Machine Learning for Antenna Array Failure Analysis

Transcription

1 Machine Learning for Antenna Array Failure Analysis Lydia de Lange Under Dr DJ Ludick and Dr TL Grobler Dept. Electrical and Electronic Engineering, Stellenbosch University MML 2019

2 Outline 15/03/2019 3

3 Introduction 15/03/2019 4

4 Antenna Arrays 5 15/03/2019

5 Reconstructed Sky Image 6 15/03/2019

6 Square Kilometer Array (SKA) 7 15/03/2019

7 Lydia s Arrays (LA) and Far-Field Patterns 8 15/03/2019

8 Problem Statement Element failure Inaccurate far-field patterns (beam patterns) Distorted results (e.g. in reconstructed sky image) Important applications: Array failure correction System health management of large antenna arrays

9 Previous work Failed antenna element detection and location possible with machine learning techniques e.g.: Feedforward neural networks Support vector models 11 15/03/2019

10 Methodology 15/03/

11 Methodology Simulate scenarios for input data Sampling methods Train FNN

12 Sampling Methods Name Number of Samples φ ( ) θ ( ) Single cut (φ = 0) θ [ 90, 90] Single cut (φ = 45) 45 Single cut (φ = 90) 90 Single cut (φ = 135) 135 Principle cuts 362 0, 90 Diagonal cuts 45, 135 All cuts 724 0, 45, 90, D pattern (182 samples) D far-field pattern sampled in a 3-D pattern (361 samples) 361 (θ, φ) grid. 3-D pattern (725 samples) 725

13 Training of FNN Multi-label feedforward neural network x Sampled far-field observation of 1 failure scenario y y = ON or OFF state of each antenna in the array multi-label 1 label for each antenna (25)

14 Training of FNN Adapt parameters β with each pass until f is as similar as possible to true relationship.

15 Results 15/03/

16 Nature of FNN*: Number of samples (S) Number of parameters (β) to be estimated Accuracy If accuracy : sampling pattern found a useful region in the 3-D far-field pattern to accurately identify failure scenarios * # training iterations = const. FNN Results 23 15/03/2019

17 FNN Results Dataset Samples Training Time (sec) Accuracy (%) 90ᵒ cut D pattern Diagonal cuts All cuts

18 Additional experiments 25

19 Additional Experiments Compared 14 other classification algorithms 1 according to accuracy using the 10 sampling method datasets. Best 4: FNN One vs Rest Classifier + Linear SVC One vs Rest Classifier + Logistic Regression One vs Rest Classifier + Logistic Regression CV x 1 Scikit-learn algorithms

20 ACCURACY (%) Additional Experiments Best: One vs Rest + Logistic Regression CV 100% accuracy achieved Number of parameters vs accuracy relationship is different 3-D sampling method contains more information than combined single cuts 27 15/03/ Classification Algorithm Comparison SAMPLING METHOD DATASETS FNN OvR+LinearSVC OvR+LogisticRegression OvR+LogisticRegressionCV

21 Conclusion 15/03/

22 Conclusion FNN used to detect and locate failed antenna elements in a bow-tie antenna array Investigated choice of training data on FNN accuracy and training time Diagonal cuts 90.91% accuracy, secs 3-D pattern (182 samples) 87.88% accuracy, secs On larger datasets with more scenarios, the difference in training time may become more significant. Additional work: Best algorithm: One vs Rest + Logistic Regression CV Best sampling method: 3-D pattern

23 Future work Manufacturing and measuring an antenna array with a spherical nearfield scanner! Look at SVMs Looking at other places in pipeline to do ML on: Power Spectral Density and Correlations

24 Acknowledgement The financial assistance of the South African SKA project (SKA SA) towards this research is hereby acknowledged ( 15/03/

25 References [1] R. J. Mailloux, Array Failure Correction With A Digitally Beamformed Array, IEEE Trans. Antennas Propag., vol. 44, no. 12, pp , [2] P. Hall, The Square Kilometre Array: An Engineering Perspective, Springer, [3] J. A. Rodrìguez, et al., A Comparison Among Several Techniques For Finding Defective Elements In Antenna Arrays, 2nd European Conference on Antennas and Propagation (EUCAP), pp. 1 8, [4] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning, MIT Press, pp , /03/2019