Multi-source heterogeneous data fusion for fault diagnosis using a hierarchically optimized LSTM-2D-CNN framework

Abstract In the context of Industry 4.0, the reliable operation of rotating machinery such as rolling bearings is critical to industrial safety and efficiency. Traditional fault diagnosis methods typically rely on single-source sensor data and manual feature extraction, which are inadequate for capturing complex fault patterns under varying operational conditions. To address these limitations, this paper proposes a novel multi-source heterogeneous data fusion framework that couples temporal dependency modeling with cross-source feature fusion to improve robustness under various working conditions. The novelty of the proposed framework lies in three key designs: (i) shared long short-term memory (LSTM)-based extractors learn temporally discriminative embeddings from each source to capture long-range dependencies in multi-source time-series signals; (ii) a structured two-dimensional representation jointly encodes the raw multi-source signals and the extracted embeddings, reducing information loss and providing an effective layout for convolutional aggregation; (iii) a hierarchical optimization strategy with an upstream classification-guiding objective decouples representation learning and deep fusion, mitigating gradient interference and improving training stability and feature separability. Validation of method effectiveness on a self-built test platform demonstrates that the proposed method achieves superior diagnostic accuracy, robustness, and noise immunity across multiple rotational speeds and fault types. With an average classification accuracy of 97.46% and stable performance under low signal-to-noise ratios, the framework significantly outperforms existing state-of-the-art methods. Ablation studies further confirm the essential roles of both the LSTM-based feature extractor and the convolutional neural network-based fusion module in achieving reliable fault diagnosis. This work provides a viable and promising solution for intelligent fault diagnosis in real-world industrial applications.