An early bearing fault warning method based on Gaussian mixture model, KL divergence, and transformer

Abstract Providing timely and accurate early warnings for bearing faults in mechanical systems under various operating conditions is with great importance for practice. However, due to the complex nonlinear nature of bearing degradation and the scarcity and difficulty in obtaining fault data, it is essential to explore a method that can provide timely and accurate early warnings for bearings in operation based solely on fault-free samples. To address this issue, this paper proposes a novel bearing fault early warning method. First, features are extracted using the Gaussian mixture model and Kullback–Leibler divergence, followed by label construction through discretization. Next, a new dataset is generated by using percentiles to extract short sequences corresponding to the labels. Finally, a modified transformer model, adapted to the dataset, is employed to predict deviations as an early warning indicator. When the deviation exceeds the warning threshold, the system is considered to be in an anomalous state. To validate the performance of this method in bearing fault prediction, experiments were conducted on both publicly available datasets and experimental data. The results demonstrate the effectiveness of the proposed method.