Institute of Integrated Sensor Systems Dept. of Electrical Engineering and Information Technology Comparison of PSO-Based Optimized Feature Computation for Automated Configuration of Multi-Sensor Systems Kuncup Iswandy and Andreas Koenig Kuncup Iswandy and Andreas Koenig
Overview • Introduction • Feature Computation Methods • Optimization Methods • Experiments and Results • Conclusions Kuncup Iswandy and Andreas Koenig
Introduction Motivation � Intelligent sensor system find more widespread application, e.g., the fields of ambient intelligence and sensor networks. � The growing spectrum of available sensor principles and implementations require: � a large variety sensor electronics � sensor signal processing techniques � The design process goes through: � the principal steps of sensor selection and scene optimization, � choice of signal and feature processing, � dimensionality reduction, and � classification Kuncup Iswandy and Andreas Koenig
Introduction Motivation � Currently, optimizing design of intelligent sensor systems is - still manually elaborated by human designer - tedious - time and labor consuming process - potentially subpotimal outcome Kuncup Iswandy and Andreas Koenig
Introduction Goals Signal Sensor Processing Dimension & & Classification Reduction Scene Feature Computation Parameter Assessment Optimization General architecture of intelligent sensor system � Goals: � contribute to the design automation activities for intelligent (multi-) sensor systems � focus on optimization of feature computation � comparison between genetic algorithms and particle swarm optimization Kuncup Iswandy and Andreas Koenig
Feature Computation Methods � The roles of feature computation techniques: • extract the meaningful information of raw data of sensor response • reduce the dimension size of variable vector of a pattern • increase the computation speed and classification accuracy 4 x 10 2.5 cycle 1 cycle 2 � In particular to the application of 2 gas sensor systems, two feature 500°C 500°C 500°C conductance [a.u.] 1.5 H2 : 7 ppm. computations have been proposed: CH4 : 1000 ppm. Ethanol : 0.8 ppm. CO : 40 ppm � Multi-level threshold 1 290°C 290°C Gaussian windowing � 0.5 (kernel technique) 900°C 90°C 23°C 23°C 0 0 50 100 150 200 250 300 350 400 time [ms] Sensor response patterns during two temperature cycles Kuncup Iswandy and Andreas Koenig
Feature Computation Methods Multi-Level Threshold (MLT) � MLT computes the features with similar to histogram and amplitude distribution. � Two methods of MLT, i.e., differential (DM) and cumulative (CM) modes. � The features of MLT can be computed as N r ∑ = δ z y T T ( , , ) • i s p q = s 1 ≤ ≤ ⎧ T y T 1 • δ = 0 p s q y T T ⎨ ( , , ) s p q otherwise ⎩ y s : magnitude value of sensor signal with s = 1, 2, ..., N r : total samples of a pattern N r i : a number of features ( i = T - 1) T : a number of thresholds used Tp and Tq : level-values with q = (2, 3, ... T ) and p = q – 1 for DM, and Feature computation of MLT for a gas stimulus with q = T and p = (1, 2, 3, ... T -1) for CM presentation of first derivative of conductance Kuncup Iswandy and Andreas Koenig
Feature Computation Methods Gaussian Windowing Function � Extract features directly from conductance curves (transient responses) � Each feature is represented by a kernel-base, i.e., a Gaussian exponential function � Parameters : mean μ and standard deviation σ � The features of Gaussian Windowing can be computed as N = ∑ r ⋅ μ σ z y G s • ( , , ) i s i i = s 1 2 ⎛ − ⎞ μ s 1 ⎜ ⎟ − i ⎜ ⎟ μ σ = • G s exp σ ⎝ ⎠ ( , , ) 2 i i i Feature computaion of Gaussian windowing (window time slicing) for a normalized conductance curve Kuncup Iswandy and Andreas Koenig
Optimization Methods Genetic Algorithms for a real-valued case � GA procedure: 1. Generate randomly a real-valued population 2. Selection for mating processess using Roulette Wheel Selection 3. Discrete recombination with rate = 0.8, example: X = (x 1 , x 2 , x 3 , x 4 , x 5 ) ------> X´ = (x 1 , x 2 , y 3 , x 4 , x 5 ) Y = (y 1 , y 2 , y 3 , y 4 , y 5 ) ------> Y´ = (y 1 , y 2 , x 3 , y 4 , y 5 ) 4. Mutation with Gaussian perturbation and rate = 0.01, example: X´ = (x 1 , x 2 , ..., x s , ..., x i ) , 1 ≤ s ≤ i x snew = x s + α s N (0, σ s ) 5. Replacement: elitism 10 % of parent population and offspring individuals 6. Repeate to step 2; otherwise stop. Kuncup Iswandy and Andreas Koenig
Optimization Methods Particle Swarm Optimization for a real-valued case � One of the evolutionary computation techniques. � Population-based search algorithm � Population of random solution is called particles. � Related to the bird flocking and swarm theory. � PSO algorithm: ( ) ( ) + = + − + − • v t wv t C rand p x C Rand p x ( 1 ) ( ) () () id id id id gd id 1 2 p i + = + + x t x t v t • ( 1 ) ( ) ( 1 ) id id id d -space w = 1 : inertia weight x i_new C 1 = C 2 = 2 : positive constants p g Rand() and rand() : random functions, [0,1] v i v i_new : best previous position of the i- th particle p i p g : best particle among all particles x i Kuncup Iswandy and Andreas Koenig
Optimization Methods Feature Assessment - Nonparametric Overlap Measure � Inspired by probability estimation in Edited-Nearest Neighbor method � Used as an optimization criterion (fitness function) � NPOM is computed as: + ∑ k ∑ k q n N L 1 1 c ∑ ∑ = = i NN i i 1 1 = q ji o ∑ k L N n 2 = = c j = i c i 1 1 1 ω = ω ⎧ n d j i = i NN q ⎨ = 1 − with n and ji ω ≠ ω NN − n i ji d ⎩ j i i NN jk � q o is well normalized in [0,1]; „1“ indicates no overlap of class regions Kuncup Iswandy and Andreas Koenig
Optimization Methods Approaches of Feature Computation / Feature Selection � Wrapper approach: Assessment and param. R Modification Classification Dimension. Raw feature Classifier Sensor result computation Train/Test Reduction � Filter approach: Assessment and Assessment and Modification Modification param. param. q o q o Classification Raw feature Dimension. Classifier Sensor result computation Reduction Train/Test Kuncup Iswandy and Andreas Koenig
Optimization Methods Genetic Algorithms for a binary case • Representation: binary (switch variables) • One point Crossover; rate = 0.8 • Mutation; rate = 0.01 • Reproduction: best 10 % parents and offsprings Crossover Mutation Random Random Random P1 P P2 Ofs Ofs1 Ofs2 : 1 : 0 Kuncup Iswandy and Andreas Koenig
Optimization Methods Particle Swarm Optimization for a binary case � Original PSO is designed for real-value problems � For binary, particles are assigned a binary value, e.g., 0 or 1 � Velocity are restricted to an interval value of [-4,4] � BPSO algorithm: ( ) ( ) + = + − + − • v t wv t C rand p x C Rand p x ( 1 ) ( ) () () id id id id gd id 1 2 ( ) < ⎧ if U sigmoid v 1 , ( 0 , 1 ) = id • x ⎨ id otherwise ⎩ 0 , ( ) 1 = • sigmoid v id − + v e 1 id � Parameters w , C 1 , and C 2 are same with original PSO Kuncup Iswandy and Andreas Koenig
Experiments and Results Data Description and Parameter Setting � Applying a benchmark data of a gas sensor system � Types of gases, i.e., H 2 , CH 4 , ethanol and CO � The data set consists of 810 measure values and 264 patterns � Separated into training (144 patterns) and testing (120 patterns) sets � Each experiment is repeated using 10 runs � Each run is limited up to 100 iterations � Population size is 20 individuals for both GA and PSO. � The number of nearest neighbors is set to five for the overlap (NPOM) measurement and the k NN voting classifier. � The classification accuracy is estimated using holdout method and leave-one-out cross-validation approach. Kuncup Iswandy and Andreas Koenig
Experiments and Results Comparison between GA and PSO MLT differential mode Method overlap Recocognition accuracy ( k NN) q o train(%) test(%) test-LOO(%) Mean / Std Mean / Std Mean / Std Mean / Std 0.9950 / 0.0035 99.44 / 0.55 99.67 / 0.58 99.17 / 0.79 GA 1.00 / 0 100 / 0 100 / 0 99.83 / 0.35 PSO MLT cumulative mode Method overlap Recocognition accuracy ( k NN) q o train(%) test(%) test-LOO(%) Mean / Std Mean / Std Mean / Std Mean / Std 0.9878 / 0.0044 98.89 / 0.36 99.50 / 6.36 98.67 / 1.48 GA 0.9953 / 0.0024 99.10 / 0.34 99.92 / 0.89 99.83 / 0.35 PSO Kuncup Iswandy and Andreas Koenig
Recommend
More recommend
