Thanks to the development of deep learning, machine abnormal sound detection (MASD) based on unsupervised learning has exhibited excellent performance. However, in the task of unsupervised MASD, there are discrepancies between the acoustic characteristics of the test set and the training set under the physical parameter changes (domain shifts) of the same machine's operating conditions. Existing methods not only struggle to stably learn the sound signal features under various domain shifts but also inevitably increase computational overhead. To address these issues, we propose an unsupervised machine abnormal sound detection model based on Transformer and Dynamic Graph Convolution (Unsuper-TDGCN) in this paper. Firstly, we design a network that models time-frequency domain features to capture both global and local spatial and time-frequency interactions, thus improving the model's stability under domain shifts. Then, we introduce a Dynamic Graph Convolutional Network (DyGCN) to model the dependencies between features under domain shifts, enhancing the model's ability to perceive changes in domain features. Finally, a Domain Self-adaptive Network (DSN) is employed to compensate for the performance decline caused by domain shifts, thereby improving the model's adaptive ability for detecting anomalous sounds in MASD tasks under domain shifts. The effectiveness of our proposed model has been validated on multiple datasets.

The repetitive transient impulses are typical symptoms of rolling bearing faults. The indicator of second-order cyclostationarity (ICS2) based blind deconvolution (CYCBD) maximizing the cyclostationary behavior of the excitation source proves to be an effective method for enhancing the periodic cyclostationarity component caused by bearing defect. However, the deviation of the fault characteristic frequenc–a common phenomenon easily caused by roller creep or measurement systems, etc., can lead to the collapse of this method in practical applications. Additionally, this method is prone to induce the generation of pseudo-cyclostationary components under the guidance of an inappropriate cyclic frequency. Thus, this work proposes a candidate fault frequencies (CFFs) based blind deconvolution, abbreviated as CFFsBD, for fault feature enhancement of bearings. This idea comes from the concept of CFFs–a collection of frequencies most likely to be associated with bearing fault, that can be identified by mining the local time-frequency features of the signal to be analyzed. A new indicator constructed by using CFFs to replace the cycle frequencies in ICS2 is utilized as the criterion to guide the solution of the deconvolution filter. This new indicator is a generalized version of ICS2, and thus CYCBD can be considered a special case of the proposed CFFsBD. The performance of the CFFsBD is demonstrated by the analysis of simulated and experimental signals of faulty bearings. The results highlight the robustness and stability of the proposed CFFsBD in extracting fault-induced cyclostationary symptoms.

The Pantograph-Catenary System (PCS) is a critical component of railway vehicles, and its performance directly affects current collection quality. The arc rate serves as an essential measurement indicator for monitoring PCS state. However, in complex railway environments—where arc sizes and shapes can vary significantly and are further influenced by factors such as reflected light, glare, and adverse weather—traditional arc detection methods are easily affected by unstable current collection and power fluctuations, resulting in increased false detection rates and reduced measurement accuracy. Deep learning methods, while promising, also face limitations when dealing with such diverse arc morphologies and strong external interference. To address these challenges, this paper proposes a Multi-Modal Imitation Learning-based Arc Detection Network (MILADNet). First, the measurement system fuses infrared and visible-light image features to enhance arc feature extraction in scenarios with strong glare or reflective interference, thereby mitigating false alarms caused by relying on a single sensor. Second, to overcome the lack of information on small arcs, an online imitation learning framework is introduced to improve the system’s detection sensitivity for small arcs. Finally, to address data bias arising from uneven arc distributions, an unsupervised transferable representation learning method is employed to reduce dependence on labeled data and enhance model generalization. Experimental results show that MILADNet exhibits outstanding detection performance for arcs of various sizes and in complex environments, demonstrating both high efficiency and accuracy during measurement and data processing. Beyond improving the precision and reliability of arc detection, this method offers a novel solution for the instrumentation and measurement field and shows significant potential for condition monitoring and anomaly detection in railway systems.