Low Power Wireless Sensor Networks

Transcription

1 Low Power Wireless Sensor Networks Siamak Aram DAUIN Department of Control and Computer Engineering Politecnico di Torino Ph.D. Dissertation Advisor: Prof. Eros Pasero February 27 th,

2 DET Neuronica LAB The Communication and Signal Processing Laboratory The Informatics Systems and Applications Group Department of Mechanical Engineering, Aristotle University of Thessaloniki (AUTh) 2

3 Taxonomy of approaches to energy savings in sensor networks (1) 3

4 Taxonomy of approaches to energy savings in sensor networks (2) Mobility- Based Duty Cycling Data Driven Less Communication Less Power Consumption The microcontroller can switch on the sensors only during the measurement, reducing the power consumption. Nevertheless, unneeded communications could sporadically happen because of transferring unnecessary data. Reducing extra communications is a way to save energy which can be followed by data driven techniques. Communication component of a sensor consumes more power than the computational unit That power consumption is minimum in the sleep state of the radio communication Combination of the approaches The rationale behind the method implies that sensors could be powered down in judiciously chosen time intervals to read required data The correlation among the points would allow prediction of sensed data during sensors idle periods.

5 Environmental Sensing using Smartphone [1,2,3] Duty Cycling Smartphones Capabilities Sensing features Communication features Bluetooth Power: 0.6mW sleep mode and 90mW during transmission This work was sponsored by: National projects AWIS (Airport Winter Information System) ITACA (Innovazione Tecnologica, Automazione e nuovi Controlli Analitici per migliorare la qualità e la sicurezza dei prodotti alimentari piemontesi) Both funded by Piedmont Authority, Italy, and by the private company Reply Sensor SHT11 Accuracy: 0.4% C and 3% Power: 80µW 5

6 90mW 26mW Low Power Acquisition System Duty Cycling 3mW 24µW Application Improvement abluesen Tokenizing Buffer Size Automatically communicating 4 Months Reducing Power Consumption Increasing number of Sensors Data Set Sensor Structure SHT21 with higher accuracy; 0.3% C and 2% Sensor Itself Bluetooth 6

7 Neural Data Driven approach (1) [4,5,6] Data Prediction Data Reduction Algorithm Uncertainty of prediction (U) = Dispersion of prediction (U1) + Prediction error (U2) Each available measurement with its uncertainty assumed to be equal sensor accuracy An additional measurement when uncertainty > selectable threshold Forward prediction by periodically using MLP Prediction computed as mean of 100 estimations Predicted and estimated uncertainty of prediction 7

8 02 Simulated Data (2) Two Signals Uncorrelated Sinusoid and Lorenz System Experimental Data (2) Indoor environmental information with three different locations in Neuronica LAB 04 Experimental Data (1) Meteorological data of Turin Caselle Airport for 100 days Simulated Data (1) Two Signals Deterministic Noise Free Quantization Resolution:

9 Result First Approach Lower slope Application of the method to non-stationary, correlated signals (A) Relation between reduction ratio and error (100 simulations with different thresholds are considered). (B) Samples for the portions of the signals with higher frequency versus those with lower frequency (same 100 simulations as in A). (C) Representative example application for the method. 9

10 Result First Approach Application of the algorithm to simulated data (A) Number of samples and mean estimation error (mean and standard deviation over ten repetitions). (B) Representative example (threshold = 0.03 for both signals). 10

11 Result First Approach Application of the algorithm to meteorological experiments Accuracy is assumed to be 0.2 C, 20 hpa, 0.1 km/h, and 1%, for the temperature, pressure, wind velocity, and humidity sensors, respectively. (A) Root mean square estimation error and reduction ratio as functions of the uncertainty threshold (20 repetitions are considered). (B) Example of application to a portion of the test set. 11

12 70% Validation 15% Training Test 15% Two Delayed feedback The network uses the temperature values at two delayed time-stamps to predict the current value y ( t ) F y ( t 1), y ( t 2),.., y ( t d ) Neural Data Driven approach (2) [7] Time Series Data Prediction Data Reduction 03 Algorithm 04 One Hidden Layer For the hidden layer, sigmoid activation function is used whereas linear function is applied at output neuron. Regularization Every iteration, regularized cost for the training data is calculated Training Algorithm 02 m 1 i i i i J ( ) - y log( P( x )) (1- y )log(1- P( x )) m 2m i 1 j 1 The Levenberg-Marquardt (LM) algorithm as the training algorithm of the classifier 01 n 2 j 06 Stopping Training 05 Training is set to be stopped if either there are six consecutive increasings in validation error or the gradient becomes less than the selected threshold Artificial NN for Time Series Prediction We used an Artificial Neural Network (ANN) by employing NAR model for time series prediction 07 Optimizing Neurons and Performance 12

13 Neural Data Driven approach (2) Time Series Data Prediction Data Reduction Block Diagram of overall Methodology 13

14 Result Second Approach Time series prediction response of the neural network with the error by varying the size of hidden layer of the network. Mean Squared Error plot for each of the sensor's data against different number of hidden layer neurons used in the network 14

15 Result Second Approach Network prediction response with 20 hidden neurons by varying the number of inputs, Temperature sensor 1, dataset 1 Mean Squared Error plot for each of the sensor's data against different number delayed outcomes used as input to the network MAPE plot for each of the sensor's data against different number of network inputs 15

16 NAR NAR was selected which performs time series prediction by using the target values at subsequent delayed time stamps as inputs, and predicts the value at the current time stamp. Result Second Approach Performance Based on MSE with the aforementioned data to estimate a good network architecture with optimum choice of hidden layer neurons. Dara Driven Time Series Optimization By optimizing number of neurons and network size Low error percentage By optimizing system specially hidden layer neurons 16

17 Overview of the Works In each step we achieved greater reduction of communication and, subsequently, lower power consumption at small error rate. Duty Cycling 01 Step Data Driven 02 Step Time Series Data Driven 03 Step 17

18 Communication 66% While 33.3% is the prediction time. It is possible to reduce the commination time. Conclusion 3.6% Error margin - Temperature 75% power can be saved within an error margin 2.6% by network + 1% of the sensor 65% Less than 4% 3.2% Error margin Humidity 66% power can be saved within an error margin 2.2% by network + 1% of the sensor 7.5% - 10% More than 4 months 1 20% - 35% Power Error 18

19 Extra Activities In AuTh, HU and Polito POLITO Representative Assistantship Howard university Reviewing Journal Paper Two Journals Co-Adviser Undergraduate students thesis Collaboration POLITO and HU 19

20 Publications [1] S. Aram, A. Troiano, E. Pasero, "Environment Sensing using Smartphone", IEEE International Conference on Sensors Applications Symposium (SAS), Brescia, Italy,2012, DOI /SAS [2] S. Aram, A. Troiano, F. Rugiano, E. Pasero, "Low Power and Bluetooth-Based Wireless Sensor Network for Environmental Sensing Using Smartphones", International Workshop on Artificial Intelligence Applications and Innovations (AIAI), Halkidiki, Greece,2012, DOI / _34 [3] S. Aram, A. Troiano, F. Rugiano, E. Pasero Mobile environmental sensing using smartphones. In: Eren, H. and Webster, J.,Measurement, Instrumentation, and Sensors Handbook, Second Edition: Electromagnetic, Optical, Radiation, Chemical, and Biomedical, Measurement. Boca Raton, FL: CRC Press, Ch. 73, th January [4] S. Aram, L. Mesin, and E. Pasero, Improving lifetime in wireless sensor networks using neural data prediction, in 2014 World Symposium on Computer Applications & Research - IEEE International Conference on Information and Intelligent Systems Jan 2014, Sousse. [5] L. Mesin, S. Aram and E. Pasero, A neural data-driven approach to increase wireless sensor networks' lifetime, in 2014 World Symposium on Computer Applications & Research IEEE International Conference on Artificial Intelligence,18-20 Jan 2014; Sousse. [6] L. Mesin, S. Aram and E. Pasero, A neural data-driven algorithm for smart sampling in wireless sensor networks, EURASIP Journal on Wireless Communications and Networking, Vol. 23 No. pp. 1-8, [7] S. Aram, I. Khosa and E. Pasero, Conserving Energy Through Neural Prediction of Sensed Data, accepted but not published yet in Journal of Wireless Mobile Networks, Ubiquitous Computing, and Dependable Applications [8] S. Aram and H. Salmani and E. Pasero, Implanted Medical Devices Networks Challenges and Vulnerabilities: Survey, finalizing for submitting. 20

21 THANK YOU 21