Deep learning methods for fault early warning of elevator traction systems with multi-source data fusion

To address the issues of significant differences in sampling frequencies, clock asynchrony among multi-source sensors, and the difficulty in detecting weak early-stage faults in elevator traction systems, a deep learning early warning framework based on multi-source data fusion is proposed. First, using the temperature signal as the time reference, resampling and operating condition segmentation are performed to align multi-source heterogeneous data, including vibration, acoustic emission, current, and displacement at a unified scale. Second, differentiated deep feature extraction methods are designed according to the physical characteristics of each signal. Then, a cross-modal spatiotemporal association network combining a bidirectional long short-term memory network with a multi-head attention mechanism is constructed to adaptively mine the temporal dependencies and spatial correlations among modalities. Finally, a strategy integrating cumulative anomaly degree with dynamic threshold is adopted to output three-level early warnings. The method is validated on six months of multi-source data collected from a real traction system experimental platform. Results show that it achieves a fault identification accuracy of 91.5% and an F1-score of 90.9%, with an average early warning lead time of 7.8 operating cycles. Moreover, under strong noise of 5 dB, the F1-score remains at 85.7%. The proposed method enables reliable early warnings of incipient faults in elevator traction systems, providing technical support for predictive maintenance.