Abstract: In recent a few years, some achievements have been obtained for artificial neural networks(ANN) in interpretable intelligent diagnosis of mechanical faults. However, the ANN model itself does not mimic the learning mechanism of biological neural networks, thus lacks biological interpretability. Spiking neural networks(SNN) can well simulate how biological signals are transmitted in the neural networks, which has good biological interpretability. However, the current spiking encoding manners have no physical interpretability. A model of wavelet kernel encoded spiking convolutional neural networks(WKE-SCNN) is proposed for bearing end-to-end interpretable intelligent diagnosis, which has both physical interpretability and biological interpretability. First, a wavelet kernel encoder is designed, in which multi-scale physical features are extracted from bearing vibration signals using wavelet kernel convolution, and spiking encoding information is obtained using spiking neurons. Then, a multi-layer spiking convolution feature extractor is constructed, which is used to extract deep-level state features from the spiking encoding information. Finally, a spiking classifier is established, which predicts the bearing health states according to fire rates of the spiking neurons in the output layer. Two groups of bearing datasets are utilized to verify the interpretability and effectiveness of the proposed model. Experimental results show that, the spiking encoding information can clearly reflect different health states of the bearings, thus has physical interpretability; the proposed WKE-SCNN can be trained in the end-to-end manner, and the fault diagnosis accuracy is comparative to the traditional convolutional neural networks(CNN), while the convergence stability of the proposed method is superior to the traditional CNN.

Key words: intelligent diagnosis, spiking neural networks, biological interpretability, physical interpretability, wavelet transform