MODELING AND CONTROLLER DESIGN FOR THE VVS-400 PILOT-SCALE HEATING AND VENTILATION SYSTEM NURUL ADILLA BT MOHD SUBHA

MODELING AND CONTROLLER DESIGN FOR THE VVS-400 PILOT-SCALE HEATING AND VENTILATION SYSTEM NURUL ADILLA BT MOHD SUBHA A project report submitted in partial fulfilment of the requirements for the award of the degree of Masters of Engineering (Electrical-Mechatronics & Automatic Control) Faculty of Electrical Engineering Universiti Teknologi Malaysia NOVEMBER 2009

iii ACKNOWLEDGEMENT I would like to express my sincere gratitude to Assoc. Prof. Dr. Mohd Fua ad bin Hj. Rahmat, who gave me this opportunity to run this project with his guidance and valuable suggestions regarding the project matters. Without his guidance and persistent help, this project would not have been possible. I would also like to thank my friends Alifa, Rozaimi, Hazrul, Zulfatman and those people who have been associated with me in this project and have helped me with it by sharing their opinion and made it a worthwhile experience. Also very special thanks to my beloved family and Norikhwan Hamzah for supports and encouragement.

iv ABSTRACT System modeling is an important task to develop a mathematical model that describes the dynamics of a system. The scope of work for this project consists of modeling and controller design for a particular system. A heating and ventilation model VVS-400 from Instrutek, Larvik, Norway is the system to be modeled and will be perturbed by pseudo random binary sequences (PRBS) signal. Parametric approach using ARX model structure will be use to estimate the mathematical model or approximated model plant of the VVS-400. The approximated plant model is estimated using System Identification approach. The conventional PID controller and artificial Fuzzy controller are designed based on the approximated plant model and also real plant model where the real plant model is developed by interfacing the Real-time Windows Target toolbox in Matlab with real VVS-plant by using data acquisition (DAQ) card PCI-1711. An artificial Fuzzy controller approach is incorporated in two ways which are conventional Fuzzy logic controller (FLC) and a replacement of conventional fuzzy controller known as Single Input Fuzzy Logic Controller (SIFLC). Simulations and experiment validate the equivalency of both controllers. Results reveal that SIFLC found to be better than FLC due to its less computation time compared to conventional FLC.

v ABSTRAK Permodelan sistem adalah langkah untuk memdapatkan model matematik yang menerangkan sifat sesuatu sistem itu. Ruang lingkup kerja dalam projek ini adalah terdiri daripada permodelan sistem dan mereka bentuk pengawal. Model VVS-400 dari Instrutek, Larvik, Norway adalah sistem yang akan di modelkan dengan menggunakan masukan Pseudo Random Binary Sequences (PRBS). Pendekatan parameter dengan struktur ARX akan digunakan untuk menerbitkan anggaran model matematik untuk sistem VVS-400. Anggaran model matematik ini boleh di terbitkan menggunakan perisian pengenalpastian sistem di dalam Matlab. Pengawal PID dan logik kabur akan di rekabentuk dengan simulasi berdasarkan anggaran model matematik dan juga berdasarkan model sebenar VVS-400 dengan menggunakan Real-time Windows Target toolbox dalam Matlab dan juga kad PCI 1711 sebagai pengantara. Di dalam kawalan logik kabur, terdapat dua jenis pengawal yang akan di reka bentuk iaitu logik kabur dan masukan tunggal logik kabur. Akhir sekali, perbandingan di antara logik kabur dan masukan tunggal logik kabur akan dibincangkan.

vi TABLE OF CONTENTS CHAPTER TITLE PAGE DECLARATION ACKNOWLEDGEMENTS ABSTRACT ABSTRAK TABLE OF CONTENTS LIST OF TABLES LIST OF FIGURES LIST OF ABBREVIATIONS ii iii iv v vi ix x xii 1 INTRODUCTION 1.1 Project background 1 1.2 Objectives of the project 2 1.3 Scope of work 2 1.4 Thesis outline 3 2 LITERATURE REVIEW 2.1 Related works 4 2.2 System description of the VVS-400 6 2.3 Pseudorandom Binary Sequences (PRBS) 10 2.4 System identification procedure 10 2.4.1 Experiments 10 2.4.1.1 Preliminary experiments 11 2.4.1.2 Main experiments 11 2.4.2 Data examination 12 2.4.3 Model structure selection 12

vii 2.4.3.1 AR model 13 2.4.3.2 ARX model 13 2.4.3.3 ARMAX model 13 2.4.3.4 Output error model 13 2.4.3.5 Box-Jenkins model 14 2.4.4 Parameter estimation 14 2.4.5 Model validation 14 2.4.5.1 Model validity criterion 15 2.4.5.2 Pole-zero plots 15 2.4.5.3 Bode diagram 15 2.4.5.4 Residual analysis 16 2.5 Data acquisition hardware (PCI-1711) 16 2.6 Real-time Windows Target toolbox 16 3 METHODOLOGY 3.1 Overview of step procedure 18 3.1.1 Design an experiment and experimental setup 19 3.1.1.1 The PRBS signal generation 20 3.1.2 Model identification using Matlab 21 3.1.2.1 Selection of model structure 28 3.1.2.2 Parameter estimation 29 3.1.2.3 Validation 31 3.1.3 Controller design 33 3.1.3.1 PID controller 33 3.1.3.2 Fuzzy logic controller 33 3.1.3.3 Single input fuzzy logic controller 37 3.1.4 Verification:Controller in a real VVS-400 39 4 RESULTS AND DISCUSSION 4.1 System modeling 40 4.2 Controller design via simulation 44 4.3 On-line control with real VVS-400 48

viii 5 CONCLUSION AND SUGGESTIONS 5.1 Conclusion 53 5.2 Future works 53 REFERENCES 55 Appendices A B 57-58

ix LIST OF TABLES TABLE NO. TITLE PAGE 2.1 Input voltage and output temperature 8 3.1 Rules table of fuzzy 36 3.2 The rule table with Toeplits structure 38 3.3 The reduced rule table of SIFLC 38

x LIST OF FIGURES FIGURE NO. TITLE PAGE 2.1 Schematic diagram of the VVS-400 heating and ventilation model 7 2.2 Local panel on VVS-400 7 2.3 Relationship between temperature and voltage 9 2.4 Simulink diagram being used to control real-time Windows Target 17 3.1 Algorithm for modeling and system identification 18 3.2 VVS-400, computer and data acquisition hardware 20 3.3 Open-loop experiment using Simulink block diagram 21 3.4 System identification tool (ident) windows 22 3.5 Import data from workspace 22 3.6 Data variable mydata icon display at Data Views 23 3.7 Input and output signals 23 3.8 Remove means of mydata 24 3.9 New data mydatad after remove means 24 3.10 Drag and drop mydatad onto Working Data 25 3.11 Select range of data 25 3.12 Data range selection for estimation and validation purpose 26 3.13 mydatade and mydatadv icons are inserted into Data Views 27 3.14 Drag the mydatade and mydatadv icons onto Working Data and Validation Data 27 3.15 Select Linear parametric models 28 3.16 Linear parametric models 29 3.17 ARX order editor 30 3.18 Icon ARX 663 is inserted into Model Views 30

xi 3.19 Model output of ARX 663 31 3.20 Arx 663 Model info 32 3.21 Residuals and pole-zero plots of the ARX 663 32 3.22 Fuzzy logic controller block 34 3.23 Fuzzy inference block 34 3.24 Membership function of error (e), derivative of error ( e ) and control input, u 35 3.25 Rule editor of fuzzy 36 3.26 Closed-loop real-time controller 39 4.1 The input-output data set 41 4.2 Measured and simulated model output 42 4.3 Pole-zero plots 43 4.4 Residuals analysis 43 4.5 Simulink block of the system and PID controller 44 4.6 Temperature process response from simulation with PID controller 45 4.7 Simulink block of the system and conventional FLC 46 4.8 Linear control surface of conventional FLC 46 4.9 Temperature process response from simulation with FLC 47 4.10 Simulink block of the system and SIFLC 47 4.11 Temperature process response from simulation with SIFLC 48 4.12 Simulink block diagram of real plant implementation with PID 49 4.13 Temperature process response from experiment with PID 49 4.14 Simulink block diagram of real plant implementation with conventional FLC 50 4.15 Temperature process response from experiment with conventional FLC 50 4.16 Simulink block diagram of real plant implementation with SIFLC 51 4.17 Temperature process response from experiment with SIFLC 51 4.18 Temperature process response from simulation 52 4.19 Temperature process response from real VVS-400 plant 52

xii LIST OF ABBREVIATIONS SISO MIMO SI PRBS ARX ARMAX AR OE BJ FPE FLC SIFLC - Single Input Singla Output - Multiple Input Multiple Output - System Identification - Pseudorandom Binary Sequences - Auto-regressive with Exogenous Input - Auto-regressive Moving Average with Exogenous Input - Auto-regressive - Output Error - Box-Jenkins - Final Prediction Error - Fuzzy Logic Controller - Single-input Fuzzy logic Controller

xiii LIST OF APPENDICES APPENDIX TITLE PAGE A Matlab program for PRBS generation 57 B Publication 58

CHAPTER 1 INTRODUCTION 1.1 Project background The heating and ventilating system is a common process in our daily life where certain desired temperature is controlled. In industries such as pharmaceutical, ability to control temperature is crucial to ensure the quality of the product always within control. However, most of heating and ventilation plants are complex with higher-order systems, which leads to unsatisfactory performance. In this project, VVS-400 pilot-scale of heating and ventilation system is selected as a model system which needs to be maintained at a certain level of temperature. Therefore, model system has to be controlled by a suitable controller to achieve its desired temperature. In order to control a system, model of that system must be created. The process of constructing models from experimental data is called system identification. System identification is a process in which experimental data is used to obtain a mathematical model for a particular system. This technique is widely used in industrial application mainly for nonlinear processes. There are several approaches for identification technique such as theoretical and empirical model. In this project, an empirical model will ne applied to the system where the system can be referred as a black-box model. A mathematical model will be developed through an experimental data by determined the input and output relationship. In this approach, the persistently excitation of input signal is crucial, since it influences data

2 sufficiency. Often, Pseudo-Random Binary Sequences (PRBS) input were chosen due to its large energy content in a large frequency range [Fazalul, 2006]. Further details in choosing the appropriate input can be found in [Barenthin, 2006]. The results from experimental data will be tested using Matlab s System Identification toolbox. From experimental data, parametric approach using Autoregressive with exogenous input (ARX) structure will be use to estimate the mathematical model (approximated model plant) of VVS-400. Controller design is also included in this project through closed loop Matlab simulation with an approximated model plant of VVS-400. There are three types of controllers that will be considered in this part which are PID controller, Fuzzy logic controller (FLC) and Single-input fuzzy logic controller (SIFLC). Then, an online control is performed with a real VVS-400 using Real-time Windows Target toolbox. Finally, discussion and conclusion are drawn. 1.2 Objectives of the project 1.2.1 To develop a mathematical model that describes the dynamics of VVS-400 using system identification approach. 1.2.2 To estimate the parameter of the VVS-400 mathematical model. 1.2.3 To design a suitable controller for the VVS-400 and to test the stability of the system. 1.3 Scope of work 1.3.1 Study the characteristics of VVS-400 1.3.2 Experimental setup and data collection 1.3.3 Study on System Identification toolbox 1.3.4 Controller design

3 1.4 Thesis outline This thesis is organized in 5 chapters. The first chapter gives an overview of the project that gives the introduction of control system and its possible application. Chapter Two covers literature review on related works, system description, system identification and controller design. Chapter Three covers the flow of methodology and description of each procedure. Chapter Four mainly discuss about the results and discussion of this project. Chapter Five includes the conclusion and recommendation of the thesis.

CHAPTER 2 LITERATURE REVIEW The project starts with literature review to build up the knowledge of system identification, control systems method and all topics that relate to the project. The key purpose of this project is to design the controller for a control system by identifying unknown parameters of the system. In order to control a system, the model s behavior and its detail understanding is very useful to provide the desired performance. 2.1 Related works In the recent years, there are many emerging control strategy approaches for controller s design of heating and ventilation systems such as robust PID controller [Masato Kahara et. al, 2001], fuzzy immune PID controller [Desheng Liu et. Al, 2009], Multiple Model Predictive Control (MMPC) [Ming He et. Al, 2005] and advanced PID auto-tuner [Qiang Bi et. Al, 2000]. For example, Kasahara [Masato Kahara et. al, 2001] propose a robust PID control system which can cope with the changes in the plant characteristic which suitable for practical applications. Another example is an auto tuner for PID controller, both for SISO and MIMO processes which developed by Bi and Cai [Qiang Bi et. Al, 2000].

5 In some cases, an artificial approach such as Fuzzy Logic Control (FLC) has gain interest in control systems design. For instance, Rafael has proposed a combination of weighted linguistic fuzzy rules together with a rule selection process in heating and ventilation system in order to maintain its indoor temperature [Rafael Alcala, 2005]. However, it is known that conventional FLC has to deal with fuzzification, rule base, inference engine and defuzzification operation. Larger sets of rules will produce longer computational time for conventional FLC. Usually, a complicated system as heating and ventilation system require many rules to perform this conventional FLC. These will results large computational time to accomplish the control algorithm. Therefore, Single Input Fuzzy Logic Controller (SIFLC) has been introduced to solve the conventional FLC problem. The SIFLC has only one input variable which significantly produce less number of rules compared to conventional FLC. Tabakova has presented the implementation of the SIFLC and its effectiveness which has less computation time in the real time application [Bilyano Tabakova, 2008]. However, in order to design very efficient controller with high quality system performance, the system must be modeled in a proper way. For unknown system which has unknown parameters, it can be called as black-box model. The mathematical modeling of this black-box model system can be obtained using System Identification (SI) technique. SI technique provides an efficient approach and proved to be very significant in practical applications. There are two methods to perform the system modeling, which are using theoretical and experimental design. The overall step of system identification procedure can be found in [Lennart Anderson et. Al, 1993 ]. Usually, system models are identified using data collected from open-loop experiment. An open-loop identification technique becomes popular because this approach is much easier compared to closed-loop identification. However, recent years, closed-loop identification has grown many interests where this approach offers more advantages compared to open-loop identification. This comparison was

6 explained detail in [Elsa et. al,2004]. In this project, only open-loop experiment is considered. 2.2 System description of the VVS-400 In this study, VVS-400 is used as a model system. The VVS-400 plant is a pilot scale of heating and ventilation system developed by Instrutek A/S, Larvik, Norway [Instrutek, 1994]. The schematic diagram of the system is shown in Figure 2.1. This plant can operate in three different modes: (i) Temperature control, (ii) Flow control and (iii) Cascade control. In this paper, only temperature control is studied (constant air flow rate). This model consists of a fan and heating element which is controlled by TRIAC. The fan blows air through the flow tube over the heating element. The temperature sensor, RTD platinum is located at the end of the tube. This plant model is also equipped by two independent local PID controllers to control the temperature and flow processes. However, in this study, local PID controller for temperature will be set as off mode which creates an open loop system for temperature while the air flow rate is fixed to a certain number and controlled by flow local PID controller. The local PID controller for temperature can be set as off by setting the switch local-pc is set to PC position at local panel on VVS-400 as shown in Figure 2.2.

7 Figure 2.1: Schematic diagram of the VVS-400 heating and ventilation model TIC (Temp.) Manipulated input 0-5V From PC To PC Process variable 0-5V from temperature sensor FIC (Flow) From PC To PC Figure 2.2: Local panel on VVS-400

8 The manipulated input can be set manually to simulate the PC 0-5 Volt output. If the manipulated variable is connect with a particular input, the feedback signal will acquire the output from the temperature sensor. The PCI-1711 card will be as an interface between the local panel on VVS-400 and the Matlab software on the computer. System is calibrated by injecting the step input signal into the system through local panel with an open-loop control. After system calibration, the relationship between voltage and temperature is obtained and plotted as shown in Figure 2.3. This is done by observing the output temperature with different input voltage as shown in Table 1. Table 2.1: Input voltage and output temperature Voltage(V) Temperature(Celcius) 2.5 50 2.8 56 3 60 3.1 62 3.3 66 3.4 68 3.5 70 3.6 72 3.8 76 4 80 4.1 82 4.2 84 4.3 86 4.4 88 4.5 90

9 100 Temperature vs Voltage Temperature, C T i mv i c 80 60 40 2 2.5 3 3.5 4 4.5 5 v i Voltage, V Figure 2.3: Relationship between temperature and voltage From Figure 2.3, it can be noted that Temperature( C) α K x Voltage(V) K = constant = gradient = 20 Hence, Temperature ( C) = 20 x Voltage (V) Ti = 20Vi where i = nth data Therefore, process output must be multiplied with constant 20, since the output from the approximated plant and data acquisition (DAQ) card is in voltage. Temperature process study of VVS-400 plant has been conducted in [Robin Mooney, 2006] which reveal the temperature process is continuously nonlinear.

10 2.3 Pseudorandom Binary Sequences (PRBS) Good parameters identification requires the usage of input signal that rich in frequencies [Mohd Nasir et. Al, 2007]. Most common type of input signal used is PRBS signal, a periodic deterministic signal with white-noise-like properties. This common choice is because of its large energy content that will guarantee the identification and to give an accurate model for a system. Other advantages of the PRBS is that its input excitation can concentrated in particular frequency ranges that correspond to the process dynamics which is very important for control system design. In this project, PRBS signal can be generated by using Matlab System Identification Toolbox by using syntax idinput (N, type, band, levels). Idinput generates input signals of different kinds, which are typically used for identification purposes. N determines the number of generated input data, argument type defines the type of input signal to be generated such as PRBS to gives a pseudorandom binary signal, argument band to the lower and upper bound of the passband and the argument levels defines the input level. 2.4 System identification procedure System identification is an iterative process and it is often necessary to go back and repeat earlier steps. 2.4.1 Experiments The experiments are done in two steps. In the first step, preliminary experiments such as step responses are performed to gain primary knowledge about important system characteristics such as time delay and dominating time constants.

11 The information obtained from the preliminary experiments are then used to determine suitable experimental conditions for the main experiments, which will give the data to be used in the System Identification Toolbox. 2.4.1.1 Preliminary experiments There are some system characteristics that can be concluded from preliminary experiments such as linearity, transient response analysis and frequency response analysis. 2.4.1.2 Main experiments In the main experiments, data are collected to be used in the System Identification Toolbox. In this part, the choice of input signal is important. The identification gives an accurate model at the frequencies where the input signal contains much energy. The input signal has good excitation at these frequencies. The frequency content of the input should therefore be concentrated to frequencies where small estimation errors are desired. A pseudo-random binary sequence (PRBS) is a common choice of input signal, since it has large energy content in a large frequency range. The experiment duration is chosen in order to get good parameter estimates. The amplitude is also chosen as large as possible in order to achieve a good signalto-noise ratio and to overcome problems with friction. However, the amplitude may not be chosen larger than the range in which the linearity assumption holds.

12 2.4.2 Data examination After an experiment has been performed, input sequence and an output sequence are represented as column vectors u and y. This available of set data will be divided into two sets, one for identification and one for validation. 2.4.3 Model structure selection In this part, only parametric models are considered. The model structure determines the set in which the model estimation is performed. By using parametric models, the model is described in terms of difference and differential equation. The complexity of the model structure, of course, affects the accuracy with which the model can approximate the real process. The most general parametric model structure used in the System Identification Toolbox is given by B( q) C( q) A( q) y( t) u( t nk ) e( t) F( q) D( q) where y and u is the output and input sequences, respectively, and e is a white noise sequence with zero mean value. The polynomials A, B, C, D, and F are defined in terms of the backward shift operator: A( q) 1 a q B( q) b b q 1 C( q) 1 c q D( q) 1 d q F( q) 1 1 1 1 f q 1 2 1 1 1 1... a 1... c na... b nc... d... f nd nf q nb q na nb 1 nc q q q nd nf However, there are some special forms, where one or more polynomial is set to identity:

13 2.4.3.1 Auto-regressive (AR) model The AR model is a special case of the ARX model with no input. AR model parameters are estimated using variant of the least-squares method which is a timeseries model with no exogenous input. A( q) y( t) e( t) 2.4.3.2 Auto-regressive with exogenous input (ARX) model The ARX model parameters are estimated based on input and output data. A( q) y( t) B( q) u( t nk ) e( t) 2.4.3.3 Auto-regressive Moving Average with exogenous input (ARMAX) model The ARMAX model parameter only supports time-domain data with single or multiple inputs and single output. A( q) y( t) B( q) u( t nk ) C( q) e( t) 2.4.3.4 Output-error (OE) model The OE model is represented by B( q) y( t) u( t nk ) e( t) F( q)

14 2.4.3.5 Box-Jenkins (BJ) model The BJ model is represented by B( q) C( q) y( t) u( t nk ) e( t) F( q) D( q) Finding the best model is a matter of choosing a suitable structure in combination with the number of parameters. 2.4.4 Parameter estimation The model estimation is the procedure of fitting a model with a specific model structure. The mathematical model of the systems can be estimated from the data using System Identification toolbox in Matlab. For the parametric models, an appropriate model order is choosing to estimate the parameters of the polynomials. For an ARX model, there are two methods for parameter estimation: which are the least squares (LS) method and the instrumental (IV) variable method. The parameters of the other model structures are estimated by use of a prediction error method. However, the Model reduction is an alternative to standard model estimation. The idea is to first estimate the parameters of a high order ARX model in order to capture most of the information in the data. Then, reduce the model order using suitable methods. The model reduction step then extracts the most significant states of this model. 2.4.5 Model validation The parametric models obtained in previous can be validated in a variety of ways. There are five model validations that are commonly used in application of

15 system identification: model validity criterion, pole zero and Bode plots, and residual analysis. In a standard identification session all of these are used to affirm an accurate model. 2.4.5.1 Model validity criterion It is possible to get an indication of a suitable model order by studying how various criteria depend on the model order. Two such criteria are the loss function and Akaike s Final Prediction Error (FPE). Akaike s FPE criterion can be used for linear and nonlinear models. It provides measure of model quality by simulating the situation where the model is tested on a different data set. FPE value is depended on the loss function value. According to Akaike s theory, the most accurate model has the smallest FPE. 2.4.5.2 Pole-zero plots A pole zero plots may indicate if the model order is too large. Then, there will be poles and zeros located close together, suggesting that model reduction is possible. 2.4.5.3 Bode diagram Stationary gain and location of dominating poles and zeros can be checked in the Bode plot.

16 2.4.5.4 Residual analysis If the residuals are computed based on the identified model and the data used for the identification, then ideally the residuals should be white and independent of the input signals. 2.5 Data acquisition hardware (Advantech PCI-1711) The PCI-1711 is a multi-function data acquisition card for the PCI bus. This card provides multiple measurement and control functions. Model PCI-1711 offers 16 12-bit single ended channels of A/D input, 16 channels of digital inputs, 16 channels of digital outputs, two 12-bit channels of analog output, and one 16-bit timer/counter with a time base of 10 MHz. 2.6 Real-time Windows Target (RTWT) toolbox Real-time Windows Target (RTWT) enables the Simulink model to run in a real time on the computer and interface with physical device. RTWT includes a set of I/O blocks (Analog Output and Analog Input) that provide connections between the physical I/O board (PCI-1711) and real-time model. The responds of Simulink model towards a real-world behavior can be quickly observe by running the hardware-in-the-loop simulations. By using Real-time Windows Target with Simulink external mode as shown in Figure 2.4, output from real-time model can be captured and display with the standard Scope blocks. There is no additional Simulink blocks is required.

Figure 2.4: Simulink diagram with Real-time Windows Target 17

CHAPTER 3 METHODOLOGY 3.1 Overview step of procedure In this project, there are a few steps need to be considered which can be illustrated in the Figure 3.1 below: Design an experiment Experimental setup Selection of model structure Parameter estimation Validation NO YES Model accepted? Controller design Figure 3.1: Algorithm for modeling and system identification

19 3.1.1 Design an experiment and experimental setup In this project, an open-loop single-input single-output (SISO) identification experiment will be considered. The purpose of conducting this experiment is to obtain the output temperature that corresponds to its input. As mentioned earlier in the previous section, Pseudo-Random Binary Sequences (PRBS) is the common input choice. The PRBS input is perturbed into the VVS-400 using idinput syntax in Matlab. The plant VVS-400, data acquisition hardware (PCI-1711 card) and computer are connected as shown in Figure 3.2 and evaluated as to whether everything is in working order. The experiment consists of a fan blowing air inside the tube, with constant air flow rate controlled by local FIC. The temperature sensor, RTD-platinum is located at the end of the tube. The PCI-1711 card that developed by Advantech is used to read and write data to and from the VVS-400.

20 Figure 3.2: VVS-400, computer and data acquisition hardware From Figure 3.3, both Analog Output and Analog Input from Real-time Windows Target (RTWT) will directly connect the Simulink Matlab to the VVS-400 plant using PCI-1711. The plant is connected to the Analog Input of PCI-1711 and input to the plant is connected to the Analog Output of PCI-1711. This is open-loop system identification. 3.1.1.1 The PRBS Signal Generation The PRBS can be generated using Matlab System Identification Toolbox. Syntax idinput (10000, prbs, [0 0.01], [0 3]) will simply generate a sequence of PRBS input data with 10000 data points, probability band, B of 0.01 and magnitude variation between 0 and 3. The probability band, B is such that the signal is constant over intervals of length 1/B (the clock period). The appropriate value of probability

21 band, B must be determined which affects the dynamic of the system under test. In selecting the suitable value of B, the step test will be very helpful. 3.1.2 Model Identification Using Matlab This section describes the procedures and steps to obtain ARX model from input-output data using Matlab System Identification Toolbox [Lennart Ljung, 2007]. Figure 3.3: Open-loop experiment using Simulink block diagram The input (u1) and output (y1) data obtained then will be imported and analyzed in System Identification toolbox by typing ident in Matlab command window. In the ident window, press the popup menu Import data and select Time domain data as shown in Figures 3.4 and 3.5.

22 Figure 3.4: System identification tool (ident) window Then, enter u1 into editable text box inside Input, y1 into the editable text box beside Output and a sampling interval of 2. In the box marked Data name is data information of u1 and y1. In this part, mydata will be the variable name of u1 and y1 data. By pressing the Import button, the mydata will be represented as an icon in the ident figure as shown in Figure 3.6. Figure 3.5: Import data from workspace

23 Figure 3.6: Data variable mydata icon display at Data Views It should be notice that the data also fills the Working Data and Validation Data icons. Click on the Time Plot checkbox beneath data Views to open the plot figure as can be seen in Figure 3.7. Figure 3.7: Input and output signals

24 In the Figure 3.8, the constant levels in the data sequences will be removed by select the Remove means from the Preprocess popup menu in the ident window. The new data set (green color) will be inserted into the data board (with a d appended to its name) and also automatically been plotted in the time plot figure as shown in Figure 3.9. Figure 3.8: Remove means of mydata Figure 3.9: New data mydatad after remove means

25 Then, this new data called mydatad is drag and drop onto the Working Data icon as shown in Figure 3.10. Now, select the Select range option from the Preprocess popup menu in the ident figure as can be seen in Figure 3.11. Figure 3.10: Drag and drop mydatad onto Working Data Figure 3.11: Select range of data

26 The new figure will open to select a portion of the data to be used for estimation and validation purposes as can be shown in Figures 3.12(a) and (b), respectively. The mydatade and mydatadv are the Data name of estimation and validation data and both data sets are also be inserted into the data board as can be seen in Figure 3.13. (a) Estimation data (b) Validation data Figure 3.12: Data range selection for estimation and validation purpose

27 Figure 3.13: mydatade and mydatadv icons are inserted into Data Views In the data board, mydatade and mydatadv are drag and drop onto Working Data and Validation Data respectively to be used for estimation and validation purposes as shown in Figure 3.14. Figure 3.14: Drag the mydatade and mydatadv icon onto Working Data and Validation Data

28 3.1.2.1 Selection of model structure Then, from the Estimate popup menu in the ident figure, step response of the system can be estimated by select the Linear parametric models as shown in Figure 3.15. This opens a new dialog window, where the selection of the model structure can be done by entering the model structure information into the Parametric models dialog and then press the Estimate button to generate a model as can be seen in Figure 3.16 (a) and (b). Figure 3.15: Select Linear parametric models

29 (a) ARX model structure (b) Model structure information Figure 3.16: Linear parametric models 3.1.2.2 Parameter estimation In Figure 3.17, the orders of na=6,nb=6 and nk=3 are select by using the popup menus in the Order editor dialog. The model will be computed and added into Model Views as an icon in the Figure 3.18.

30 Figure 3.17: ARX Order editor The result in the Figure 3.19 can be viewed by clicking on the Model output checkbox beneath Model Views in the ident figure. Figure 3.18: Icon ARX663 is inserted into Model Views

31 Figure 3.19: Model output of ARX 663 3.1.2.3 Validation The parametric models can be validated in various ways such as model validity criterion, pole-zero plots and residual analysis. In the model validity criterion, two such criteria are considered which are the loss function and Akaike s Final Prediction Error (FPE). These two criteria is given by double-click the model icon in the Model Views to open the Data/model info window as shown in Figure 3.20.

32 Figure 3.20: ARX 663 Model info Then, for pole-zero plots and residuals, click on the Model resids and Zeros and poles checkbox beneath Model Views in the ident figure as can be seen in Figures 3.21. Figure 3.21: Residuals and Pole-zero plots of the ARX 663 model

33 3.1.3 Controller design There are three types of controller are designed via simulation for this system which is PID controller, Fuzzy logic controller (FLC) and Single input fuzzy logic controller (SIFLC). 3.1.3.1 PID controller The PID controller is often implemented for industrial practice since it has a simple structure, straightforward implementation and easy to tune. In this paper, the PID controller is designed using the parameters of K p (proportional gain), K i (integral gain) and K d (derivative gain) tuned by Ziegler-Nichols method. The discrete-time expression of PID controller has the following form: u( k) K pe( k) KiTs n i 1 K e( i) T d s e( k) where u(k) is the control signal, e(k) is the error between the reference input and the process output and T s is the sampling time for the controller. However, finding an optimum adjustment for this system is not trivial. Fine tuning is required for an optimum result. 3.1.3.2 Fuzzy Logic Controller (FLC) For the FLC control design structure, it involves three main stages:(i) fuzzification, (ii) rule base, and (iii) defuzzification as can be shown in Figure 3.22

34 [Kevin, 1998]. The rule base is extracted from the knowledge or experience about the system itself. FUZZIFICATION RULE BASE DEFUZZIFICATION INFERENCE ENGINE Figure 3.22: Fuzzy logic controller block The conventional FLC has two inputs which are error, e and derivative error, e and only one control input, u as represented in Figure 3.23. In fuzzy control, the membership function, rules and scaling factor (gain) are tuning parameter. The membership function of error, e, derivative error, e and control input, u are assigned as NL: Negative large, NM: Negative medium, NS: Negative small, Z: Zero, PS: Positive small, PM: Positive medium, and PL: Positive large as can be seen in Figure 3.24. The ranges of this membership function are -10 to 10. Figure 3.23: Fuzzy inference block

35 NL NM NS Z PS PM PL -10-6.67-3.33 0 3.33 6.67 10 Figure 3.24: Membership function of error (e), derivative of error ( e ) and control input, u Since we have 7 variables for each fuzzy input, it gives 49 fuzzy rules as illustrated in Table 3.1. The rules are written as; IF error, e is PL AND derivative error, e is NL, THEN control input, u is Z Therefore, 49 fuzzy rules in Table 3.1 must be reads as mentioned and be performed in rule viewer of FIS editor in Fuzzy Matlab as shown in Figure 3.25.

36 Table 3.1: Rules table of fuzzy e e PL PM PS Z NS NM NL NL Z NS NM NL NL NL NL NM PS Z NS NM NL NL NL NS PM PS Z NS NM NL NL Z PL PM PS Z NS NM NL PS PL PL PM PS Z NS NM PM PL PL PL PM PS Z NS PL PL PL PL PL PM PS Z Figure 3.25: Rule editor of fuzzy

37 3.1.3.3 Single input fuzzy logic controller (SIFLC) The designed of SIFLC for this system employed Signed Distance method [Kashif, 2008]. From Table 3.1, it is common to have same output membership function in a diagonal direction. Then, each diagonal line has a magnitude which proportional to the distance from its main diagonal line. Instead of using two inputs (e, e ) in the conventional FLC, this method simplifies the number of input into one single input known as distance, d. The distance represents the absolute distance magnitude of the parallel diagonal lines (in which the input set of e and e lies) from the main diagonal which can be written as follows, d e e 2 1 with slope of diagonal line, is equal to 1. In order to obtain the distance, d value, the diagonal lines need to be calculated. The output of rule table for conventional FLC as shown in Table 3.1 can be represented in the constant number as follows, e e 0 e e 0 Then, this equation will results seven diagonal lines correspond to seven input values that can be seen in Table 3.2. Therefore, d can have positive or negative value. The diagonal line that result 0 is called main diagonal line.

38 e e NL -10 NM -6.67 NS -3.33 Z 0 10 Table 3.2: The rule table with Toeplits structure PL PM PS Z NS NM NL 10 6.67 3.33 0-3.33-6.67-10 0-3.33-6.67-10 -10-10 -10 3.33 0-3.33-6.67-10 -10-10 6.67 3.33 0-3.33-6.67-10 -10 6.67 3.33 0-3.33-6.67-10 PS 3.33 10 10 6.67 3.33 0-3.33-6.67 PM 6.67 10 10 10 6.67 3.33 0-3.33 PL 10 10 10 10 10 6.67 3.33 0 The derivation of distance, d input variable resulted in one dimension rule table compared to the conventional FLC which have many rules. The rule table is depicted in Table 3.3 with the output of corresponding diagonal lines, u o. Table 3.3: The reduce rule table of SIFLC d e e 2 1-10 -7-4. 66-2. 33 0 2. 33 4. 66 7 10 u o = e e -9. 9-9. 9-6. 6-3. 3 0 3. 3 6. 6 9. 9 9. 9

39 3.1.4 Verification: Controller in a real VVS-400 Simple closed-loop real-time controller using Real-time Windows Target (RTWT) toolbox in Matlab is developed. The controlled plant is connected to the Analog Input of PCI-1711, subtracted from set point, processed by controller as can be seen in Figure 3.26. The output of controller drives the input of the plant using Analog Output of PCI-1711. Both Analog Output and Analog Input from RTWT will directly connect the Simulink Matlab to the VVS-400 plant using PCI-1711. Set point + Controller output ANALOG OUTPUT REAL VVS-400 PLANT ANALOG INPUT Process output Figure 3.26: Closed-loop real-time controller

CHAPTER 4 RESULTS AND DISCUSSION 4.1 System modeling Initially, system model must be determined before control technique is applied. The system modeling part is the most challenging and vital part in designing the control system of VVS-400 due to its large time constant and slow process response. In order to obtain a particular model for this system, the open loop identification experiment has been done using parametric approach. In this experiment, a system model is identified using data collected when the Pseudo Random Binary Sequence (PRBS) is perturbed into the system as can be seen in Figure 4.1. From Figure 4.1, there are 2297 samples of data with 2 seconds sampling interval. The PRBS input is generated in Matlab. The collection of data was performed by PCI-1711 interface card. The input-output data is then be analyzed by System Identification toolbox in Matlab.

41 4 Input and output signals 3 y1 2 1 0 0 500 1000 1500 2000 4 3 u1 2 1 0 0 500 1000 1500 2000 Time Figure 4.1: The input-output data set From the set of input-output data in Figure 4.1, it was divided into two parts. The first part is the training data and the second is for testing or validation data. In this project, the VVS-400 system is modeled based on Autoregressive with exogenous input -ARX model structure with sixth order. The best fit of output model is 82.84% as depicted in Figure 4.2. Its polynomial structure can be written as A( q) y( t) B( q) u( t) e( t) A( q) 1 0.4776q 1 0.441q 2 0.774q 3 0.4322q 4 0.1352q 5 0.1308q 6 B( q) 0.0002502q 0.000266q 8 3 0.0008348q 4 0.0003908q 5 0.0003052q 6 0.0006835q 7

42 0.6 0.4 Measured and simulated model output measured1 estimated 0.2 0-0.2-0.4 1000 1500 2000 2500 Time Figure 4.2: Measured and simulated model output Then, Loss function = 0. 0000123078 and Akaike s Final Prediction Error(FPE) = 0. 000012567. Therefore, the pilot scale heating and ventilation VVS- 400 plant can be approximated modeled by this following equation 3 4 B( q) 0.0002502q 0.0008348q 0.0003908q 1 2 A( q) 1 0.4776q 0.441q 0.774q 5 3 6 7 0.0003052q 0.0006835q 0.000266q 4 5 6 0.4322q 0.1352q 0.1308q 8 Next, by observing the pole-zero plot of the model, there is one zero outside the unit circle of the z-domain as shown in Figure 4.3. This specific zero is called non-minimum phase model. For a non-minimum phase process the converse is true, a non-minimum phase pole will tend to cause a +90º phase shift, and a non-minimum phase zero will tend to cause a -90º phase shift. Since the system is assumed to be stable, since all the poles have negative real parts.

43 1 Poles (x) and Zeros (o) 0.5 0-0.5-1 -3-2 -1 0 1 Figure 4.3: Pole-zero plot In Figure 4.4, it computes the residuals (prediction error) from the model when input data is applied. From this figure, an autocorrelation performed a whiteness test while cross-correlation performed independence test. Both correlation shows the residuals within the confidence interval where the autocorrelation indicates the residuals are uncorrelated and cross-correlation indicates that the residuals are uncorrelated with the past inputs. 0.5 Autocorrelation of residuals for output y1 0-0.5-20 -15-10 -5 0 5 10 15 20 0.1 Cross corr for input u1 and output y1 resids 0-0.1-20 -15-10 -5 0 5 10 15 20 Samples Figure 4.4: Residuals analysis

44 Hence, based on this approximated plant model, conventional PID and artificial Fuzzy logic controller will be designed to perform the closed loop system simulation. The approximated plant gives a higher order model where an excess model order is usually represent the noise. Since the ARX model incorporate with noise in the system model, the model might be influenced by this noise [Fazalul, 2007]. 4. 2 Controller design via simulation Before the real process implementation, a simulation is carried out for each controllers to verify the propose controllers design. The aim of simulation is to give emphasis to the designing of the conventional proportional-integral-derivative (PID) and artificial Fuzzy Logic controller. To insure stability, only closed loop controller is considered in this control system. The step input is applied to the system as a reference input with set point of 60. Figure 4.5: Simulink block of the system and PID controller Figure 4.5 shows the simulink block diagram with PID controller and the performance of output response can be represented in Figure 4.6. From Figure 4.6, it can be seen that the overshoot of the system output is a quite high with settling time is 100 seconds.

45 90 80 70 Temperature(Celcius) 60 50 40 30 20 10 0 0 50 100 150 200 250 Time(sec) Figure 4.6: Temperature process response from simulation with PID controller Even though the PID controller is widely used in industrial process, the tuning of PID parameters is a crucial issue in particular for the system s characteristic which has large time delay and high order system [Underwood, 2000]. Commonly in industrial process, only an expert or experience workers are able to monitor and tune the PID parameters based on their experience. Therefore, in certain cases where there is deficient of experience with the processes, it is sometimes quite impossible to achieve a satisfactory performance. For these reason, it is desirable to introduce other types of controller such as an artificial conventional Fuzzy logic controller (FLC) or Single input fuzzy logic controller (SIFLC). Figure 4.7 shows the Simulink block diagram of the system with fuzzy controller. There are two scaling factors at the input and one scaling factor at the output of conventional FLC. Step input is performed in order to obtain the output response of the system.

46 Figure 4.7: Simulink block of the system and conventional FLC The control surface of the conventional FLC is shown in Figure 4.8. This control surface represents the correlation between input and output in threedimensional plot. From Figure 4.8, it is clearly shown that conventional FLC is behaves as linear controller. 8 6 4 2 output1 0-2 -4-6 -8-10 -5 0 5 10-10 0 10 input1 input2 Figure 4.8: Linear control surface of Conventional FLC

47 70 60 50 Temperature(Celcius) 40 30 20 10 0 0 50 100 150 200 250 300 350 400 450 500 Time(sec) Figure 4.9: Temperature process response from simulation with FLC Figure 4.9 shows the output response of conventional FLC with small overshoot. Although the output response has less overshoot, this approach take a longer computation time (95 seconds) to accomplish the controller algorithm. In fuzzy control, the computation time is depends on the number of rules used. More rules will result the longer computation time. This problem can be solved by replacing the conventional FLC into Single-input FLC (SIFLC), where there are no rules at all. In this approach, the rules are computed into constant number using a specific equation and will be performed using Look-up Table as shown in Figure 4.10. Figure 4.10: Simulink block of the system and SIFLC

48 Figure 4.11 shows the output of the system with SIFLC where there is small overshoot. As shown in Figure 4.11, the SIFLC control performance (in terms of output results) is almost the same as the FLC controller in Figure 4.9. However, it is obvious that SIFLC provides much better performance in computation time, which is less than 1 second for the same computation that took FLC 95 seconds. This comparable performance is achieved by reducing the number of rules from 49 rules in FLC to 7 rules in SIFLC. 70 60 50 Temperature(Celcius) 40 30 20 10 0 0 50 100 150 200 250 300 350 400 450 500 Time(sec) Figure 4.11: Temperature process response from simulation with SIFLC 4.3 On-line control with real VVS-400 In the previous section, three types of controller have been designed via simulation. However, it was not enough to ensure that all the design controllers are exactly capable to control the VVS-400 system model until it was implemented to perform an online control. This real system implementation is done using Real Time Windows Target (RTWT) toolbox in Matlab. Two blocks called Analog Output and Analog Input from RTWT connect the Simulink Matlab to the VVS-400 plant using data acquisition (DAQ) card PCI-1711. The controller will respond to the online

49 process with 2 seconds sampling interval. The output of the controller will be fed into the Analog Output and the process output is generated from the Analog Input. Since only voltage is applicable in this RTWT toolbox, the output from the Analog Output need to be converted into temperature by multiply with constant, 20 as given in the previous section. The simulink block diagram of the system with PID, conventional FLC and SIFLC controller are represented in Figure 4.12, 4.14, and 4.16, respectively. The system output with PID, conventional FLC and SIFLC controllers are shown in Figure 4.13, 4.15 and 4.17, respectively. However, to satisfy the output, tuning parameter requires a little adjustment since the simulation tuning parameter is designed based on the approximated plant. Figure 4.12: Simulink block diagram of real plant implementation with PID 70 60 50 Temperature(Celcius) 40 30 20 10 0 0 50 100 150 200 250 300 350 400 450 500 Time(sec) Figure 4.13: Temperature process response from experiment with PID

50 Figure 4.14: Simulink block diagram of real plant implementation with conventional FLC 70 60 50 Temperature(Celcius) 40 30 20 10 0 Figure 4.15: Temperature process response from experiment with conventional FLC

51 Figure 4.16: Simulink block diagram of real plant implementation with SIFLC 70 60 50 Temperature(Celcius) 40 30 20 10 0 0 50 100 150 200 250 300 350 400 450 500 Time(sec) Figure 4.17: Temperature process response from experiment with SIFLC By comparing the Figures 4.15 and 4.17, it shows that SIFLC capable to provide almost similar result as conventional FLC with less number of rules. The computation time for SIFLC is 974 seconds which is less than FLC (1002 seconds).

52 90 80 70 FLC PID Input SIFLC Temperature (Celcius) 60 50 40 30 20 10 0 0 50 100 150 200 250 300 350 400 450 500 Time(Sec) Figure 4.18: Temperature process response from simulation 70 60 PID SIFLC FLC Input 50 Temperature(Celcius) 40 30 20 10 0 0 50 100 150 200 250 300 350 400 450 500 Time(sec) Figure 4.19: Temperature process response from real VVS-400 plant The overall system outputs are shown in Figures 4.18 and 4.19. The PID controller gives high overshoot in the simulation result compare to FLC and SIFLC. In contrast, in the online implementation with a real VVS-400, PID controller has less overshoot compare to FLC and SIFLC after re-tuning. The FLC and SIFLC produced almost similar result with SIFLC has less computation time than FLC.

CHAPTER 5 CONCLUSION AND SUGGESTIONS 5.1 Conclusion In this project, several important ideas has been materialize such as the model of the system (VVS-400), the design of PID, FLC and SIFLC controllers for the control of Single-input single-output (SISO) system and the experiment design for open-loop system identification. The pilot scale of heating and ventilation VVS-400 plant has been successfully modeled by ARX model structure using System Identification toolbox in Matlab. The PID, conventional FLC and SIFLC controllers are developed on this plant which are not only designed by an approximated model plant but also have been implemented to the VVS-400 plant. From this study, it can be clearly seen that SIFLC is better than FLC with respect to the computation time due to the number of rules that can be significantly reduced. Though, both controllers produced almost similar results, the computation time is also considered as vital part of choosing suitable controller. 5.2 Future works 5.2.1 The closed-loop system identification is suggested for further improvement since it may offers a few advantages compared to open-loop system identification. The model obtained in this thesis from open-loop data system identification can be used as reference to evaluate the closed-loop system identification approach.