since a fixed amount of bits in the order of 10 7 for both experiments was transmitted and measured. The same experiment with the transmission signal boosted exhibits the opposite trait in that its bit error rate increased (as one may expect) from with no engine running to with the engine running. Other than that anomaly, the distribution graphs do not show much change in shape or form between the engine being on or off. Figures 4.9 and 4.10 show typical error vector patterns for the non-boosted and boosted transmission runs respectively. Figure 4.11 shows the distributions. 56

76 Section 4.4: Experimental Results Chapter 4 Figure 4.9: Typical Error Vector Pattern for Transmission with Cell Phone Charger Plugged In - Engine On Figure 4.10: Typical Error Vector Pattern for Transmission with Cell Phone Charger Plugged In - Engine On - Boosted Transmission 57

77 Section 4.4: Experimental Results Chapter 4 Figure 4.11: Error Distribution Statistics for the In-Vehicle Power Line Experiments - Cell Phone Charger Plugged In - Engine On Windscreen Wipers On Engine Off For this set of results, the car key was moved to the on position - the last position before ignition. This was to enable the use of the windscreen wipers. The wipers were put on the slowest setting, which gives them a full cycle period (wipe up and then back down again) of about 1.5 seconds. The typical error vector patterns from the non-boosted and boosted transmissions are shown in Figures 4.12 and 4.13 respectively. The first one is spanned over 7000 bits, while the second is over 10 6 bits. The BER of the experiment without boosted transmission was and with boosted transmission it was Figure 4.14 shows the distributions. Again, there is the sharp increase in the burst length frequency between 8 and 9 bits. An interesting point is that the cluster length fre- 58

78 Section 4.4: Experimental Results Chapter 4 quency distributions are almost exactly the same, whether or not the signal is boosted. We also observe a sharp step in the burst-interval length frequency at just under bits for the boosted transmission. At a baud rate of , this corresponds to the 1.5 second cycle period of the windscreen wipers. This is evidence towards the preposition that the wiper pulls extra current and causes extra noise when its motors/servos start to pull the wipers upwards to start the cycle. Figure 4.12: Typical Error Vector Pattern for Transmission with Windscreen Wipers On - Engine Off Figure 4.13: Typical Error Vector Pattern for Transmission with Windscreen Wipers On - Engine Off - Boosted Transmission 59

79 Section 4.4: Experimental Results Chapter 4 Figure 4.14: Error Distribution Statistics for the In-Vehicle Power Line Experiments - Windscreen Wipers On - Engine Off Engine On With the wipers as well as the engine on, the BER for the non-boosted transmission was ; lower than when the engine was off, which shows a similar situation to the experiment with the cell phone charger plugged in. Just as interestingly, the BER with boosted transmission also decreased; from with the engine off to with the engine on. Typical error vector patterns for the non-boosted and boosted conditions are shown in Figures 4.15 and 4.16, respectively. The distributions are shown in Figure Again, in the burst-interval length frequency distribution for the boosted signal, we observe the step at just under bits. However, the fact that the step is not as high, and that the BER is less, the engine being on seems to cause some sort of noise suppression to an extent. This is discussed later in the discussion section below. 60

80 Section 4.4: Experimental Results Chapter 4 Figure 4.15: Typical Error Vector Pattern for Transmission with Windscreen Wipers On - Engine On Figure 4.16: Typical Error Vector Pattern for Transmission with Windscreen Wipers On - Engine On - Boosted Transmission 61

81 Section 4.5: Results Summary and Discussion Chapter 4 Figure 4.17: Error Distribution Statistics for the In-Vehicle Power Line Experiments - Windscreen Wipers On - Engine On 4.5 Results Summary and Discussion Table 4.1 shows all the measured bit error rates for all the performed experiments. Probably the most interesting result from this table is the decrease in BER when turning the engine on in some of the experiments. A possible explanation is that when the engine is on, the alternator is a source that will cause the transceivers to see a lower impedance on the line, which would help decrease the amount of attenuation of the transmission. All of the burst length frequency graphs show a step in the 8-9 bit region. This indicates that the channel (or the transceivers themselves) has a general characteristic of this length bursts of noise/attenuation. The graphs show that the bursts and clusters do not last much longer than these 8-9 bits, which is an important characteristic of the channel as it indicates impulsive noise. Another interesting characteristic is that the burst-interval lengths for the non-boosted 62

82 Section 4.6: Conclusion Chapter 4 Operating Condition No Devices On Cell Phone Charger Windscreen Wipers Engine Off Engine Off - Boosted Engine On Engine On - Boosted Table 4.1: Summary of BERs for Experimental Results transmission experiments for when the cell phone charger is plugged in also have large steps around the 8-9 bit region. This means that the burst noise seems to be on for 442 µs and then off for the same amount of time; giving a frequency of about 1131 Hz. These kinds of bursts can be considered the result periodic impulsive noise. The source of this noise must be due to something in the charger itself. 4.6 Conclusion In this chapter we present the results from multiple experiments for transmission over the vehicle s battery bus under different conditions. The experiments were performed using the Picoscope Sampling method developed in the previous chapter. The results produced by the experiments were discussed in the last section of this chapter which also points out some interesting characteristics of the channel. 63

83 CHAPTER 5 Markov Modelling Research is what I m doing when I don t know what I m doing. - Wernher von Braun 5.1 Introduction Chapter 3 presents the development of the experimental method. In this chapter, we investigate the development of methods to create Markov models based on the experimental results presented in Chapter 3. We first parametrise the classical models presented in Chapter 2 and then move on to propose two new models as well as parametrise them. An analysis of each model and comparisons between them is also provided. A number of tools is used to visually and numerically compare the simulation outputs of each model with the experimental results; including the computed expected BER (as described in Chapter 2 under the Transition Matrix subsection), the actual BER of simulations, gap distributions, error distributions, block error probabilities and other comparison mechanisms, all of which are described within the chapter. Using the experience and methods developed throughout this chapter, we can then go on to model the channel for the different operating conditions presented in Chapter Gilbert Markov Model The Gilbert Model is the quickest and easiest model to parametrise. Following the method and equations presented in Chapter 2 and the experimental data we get the 64

84 Section 5.3: Fritchman Markov Model Chapter 5 following parameters: q = h = P = (5.1) Statistically the expected BER of this model would be given by the value of P (1), which is equal to the BER rate of the experimental data. One hundred simulations were run and the error free bit run distribution was collected for each simulation and the average of this data was calculated. Figure 5.1 shows a typical simulation run based on the Gilbert model and by inspection it looks very similar to the experimental results. The average distribution of error free bit runs as well as error bit runs of the simulations as compared with that of the experiment are shown under the comparison section at the end of this chapter. The average BER over these 100 simulations was Figure 5.1: Typical Simulation run using Gilbert s Model 65

85 Section 5.3: Fritchman Markov Model Chapter Fritchman Markov Model We use the method presented in Chapter 2 to parametrise the Fritchman model. It is common to convert the error free run distribution into its logarithmic scale to make the curve fitting easier and more accurate. Because of this our curve will be of the following form (as done in [33]): log (P (0 m 1)) = log ( = log ( k i=1 P N,i P i,i (P i,i ) m ) k α i βi m ). (5.2) i=1 The algorithm used for curve fitting was an algorithm provided by Matlab s curve fitting toolbox called Trust-Region, which is a least square optimization technique [46]. There is no definitive way of choosing the number of terms to best fit the curve, hence Matlab was used to fit a number of curves where each had a different number of parameters to see which one fits the best. The R-squared values of the fits was used to determine which fit was best. Refer to Appendix C for an explanation regarding the R-squared definition and calculation. The best fit was found to be the curve with 7 terms (with an R-squared value of ), which gives us N = 8 (since the number of terms in the curve is equal to the number of good states). The parameters of the curve are: α 1 = β 1 = α 2 = β 2 = α 3 = β 3 = α 4 = β 4 = α 5 = β 5 = α 6 = β 6 = α 7 = β 7 = (5.3) 66

86 Section 5.3: Fritchman Markov Model Chapter 5 Figure 5.2 shows the curve of the form presented in equation 5.2 with the parameters presented by equation 5.3. Figure 5.2: Experimental Error Free Run Distribution with the Fitted Curve Using these parameters, we create the transition matrix for the model: (5.4) The model parameters were rounded to fit the matrix on the page. With this transition matrix we can compute the expected BER by using the method described in Chapter 2. This is done using Matlab and is found to be

87 Section 5.4: Finite State Spreading Chain Chapter 5 As with the Gilbert model, we run 100 simulations and gather the average data for the distribution graphs which will be presented under the comparison section. The average bit error rate of these simulations was Finite State Spreading Chain This section introduces a new model to the BER analysis arena. Technically, any Markov model that has a finite number of states partitioned into non-error states and error states is considered a Fritchman model. The Fritchman model with the particular transition characteristics used in the previous section, however, is the most commonly used Fritchman model (and is thus commonly referred to as the Fritchman model ), and most other models with this same base concept but with different transition characteristics are referred to by other names. The Finite State Spreading chain model that we introduce in this section can thus be technically defined as a Fritchman model, but its transition characteristics makes it very different from the one used in the previous section. Model Theory This model is only similar to the common Fritchman model in that it will have k good states, one bad state and N = k + 1 states in total. The difference lies in the allowable transitions. Figure 5.3 shows how the model works. The model can only transition from one good state to the next good state, or to the bad state. Once in the bad state, the channel can either stay in the bad state or transition to the first good state (not any of the other good states). When the channel reaches the last good state, it can only stay in that state or transition to the bad state. This model has not yet been presented in current literature but has a very similar transition structure to that of the infinite-state model presented in [47] and [48], which is referred to as the Slowly Spreading Chain of the first kind. Hence the name Finite State Spreading Chain is coined for the model proposed in this section. 68

88 Section 5.4: Finite State Spreading Chain Chapter 5 Figure 5.3: Transition Diagram for the Finite State Spreading Chain Model The main feature of this model is that the probability of the channel going into the bad state will become less as the channel transitions through each good state. The model can be considered a dynamic Gilbert model in that the probability of an error occurring changes after almost every time step. To get the parameters of this model, we plot P (0 m 1 1) with respect to m. We then fit an appropriate curve to the data which is usually of the general formula ae bm + c. The expression P (0 m 1 1) refers to the probability of the channel returning to the bad state after an error-free gap of length m. Using simple statistical reasoning, we arrive at the following equations: P ( ) = P N,N, (5.5) P ( ) = P N,1 P 1,N, (5.6) P (0 m 1 1) = P N,1 P m,n m P i 1,i i=2 (m = [2, 3, 4,..., k 1]), (5.7) 69

89 Section 5.4: Finite State Spreading Chain Chapter 5 P (0 m 1 1) = P N,1 P k,n (P k,k ) m k k i=2 P i 1,i (m = [k, k + 1, k + 2,...)). (5.8) The probabilities would be given by the curve that was given by the fitting of the gap length probability data (i.e. P (0 m 1 1) = ae bm + c (m = [0, 1, 2,...)). These equations enable us to calculate the parameters of the model and easily select the number of states to have. Below are the equations to calculate the parameters for each state: P N,N = P ( ) = ae b0 + c = a + c, (5.9) P N,1 = 1 P N,N. (5.10) We define ψ(j): P N,1 if j = 1 ψ(j) = j P N,1 P i 1,i if j > 1, i=2 (5.11) P j,n = P (0j 1 1) ψ(j) = ae bj + c, (5.12) ψ(j) P j,j+1 = 1 P j,n, (5.13) 70

90 Section 5.4: Finite State Spreading Chain Chapter 5 P k,k = 1 P k,n. (5.14) For the bad state (state N) we use equations 5.9 and For states 1 through to k 1 we use equations 5.11, 5.12 and For the last good state (state k) we use equation 5.14 in place of equation These equations are easy to implement in an algorithm in a programming language such as Matlab and determining the parameters for the same model but with more states is a simple case of performing the algorithm as many times as there are states. This is very simple compared to the previous Fritchman model where one would have to refit a curve of more complexity and recalculate all the parameters to remodel the channel with more states. Model Parametrising and Simulation As described above, to get the parameters we must first plot P (0 m 1 1) and fit a curve to it. After a number of different fits, it was found a formula of ae b + c produced the best results. Figure 5.4 shows the experimental data and the fit. The equation for the fit is: e x Figure 5.4: Experimental Data for P (0 m 1 1) and the Fitted Curve Using this fitted curve and equations above, we can now create the model to any degree 71

91 Section 5.4: Finite State Spreading Chain Chapter 5 of states. Using Matlab, this was done for N = 3 up to N = 50 and the expected BER (as calculated according the method presented in Chapter 2) for each transition matrix was calculated. Figure 5.5 plots the expected BER of the transition matrix against the size of the matrix. As can be seen, the expected BER settles at a final value which is about , which is very close to the experimental BER of Figure 5.5: Expected BERs of the Different Size Transition Matrices The best matrix then would be the one that has the fewest number of states that has this settling value which is found to be N = 27. Due to space constraints, we do not present the whole matrix, but present enough parameters to fully define it: P 1,2 = P 2,3 = P 3,4 = P 4,5 = P 5,6 = P 6,7 = P 7,8 = P 8,9 = P 9,10 = P 10,11 = P 11,12 = P 12,13 = P 13,14 = P 14,15 = P 15,16 = P 16,17 = P 17,18 = P 18,19 = P 19,20 = P 20,21 = P 21,22 = P 22,23 = P 23,24 = P 24,25 = P 25,26 = P 26,26 = P 27,27 =

92 Section 5.5: Error Spreading Chain Chapter 5 As with the other models, we run 100 simulations with this model and store the average results for the comparison section. The average BER over these simulations was Error Spreading Chain Model Theory This next model is arguably a mirrored version of the model presented above, where instead of a chain of error free states and one error state, we have a chain of error states and one error free state. Figure 5.6 illustrates the transition diagram of this model. Figure 5.6: The Mirrored Spreading Chain Transition Diagram The method to parametrising this model is very similar to the method for the Finite State Spreading Chain model but with very important differences. The model is based on the error run distribution of the channel. I.e. P (1 m 0). We plot the experimental data again and as above, fit an exponential function to it. The following equations are based on the model: P (1 1 0) = P N,1, (5.15) 73

93 Section 5.5: Error Spreading Chain Chapter 5 P (1 m 0) = P N,1 m P i 1,i i=2 (m = [2, 3, 4,..., k 1]), (5.16) P (1 m 0) = P N,1 (P k,k ) m k k i=2 P i 1,i (m = [k, k + 1, k + 2,...)). (5.17) To extract the parameters we follow a very similar procedure: P N,1 = P (1 1 0) = ae 1, (5.18) P N,N = 1 P N,1. (5.19) We define ψ(j): P N,1 if j = 1 ψ(j) = j P N,1 P i 1,i if j > 1, i=2 (5.20) P j,j+1 = P (1j 0) ψ(j) = ae bj ψ(j), (5.21) 74

94 Section 5.5: Error Spreading Chain Chapter 5 P j,n = 1 P j,j+1, (5.22) P k,k = 1 P k,n. (5.23) Following the same procedure as the previous model but with these modified equations we can get the parameters for this model. Model Parametrising and Simulation We follow a very similar procedure as to the previous model to get the model parameters. We first fit a curve of the form ae bm to the experimental data of the error run distribution graph (P (1 m 0)). Figure 5.7: Fitted Curve for the Error Run Probability Distribution The equation of this curve is: e m. As was done with the previous model, we get the transition matrices for different state numbers (N = 3 up to N = 50) and compute and record each one s expected BER. It is important to remember that for this model there are k bad states and one good state when considering the transition 75

95 Section 5.6: Model Comparisons Chapter 5 matrix. This requires slight changes in the algorithms that calculate the expected BER as well as the simulation algorithms. Figure 5.8: Expected BER for Different Sized Transition Matrices We can see that the expected BER settles after a certain size, as with the previous model. This time the settling value is which is fairly close to the experimental BER. Again, we pick the matrix with the fewest number of states with this settling value. This is N = 13 states. As previously, we list the defining parameters of the matrix: P 1,2 = P 2,3 = P 3,4 = P 4,5 = P 5,6 = P 6,7 = P 7,8 = P 8,9 = P 9,10 = P 10,11 = P 11,12 = P 12,12 = P 13,13 = With these parameters we simulate the channel using this model 100 times and store the average results for the comparison section. The average BER for these simulations was

96 Section 5.6: Model Comparisons Chapter Model Comparisons After parametrising and simulating each model, we can now compare them against the experimental data. As a preliminary comparison tool, we can take a look at what the error vector patterns look like for a simulation run of each of the models and compare it with the experimental data by visual inspection. Figure 5.9 presents the experimental error vector pattern (as presented in the previous chapter) followed by typical error vector patterns from simulations from each of the models. From visual inspection, one can argue that the two best models are the Finite State Spreading Chain model and the Fritchman model. The most noticeable aspect of the figure, however, is the distribution of errors of the Error Spreading Chain model. It looks like it has a lot more errors, but the BER is not too much higher than the experimental BER. This simply means that the errors are not happening in bursts, but rather as single errors throughout the simulation. This shows how important it is that the model accommodates for the channel s error distribution characteristics and not just it s BER. Figures 5.10 and 5.11 shows the error free run distribution and the error run distribution of the different models as compared to the experimental results. Figure 5.12 shows the other distributions compared with the experimental results. These were all obtained by taking the average over 100 simulations of each model. By visual inspection it can be seen that the only model that does not match the experimental error free run distribution closely is that of the error spreading chain model (as expected from the preliminary comparison). This is probably because the model s transitions do not easily allow for the same kind of error bursts that the channel exhibits. The burst-interval and cluster length distributions are simulated well by all the models (except the error spreading chain model). The burst length distribution, however, was not represented by the simulations as well as the other distributions. This is a common issue when modelling different channels [43] [49]. However, the simulated distributions are still considered good fits to the experimental data. 77

97 Section 5.6: Model Comparisons Chapter 5 Table 5.1 compares the different models with different values. Below is the explanation of each heading in the table. Model Expected BER Simulated BER Average Length of Error Bursts R-Squared Value N Gilbert * Fritchman Finite State Spreading Chain Error Spreading Chain Experimental Values Table 5.1: Comparison Table for the Four Different Models Expected BER This is the expected BER calculated using the transition matrix as described in Chapter 2. With the exception of the Gilbert model, where there is no transition matrix. In this case the expected BER is equal to the experimental BER because the model is based directly on the experimental BER. The values under this heading should be compared with the experiment s value of Simulated BER This is the average BER over 100 simulations. The values under this heading should be compared with the experiment s value of Both the expected BER and 78

98 Section 5.6: Model Comparisons Chapter 5 simulated BER are presented because the expected BER statistically tells us what the BER would be if the channel transmitted indefinitely. However, it is still interesting to take note of the average BERs of actual simulations Average Length of Error Bursts This is the average length of the error bursts of the simulations. This is important as it also helps characterise the nature of the errors. The values under this heading should be compared with the experiment s value of R-Squared Value This is the R-squared value (as described in Appendix C) of the averaged simulated error free run distribution data to the experimental data (i.e. how well the simulated curves fit the experimental curve in Figure 5.10). One will notice a negative value for the error chain model, which indicates its data does not follow the trend of the experimental data Number of States (N) This is the number of states N the model consists of. *It is important to keep in mind the structure of the Gilbert model when considering this aspect of the model because it may have 2 states but each state represents a binary symmetric channel (as described in Chapter 2) Block Error Probabilities A final method of comparison will be to observe the block error probabilities. The block error probabilities is the probabilities of finding a number of error bits in a block of bits with a particular length. We use the notation described in [43] which is: P (m, n). P (m, n) is the probability that there will be m errors in a random block of n bits in the 79

99 Section 5.6: Model Comparisons Chapter 5 transmission. We do this for block lengths 8, 16 and 32. Figures 5.13, 5.14 and 5.15 are bar graphs showing the comparisons of each simulation with the experimental data respectively. These bar graphs show how well the models represent the model in terms of block error probabilities. It can be seen visually that the finite state model contains probabilities for higher number of errors in the larger blocks than the other models and the experimental data. However, using the R-Squared value to determine the closeness of the simulation data sets to the experimental data set, it is determined that the Finite State model provides the closest resemblance to the channel for all three block sizes. 80

100 Section 5.6: Model Comparisons Chapter 5 Figure 5.9: Experimental Error Vector Pattern and Typical Simulation Run Error Vector Patterns 81

101 Section 5.6: Model Comparisons Chapter 5 Figure 5.10: Error Free Run Distribution Graph of the Different Models and the Experiment Figure 5.11: Error Run Distribution Graph of the Different Models and the Experiment 82

102 Section 5.6: Model Comparisons Chapter 5 Figure 5.12: Distributions of the Different Models and the Experiment Figure 5.13: Block Error Probabilities for Blocks of Length 8 83

103 Section 5.6: Model Comparisons Chapter 5 Figure 5.14: Block Error Probabilities for Blocks of Length 16 Figure 5.15: Block Error Probabilities for Blocks of Length 32 84

104 Section 5.7: Choosing the Best Model Chapter Choosing the Best Model This section uses the comparisons presented in the previous section to decide which would be the best model to represent the channel Error Spreading Chain Model Firstly, we can rule out the Error Spreading Chain model due to its error distribution clearly not matching that of the channel. This model demonstrates the importance of keeping in mind the characteristics of the channel s error events when creating a model. This model may be used more successfully to model channels of a different kind, perhaps where the errors are more evenly distributed throughout the transmission Gilbert Model Next, we can argue that the Gilbert model is not as good as the Finite State Spreading Chain and Fritchman models because its distribution is not quite as close to that of the channel. The model is directly built on three experimental values: the channel s BER, the probability of an error occurring directly after another error and the probability of there being an error between two other errors. This makes the model simulate the BER as well as the short term burst characteristics but does not accommodate for error bursts with longer error free gaps within the bursts Fritchman and Finite State Spreading Chain Models Both of these models simulate the channel very well. The Finite State Spreading Chain model has a slightly better chance of producing simulations with the same BER as the channel s, as well as the same block error probabilities. The main difference lies in the average length of the clusters of errors. This is a very important aspect that contributes to the characterisation of the models and the actual channel. The Fritchman model produced an average error length error of , which is relatively higher than that of the experiment (1.6887). Whereas the Finite State Spreading Chain model produced a 85

105 Section 5.8: Models for the Channel Under Different Conditions Chapter 5 closer value to that of the experiment of In this way, the Finite State Spreading Chain model is a better model. The number of states is also important as this would directly affect the time it takes to simulate the channel. This is where the Fritchman model is better The Best Model The Gilbert model suits the experimental data better when it comes to the BER but does not simulate the actual error and gap distribution as well as the Fritchman and Spreading Chain models do. The Fritchman model simulates the error and gap distributions closely to that of the experimental data, but does not yield BERs as close as the Gilbert and Spreading Chain models do. The Finite State Spreading Chain model simulates both these two very important aspects (as well as all the other aspects discussed) of the experimental data very well, and is a fairly easier model to create, and thus can be concluded as the best model for the channel. The model only faults with its large number of states, which would make any computations based on the model more complex. 5.8 Models for the Channel Under Different Conditions Throughout this chapter we have investigated and developed the methods for producing Markov models based on experimental results. Now that we have developed a method, and decided upon the best model, we can move on to model the channel under the different test conditions presented in Chapter 4. We will only model the non-boosted transmission experiments, with them being the more noisy and more interesting results. We follow the same procedure from Section 5.4 to create Finite State Spreading Chain models for each set of the experimental results. The results of this exercise are shown below. Note that each condition had a different number of optimal states. 86

106 Section 5.8: Models for the Channel Under Different Conditions Chapter Cell Phone Charger Plugged In - Engine Off The defining probability matrix elements for the Finite State Spreading Chain model are: P 1,2 = P 2,3 = P 3,4 = P 4,5 = P 5,6 = P 6,7 = P 7,8 = P 8,9 = P 9,10 = P 10,11 = P 11,12 = P 12,13 = P 13,13 = The model was simulated for bits and the distributions are presented with the experimental distributions in Figure Figure 5.16: Distribution Graphs for In-Vehicle Experiment (Solid Line) and Corresponding Simulation (Dashed Line) for Cell Phone Charger Plugged In with Engine Off 87

107 Section 5.8: Models for the Channel Under Different Conditions Chapter Windscreen Wipers On - Engine Off The defining probability matrix elements for the Finite State Spreading Chain model are: P 1,2 = P 2,3 = P 3,4 = P 4,5 = P 5,6 = P 6,7 = P 7,8 = P 8,9 = P 9,10 = P 10,11 = P 11,12 = P 12,13 = P 13,14 = P 14,15 = P 15,16 = P 16,17 = P 17,18 = P 18,19 = P 19,20 = P 20,21 = P 21,22 = P 22,23 = P 23,24 = P 24,25 = P 25,26 = P 26,27 = P 27,28 = P 28,29 = P 29,30 = P 30,31 = P 31,32 = P 32,33 = P 33,34 = P 34,35 = P 35,36 = P 36,37 = P 37,38 = P 38,38 = P 39,39 = The model was simulated for bits and the distributions are presented with the experimental distributions in Figure

108 Section 5.8: Models for the Channel Under Different Conditions Chapter 5 Figure 5.17: Distribution Graphs for In-Vehicle Experiment (Solid Line) and Corresponding Simulation (Dashed Line) for Windscreen Wipers On with Engine Off No Devices On - Engine On The defining probability matrix elements for the Finite State Spreading Chain model are: P 1,2 = P 2,3 = P 3,4 = P 4,5 = P 5,6 = P 6,7 = P 7,8 = P 8,9 = P 9,10 = P 10,11 = P 11,12 = P 12,13 = P 13,14 = P 14,15 = P 15,16 = P 16,17 = P 17,18 = P 18,19 = P 19,20 = P 20,21 = P 21,22 = P 22,23 = P 23,23 = P 24,24 = The model was simulated for bits and the distributions are presented with the experimental distributions in Figure

109 Section 5.8: Models for the Channel Under Different Conditions Chapter 5 Figure 5.18: Distribution Graphs for In-Vehicle Experiment (Solid Line) and Corresponding Simulation (Dashed Line) for No Devices On with Engine On Cell Phone Charger Plugged In - Engine On The defining probability matrix elements for the Finite State Spreading Chain model are: P 1,2 = P 2,3 = P 3,4 = P 4,5 = P 5,6 = P 6,7 = P 7,8 = P 8,9 = P 9,10 = P 10,11 = P 11,12 = P 12,13 = P 13,14 = P 14,15 = P 15,16 = P 16,17 = P 17,18 = P 18,19 = P 19,20 = P 20,21 = P 21,22 = P 22,23 = P 23,24 = P 24,25 = P 25,26 = P 26,27 = P 27,28 = P 28,29 = P 29,30 = P 30,31 = P 31,32 = P 32,33 = P 33,33 = P 34,34 =

110 Section 5.8: Models for the Channel Under Different Conditions Chapter 5 The model was simulated for bits and the distributions are presented with the experimental distributions in Figure Figure 5.19: Distribution Graphs for In-Vehicle Experiment (Solid Line) and Corresponding Simulation (Dashed Line) for Cell Phone Charger Plugged In with Engine On 91

111 Section 5.8: Models for the Channel Under Different Conditions Chapter Windscreen Wipers On - Engine On The defining probability matrix elements for the Finite State Spreading Chain model are: P 1,2 = P 2,3 = P 3,4 = P 4,5 = P 5,6 = P 6,7 = P 7,8 = P 8,9 = P 9,10 = P 10,11 = P 11,12 = P 12,13 = P 13,14 = P 14,15 = P 15,16 = P 16,17 = P 17,18 = P 18,19 = P 19,20 = P 20,21 = P 21,22 = P 22,23 = P 23,24 = P 24,25 = P 25,26 = P 26,27 = P 27,28 = P 28,29 = P 29,30 = P 30,30 = P 31,31 = The model was simulated for bits and the distributions are presented with the experimental distributions in Figure

112 Section 5.8: Models for the Channel Under Different Conditions Chapter 5 Figure 5.20: Distribution Graphs for In-Vehicle Experiment (Solid Line) and Corresponding Simulation (Dashed Line) for Windscreen Wipers On with Engine On 93

113 Section 5.9: Conclusion Chapter Modelling Summary Table 5.2 summarises the experimental BERs of the experiments performed in Chapter 4, and the simulated model BERs for the models created above. It also presents the number of states each model has. Operating Condition Experimental BER Simulated BER N No Devices On - Engine Off Cell Phone Charger Plugged In - Engine Off Windscreen Wipers On - Engine Off No Devices On - Engine On Cell Phone Charger Plugged In - Engine On Windscreen Wipers On - Engine On Table 5.2: Summary of Models Based on In-Vehicle Experimental Results It is interesting to note that the simulations are not able to replicate the sharp steps in the distributions. These simulated distributions, however, can be considered good enough fits, as they are of similar standard to previous work done in [42], [43] and [49]. 5.9 Conclusion This chapter presented the channel s initial experimental results (with no devices on and the engine switched off) as modelled by two classic Markov Models (Gilbert and 94

114 Section 5.9: Conclusion Chapter 5 Fritchman) and two newly proposed models (Finite State Spreading Chain and Error Spreading Chain). After comparing simulations of all the models with each other and the experimental results using several different comparison aspects, it is clear that both the Finite State Spreading Chain model as well as the Fritchman model resemble the channel very closely. The Finite State Spreading Chain model did so slightly better, and is much easier to parametrise. This model comes at the cost of having more states. Using this conclusion from the comparisons of the different models, we went on to model the channel under the non-boosted conditions of the experiments performed in Chapter 4. 95

115 CHAPTER 6 Channel Error Coding But slight mistakes accumulate, and grow to gross errors if unchecked. - Jacqueline Carey 6.1 Introduction Chapter 5 provides us with Markov models for the channel under different conditions. As a complementary exercise to this dissertation to show how these models can be used to analyse coding schemes, we now use these models to analyse the effectiveness of simple error detection techniques employed by the common LIN in-vehicle communication protocol (a description of the protocol is provided in Chapter 2). We investigate the error detection techniques specifically from the LIN protocol because the SIG60 transceivers are specifically designed to be LIN compatible and most in-vehicle devices will have built-in LIN functionality. We will then go on to investigate some error correction coding that we can implement on top of the LIN protocol requirements. Using the results from these exercises, one can determine if it is worth it to rather implement error control coding rather than having the extra components required for boosted transmission. 6.2 LIN Error Detection Error detection is the simple process of detecting an error in the received transmission by adding redundant bits that can be used to check the other bits. There are a number 96

116 Section 6.2: LIN Error Detection Chapter 6 of ways of doing this and we will do it by using the same techniques described in the LIN protocol. If an error is detected, then the received message can be discarded and then either re-transmission must be requested or the message is simply ignored, depending on the network design Single Bit Parity Check We first take a look at a simple single parity check bit (as a build up to the 2 parity bit check used by the LIN protocol). A single parity bit is a popular way to detect errors mainly in mostly error-free environments where the probability of two errors occurring in one byte is very small, and we only expect one error now and then. To stay to the theme of 8-bit bytes (which is what the LIN protocol makes use of), we will investigate the case where there are 7 data bits and 1 parity bit. This gives us an information rate of 7 /8 (see Chapter 2 for description of information rate). A single parity bit is only able to detect an odd number of errors (since only an odd number of errors will change the parity), as mentioned in Chapter 2. If an error is detected then retransmission is requested or the message is ignored (depending on the network setup) and no error is passed on to the application. So if we know the probability of an even number of errors occurring in a block of 8 bits, then we know the probability of an error being passed on to the application. This is the basis of the algorithm used to calculate the effective bit error rate (i.e. the number of bits passed onto the application). Table 6.1 presents the results of this exercise. The first column shows us the initial BER (with no parity checks) of the simulation. The second column ( Effective BER ) is the BER after the erroneous bytes detected by the parity bit are taken away. This is calculated by taking the number of errors in the undetected bytes divided by the number of received bits. The number of received bits is the number of transmitted bits minus the number of bits discarded due to the detected erroneous bytes. The next 97

117 Section 6.2: LIN Error Detection Chapter 6 column shows the BER of the experimented with boosted transmission, for comparison. The last column is the decrease in order of magnitude of the BER (i.e. the logarithm of the initial BER divided by the effective BER). Simulation Model BER Effective BER Boosted BER BER Order of Magnitude Decrease No Device On - Engine Off Cell Phone Charger - Engine Off Windscreen Wipers On - Engine Off No Devices On - Engine On Cell Phone Charger - Engine On Windscreen Wipers On - Engine On Table 6.1: Bit Error Rate Decrease for 1-bit Parity Check This table shows that we can half the bit error rate in most cases by just converting one of the data bits into a parity check bit (the downfall, of course, being that we lose out on one bit of useful information for this parity check). We cannot however, reach the BER performance of the boosted signals with a single parity bit check. In the next section we see how using a second parity bit greatly improves these results. 98

118 Section 6.2: LIN Error Detection Chapter Two Bit Parity Check The first error detection mechanism in the popular in-vehicle LIN protocol is the 2 parity bits in the identifier field, which is 1 byte long. These parity bits protect the identifier byte, which contains the addressing and data length of the expected communication. The construction of the identifier byte, as well as how the parity bits are calculated are explained in Chapter 2. Having two parity bits creates more of a chance to detect errors (and thus less errors are passed through to the application), but at the cost of changing a data bit to a parity bit to have only six data bits instead of seven. This gives us an information rate of 6 /8 as opposed to 7 /8. Calculating the probability of errors going through to the application is a bit more complicated when two parity bits are involved, compared to the single parity bit (which is just a case of counting how many bytes have an even or odd number of errors). The easiest way to do it is to find out exactly what condition is required for error(s) to occur with both parity bits still satisfied by their equations shown in Chapter 2, by equation set 6.5. It turns out we can do this by modulo 2 summation and multiplication using the following formula: C = (P 0 D 0 D 1 D 2 D 4 ) (P 1 D 1 D 3 D 4 D 5 ) (6.1) C in equation 6.1 represents a value which will indicate if the byte will go through to the application or not. If C is 1, the byte is accepted. If C is 0, errors are detected and the byte is rejected. Thus, to calculate the BER of the bits being passed to the application, we take a look at 8-bit blocks in the error vector and calculate C for each block and we count the number of errors that are in the blocks that produce a C value of 1. This will tell us the number of errors that have passed through to the application. It is very important to note that this calculation for C is created on the basis that we 99

119 Section 6.2: LIN Error Detection Chapter 6 are now dealing with bit error vectors and not the bit values themselves. I.e. if a bit is not an error we put a 0, and if it is an error we put a 1 (just like the bit error vectors presented in the previous chapters). This is also advantageous because our simulations output the error vector anyway, where 0 s and 1 s represent non error bits and error bits respectively. From this equation, we can see that only very particular combinations of errors will be passed through to the application. For example, if D 0 and D 1 are errors we get: C = (P 0 D 0 D 1 D 2 D 4 ) (P 1 D 1 D 3 D 4 D 5 ) = ( ) ( ) = (0) (1) = 1 0 = 0 (6.2) Example 6.2 shows us that not all blocks with an even number of errors are passed through to the application. Conversely, not all blocks with an odd number of errors are detected as erroneous. For example, if D 1, D 2 and P 1 are errors: C = (P 0 D 0 D 1 D 2 D 4 ) (P 1 D 1 D 3 D 4 D 5 ) = ( ) ( ) = (0) (0) = 1 1 = 1 (6.3) As done with the single parity check, Table 6.2 shows how the BER improves with 2 parity bits. From these results we see a substantial improvement in the effective BER (number of errors passed through to the application). More than half an order of magnitude 100

120 Section 6.2: LIN Error Detection Chapter 6 Simulation Model BER Effective BER Boosted BER BER Order of Magnitude Decrease No Device On - Engine Off Cell Phone Charger - Engine Off Windscreen Wipers On - Engine Off No Devices On - Engine On Cell Phone Charger - Engine On Windscreen Wipers On - Engine On Table 6.2: Bit Error Rate Decrease for 2-bit Parity Check improvement in BER for all the simulations except one. An interesting result is that the improvement is similar for the different conditions, except when the line is too noisy (with the cell phone charger plugged in and engine off). The small sacrifice in information rate can be considered well worth the fairly large decrease in BER for a 2-bit parity check Checksum The next error detection mechanism in the LIN protocol is the checksum that is calculated on the data bytes. As described in Chapter 2, it is the inverted sum of all the data 101

121 Section 6.2: LIN Error Detection Chapter 6 bytes in modulo 256. The number of data bytes in a transmission, as specified by the protocol, is 2, 4 or 8. This means an information rate of 16 /24, 32 /40 or 64 /72 respectively. This is much more difficult to evaluate than the parity bits because the combinations of errors that are passed through to the application are dependent on what the bit values actually are, and cannot be assessed by using the error vector alone. We thus create a Matlab algorithm to do the following: ˆ Simulate the channel using the appropriate model. This will output a simulated error vector. ˆ Create a random sequence of bits (with equal probability of 1 s and 0 s) that has the required length for the number of data bytes (e.g. 16 bits if 2 data bytes are to be sent). ˆ Compute the necessary checksum byte on the randomly generated data bytes and add it to the block (e.g. for 2 data bytes, we would now have blocks of 3 bytes). ˆ Do this for as many blocks as there are in the simulated error vector, so that the sequence of data bytes with their corresponding checksum bytes has an equal amount of bits as the simulated error vector. ˆ For every 1 in the simulated error vector, we invert the corresponding bit in the sequence. ˆ Now check if the checksum for each block still corresponds to their respective data bytes. ˆ If the checksum passes the check, record the number of errors in that block (if any). Using this method, we are able to obtain the effective BER for each case. Tables 6.3, 6.4 and 6.5 present these results. From these tables we can deduce that having the checksum check 2 data bytes has a slightly better effect on the BER than a 2 bit parity check. There is a very large improvement when the checksum checks 4 and 8 data bytes (with the exception of the 102

122 Section 6.2: LIN Error Detection Chapter 6 Simulation Model BER Effective BER Boosted BER BER Order of Magnitude Decrease No Device On - Engine Off Cell Phone Charger - Engine Off Windscreen Wipers On - Engine Off No Devices On - Engine On Cell Phone Charger - Engine On Windscreen Wipers On - Engine On Table 6.3: Bit Error Rate Decrease for Checksum with 2 Data Bytes cell phone charger with engine off simulation), however, and is even more appealing when one considers that there are more data bits being checked by the same amount of check bits. The drawback here is that if there is an error detected then the whole block has to be re-sent, possibly resulting in slower communication. With the checksum byte method we are able to get more than an order of magnitude improvement. However, we are still unable to able to match the BER performance of the boosted signal. 103

123 Section 6.2: LIN Error Detection Chapter 6 Simulation Model BER Effective BER Boosted BER BER Order of Magnitude Decrease No Device On - Engine Off Cell Phone Charger - Engine Off Windscreen Wipers On - Engine Off No Devices On - Engine On Cell Phone Charger - Engine On Windscreen Wipers On - Engine On Table 6.4: Bit Error Rate Decrease for Checksum with 4 Data Bytes 104

124 Section 6.2: LIN Error Detection Chapter 6 Simulation Model BER Effective BER Boosted BER BER Order of Magnitude Decrease No Device On - Engine Off Cell Phone Charger - Engine Off Windscreen Wipers On - Engine Off No Devices On - Engine On Cell Phone Charger - Engine On Windscreen Wipers On - Engine On Table 6.5: Bit Error Rate Decrease for Checksum with 8 Data Bytes 105

125 Section 6.3: Error Correction Chapter Error Correction The performance of sifting out errors offered by the error detection techniques from the LIN protocol may not be strong enough for some in-vehicle communication network applications. Hence, we now evaluate the use of simple error correction techniques to help decrease the number of errors passed on to the application layer of the network. To stay within the limits of the LIN protocol, we will evaluate error correction techniques that can be implemented on top of the checksum byte error detection. To maximise the decrease in effective BER to the application as well as to keep the information rate as high as possible, we will make use of the checksum over 8 bytes. Thus, we will limit the constructions of the error correcting codes such that we will always have 8 bytes (64 bits) that the checksum will be calculated on, so that we still comply with the protocol requirements. The number of actual source data bits will depend on the coding scheme. Figure 6.1 illustrates this concept, where k in this figure is the number of actual information bits. The algorithm used in MATLAB to simulate the effects of these error correcting codes on the BER performance is similar to that of the one used for the checksum byte: ˆ Simulate the channel using the appropriate model. This will output a simulated error vector. ˆ Create a random sequence of k bits (with equal probability of 1 s and 0 s) to be used as the source data. ˆ Encode the source data such that the length of the sequence is now 64 bits (8 bytes). ˆ Compute the necessary checksum byte on the coded sequence. ˆ Do this for as many blocks as there are in the simulated error vector, so that the sequence of encoded bytes with their corresponding checksum bytes has an equal amount of bits as the simulated error vector. ˆ For every 1 in the simulated error vector, we invert the corresponding bit in the sequence. 106

126 Section 6.3: Error Correction Chapter 6 Figure 6.1: Coding Scheme Overview Diagram ˆ Now check if the checksum for each block still corresponds to their respective encoded bytes. ˆ If the checksum passes the check, decode the rest of the sequence according to the error correction coding scheme used to encode it. ˆ Compare the decoded data with the original data and count the number of errors Repetition Coding Often, the very first error correcting technique used to help explain error correction is the method of simply repeating the data. The most common example is to repeat the message 3 times. If there is an error in one of the bits, then one can use the other two decide that there was an error. Due to the constraint on the number of bytes, which is 8, we will repeat 2 bytes of data 4 times. This way, we can correct up to 1 error, but still detect if there are 2 errors. 107

127 Section 6.3: Error Correction Chapter 6 Although this is a very robust and simple error correction technique, it is also an extremely redundant one. Encoding: Encoding is very obvious and simple. We have 2 bytes of source data (16 bits), and we repeat it 4 times to get our 8 byte coded sequence, on which we can calculate our checksum byte. The information rate is then 16 /72. Decoding: Decoding is a simple case of checking each bit s repetition. If there are at least 3 of them of the same value, then that value is the decoded bit. If 2 of them have one value, and 2 of them another, then an error has been detected and the message must be discarded. BER Performance: Table 6.6 shows how this error correction scheme on top of the checksum error detection decreased the amount of errors going through to the application. As we can see from the table, the repetitive coding scheme is very effective, at the cost of a very high redundancy. We are able to eliminate (through detection and correction) all errors in some cases. In other cases we are at least able to reach the BER performance of the boosted case (with exception of the case with the engine off and cell phone charger plugged in). 108

128 Section 6.3: Error Correction Chapter 6 Simulation Model BER Effective BER Boosted BER BER Order of Magnitude Decrease No Device On - Engine Off Cell Phone Charger - Engine Off Windscreen Wipers On - Engine Off No Devices On - Engine On Cell Phone Charger - Engine On Windscreen Wipers On - Engine On Table 6.6: Bit Error Rate Decrease for Repetition Coding on Top of Checksum 109

129 Section 6.3: Error Correction Chapter Rectangular Coding Rectangular coding is the next simplest technique in error correction coding. This technique involves taking the information and arranging it in a n 1 by m 1 rectangle. Then, for each row of the rectangle you calculate the even parity and add it to the end of the row. Then, you do the same for each column so that you end up with an n by m rectangle of bits which is the encoded sequence. Figure 6.2 illustrates how the encoded rectangle is constructed. In this figure, the o s represent source data bits and the x s represent the check bits. If an error occurs in position (i, j) then the parity check in row i would be incorrect and as well as that for column j. Thus, we are able to locate and correct the error. o o... o x o o... o x o o... o x x x... x x (6.4) Figure 6.2: Rectangular Code Illustration Encoding: Since we would like to land up with an 8 byte encoded sequence, we will use 7 lines of information bits, each with 7 bits. Then, as described above, we create an 8 by 8 square by adding the parity row and column. This gives us the 64 bits that we can calculate our checksum byte on. This gives us an information rate of 49 /72. Decoding: To decode the sequence, we extract the 7 by 7 block of information bits. We then re-encode it and then compare this newly created 8 by 8 block with the received 8 by 8 block. This comparison will indicate which parity bits do not add up. Then, we can 110

130 Section 6.3: Error Correction Chapter 6 correct the necessary errors. BER Performance: Table 6.7 shows how this error correction scheme on top of the checksum error detection decreased the amount of errors going through to the application. Simulation Model BER Effective BER Boosted BER BER Order of Magnitude Decrease No Device On - Engine Off Cell Phone Charger - Engine Off Windscreen Wipers On - Engine Off No Devices On - Engine On Cell Phone Charger - Engine On Windscreen Wipers On - Engine On Table 6.7: Bit Error Rate Decrease for Rectangular Coding on Top of Checksum Despite having a much higher information rate than the repetition coding, the rectangular code still performs fairly well. We are able to match the BER performance of the boosted cases under most conditions. 111

131 Section 6.3: Error Correction Chapter 6 The cell phone charger plugged in with the engine off simulation, however, gave us an effective BER that is higher than the initial simulated BER (without error control coding). This means that the bits that were passed to the application by the checksum and the error correction coding contained more errors per bit than that of the transmitted bits. This shows an extremely poor performance of this error control method for this situation. 112

132 Section 6.3: Error Correction Chapter Hamming Coding Hamming coding also involves adding parity check bits in such a way that errors can be corrected. The simplest form of Hamming coding is encoding 4 bits of data into a 7 bit encoded word. The 7 bit encoded words are constructed as follows [37]: P 1 P 2 D 1 P 3 D 2 D 3 D 4 The parity bits are calculated as follows: P 1 = D 1 D 2 D 4, P 2 = D 1 D 3 D 4 D 5, P 3 = D 2 D 3 D 4. (6.5) Notice that each parity bit is linearly independent of each other. This enables us to correct an error, or detect two errors (for more detail on the general construction and amount of error detection/correction for Hamming codes, refer to [37], [38] and [39]). Encoding: The Hamming code converts 4 bits of information into a 7 bit sequence. However 7 bits does not divide into our required 64 bit sequence that we need. Thus, we implement the rectangular coding on top of the Hamming coding to achieve our required 8 bytes that we calculate our checksum over. To encode, we just calculate the appropriate parity bits and construct the 7 bit code as according to the construction described above. We do this on 7 4-bit data sequences, which will give us 7 Hamming coded sequences, each being 7 bits long. We then encode this using the rectangular coding to get our required 64 bits. This gives us an information rate of 28 /

133 Section 6.3: Error Correction Chapter 6 Decoding: To decode, we first decode the rectangular code. This will give us the 7 bit Hamming sequences, which we then decode by seeing which parity bits do not add up, and correct/detect the errors accordingly. BER Performance: Table 6.8 shows how this error correction scheme on top of the checksum error detection decreased the amount of errors going through to the application. From this table we can see that in some cases this scheme works better than the Rectangular coding; but it does not seem to be worth the sacrifice in information rate. We also have the very poor performance for the cell phone charger, engine off condition simulation again. 114

134 Section 6.3: Error Correction Chapter 6 Simulation Model BER Effective BER Boosted BER BER Order of Magnitude Decrease No Device On - Engine Off Cell Phone Charger - Engine Off Windscreen Wipers On - Engine Off No Devices On - Engine On Cell Phone Charger - Engine On Windscreen Wipers On - Engine On Table 6.8: Bit Error Rate Decrease for Hamming Coding on Top of Checksum 115

135 Section 6.4: Error Detection/Correction Comparison Chapter Error Detection/Correction Comparison Table 6.9 shows a comparison of the performance of the different error detection and correction techniques. The table presents the averaged improvement in order of magnitude for the BER performance for each scheme. The information rate for each scheme is also presented, and made to have the same denominator for easier comparison. Error Control Scheme Single Parity Bit Double Parity Bit Checksum Over 2 Bytes Checksum Over 4 Bytes Checksum Over 8 Bytes Checksum with Repetition Checksum with Rectangular Checksum with Hamming Average Decrease in Order of Magnitude of BER Information Rate / / / / / / / /72 Table 6.9: Bit Error Rate Decrease Performance for Bit Error Correction Techniques Applied on Top of Checksum Error Detection *NOTE: We limit the infinity results to 15 to calculate the averages. 116

136 Section 6.5: Conclusion Chapter 6 From this table we can deduce that for the most reliable communication we can choose the repetition coding on top of the LIN checksum. Using this technique, however, will mean having a very low information rate. A very well balanced technique would be to use rectangular coding on top of the LIN checksum; where the reliability is relatively good and the information rate is still decent (at least above 0.5). This technique is also often able to reach the BERs of the boosted transmissions which means, at the sacrifice of data speed transfer, one can implement this coding scheme instead of spending money on a booster circuit. One exception to this is the condition of the vehicle when a cell phone charger is plugged in and the engine is not running. In this case, the bit error rate is too high and the nature of the errors are such that coding cannot handle it to get decent transmission reliability. 6.5 Conclusion In this chapter we investigated the effect of error detection techniques employed by the LIN protocol (2 parity bit check and the checksum) as well as the very simple single bit parity check. We then offer an analysis of the use of simple error correction techniques that can be implemented on top of the LIN protocol. We find that checksums are a fairly effective method for sifting out errors, where the more data bytes the check byte is checking, the more effective it is. The checksum detection may be good enough for some in-vehicle network applications, but for others it may not provide for reliable enough communication. Thus, we implement simple error correction techniques on top of the checksum detection to further reduce the amount of effective errors going through to the application. It is shown that the rectangular coding scheme has a good trade-off between information rate and reliability. This chapter acts as a complementary exercise to the bit error recording and channel 117

137 Section 6.5: Conclusion Chapter 6 modelling from the other chapters to show how the created models can be used to analyse different coding techniques. We focused on the LIN protocol and stayed in the scope of the LIN protocol because most devices in vehicles would have the this protocol built into its application layer (and the SIG60 PLC transceiver was designed to be compatible with LIN), and to demonstrate how the channel models can be used to analyse the performance of a protocol s error control techniques. The analysis can easily be performed for many other coding techniques. And since the SIG60 IC is not necessarily limited to LIN, its user has the advantage of being able to implement any kind of binary coding to adapt to the channel and the conditions it may be under (considering one has a model for that particular condition). 118

138 CHAPTER 7 Conclusion A conclusion is the place where you got tired thinking. - Martin H. Fischer 7.1 Thesis Summary Literature Study The literature study starts by providing a general overview of current in-vehicle communication technology, focusing on the two most commonly used protocols (CAN and LIN). It is important to understand the current protocols in place, because creating a new technology to enter the same arena would have to utilise or compete with these protocols. We then investigate the feasibility of in-vehicle power line communication, followed by the conclusion that it is a concept which is already in use industrially and thus considered feasible. This is followed by further research into various aspects of invehicle power line communication, such as regulations, channel characterisation and other possible issues. Markov modelling is then discussed. Two model structures are discussed: Gilbert Model and Fritchman Model. These two models are commonly found in literature and are often used to simulate the BER characteristics of communication channels. How to obtain the parameters for each model is explained. These two models, together with two new models, were used in Chapter 4 to model the channel s BER characteristics. Next, the SIG60 was discussed. This very important integrated circuit is the main 119

139 Section 7.1: Thesis Summary Chapter 7 device on which the experiments were based on. Other devices would probably produce different results. As explained in the chapter, however, this is the only readily available device on the market that is specifically designed for in-vehicle power line communication, and thus the most likely to be used. They are thus chosen to be used to gain experimental results for a channel analysis. Following the discussion of the SIG60 is the discussion of two important devices that were used to attempt and finally achieve bit error recording; that is the Arduino microcontroller platform, and the Picoscope PC based oscilloscope Bit Error Recording In Chapter 3, we present the evolution of the bit error recording method to get useful results. It seems like a simple task, but turned out to be a time consuming endeavour that lead us through 3 different methods to finally achieve proper bit error recording. Albeit time consuming, lots was learnt about serial communication timing, microprocessor programming, microprocessor limitations, C application programming and driver interaction with the PicoScope PC-based oscilloscope. The chapter goes into detailed discussions of how each method was performed and why it was successful or not. At the end the experimental results were presented Experimental Results In Chapter 4, we use the bit error recording method developed in Chapter 3 to gain further experimental results for the vehicle under different conditions. The results are presented in a few forms, including various cumulative frequency distributions and the error vector patterns. 120

140 Section 7.1: Thesis Summary Chapter Markov Modelling In Chapter 5, we use the results from Chapter 3 to obtain the parameters required for the two Markov models discussed in the literature study. We then moved on to describe the structure and parametrisation of two new models (not found in literature at the time of the writing of this dissertation). The names coined for these models were: Finite State Spreading Chain Model and Error State Spreading Chain Model. All the models were used to simulate the channel and these results were compared and a discussion of the comparisons is provided. All the models except the Error State Spreading Chain model simulated the channel very well. However, the Finite State Spreading Chain model was ultimately the better model, according to the numerical analysis provided (with the drawback of having many states). Using this conclusion, we create Finite State Spreading Chain models for each of the set of experimental results presented in Chapter Channel Error Coding In Chapter 6, we use the models created in Chapter 5 to simulate the channel under the different conditions that the experiments in Chapter 4 were performed at to analyse the effectiveness of certain error detection and correction methods. The first method is a simple single parity bit. The other two methods (2 bit parity and checksum) are the error detection mechanisms for the LIN protocol. We then investigate the use of three different error correction schemes that can be used on top of the checksum error detection; namely Repetition coding, Rectangular coding and Hamming Coding. The analysis shows how each of these methods improves the effective BER that the application will see. The checksum by itself may be reliable enough for some applications, but implementing rectangular coding on top of the checksum is an effective method and 121

141 Section 7.2: Contributions Chapter 7 can match the BER performance of the boosted BERs (in most cases). An important result is that of the cell phone charger plugged in with the engine off condition, where none of the coding methods are really effective. There are many other coding techniques and schemes that can be implemented on a simple UART based communication network, and Chapter 6 demonstrates how one can use the Markov models (that are parametrised using experimental results) to test the effectiveness of different error control mechanisms without having to actually implement them. 7.2 Contributions The main contributions of this dissertation are: ˆ We have shown that there is a large potential for Power Line Communication in the In-Vehicle Communication arena, as it can be used to eliminate the need for extra wiring (which leads to lower cost, weight and complexity). ˆ We developed a method for bit error recording that could be used for any UART application. ˆ We performed several in-vehicle experiments to get the bit error characteristics of the channel under different conditions. ˆ We show how to get the parameters for two Markov models that are often presented in literature. ˆ We introduce two new Markov models, and provide both explanation of concept and how to get their parameters. ˆ We provide methods for numerically assessing and comparing the Markov models with the experimental data to determine which one simulates the channel the best. ˆ We created Markov models for each of the set of in-vehicle experiments performed. ˆ We provide an analysis for current in-vehicle communication protocol (specifically LIN) error detection methods used for power line communication, as well as the 122

142 Section 7.3: Further Research Chapter 7 use of simple error correction coding. 7.3 Further Research There is still plenty of research to be done on the topic. Having the methods developed for bit error recording and Markov modelling, it is now easier to model the channel under different conditions; such as transmission at different frequencies, different transmission rates and different channel parameters (different loads, noise levels etc.). This further research can provide a more complete model of the power line and can contribute towards the creation of regulations and protocols specifically for in-vehicle power line communication. This would include developing new coding techniques specifically for the channel. 123

143 References [1] Midas Letter. (2011) Copper Prices to Continue Rising. [Online]. Available: [2] Sumitomo Electric. (2012) Automotive Wiring Harness. [Online]. Available: [3] E. Wade and H. Asada, Design of a broadcasting modem for a dc plc scheme, Mechatronics, IEEE/ASME Transactions on, vol. 11, no. 5, pp , October [4] Y. Maryanka, D. O. Amrani, and A. Rubin, The Vehicle Power Line as a Redundant Channel for CAN Communication, SAE World Congress and Exhibition, [5] M. Wilson, H. Ferreira, and R. Heymann, Markov modelling of in-vehicle power line communication, in Africon 2013, Sept. 2013, pp [6] U. Keskin, In-Vehicle Communication Networks: A Literature Survey, Technische Universiteit Eindhoven, July [7] J. H. Park, M. H. Kim, S. Lee, and K. C. Lee, Implementation of a can system with dual communication channel to enhance the network capacity, in IECON th Annual Conference on IEEE Industrial Electronics Society, Nov. 2010, pp [8] CiA. (2012) CAN History. [Online]. Available: [9] Computer Solutions Ltd. (2012) CAN Tutorial. [Online]. Available: tutorials/can tutorial.htm 124

144 References [10] Bosch. (1991) CAN Specification: Version 2.0. [Online]. Available: [11] Oliver Bremicker (Siemens). (2002) Local Interconnect Network Training. [Online]. Available: Training.pdf [12] S. REY. (2003) Introduction to LIN (Local Interconnect Network). [Online]. Available: ressources/ressources/ Networks/LIN/LIN bus.pdf [13] LIN Consortium. (2012) Local Interconnect Network: Technical Overview. [Online]. Available: 86b8ef40dce6d1ae9cf73f8bdb0d1f89 [14] Franoise Cacciaguerra. (2003) Introduction to Power Line Communication. [Online]. Available: [15] F. Benzi, T. Facchinetti, T. Nolte, and L. Almeida, Towards the powerline alternative in automotive applications, in Factory Communication Systems, WFCS IEEE International Workshop on, May 2008, pp [16] UNECE. (2013) UN Vehicle Regulations. [Online]. Available: [17] F. Rouissi, V. Degardin, M. Lienard, and P. Degauque, Low amplitude impulsive noise in vehicular power line network, in Intelligent Transport Systems Telecommunications,(ITST),2009 9th International Conference on, Oct. 2009, pp [18] P. van Rensburg, H. Ferreira, and A. Snyders, An experimental setup for in-circuit optimization of broadband automotive power-line communications, in Power Line Communications and Its Applications, 2005 International Symposium on, April 2005, pp

145 References [19] M. Carrion, M. Lienard, and P. Degauque, Communication over vehicular dc lines: Propagation channel characteristics, in Power Line Communications and Its Applications, 2006 IEEE International Symposium on, 2006, pp [20] A. Schiffer, Statistical channel and noise modeling of vehicular dc-lines for data communication, in Vehicular Technology Conference Proceedings, VTC 2000-Spring Tokyo IEEE 51st, vol. 1, 2000, pp vol.1. [21] M. Mohammadi, L. Lampe, M. Lok, S. Mirabbasi, M. Mirvakili, R. Rosales, and P. van Veen, Measurement study and transmission for in-vehicle power line communication, in Power Line Communications and Its Applications, ISPLC IEEE International Symposium on, April 2009, pp [22] P. Tanguy, F. Nouvel, and P. Maziearo, Power line communication standards for in-vehicule networks, in Intelligent Transport Systems Telecommunications,(ITST),2009 9th International Conference on, Oct. 2009, pp [23] F. Nouvel and P. Tanguy, What is about future high speed power line communication systems for in-vehicles networks? in Information, Communications and Signal Processing, ICICS th International Conference on, Dec. 2009, pp [24] A. Vallejo-Mora, J. Sa andnchez Marti andnez, F. Ca andete, J. Corte ands, and L. Di andez, Characterization and evaluation of in-vehicle power line channels, in Global Telecommunications Conference (GLOBECOM 2010), 2010 IEEE, Dec. 2010, pp [25] F. Nouvel and P. Tanguy, Towards power line communication in vehicle, in General Assembly and Scientific Symposium, 2011 XXXth URSI, Aug. 2011, pp [26] S. Barmada, M. Raugi, M. Tucci, and T. Zheng, Power line communication in a full electric vehicle: Measurements, modelling and analysis, in Power Line 126

146 References Communications and Its Applications (ISPLC), 2010 IEEE International Symposium on, March 2010, pp [27] M. Lienard, M. Carrion, V. Degardin, and P. Degauque, Modeling and analysis of in-vehicle power line communication channels, Vehicular Technology, IEEE Transactions on, vol. 57, no. 2, pp , March [28] Interface Bus. (2012) CAN Bus Description. [Online]. Available: http: // [29] R. Manning and J. Ewing, Temperature Survey in Cars, Royal Automobile Club of Queensland, February. [30] Arduino Team. (2013) Arduino. [Online]. Available: [31] Pico Technology. (2013) PicoScope 2000 Series EntryLevel Oscilloscopes. [Online]. Available: [32] E. N. Gilbert, Capacity of a burst-noise channel, in The Bell System Technical Journal, September 1960, pp [33] J.-Y. Chouinard, M. Lecours, and G. Delisle, Estimation of gilbert s and fritchman s models parameters using the gradient method for digital mobile radio channels, Vehicular Technology, IEEE Transactions on, vol. 37, no. 3, pp , Aug [34] C. Pimentel and I. Blake, Modeling burst channels using partitioned fritchman s markov models, Vehicular Technology, IEEE Transactions on, vol. 47, no. 3, pp , Aug [35] B. Fritchman, A binary channel characterization using partitioned markov chains, Information Theory, IEEE Transactions on, vol. 13, no. 2, pp , April [36] D. Poole, Linear Algebra: A Modern Introduction. Belmont, USA: Thomson,

147 References [37] R. W. Hamming, Coding and Information Theory. Englewood Cliffs, N.J.: Prentice-Hall, [38] D. Hoffman, D. Leonard, C. Lindner, K. Phelps, C. Rodger, and J. Wall, Coding Theory: The Essentials. 270 Madison Avenue, New York, New York 10016: Marcel Dekker, Inc., [39] S. Ling and C. Xing, Coding Theory: A First Course. The Edinburgh Building, Cambridge CB2 2RU, UK: Cambridge University Press, [40] Yamar Electronics. (2013) Yamar Electronics Website. [Online]. Available: [41] S. Tsai, Markov characterization of the hf channel, Communication Technology, IEEE Transactions on, vol. 17, no. 1, pp , [42] D. Oosthuizen, Markov models for mobile radio data communication systems, Master s thesis, Rand Afrikaans University. [43] W. Zhou, An experimental evaluation of markov channel models, Master s thesis, Rand Afrikaans University. [44] Firuz Zare. (2009) EMI Issues in Modern Power Electronic Systems. [Online]. Available: [45] Crane Aerospace and Electronics. (2006) Measurement and Filtering of Output Noise of DC/DC Converters. [Online]. Available: product documents/dc DC Converters Output Noise.pdf [46] Matlab. (2013) Least-squares (model fitting) algorithms. [Online]. Available: least-squares-model-fitting-algorithms.html [47] L. Kanal and A. R. K. Sastry, Models for channels with memory and their applications to error control, Proceedings of the IEEE, vol. 66, no. 7, pp , July. 128

148 References [48] J. Metzner, An interesting property of some infinite-state channels (corresp.), Information Theory, IEEE Transactions on, vol. 11, no. 2, pp , Apr. [49] F. Swarts, Markov characterization of fading channels, Master s thesis, Rand Afrikaans University. [50] Yamar Electronics, SIG60 - UART Over Powerline, for AC/DC-BUS Network, [51] Adelle Coster. (2013) Goodness-of-Fit Statistics. [Online]. Available: web.maths.unsw.edu.au/ adelle/garvan/assays/goodnessoffit.html [52] Robert F. Nau. (2013) What s a good value for R-squared? [Online]. Available: people.duke.edu/ rnau/rsquared.htm 129

149 APPENDIX A The Yamar SIG60 Integrated Circuit Yamar Electronics is a company in Israel that specialises in developing hardware for in-vehicle DC power line communication. Yamar s SIG60 IC is a DC PLC transceiver that is able to communicate up to speeds of Kbps [50]. It is the only readily available and relatively cheap PLC transceiver on the market that is designed specifically for in-vehicle (12 V DC-bus) use. It can, however, operate on a wide variety of different power lines; even AC lines. It also has a selectable frequency range up to 13.5 MHz, which is within the band recommended by the literature presented in this chapter. The SIG60 makes use of a special modulating technique developed by Yamar called multiple phase modulation. It is based on a combination of phase shifts which determines if a received signal is a 1, 0 or noise. It uses an UART to communicate with the host controller, and essentially creates a seemless UART link between hosts through the power line. This makes the SIG60 ideal for bit error analysis and coding experimentation on power lines. It also makes the SIG60 a very easy device to interface with, where only UART capability is needed to make use of the SIG60 s ability to communicate over the power line. Figure A.1 shows how the SIG60 would fit into a power line communication network. It is the physical link between the host controller s UART communication and the actual signal on the power line (the host controller could be a computer, micro-controller, ECU or any other device that has UART capabilities). The host controller would be programmed to either be a master or slave, its buffer size would be defined by its own hardware capabilities and it would be responsible for the encoding and decoding of messages. 130

150 Appendices Figure A.1: Block Diagram illustrating the function of the SIG60 in a Power Line Network The SIG60 requires one low start bit and one high stop bit. The start bit indicates to the SIG60 to start transmitting. The next 8 bits sent to the SIG60 are then transmitted over the power line. The stop bit indicates to the SIG60 to stop transmitting. Thus, it is possible to send a custom bit stream through the power line using the SIG60 as long as there are start and stop bits in the necessary positions to keep the stream continuous. Issues concerning sending a custom bit stream continuously are discussed in detail in Chapter 3. Figure A.2 shows the SIG60 multiple phase modulated signal on a 12 V car battery as well as the serial input into the transceiver. Note that as soon as the serial line drops to zero, the SIG60 starts transmitting. Then there are 8 bits of data, followed by a high stop bit. If the SIG60 does not receive a high stop bit, then it would continue to transmit zeros until the line goes high again (this is so that the SIG60 can send a synchronisation break as specified in the LIN protocol, as described earlier in this chapter). The duration of the transmission shown in the figure is 10 bits and was at a baud rate of The receiving SIG60 has an output delay of about three bits, as demonstrated in Figure A.3. The main features of the SIG60 are: 131

151 Appendices ˆ Has selectable bit rates between 9.6 and kbps (almost three times faster than specified by the LIN protocol). ˆ Has selectable carrier frequencies between 1.75 and 13 MHz. ˆ Communicates to the host controller via UART. ˆ Specifically designed to operate on a vehicle power line. ˆ Uses multiple phase modulation. Figure A.2: The SIG60 Multiple Phase Modulated Signal on a 12 V Car Battery and the Serial Input Into the SIG60 Figure A.3: The SIG60 Signal on a Power Line Versus the Output of the Receiving Device 132