An axle box bearing is one of the most important components of high-speed EMUs (electric multiple units), which runs at a very fast speed, suffers a heavy load, and operates under various complex working conditions. Once a bearing fault occurs, it not only has an enormous impact on the railway system, but also poses a threat to personal safety. Therefore, there is significant value in studying a real-time fault early warning of a high-speed EMU axle box bearing. However, to our best knowledge, there are three obvious defects in the existing fault early warning methods used for high-speed EMU axle box bearings: (1) these methods based on vibration are extremely mature, but there are no vibration sensors installed in high-speed EMU axle box because it will greatly increase the manufacturing cost; (2) a TADS (trackside acoustic device system) can effectively detect early failures, but only a portion of railways are equipped with such a facility; and (3) an EMU-ODS (electric multiple unit onboard detection system) has reported numerous untimely warnings, along with warnings of frequent occurrence being missed. Whereupon, a method is proposed to realize the fault early warning of an axle box bearing without installing a vibration sensor on the high-speed EMU in service, namely a MLSTM-iForest (multilayer long short-term memory–isolation forest). First, the time-series data of the temperature-related variables of the axle box bearing is used as the input of MLSTM to predict the axle box bearing temperature in the future. Then, the deviation index of the predicted axle box bearing temperature is calculated. Finally, the deviation index is input into an iForest algorithm for unsupervised classification to realize the fault early warning of an axle box bearing. Experimental results on high-speed EMU operation data sets demonstrated the availability and feasibility of the presented method toward achieving early fault warnings of a high-speed EMU axle box bearing.

As critical components of a machine, bearings are fundamental to maintaining system reliability, which emphasizes the importance of intelligent fault diagnosis. In industrial applications, bearing failures are episodic. As a result, a large amount of samples collected is normal, and the extreme imbalance of samples poses a challenge for bearing fault diagnosis (BFD) in industrial applications. In addition, industrial applications impose higher requirements for model lightweight due to the limitations in computational capability of the device. To address these challenges, a novel knowledge distillation framework with improved activation function and adaptive feature selection strategy was proposed for imbalanced BFD. Firstly, an enhanced activation function was introduced to improve the model’s nonlinear feature extraction capabilities, thereby boosting its performance when handling imbalanced datasets. Secondly, an adaptive feature selection strategy was proposed to ensure the extraction of sufficient information, thereby enhancing the model’s generalization capability across diverse devices. Finally, a novel knowledge distillation framework was proposed to downsize the model, which makes the model lightweight to meet the requirements of industrial applications. The effectiveness of the proposed methodology was experimentally validated on two imbalanced bearing datasets. The experiments demonstrate that the proposed methodology outperforms existing models in diagnostic performance.

Edge intelligence enables significant latency reduction and enhances the timeliness of model-based fault diagnosis. However, existing deep learning-driven bearing fault diagnosis models are ill-suited for deployment on edge devices, primarily due to three critical limitations: (1) Lightweight models typically exhibit inadequate anti-noise performance, failing to meet the reliability requirements of real-world engineering scenarios. (2) Models with superior anti-noise capabilities often demand high-performance hardware for operation, thereby restricting their deployment on resource-constrained edge devices. (3) These models adopt a fixed input length, which makes it difficult to guarantee diagnostic accuracy across diverse application scenarios—attributed to variations in sampling frequencies, bearing parameters, and other relevant factors. To address these challenges, this paper proposes a lightweight anti-noise diagnostic framework (LADF) for edge-intelligent bearing fault diagnosis in complex engineering environments. The LADF comprises three core modules: a dynamic input module (DIM), a lightweight network module (LNM), and a denoising branch. Specifically, the DIM is designed based on the envelope spectrum, leveraging its inherent demodulation characteristics to dynamically adapt to input signals across diverse scenarios. Group convolution and layer normalization are employed to construct the LNM, ensuring robust diagnostic performance while achieving efficient computation. The denoising branch constrains the feature extractor via a loss function, enabling it to learn generalized fault features under varying noise environments and thereby enhancing the anti-noise capability of the framework. Finally, the proposed LADF is validated through test rig experiments on two datasets of train axle box bearings. Comparative analysis with state-of-the-art models demonstrates that the LADF achieves superior diagnostic stability and anti-noise performance while maintaining a more lightweight architecture, making it well-suited for edge deployment in railway bearing fault diagnosis.

With the rapid development of deep learning, edge intelligence applications (EIA) have achieved numerous results. However, redundant parameters of model and strong noise pollution pose challenges to EIA for bearing fault diagnosis. To solve these challenges, a model with lightweight network and antinoise ability was proposed for bearing fault diagnosis. First, a novel pluggable channel slimming module was designed to make the model lightweight, which can effectively reduce the parameters and computation of the model. Second, an antinoise learning network is proposed, which has a noise discriminator to enhance the network’s feature extraction capability through supervised learning. Finally, an adaptive input module was proposed to enhance the generalization ability of the model, which can adaptively adjust the input information under different application environments to improve the stability and accuracy of the model. The performance of the proposed model was verified through the test rig experiments on two types of train axle box bearings datasets, which indicated the proposed model achieves more than 89% diagnostic accuracy at −10 dB.

Intelligent bearing fault diagnosis methods based on deep learning have achieved massive results. However, acquiring sufficient labeled samples in engineering is a time-consuming and laborious task, and the signals are often contaminated by strong noise. Therefore, two main strategies are integrated to address these challenges. First, to address insufficient labeled samples, an improved dynamic pseudo-label learning strategy is developed, improving model generalization by involving more unlabeled data in training. Then, to solve the problem of strong noise faced in engineering, a noise reduction learning strategy based on self-supervised learning is proposed, extracting invariant features under noise conditions by a new optimization objective. In addition, a novel dynamic channel pruning strategy is proposed, which prunes channels that play a smaller role in the model, reducing the size of the model and decreasing the model hardware requirements for engineering applications. Finally, the effectiveness and robustness of the proposed strategies are verified on the two high-speed train wheelset-bearing datasets.