Abstract With the rapid development of rail transit systems, daily operation and maintenance tasks for trains are becoming increasingly demanding. The reliable operation of the air-conditioning system is directly related to running safety and passenger comfort. However, existing monitoring methods mostly rely on traditional threshold-based approaches for anomaly warning, which suffer from delayed detection of early anomalies and a high false alarm rate. To address these issues, this paper proposes a condition monitoring and anomaly early warning method that integrates a convolutional neural network (CNN), a long short-term memory network (LSTM), and an attention mechanism. The method uses the CNN to extract local temporal features from multi-sensor data, employs the LSTM to model long-term dependencies, and introduces the attention mechanism to dynamically focus on critical time steps, thereby improving overall prediction accuracy. Furthermore, an adaptive recursive status index (ARSI) combined with statistical process control (SPC) is used to determine a statistical control threshold, enabling real-time system status monitoring and anomaly early warning. Validation is performed using TCMS data from a specific model of high-speed train. Experimental results show that the proposed model achieves significantly higher prediction accuracy on healthy samples than baseline models such as LSTM, MSET, and TCN. In anomaly detection tasks, the model attains an accuracy of 98.63% with a false alarm rate of only 1.42%, and issues a warning 31 minutes before the TCMS-reported fault alarm of a compressor pressure fault. Moreover, on a cooling performance degradation fault caused by fan performance deterioration, the model achieves an accuracy of 98.65% with a FAR of only 1.04%, and issues a warning approximately 73 minutes before the fault alarm, confirming its potential applicability across different fault modes. Ablation studies further confirm the necessity of the synergistic collaboration of CNN, LSTM, and the attention mechanism.
Abstract This paper presents a data-driven framework for geometric parameter design of a six-degree-of-freedom Stewart–Gough platform under prescribed motion requirements. Actuator stroke characteristics are strongly coupled with hinge distribution geometry, making conventional design methods inefficient and task-specific due to repeated kinematic optimization. To overcome these limitations, a neural network–assisted optimization framework is proposed to learn the nonlinear mapping from motion parameters to optimal geometric configuration. A large-scale dataset is generated through kinematic simulations, associating motion requirements with hinge radii, adjacent hinge distances, and resulting actuator strokes. A neural network is trained to predict hinge configurations and auxiliary structural parameters that provide large stroke margins. To enhance optimization reliability, a hybrid strategy combining swarm intelligence for global exploration and backpropagation-based refinement is employed, together with an efficient stroke evaluation scheme using representative extreme motion cases. \rev{Simulation and experimental results on a physical platform} demonstrate that the proposed framework significantly improves stroke consistency between theoretical design and actual motion realization, achieving a mean IoU of 0.9930 for stroke overlap validation with a standard deviation of 0.0032, while reducing the mean geometric deviation to 1.6185 and maintaining a mean execution time of 0.118 ms. Compared with competing methods, the proposed approach demonstrates improved statistical stability and computational efficiency for intelligent structural design of motion platforms and related parallel mechanisms. Physical experiments conducted on a real six-DOF Stewart platform under six representative synchronized motion patterns confirm that the predicted actuator stroke intervals agree closely with the measured ones, achieving a mean interval IoU of 0.9831 and a mean boundary discrepancy of 2.10 mm.
Abstract In subspace-based enhancement of signals, covariance eigenspace information is used to separate signal and noise dominated components. Traditional approaches like PCA and EWF achieve good noise reduction, although the fixed or linear weighted approach restricts the adaptability under diverse signal and noise situations. In this paper, an approach named as Adaptive Exponential Quadratic Root Filter (AEQRF) is developed for adaptive enhancement of signals based on the information from eigenspace. The framework of the algorithm estimates the characteristics of signal and noise eigenspaces by means of covariance-matrix eigenvalue decomposition. Then Eigen Separation Index (ESI) and bounded Eigen Quality Factor (EQF) are used to design filter parameters. Eigenspace quality measure is mapped into pole radius with the aid of an exponential relationship, while two poles forming a complex-conjugate pole pair constitute a quadratic second-order IIR filter section. The AEQRF can be implemented in two versions of filter namely AEQRF-Fixed and AEQRF-Variable. Input SNR ranging from −10 dB to 20 dB is chosen for numerical simulation under different types of noises: additive white Gaussian noise, colored noise and impulsive noise. Performance is measured in terms of output SNR, mean square error, relative performance gain, eigenvalue domain weights, convergence properties and Monte-Carlo statistical analysis. The robustness of AEQRF-Fixed is enhanced in very poor conditions, while that of AEQRF-Variable is enhanced by the adaptation capability when there is a change in the signal conditions. The experimental study reveals that the proposed pole adaptation scheme using eigenspace quality is quite efficient in terms of signal enhancement performance in the considered noise environments. However, the relative performance of the two order variants is dependent upon the input SNR and type of noise environment.
Abstract The rapid development of unmanned aerial vehicles (UAVs) has raised increasing concerns regarding low-altitude public safety and privacy protection. Reliable UAV detection remains challenging because UAV targets are often small, distant, and easily affected by complex background interference. In this paper, the UAV-Oriented Multi-Scale Receptive-Field Enhancement Network (UAV-MSRFENet) is proposed as a parameter-efficient and effective UAV detection framework developed upon YOLO11n, aiming to improve detection robustness while preserving real-time performance. First, a Multi-Scale Depthwise Residual Block is designed to enhance multi-scale feature representation and capture contextual information for UAVs with different apparent scales. Second, Receptive-Field Attention Convolution is introduced to strengthen feature diversity and improve the expressiveness of backbone features. Third, a Shared-Convolution Decoupled Head is developed to reduce parameter redundancy and improve inference efficiency. Comprehensive experiments on two publicly available UAV detection benchmarks demonstrate that the proposed UAV-MSRFENet achieves competitive detection performance with a substantially reduced parameter footprint, particularly for small UAV targets. Embedded platform deployment further verifies its real-time inference capability. These results show that the proposed method provides an effective balance between detection accuracy and computational efficiency for practical low-altitude UAV monitoring. The relevant data are available at: https://github.com/MYM-sudo88/MSRFENET
Abstract To mitigate fluctuations in head and efficiency during the start-up process of gas-liquid two-phase centrifugal pumps and enhance operational stability, this study conducted a multi-objective optimization of transient performance based on CFD simulations and experimental validation. Using the Q5H26 centrifugal pump as the research object, radial basis function (RBF) surrogate models were established for average head and average efficiency ( H avg / η avg ) and their corresponding fluctuations (Δ H avg /Δ η avg ); the predictive accuracy of these surrogate models was evaluated using quantitative performance metrics. The NSGA-II multi-objective genetic algorithm was employed to optimize five key design variables: impeller inlet diameter, impeller outlet width, impeller outlet diameter, blade wrap angle, and blade outlet angle. Response analysis was performed on the optimization results, and the performance of the pump model before and after optimization was compared via numerical simulation. Results demonstrate that the RBF surrogate model exhibits high predictive accuracy and can be effectively used to predict and optimize the transient characteristics of centrifugal pumps. After optimization, mean fluctuations in head and efficiency of the gas-liquid two-phase centrifugal pump decreased by 9.33% and 19.05%, respectively, while average head and average efficiency increased by 5.26% and 3.64%, respectively. The optimized internal flow field exhibits enhanced uniformity and stability, thereby facilitating more efficient fluid energy conversion.
