MSIF-Convformer: a novel end-to-end fault diagnosis framework with multi-source sensors under strong noise

In practical applications, reliable fault diagnosis of rotating machinery faces considerable challenges due to strong noise interference and the absence of effective multi-source information fusion strategies. To address these issues, this paper proposes MSIF-Convformer, a novel end-to-end multi-source information fusion framework that integrates Convolutional Neural Networks (CNNs) and Transformers, specifically designed for multi-source sensor data in noisy environments. At the data level, parallel CNNs feature extractors generate and encode time-frequency representations of vibration and acoustic signals. At the algorithmic level, a cross-attention mechanism enables efficient multi-source fusion, while the integrated CNNs-Transformer architecture captures both local and global dependencies. Global pooling further refines feature representations for robust classification. The proposed MSIF-Convformer is validated on two multi-source datasets: a bearing test bench and a rolling mill platform. Experimental results demonstrate that MSIF-Convformer achieves superior accuracy and robustness compared with state-of-the-art methods, particularly under low signal-to-noise ratio conditions. Visual analyses further confirm its effectiveness in distinguishing fault categories. Overall, this work provides an innovative and reliable solution for mechanical fault diagnosis in highly noisy environments and advances multi-source fusion strategies in intelligent health monitoring.