Fault Diagnosis of Rolling Bearing Using Convolutional Denoising Autoencoder and Siamese Neural Network With Small Sample

Abstract

Bearing fault diagnosis is critical for ensuring mechanical reliability and operational safety. Industrial Internet of Things (IIoT) sensors provide real-time monitoring data, advancing research in data-driven approaches to bearing fault diagnosis. However, current studies overlook two key challenges: 1) susceptibility to noise interference during fault signal acquisition and 2) the scarcity of fault data for effective diagnostic tasks in practical scenarios. To address these issues, this article proposes a novel method termed convolutional denoising autoencoder and siamese neural network (CDAE-SNN) for fault diagnosis in rolling bearings. This method is designed to be robust against noise and applicable in scenarios with limited data. Initially, Gaussian white noise is added to raw signals to simulate noisy signals encountered in real operating conditions. Subsequently, a convolutional denoising autoencoder (DAE) is constructed and optimized. The encoder in CDAE compresses feature information from samples into a lower dimensional space, while the decoder reconstructs signals to mitigate noise effects. Denoised signal sample pairs are then fed into a 2-D convolutional neural network-based siamese network to generate embedding vectors. Fault classification of rolling bearings is performed based on similarity metrics between sample pairs. Experimental results confirm the enhanced diagnostic accuracy of our proposed model across various signal-to-noise ratios and sample sizes. Furthermore, the model exhibits superior performance in classifying faults across diverse proportion of new categories.