MAC‐MLSTM: A Dual‐Branch Spatiotemporal Model With Adaptive Fusion and Uncertainty Quantification for RUL Prediction

Abstract

Remaining useful life (RUL) prediction is a key component in ensuring equipment reliability and optimizing maintenance strategies. To address the limitations of existing models that overlook the dynamic correlations and heterogeneous contributions of spatiotemporal features in multi‐sensor data, as well as the lack of quantitative assessment of prediction uncertainty, this study proposes a dual‐branch spatiotemporal network with an adaptive weighting unit (MAC‐MLSTM). The model integrates a multi‐scale attention convolutional neural network (MA‐CNN), which combines a mogrifier long short‐term memory (MLSTM) network and a convolutional block attention module (CBAM) to achieve temporal feature extraction and spatial adaptive feature selection, respectively. The adaptive weighting unit dynamically fuses these two types of features based on task relevance, while the integrated uncertainty‐quantification module provides reliable confidence intervals for RUL prediction. Extensive experimental results on the C‐MAPSS and N‐CMAPSS datasets demonstrate that MAC‐MLSTM not only surpasses existing state‐of‐the‐art methods in prediction accuracy but also provides uncertainty intervals for the predicted results.