Structured multimodal feature fusion for rolling bearing RUL prediction using a lightweight TCN-CMV network

Abstract Accurate prediction of the remaining useful life (RUL) of rolling bearings is critical for the reliable and safe operation of rotating machinery. Existing methods, however, often suffer from insufficient multimodal feature fusion, limited ability to model long-term temporal dependencies, and reduced robustness under noisy or variable operating conditions, making them inadequate for real-time industrial monitoring. To address these limitations, this study proposes a novel lightweight deep learning framework—TCN-CMV—that introduces a parallel dual-branch architecture to simultaneously capture local multi-scale degradation features and long-term temporal dependencies. The model integrates structured multimodal feature fusion, combining manually extracted statistical, frequency-domain, and complexity features with deep representations, thereby enhancing both prediction accuracy and interpretability. Comparative experiments on the XJTU-SY and FEMTO bearing datasets demonstrate that TCN-CMV significantly outperforms conventional CNN-, LSTM-, and Transformer-based approaches in terms of accuracy, robustness, and computational efficiency, providing an effective solution for real-time health management of industrial equipment.