In the process of structural health monitoring, effective fault diagnosis of the rotary machinery is crucial to ensure safe and reliable operations. The majority of rotary machinery, however, typically operates under varying operational conditions and is prohibited from operating in the fault state. The fluctuation in these conditions results in a corresponding variation in the distribution of samples, and the difficulty of collecting fault data leads to an imbalance in the number of health and fault samples. Transfer learning provides a promising tool for handling the cross-domain diagnosis problems. However, the potential of transfer learning techniques in motor current analysis as a cost-effective alternative has not yet received sufficient attention." In this article, a novel deep domain adaptation approach called Fault Feature Proxy Transfer (FFPT) is implemented to facilitate knowledge transfer across different operating condition domains in the presence of imbalanced class distribution. The proposed end-to-end scheme automatically learns the fault features from the raw current signals and transfers the fault feature explicitly both considering marginal and conditional distribution discrepancies. In FFPT, a set of proxies is employed to facilitate the transfer of features by establishing an intermediate domain. We introduce the pseudo-labeling-based metric learning to alleviate the problem of class imbalance between health and damage samples. A high accuracy for single and combined faults, a great robustness against class imbalance, and a significant gain in performing knowledge transfer prove the competitive performance of the proposed domain adaptation approach.

Reaching reliable decisions on equipment maintenance is facilitated by the implementation of intelligent fault diagnosis techniques for rotating machineries. Recently, the Transformer model has demonstrated exceptional capabilities in global feature modeling for fault diagnosis tasks, garnering significant attention from the academic community. However, it lacks sufficient prior knowledge regarding rotation invariance, scale, and shift, necessitating pre-training on extensive datasets. In comparison, contemporary convolutional neural networks exhibit greater ease of optimization. This limitation becomes particularly evident when applying the Transformer model in fault diagnosis scenarios with limited data availability. Moreover, the increasing the number of parameters and FLOPs. Pose a challenge to its suitability for mobile services due to the limited computational resources available on edge devices. To mitigate these issues, this paper introduces a novel lightweight Transformer (SepFormer) based on separable linear self-attention (LSA) for fault diagnosis task. The SepFormer performs a novel sequence-level feature embedding to better leverage the inductive bias inherent in the convolutional layers. Furthermore, it integrate a novel separable LSA mechanism into the Transformer architecture, effectively mitigating the computational burden concerns and significantly enhancing the training convergence speed. Extensive experiments are conducted extensively on a bearing fault dataset and gear fault dataset. The experimental results demonstrate that the SepFormer achieves a top-1 accuracy exceeding state-of-the-art approaches by more than 5%, while utilizing the fewest FLOPs. Moreover, the optimizability of SepFormer surpasses that of CNN, ensuring its superior preservation of inductive bias.

Bearing fault diagnosis is essential for ensuring the reliability and operational efficiency of rotating machinery. Traditional diagnostic methods often rely on vibration acceleration signals, which require intrusive hardware for signal acquisition and struggle to detect specific fault types, such as cage faults. Because bearings are often located deep within mechanical transmission systems, their vibration characteristics do not propagate effectively to external monitoring points, making vibration acceleration sensors less sensitive to certain faults. To address these challenges, we propose CAFormer, a complex attention Transformer for bearing fault diagnosis is based on motor current signals. In traction motor control, the Clarke and Park transforms are commonly used to convert three-phase ac signals into the dq coordinate system, where D-axis and Q-axis currents are represented as orthogonal signals. In this article, we first propose a complex attention to process dual-channel orthogonal data and enhance fault feature separation. Second, to address the imbalance between fault and healthy samples, we propose a new loss function that assigns adaptive optimization gradients to each sample, thereby accelerating model convergence. We conducted extensive experiments on two datasets: the Paderborn University dataset and a bearing dataset generated using a mechanical transmission simulator. Our experiments demonstrate that CAFormer significantly reduces computational costs while achieving state-of-the-art performance, surpassing existing methods in both fault classification accuracy and adaptability to imbalanced datasets.

The accurate diagnosis of bearing faults is crucial for ensuring the safety and reliability of rotary mechanical systems. In practice, the data size of healthy bearing samples is significantly larger than that of faulty samples. As a result, numerous data-driven models encounter challenges in effectively implementing fault diagnosis. The majority of scholars approach this situation primarily from the perspective of class imbalance, while overlooking the effective utilization of these extensive healthy samples. To effectively harness healthy samples which can be easily collected, this article proposes a new self-supervised learner Wavelet Packet Masked Transformer (WPMFormer) for fault diagnosis. In the generative self-supervised learning paradigm introduced in this article, the utilization of abundant healthy samples facilitates the extraction of robust signal representations and enables analysis of potential relationships among their sub-sequences. This leads to obtaining a pre-trained model that possesses initialization parameters with strong generalization capabilities. To maintain the original neighborhood structure and provide more details for the signal reconstruction process, we first perform the wavelet packet decomposition in the signal pre-processing stage, enabling the discretization of the original signal into multiple sub-sequences. Second, we employ a substantial proportion of sub-sequence masking. It results in a significant and meaningful self-supervised task. Third, we propose an innovative reconstruction loss function that prioritizes training on data pairs with higher significance and mitigates the potential dominance of larger-value data points during the restructuring. The effectiveness of the proposed model and other tricks are fully validated on the bearing dataset generated by railway motors experiment rig and a public bearing dataset. The experimental results demonstrate that the WPMFormer allows for learning a high-capacity model that generalizes well and outperforms other state-of-the-art methods in downstream tasks.