The development of Internet of Things (IoT) technology promotes the application of fault diagnosis method based on data-driven in power transformer. However, collecting a large number of labeled fault samples in real-world engineering scenarios is often costly and labor-intensive. To address this challenge, a novel graph-based spatiotemporal two-stream network (GSTSN) is proposed to accurately identify transformer fault conditions under limited labeled data. Specifically, the designed two-stream graph module constructs a spatiotemporal feature graph incorporating both real and pseudo labels via a label propagation strategy. Subsequently, a spatiotemporal feature augmentation graph convolutional network (SFA-GCN) is designed to capture both specific and shared information embedded in the spatiotemporal graph. Finally, an adaptive feature fusion module based on the minimum redundancy maximum relevance (mRMR) principle is developed to dynamically and selectively integrate temporal, spatial, and common feature embeddings, achieving reliable diagnostic results with limited supervision. Real vibration data from a 693/400 V test transformer are employed to assess the performance of the GSTSN framework. Experimental results show that under a 5% labeling rate, an average diagnostic accuracy of 96.14% is attained by the proposed method, highlighting its significant applicability in real-world engineering scenarios.

Electrothermal modeling is critical for accurate battery state estimation under dynamically changing loads, especially in real-time digital twin (DT) applications. To address the limitations of traditional decoupled or unidirectional electro-thermal frameworks, this study proposes a cross-task attention-based multitask learning (CTA-MTL) model that enables bidirectional electro-thermal coupling and joint residual correction of voltage and temperature predictions. The model integrates a shared temporal encoder based on a quasi-recurrent neural network, a two-dimensional convolutional neural network enhancement module for spatiotemporal coupling, and a sparse cross-task attention mechanism to capture task interdependencies. Built upon an equivalent circuit model and a lumped thermal model, the framework provides a lightweight and physically interpretable structure suitable for embedded deployment. Experimental evaluations across five discharge scenarios—including high C-rate and pulse conditions—demonstrate superior prediction accuracy and generalization compared to representative soft parameter sharing and single-task correction baselines. Ablation studies further confirm the effectiveness of shared representation learning, task-aware attention, and sparse selection. These results validate CTA-MTL as a robust and efficient solution for real-time electro-thermal state estimation in battery DTs.

The complexity of data and limited model generalization significantly hinder prediction accuracy. A physics-informed long short-term memory model with adaptive weight assignment (PILSTM-AWA) is proposed. First, PILSTM-AWA employs segmented feature extraction to enhance local information capture and improve feature extraction capabilities for nonlinear data. Then, the physical distribution and dynamic changes law of dissolved gas in oil are embedded into the LSTM framework. PILSTM is designed to constrain data fluctuations and predict dissolved gas in oil. Finally, an adaptive dynamic weighting strategy is introduced to balance physical and data information, enhancing forecast accuracy. The study utilizes online monitoring data from a 1000 kV transformer. Experimental results demonstrate that the model outperforms comparative algorithms in evaluation metrics. Notably, it achieves a coefficient of determination exceeding 91% under both normal and abnormal conditions, surpassing the predictive performance of conventional models.

State of health (SOH) estimation is essential for battery health monitoring, particularly in cross-domain scenarios where data variability and domain shifts present significant challenges. To address these issues, this study proposes the adversarial dynamic spatiotemporal attention network (ADSTAN), which integrates a graph attention network for spatial feature extraction, a gated recurrent unit for temporal dependency modeling, and a gradient reversal layer-based domain alignment module for unsupervised domain adaptation. Representing battery health data as dynamic graphs, with each cycle serving as a node, ADSTAN effectively captures spatiotemporal dependencies and dynamically aligns feature distributions between source and target domains. Experiments on cross-domain datasets, including the CALCE and NASA battery datasets, demonstrate the model’s effectiveness. Using data from 10 cycles to predict SOH for 5, 10, and 15 horizons, ADSTAN achieved RMSE values of 2.49%, 2.61%, and 2.94%, respectively. Ablation experiments validated the model’s design, highlighting the superiority of its spatial, temporal, and alignment modules. These results underscore ADSTAN’s robust performance and its suitability for accurate and generalizable SOH estimation in diverse cross-domain settings.

Early fault warning is critical for long-term stable operation of power transformer. However, slight mechanical fault of the power transformer does not lead to obvious changes in the vibration signal, which undoubtedly increases the difficulty of early warning. To solve this problem, this article proposes an early fault warning scheme for power transformer using vibration signal. Specifically, a new unsupervised learning framework, parallel multiscale convolutional autoencoder network (PMCAEN), is first developed to capture multidimensional and multiscale feature information of power transformer. Subsequently, support vector data description is established, and the reconstruction error obtained by PMCAEN under normal conditions is used as new observation information to construct the judgment domain of degraded state. The effectiveness of the proposed method is verified by the data collected from a 220-kV power transformer. The experimental results show that the proposed method achieves higher early warning accuracy with respect to the compared algorithms.

This article focuses on data-driven approaches for icing detection (ID) on wind turbine blades. In light of the widespread application of sensor technologies in wind turbines, such data-driven ID methods have become increasingly prominent. However, current methods have deficiencies, particularly in acknowledging the structural properties of multivariate sensor data and in differentiating icing stages, both critical for the identification of failure patterns. To bridge these gaps, we propose a spatiotemporal attention Siamese network (STASN) for blade ID. This model employs a Siamese network architecture for efficient few-shot learning amidst class imbalance. It uniquely incorporates a graph attention network and gated recurrent unit for extracting spatiotemporal features from sensor data. This design not only acknowledges the spatial structure of the data but also distinctly identifies features pertinent to various icing stages. The efficacy of STASN was validated using actual sensor data from supervisory control and data acquisition systems. The results demonstrate STASN's capability in discerning distinct icing stage features and its potential in early icing prediction. This research underscores STASN's utility in providing advanced, flexible fault alarms for blade icing, representing a significant stride in wind turbine maintenance and safety.

Extracting sensitive information from vibration signal has become a frequently adopted way in fault diagnosis. However, most previous methods fragmented the relationship between quantification and visualization analysis, which affects the interpretability, accuracy and comprehensiveness of the extracted information. To this end, this paper proposes distribution recurrence plots (DRP) and measures (DRM) to realize the unity of visualization and quantification analysis of the signals. Specifically, DRP is a novel feature graphical representation method following the thought of symbolic dynamics. Derived from DRP, DRM is developed containing four quantifiers for extracting comprehensive fault features that allows a multiclass support vector machine (SVM) to identify the fault types of wind turbine drivetrain system (WTDS). Specially in DRM, pattern entropy is a newly designed quantifier by considering pattern distribution to obtained more accurate quantitative representation of the signals. Using simulated data, DRP and DRM are validated to reveal the intrinsic structural changes for different dynamic systems and robustness to noise. Applications on wind turbine gearbox illustrate that the proposed method has favorable diagnosis performance and stability compared with other competitors. This approach is easy to interpret, is robust to noise, and has a low computational burden, becoming viable for WTDS fault diagnosis.