A feature extraction interpretability research of one-dimensional convolutional neural networks in mechanical fault diagnosis

Abstract

To address the “black-box” nature of one-dimensional convolutional neural networks (1D-CNNs) in mechanical fault diagnosis from a signal-processing perspective, a correspondence between the network and classical spectral analysis has been established. First, visualization of the convolutional layer weights before and after training revealed the evolution of their frequency responses, demonstrating that each convolutional kernel can be interpreted as exhibiting finite-impulse-response (FIR)-like temporal filtering behavior under fixed weights. Next, the nonlinear operation of the max-pooling layer was analyzed and found to satisfy the discriminative requirements of a binary classification task, allowing the activation function to be omitted after convolution. The influence of network depth on frequency resolution was then investigated, showing that layer-by-layer stacking of convolution and pooling effectively focuses on signal components near theoretical fault-feature frequencies-a mechanism highly analogous to the Fourier transform. Finally, by adjusting the number of layers such that the spectral decomposition range covers one to three times the fault-feature frequency, both feature-extraction precision and classification performance were significantly improved. Validation on the benchmark of rolling-bearing dataset and demonstrated that the proposed method not only maintains high accuracy but also markedly enhances model interpretability, offering a new approach for studying the “black-box” mechanism of 1D-CNNs and designing explainable intelligent diagnostic models.