High-speed railroads in the central and western regions of China are widely used in the design of long ramps. The train brake pads are frequently exposed to high temperatures and heavy loads, which leads to the aggravation of brake pad surface damage, and significantly reduces the braking efficiency of the train, posing a major potential safety hazard. Data-driven fault diagnosis of brake pad surface damage is a timely and effective method, but because of the mutual coupling and mutual interference between different surface damage, the frequency characteristics of compound fault signals are complex and difficult to recognize. To address the above problems, this paper proposes a multi-source self-attention deep reinforcement learning (DRL) network to realize the compound fault diagnosis in high-speed train brake pads on long ramps. First, the wavelet packet decomposition technique is utilized to deconstruct the features of fault data across different frequency bands, enhancing the representational separability between different fault modes, and then constructing the multi-source monitoring data into three-dimensional time-frequency feature tensors. Then, the multi-source information fusion features are extracted by the position-channel dual branching self-attention mechanism, sufficiently capturing the internal time-frequency features of various data sources and the information discrepancies between different data sources. Subsequently, the dynamic sensing and adaptive decision-making capabilities of the improved DRL network are utilized to achieve multi-label decoupling of compound faults. Finally, the proposed method is validated by the experimental measured data collected from the braking tests and public datasets. The results demonstrate that the proposed method significantly improves the recognition accuracy and robustness compared to other advanced algorithms, providing reliable assurance for the safe operation of high-speed trains.
Zhang et al. (Sat,) studied this question.