Bayesian Deep Learning for Fault Diagnosis of Induction Motors With Reduced Data Reliance and Improved Interpretability

Abstract

Fault diagnosis holds significant practical importance for high performance and reliable control of induction motors. However, existing deep learning-based fault diagnosis methods demand a large amount of training data and lack interpretability, which can limit their reliability and practical usability in real-world applications. To address these challenges, this paper proposes a novel fault diagnosis method that equips a multi-input convolutional neural network (MICNN) with a feature extraction model and further enhances it with Bayesian deep learning. MICNN simultaneously extracts features from two distinct signal domains, enhancing the model's comprehensive understanding of vibration signals. Additionally, combining features from these two channels is analogous to acquiring information from two different data sources, effectively enriching the data and reducing the model's reliance on data. Bayesian deep learning treats model parameters as random variables, enabling quantification of diagnostic result uncertainty and enhancing model interpretability. The proposed method is capable of quantifying and decomposing the uncertainty under noisy environments and variable operating conditions, assisting decision-makers in understanding the uncertainty in diagnosis. Various validation experiments are conducted on an induction motor, and the results indicate that this method improves diagnostic accuracy by 15% compared to the baseline model under the condition of limited samples.