Bearing RUL prediction and fault diagnosis system based on parallel multi-scale MIMT lightweight model

In contemporary industrial production, the significance of safety is increasingly evident, as the degradation and failure of machinery and equipment pose potential safety hazards. Consequently, there is a growing trend toward real-time monitoring, prediction, and diagnosis of industrial equipment to mitigate the unpredictable impacts on life and property safety caused by sudden failures. To address this issue, this paper proposes a real-time degradation anomaly detection system based on parallel multiscale autoencoders, along with a lightweight model for parallel multiscale multi-input multi-task applications in bearing Remaining Useful Life (RUL) prediction and fault diagnosis. Firstly, the multiscale autoencoder method is employed to simulate actual working conditions and reconstruct original vibration signals, thereby establishing intervals for detecting abnormal degradation. The interval [0, µ +3 σ ] interval is utilized to assess abnormal degradation based on reconstruction errors, with the first Prediction Time determined adaptively. Secondly, a method for constructing dimensionless auxiliary datasets is introduced, which adopts a multi-input format based on deep separable convolution for feature extraction from original vibration signals, kurtosis, and peak values. This approach enhances the prediction and diagnostic performance of the lightweight model. Finally, a multi-task output method that combines clustering and regression is implemented to achieve RUL prediction and fault diagnosis for bearings. The proposed method addresses the limitations of traditional bearing RUL prediction and diagnosis techniques, demonstrating both theoretical innovation and practical engineering applications. Validation on two bearing datasets confirms the effectiveness of the proposed method.