Interpretable Neural Networks for Mechanical Fault Diagnosis in Strong Noise Condition

Deep learning has made significant progress in the field of mechanical fault diagnosis, but in real industrial scenarios, noise poses challenges to the accuracy and generalization of these models. Currently, some deep learning-based fault diagnosis methods take into account the effect of noise, but complex network structures and poor interpretability limit their practical application. Therefore, this work proposes a novel interpretable neural network for mechanical fault diagnosis in strong noise condition. First, considering the poor interpretability due to the complex structure of traditional deep neural networks, a feature mapping layer based on hybrid kernel function is proposed for mapping raw monitoring signals into high-dimensional space. Specifically, forward propagation and back propagation algorithms are designed for weight updating of the hybrid kernel function-based feature mapping layer. The exponential function and segmentation function are also introduced to constrain the boundaries of the weights. Secondly, a neural network based on multi-head attention mechanism is developed for high-dimensional feature learning. It can help the network adaptively select effective features and reduce the impact of noise on fault diagnosis performance. Experimental results show that the proposed method can adapt to mechanical fault diagnosis in strong noise condition, providing favorable accuracy and robustness. Notably, the proposed method has favorable feature learning transparency and interpretability.