Integration of Bayesian optimization into hyperparameter tuning of the particle swarm optimization algorithm to enhance neural networks in bearing failure classification

Bearings are a primary source of defects in induction motors (IM), requiring effective diagnostic methods. Neural Networks (NNs) are useful for this purpose, but optimizing their parameters is challenging. This study presents an alternative approach that modifies Particle Swarm Optimization (PSO) by integrating Bayesian optimization to dynamically tune PSO hyperparameters, enhancing the NN's ability to detect defects in IM bearings. Unlike traditional methods with empirically determined hyperparameters, this approach adapts to varying data conditions for better performance. The method was tested using vibration and current signals of different durations (2s, 1s, 0.5s, 0.25s) and torque ranges (0Nm to 22Nm) from a laboratory-generated dataset, with results compared to those obtained using other optimizers. The accuracy achieved was 92.57% for vibration signals at 0.25s and 97.23% across torque ranges. For current signals, the accuracy was 91.29% for 0.25s samples and 97.5% across torque ranges. © 2024 Elsevier Ltd