Weighted ANN input layer for adaptive features selection for robust fault classification

Abstract

Model based feature selection for identification of diverse faults in rotary machines can significantly cost time and money and it is nearly impossible to model all faults under different operating environments. In this paper, feedforward ANN input-layer-weights have been used for the adaptive selection of the least number of features, without fault model information, reducing the computations significantly but assuring the required accuracy by mitigating the noise. In the proposed approach, under the assumption that presented features should be translation invariant, ANN uses entire set of spectral features from raw input vibration signal for training. Dominant features are then selected using input-layer-weights relative to a threshold value vector. Different instances of ANN are then trained and tested to calculate F1_score with the reduced dominant features at different SNRs for each threshold value. Trained ANN with best average classification accuracy among all ANN instances gives us required number of dominant features.