Multi-Information Fusion Fault Diagnosis of Bogie Bearing Under Small Samples via Unsupervised Representation Alignment Deep Q-Learning

With the ever-accelerating development of information and sensor technology, plenty of data-driven fault diagnosis algorithms have shown impressive performance. However, in practical engineering scenarios, it is often difficult to obtain sufficient labeled data, and the signal fault characteristics of the bogie bearing are not obvious due to the influence of environmental noise and track excitation, which would seriously reduce the performance of the model. In this article, we proposed a diagnostic algorithm based on a multi-information fusion technique and unsupervised representation alignment deep Q network (URADQN), to make full use of the fault information contained in each sensor and address the issue of the small sample size. First, the acquired samples from several sensors are processed by a dual fusion module to dig the fault feature information from the original data; then through unsupervised contrastive learning of similar feature representations, the URADQN is pre-trained to optimize the feature representation ability, to shorten the training time of deep Q-learning; finally, the fault classification is realized by using the well-trained URADQN through reinforcement learning (RL) strategy. The effectiveness of the proposed algorithm is validated through two multisensor-bearing datasets on the bogie. In comparison with other advanced methods under various data setups, the proposed algorithm demonstrates superior performance for small-sample fault diagnosis.