Integrating DWT and Bayesian Neural Networks for Effective Bearing Fault Detection With Uncertainty Evaluation in Induction Machines

Proper maintenance of Induction Machines (IM) is essential to ensure suitable performance, extend their service lifespan, and minimize the risk of catastrophic failures. Bearing failures are the most common failure mode in induction machines reported in industrial applications, and vibration signal monitoring has become an effective technique for their detection. However, accurate interpretation of these signals requires high expertise to differentiate between normal and faulty operating conditions. In this context, supervised artificial intelligence algorithms have emerged as promising tools for pattern classification in vibration signals. This study presents an innovative approach that integrates the energy levels of Discrete Wavelet Transform coefficients as inputs to train two classification models, a Fully Connected Neural Network (FCNN) and a Bayesian Neural Network (BNN), together with the uncertainty assessment of the inferences. The BNN, in particular, allows for the quantification of epistemic uncertainty, providing a measure of confidence in the predictions. The FCNN architectures were optimized using a random search algorithm in Python, while the BNN parameters were tuned through variational inference with normal distributions for the weights. Results demonstrated high accuracy performance, but the BNN model shows the lowest epistemic and aleatory uncertainty compared to the FCNN method. The results prove the ability to use BNNm models with high confidence in diagnostic and condition monitoring for IM applications.