A multi-source data fusion anomaly detection method for rolling bearings based on heterogeneous feature collaborative learning

Abstract Rolling bearings are key components in rotating machinery, and their health directly affects operational safety and productivity. Traditional single-source vibration analysis is constrained by limited information capture during early fault stages, reduced diagnostic accuracy under variable operating loads, and insufficient cross-modal validation for reliable decision-making. To address these limitations of single-source data and challenges in heterogeneous data fusion, we propose a heterogeneous feature collaborative learning framework (HFCLF). This framework uses a dual-stream architecture to process time–frequency spectrograms and multivariate time series separately. The image stream includes a ResNeXt-based multi-scale spatial feature extractor and a pyramid feature fusion module. The time-series stream uses a hierarchical long short-term memory (LSTM) encoder–decoder with a time-aware mechanism to assess temporal importance. A cross-modal alignment mechanism links semantic features across streams, enabling deep fusion of heterogeneous data. A multi-task learning strategy jointly optimizes fault classification and severity estimation. Experiments on the Case Western Reserve University (CWRU) benchmark dataset, the Paderborn University bearing dataset, and real industrial datasets show 95.8% fault detection accuracy and 86.7% early fault detection rate, outperforming existing deep models by 3.1% and 8.5%, respectively. Ablation studies with statistical significance testing confirm the effectiveness of each module, especially the impact of cross-modal alignment.