A data-driven fault diagnosis method for rolling bearings: a model based on MorletConv CNN interpretability
TL;DR
The validation results on multiple typical bearing fault datasets show that the proposed MorletConv CNN model is characterised by enhanced physical interpretability and generalisation ability while maintaining high diagnostic accuracy, providing new ideas and method support for achieving highly reliable rolling bearing fault diagnosis.
Abstract
As a key component in mechanical equipment, the operating status of rolling bearings directly affects the stability and performance of the entire system. Existing mainstream fault diagnosis methods using vibration signals are primarily based on ‘black box’ models. Although they are characterised by high recognition accuracy, the decision-making basis and interpretability are insufficient, making it challenging to meet the needs of high-reliability application scenarios. Therefore, an interpretable Morlet convolutional neural network (MorletConv CNN) model based on Morlet convolution is proposed in this paper to reveal the inherent mechanism by which the model identifies fault features. It is difficult to extract features adaptively with traditional signal processing methods. Furthermore, the decision logic of convolutional neural networks (CNNs) is opaque. As a result, a Morlet preprocessing layer, embedded into the CNN structure, is designed in this paper. Combined with prior knowledge of time-frequency analysis, the proposed layer enables the model to learn key structural features in vibration signals more effectively, improving the accuracy and reliability of fault recognition. Specifically, the proposed model utilises two convolutional kernel branches, one for the real part and one for the imaginary part, to enhance sensitivity to minor signal changes and improve its ability to detect early faults. Simultaneously, the control parameter θ of the Morlet kernel function is set as a trainable variable to reduce model complexity and accelerate training. This function automatically aligns key frequency components during the optimisation process, improving the time-frequency feature extraction ability of the model. The validation results on multiple typical bearing fault datasets show that the proposed MorletConv CNN model is characterised by enhanced physical interpretability and generalisation ability while maintaining high diagnostic accuracy, providing new ideas and method support for achieving highly reliable rolling bearing fault diagnosis.