A Bayesian Multi-input CNN for Induction Motors Fault Diagnosis

Fault diagnosis is critical for ensuring the high performance and reliability of induction motors. However, existing deep learning-based fault diagnosis methods require substantial amounts of training data which is impractical in real industrial scenarios. Additionally, as with classification problems, traditional deep learning-based fault diagnosis methods typically provide point-estimate predictions without offering confidence levels, leading to situations of overconfidence and underconfidence. To address these challenges, this paper proposes a novel induction motor fault diagnosis method that combines a multi-input convolutional neural net (MICNN) with a feature extraction model, further enhanced by Bayesian deep learning. The MICNN extracts features from two different signal domains, providing a more comprehensive understanding of vibration signals. Bayesian deep learning treats model parameters as random variables, enabling the estimation of a probability distribution over the model predictions. Experimental results demonstrate the effectiveness of the proposed method, achieving $15 \%$ higher diagnostic accuracy compared to traditional fault diagnosis methods under the conditions of limited data. Additionally, in noisy environments, the proposed method reflects the confidence of diagnostic results, enhancing model trustworthiness.