Engine Multiple Faults Detection base on Bispectrum and Convolutional Neural Network

A pattern recognition model is proposed based on bispectrum and convolutional neural network (CNN) to detect engine faults in multiple working conditions end-to-end. The vibration signal collected from the engine directly contains much noise, making it challenging to extract deep information. Bispectrum is employed to transform the one-dimensional time-dependent vibration signal into a two-dimensional time-frequency matrix. For efficiency, the bispectrum matrix size is optimized by interpolation. Four interpolation kernels are analyzed to hold more information, in which the triangle kernel shows the best performance. The batch normalization (BN) is introduced to optimize the over-fitting in the CNN with a small and complex dataset. Based on that, five CNN models with different structures are designed, in which the influences of layer number, channels, kernel size, and padding on recognition rate are analyzed synthetically. The best model with a recognition rate of 97.65% is obtained. The method researched in this paper could benefit the engineering application of deep learning in engine faults detection.