A Transferable Multi-Scale Convolution Domain Adaptation Network for Bearings RUL Prediction

Transfer learning methodologies predominantly concentrate on adapting to varied operational conditions. However, disparities in probability distributions and the deterioration of performance characteristics within identical operational environments, coupled with the gradient vanishing problem caused by minimal or absent overlap between source and target probability distributions in high-dimensional spaces also detrimentally influence the accuracy of remaining useful life (RUL) predictions. A transferable multi-scale convolution domain adaptation network (TMSCDAN) is proposed to achieve feature distribution alignment and extract domain-invariant features under the same working conditions to mitigate this challenge. Furthermore, a health indicator (HI) is effectively derived through the utilization of a stacked convolutional auto-encoder (SCAE) in conjunction with the Pearson correlation coefficient. Subsequently, a multi-channel convolutional neural network is introduced to extract multi-scale degradation features. An improved domain adversarial structure based on Earth-Mover (EM) distance is applied to eliminate the gradient vanishing problem, learn cross-domain invariant features, and maximize the distance between the source and target probability distributions instead of the binary classification task in the traditional domain adversarial module. Finally, the experimental results, including ablation studies and comparative analyses, confirm the effectiveness, superiority, and generalizability of the proposed method.