Supervisory Control and Data Acquisition (SCADA) systems collect vast amounts of multi-sensor monitoring data, which is widely used in the intelligent fault diagnosis of wind turbines with the fast development of deep learning technologies. However, the complex structure of wind turbines and their time-varying operating conditions result in intricate spatio-temporal correlations within SCADA data, presenting significant challenges for feature extraction and accurate fault diagnosis. Current spatio-temporal fusion methods often treat SCADA data as Euclidean data, limiting their ability to capture the complex spatio-temporal coupling characteristics, which leads to reduced diagnostic accuracy. To solve abovementioned problems, a novel deep learning based spatio-temporal graph convolutional neural network (STGCN) is developed for intelligent fault diagnosis of wind turbines in this paper. Firstly, an adjacency matrix is constructed based on the Gaussian kernel function to graphically represent the SCADA data, so as to improve the representation capacities for spatial characteristics. Then, the spatial and temporal fault features are extracted using the graph convolutional network (GCN) and the one-dimensional convolutional network (1D-CNN) respectively. Finally, a spatio-temporal feature fusion module is developed as the sandwich structure to construct the proposed STGCN. The feasibility and effectiveness of the proposed method are verified by two cases of blade icing detection and main bearing wear diagnosis. The results show that the proposed method is able to accurately describe the spatio-temporal correlation of SCADA data, and improve the diagnostic accuracy.

Compared with high-speed trains, the operation lines of metro are more complicated and diversified. The frequent start-up and braking of metro trains will aggravate rail wear, especially in the small-radius curved section of the line, which is more likely to produce a track irregularity state. To solve this problem, this article mainly proposes a new intelligent detection method for track irregularity. First, a new residual attention multistate gate control unit (RMGCU) is presented, featuring an innovative multistate hierarchical division mechanism in memory cells to improve the feature perception ability. Second, the elite opposition-based learning (EOBL) and Gaussian mutation strategy are introduced into the traditional Osprey optimization algorithm (OOA) to improve its performance, and the improved OOA is adopted to optimize the hyperparameters of CNN and RMGCU. Finally, a novel residual attention multistage spatial-temporal network (RAMSTN) is developed by fusing a CNN for multiband spatial feature extraction and RMGCU for enhanced temporal sequence modeling with hierarchical memory gating. The proposed RAMSTN effectively captures the spatial-temporal characteristics of vehicle body acceleration, enabling accurate end-to-end detection of vertical track irregularities. Experimental results demonstrate that the RAMSTN model achieves high accuracy in identifying the track irregularity value. This study provides both theoretical insights and practical tools for metro track condition evaluation and maintenance planning.

SCADA data-driven fault diagnosis is a research hotspot in the field of wind turbine intelligent fault diagnosis. However, most existing methods rely on labeled data and neglect to mine fault information from a large amount of unlabeled data. Moreover, the common distribution information between labeled and unlabeled data is not well considered. To solve these problems, a semi-supervised broad learning generative adversarial network (SSBAN) is proposed for wind turbine intelligent fault diagnosis. First, the SSBAN is established through an integrated combination of BLS and GAN discriminator. Second, labeled data participates in multi-classification task for classification, while unlabeled data and simulated data generated based on labeled data are used for binary-classification task to distinguish between real data and generated data. Finally, two classification tasks share parameters to extract the shared distribution information from both labeled and unlabeled data, reduce over-reliance on labeled data, and enhance model performance. Two experiments are carried out to verify the effectiveness of the proposed method, and the results show that the proposed SSBAN can leverage the role of unlabeled data to improve diagnostic performance.