A multi-scale cross-dimension interaction approach with adaptive dilated TCN for RUL prediction.

In the domain of Prognostics and Health Management (PHM) technologies, the focus of Remaining Useful Life (RUL) prediction is on the forecasting of the time to failure by uncovering the complex correlations between equipment degradation features and RUL labels, thereby enabling effective support for predictive maintenance strategies. However, extant research primarily emphasizes single-scale and single-dimensional feature extraction, which fails to adequately capture both long- and short-term dependencies as well as the interrelationships among sensor feature dimensions. This limitation has a detrimental effect on the accuracy and robustness of RUL prediction. To address the aforementioned issues, this paper proposes an Adaptive Dilated Temporal Convolutional Network (AD-TCN) approach, incorporating a Multi-Scale Cross-Dimension Interaction Module (MSCDIM) to enhance feature extraction and interaction. First, a dynamic adaptive dilation factor is incorporated into the TCN, thereby enabling the model to adjust its receptive field dynamically, which facilitates the capture of long- and short-term dependencies across different scales, allowing a more comprehensive representation of equipment degradation patterns. Second, the MSCDIM module has been designed to perform multi-scale feature extraction, interaction, and fusion across temporal and sensor dimensions, dynamically adjusting feature weights in order to suppress redundant information and enhance the representation of critical features. Finally, contrastive and ablation experiments are conducted on the widely used C-MAPSS dataset and N-CMAPSS dataset. And the experimental results demonstrate that the proposed method achieves high performance, with the MSCDIM module exhibiting strong adaptability and significant potential for broad applications.