An Uncertainty-Aware Combined Multistable Stochastic Resonance Bayesian Neural Network for Gearbox Fault Diagnosis in Noisy Envi

    In industrial transmission systems, gearboxes often operate under strong noise and complex conditions, where weak fault information is frequently buried in background noise, directly affecting the safe operation and life-cycle management of equipment. Although traditional stochastic resonance methods amplify fault features by exploiting noise gain effects, existing work is often limited to single potential well models or ignores model uncertainty, resulting in narrow applicability and poor robustness. To address the difficulty of identifying weak gearbox faults in noisy environments, this paper proposes an uncertainty-aware combined multistable stochastic resonance\u2013Bayesian neural network (MSSR\u2013BNN) method. First, a stochastic resonance model containing triple-well and multi-well combinations is constructed to amplify weak fault features by adaptively adjusting the noise intensity and well depth. Second, a Bayesian neural network is introduced to quantify uncertainty by approximating the posterior through Monte Carlo dropout, and a gating mechanism is designed to fuse multiple model predictions to enhance generalization performance. Finally, comparative experiments and simulations are conducted on a public gearbox dataset and self-collected industrial data to verify the superior performance of the proposed method under noisy conditions. Experimental results show that the proposed MSSR\u2013BNN improves the average accuracy by $2\\text{--}7\\%$ and reduces the calibration error by $30\\%$ at different noise levels, while providing reliable confidence estimates, demonstrating better advantages than the comparison methods.