A Distributed Neurocomputing Approach for Infrasound Event Classification

Transcription

1 A Distributed Neurocomputing Approach for Infrasound Event Classification Fredric M. Ham, Ph.D., FIEEE Harris Professor of Electrical Engineering Director of the Information Processing Laboratory Florida Institute of Technology Melbourne, Florida (USA) January 5, 29

2 Acknowledgements: US Army Research Laboratory (ARL) Infrasound Laboratory University of Hawaii Dr. Milton Garces David Fee - For Providing Infrasound Data - - All the Students Working in the Information Processing Laboratory-

3 Classification of Infrasound Events Using Neural Networks F.M. Ham and I. Kostanic, Principles of Neurocomputing for Science and Engineering, McGraw-Hill, 2.

4 ARL Infrasound Signals: Artillery (Impulse) Rocket Shuttle

5 Three Infrasound Classes Rocket Class number 2 3 Event Impulse Rocket* Shuttle* Amplitude No. of SOI (Total of 53) No. of SOI used for Training (Total of 326) No. of SOI used for Testing (Total of 24) f s = Hz and 5 Hz for Impulse signals (3.27 sec/6.54 sec) f s = Hz for Rocket signals (25.9 sec) f s = Hz and 5 Hz for Shuttle signals ( sec/2 sec) * The number of training and test signals can vary because the Rocket and Shuttle signals were collected from 4- sensor and 5-sensor arrays

6 Impulse Three Infrasound Classes Shuttle Rocket Class number 2 3 Event Impulse Rocket* Shuttle* No. of SOI (Total of 53) No. of SOI used for Training (Total of 326) Rocket infrasound signals from: Nike-Orion, Orion, Black Brant, Terrier Orin, Nike Black Brant, Terrier MK7-Improved Orion No. of SOI used for Testing (Total of 24) * The number of training and test signals can vary because the Rocket and Shuttle signals were collected from 4- sensor and 5-sensor arrays

7 Sensor Array Geometries

8 Simulation Results Using the RBF Neural Network as the Classifier

9 Question: How Can Infrasound Signals be Represented as Feature Vectors, and then these Feature Vectors be used to Train and Test a Neural Network? Answer: Using Cepstral Coefficients and Associated Derivatives Cepstral Feature Vector Cepstral Feature Vector 2 Cepstral Feature Vector 3 Cepstral Feature Vector 4

10 Three Infrasound Classes.4.2 Shuttle

11 Three Classes Cepstral Feature Vectors (53 features) Amplitude Amplitude Amplitude

12 Cepstral Feature Vectors of Speech From: Mammone et al., Robust Speaker Recognition-A Feature-Based Approach, IEEE Signal Processing Magazine, Sept. 996, pp

13 Receiver Operating Characteristic (ROC) curve The ROC curve yields the optimal threshold value for each RBF NN. True positive (TP) (sensitivity) Reference point Closet point False Positive (FP) (-specificity) 2-D ROC curve 3-D ROC curve

14 3-D ROC Curves (RBF) Optimal threshold=.432 Optimal threshold= Optimal threshold=

15 Final Results (RBF): CCR: 97.55% ACC: 98.4% Correct Classification Rate (CCR) No. of correct predictions - No. of multiple classifications CCR = Total No. of predictions = p + s -No. of multiple classifications p + q + r + s Accuracy (ACC) No. of correct predictions ACC= Total No. of predictions p+ s = p + q + r + s Actual Confusion Matrix Positive Negative Prediction Positive p r Negativ e Where: p: is the number of correct predictions that an occurrence is positive (true positive) q: is the number of incorrect predictions that an occurrence is positive (false negative) r: is the number of incorrect of predictions that an occurrence is negative (false positive) s: is the number of correct predictions that an occurrence is negative (true negative) q s

16 Cluster Plot for RBF Results Impulse.5 Rocket Shuttle

17 Confusion Matrix for RBF Results Prediction Actual Impulse Rocket Shuttle unknown Total Impulse Rocket Shuttle ()* * Double classification with Impulse and Shuttle

18 Confidence Intervals (95%) for RBF Results

19 Simulation Results Using the PLS Calibration Model as the Classifier

20 Three Classes Cepstral Feature Vectors (44 features) 9 8 Shuttle Rocket

21 Selection of the Number of PLS Factors Optimal PLS Factors for Shuttle

22 3-D ROC Curves (PLS) 2 3-D ROC Curve for Impulse Optimal threshold= Optimal threshold=.448 Optimal threshold=.6995

23 Final Results (PLS): CCR: 97.56% ACC: 98.54% Cluster Plot for PLS Results.5 Impulse Rocket Shuttle

24 Confusion Matrix for PLS Results Prediction Actual Impulse Rocket Shuttle unknown Total Impulse 3 ()* 32 Rocket Shuttle ()* * Double classification

25 Confidence Intervals (95%) for PLS Results

26 Summary of Performance Results Final Results (RBF) CCR: 97.55% ACC: 98.4% Final Results (PLS) CCR: 97.56% ACC: 98.54% Impulse.5 Rocket Shuttle RBF

27 Summary of Performance Results-continued Mean for PLS 96.% Mean for RBF 95.7% PLS RBF

28 ISLA Infrasound Signals: Ocean surf at three locations on the North Shore of Oahu

29 Oahu Sensor Array

30 Parallel Classifier Bank For Classifying Ocean Waves

31 time time time generate cepstral feature vectors

32 Cepstral feature vectors (44 features)

33 3-D ROC Curves (RBF) Optimal threshold=.4533 Optimal threshold=.444 Optimal threshold=.5322

34 Final Classification Results: Cluster Plot CCR: 87.% ACC: 89.3 %

35 Confusion Matrix 95% Confidence Intervals CCR ACC % Result % Confidence Interval (min, max) (.8322,.96) (.8566,.929)

36 Fredric M. Ham, Ph.D., FIEEE Information Processing Laboratory Florida Institute of Technology Melbourne, Florida (USA)