To address the challenge of anomaly detection in vibration signals from high-speed electric multiple unit (EMU) motor bearings, characterized by strong non-stationarity and multi-component coupling, this study proposes a synergistic approach integrating variational mode decomposition (VMD) and deep learning. Unlike datasets focused on fault diagnosis (identifying known fault types), anomaly detection identifies deviations into unknown states. The method utilizes real-world, non-real-time vibration data from ground monitoring systems to detect anomalies from early signs to significant deviations. Firstly, adaptive VMD parameter selection, guided by power spectral density (PSD), optimizes the number of modes and penalty factors to overcome mode mixing and bandwidth constraints. Secondly, a hybrid deep learning model integrates convolutional neural networks (CNNs), bidirectional long- and short-term memory (BiLSTM), and residual network (ResNet), enabling precise modal component prediction and signal reconstruction through multi-scale feature extraction and temporal modeling. Finally, the root mean square (RMS) features of prediction errors from normal operational data train a one-class support vector machine (OC-SVM), establishing a normal-state decision boundary for anomaly identification. Validation using CR400AF EMU motor bearing data demonstrates exceptional performance: under normal conditions, root mean square error , Mean Absolute Error , and Coefficient of Determination ; for anomaly detection, accuracy = 95.2% and F1-score = 0.909, significantly outperforming traditional methods like Isolation Forest (F1-score = 0.389). This provides a reliable technical solution for intelligent operation and maintenance of EMU motor bearings in complex conditions.
With the widespread adoption of wireless communication technologies in modern high-speed rail systems, the Train-to-Ground (T2G) communication system for Electric/Diesel Multiple Units (EMU/DMU) has become essential for train operation monitoring and fault diagnosis. However, this system is increasingly vulnerable to various cyber-physical threats, necessitating more intelligent and adaptive security protection mechanisms. This paper presents an intelligent security defense framework that integrates intrusion detection, risk scoring, and response mechanisms to enhance the security and responsiveness of the T2G communication system. First, feature selection is performed on the TON_IoT dataset to develop a Dream Optimization Algorithm (DOA)-optimized backpropagation neural network (DOA-BPNN) model for efficient anomaly detection. A Bayesian risk scoring module then quantifies detection outcomes and classifies risk levels, improving threat detection accuracy. Finally, a Q-learning-based reinforcement learning (RL) module dynamically selects optimal defense actions based on identified risk levels and attack patterns to mitigate system threats. Experimental results demonstrate improved performance in both multi-class and binary classification tasks compared to conventional methods. The implementation of the Bayesian risk scoring and decision-making modules leads to a 63.56% reduction in system risk scores, confirming the effectiveness and robustness of the proposed approach in an experimental environment.
Axle box bearings are critical rotating components in the running gear of high-speed trains, and their damage condition directly affects the operational safety of trains. To address the problems of non-stationary vibration signals, limited fault samples, and the ordinal degradation characteristics of bearing damage levels, this paper proposes a bearing damage severity assessment method based on S-transform (ST) and a Siamese ordinal prototypical network (ProtoNet). First, the ST is used to convert one-dimensional vibration signals into two-dimensional time–frequency images to enhance fault feature representation. Then, a Swin Transformer is employed to extract features from the time–frequency images, and a Siamese network is constructed to optimize the feature space distribution through contrastive loss. On this basis, an ordinal ProtoNet combined with the Wasserstein distance loss is introduced to impose ordinal constraints on different damage levels, enabling effective identification of bearing damage severity. Experiments are conducted on a high-speed train bearing vibration dataset under a few-shot setting, and comparisons are made with convolutional neural network (CNN), ProtoNet, Siamese CNN, ResNet18, Swin Transformer, and convolutional block attention module-residual network-Siamese models. The results show that the proposed method achieves a classification accuracy of 93.8% in a four-class bearing condition recognition task, outperforming the compared models and demonstrating its effectiveness in small-sample bearing damage severity assessment.
Axle-box bearings are safety-critical components in high-speed electric multiple units, and their vibration responses are affected by wheel–rail excitation, load variation, and environmental noise. Existing deep-learning diagnosis methods often rely on either raw time-series signals or time-frequency images, which limits their ability to exploit complementary temporal and spectral fault cues. This paper proposes Time-Frequency Dual-Stream Contrastive BearingNet (TFDC-BearingNet). The core idea is that a single vibration signal can be represented in two perceptually complementary forms, a global composite time- and frequency-domain visual view and a local temporal waveform view, and the global representation can serve as a query to guide the selection of fault-discriminative local signal components. Specifically, a visual branch extracts global representations from composite time-domain and frequency-domain images with a pre-trained Vision Transformer, whereas a signal branch uses a multi-scale tokenizer and Transformer encoder to preserve local temporal dependencies from raw sequences. A global-guided local selection module further uses cross-attention and gated fusion to select diagnosis-relevant signal components under visual guidance, and a contrastive objective aligns the two views in a shared semantic space. Experiments on a high-speed train rolling test-rig dataset and the public Southeast University bearing dataset achieve accuracies of 98.78% and 99.95%, respectively. Bidirectional cross-condition experiments on the SEU dataset show target accuracies of 94.77% and 93.46%. Comparative, ablation, cross-condition, noise-robustness, feature-visualization and statistical analyses indicate that the proposed dual-stream alignment and global-guided local selection improve diagnostic accuracy.