Accurate wind power forecasting is vital for the stable and safe operation of power systems. Wind power data typically exhibit multi-scale characteristics, manifested as short-term fluctuations and long-term trends. Traditional forecasting methods often neglect these characteristics, making it difficult to capture complex nonlinear relationships within the data. This limitation reduces prediction accuracy. To address this issue, we propose a Multi-scale Hierarchical Interaction Transformer (MHIT) model for wind power forecasting. The model combines a multi-scale non-overlapping segmentation module and a hierarchical interaction encoder. Through time series tokenization, the input sequence is decomposed into tokens of varying scales—shorter tokens specialize in capturing local high-frequency variations, while longer tokens effectively extract global trends and periodic patterns. The hierarchical interaction encoder employs multi-scale sub-encoders to extract features across different scales, with cross-scale interaction achieved by embedding small-scale latent representations into large-scale encoding processes. This design enhances the model's ability to simultaneously capture localized details and macroscopic temporal dynamics. Experiments on real wind farm datasets demonstrate that the proposed multi-scale model effectively captures the temporal characteristics of wind power data, significantly improving prediction accuracy compared to existing methods.

Unsupervised Domain Adaptation (UDA) technology has shown significant potential in multi-condition mechanical fault diagnosis. Its core lies in utilizing class labels, domain labels, and data structure information to achieve cross-domain knowledge transfer. However, existing methods generally suffer from two key issues: they only model the first two types of information through classification loss and domain adversarial loss while neglecting data structure information, resulting in insufficient feature discriminative capability; meanwhile, they fail to effectively integrate multi-view signals, limiting the comprehensive characterization of fault features. To address these issues, this paper proposes a Time-Frequency Feature Fusion Graph Convolutional Adversarial Network (TFFGCAN). First, a dual-channel graph structure modeling module is constructed to build graph structures for time-frequency information separately. Graph Convolutional Network (GCN) are employed to extract features from both graph data, and time-frequency features are fused through a cross-attention mechanism. Second, joint modeling of the three types of information is achieved in a unified framework, where multi-kernel maximum mean discrepancy is used to measure differences in graph structures across domains. Finally, the effectiveness of TFFGCAN is verified on the XJTUSpurgear dataset. The results show that this method achieves the best performance among comparative methods.

In complex industrial processes, digital twin technology faces modeling difficulties due to the absence of key quality variables, data driven soft sensing demonstrates significant potential in predicting quality variables due to its advantages of low cost and rapid response. However, data driven soft sensing models often suffer from data scarcity due to difficulties in obtaining labeled samples and high data repetition rates, which in turn affects prediction accuracy. Although virtual sample generation methods have been proposed to alleviate this issue, existing methods generally overlook the interdependencies among features and lack effective evaluation of the quality of generated samples, resulting in all samples participating in model training with equal weights. To address these shortcomings, this paper proposed a data augmentation and sample weighting soft sensing modeling method based on graph feature learning, referred to as GFSW-VSG, to mitigate the constraints imposed by insufficient samples on model performance. This model introduces a Graph Convolutional Neural Network (GCN) into the discriminator to maintain consistent interdependencies between the features of generated and original data. Subsequently, sample confidence is calculated to assign weights to virtual samples for weighted learning in the soft sensing model. The effectiveness of the proposed method is validated through an industrial chlortetracycline (CTC) fermentation process.

Deep learning techniques have achieved impressive results in intelligent fault diagnosis by virtue of their powerful feature extraction and pattern recognition capabilities. However, frequent changes in industrial operating conditions can lead to differences in sample distribution and diversity, thus affecting the accuracy of the model. In order to improve the diagnostic performance, this paper proposes a dual weighted adversarial multi-source domain adaptation network (DWAMDA). Firstly, sample weighting is combined to prioritise the importance of samples, and the contribution of cross-domain samples is adaptively adjusted based on the feature similarity metric. Secondly, a domain weighting strategy is also introduced to consider the influence of different domains in the process of multi-source domain adaptation, which enhances the generalisation capability by accurately assessing the contribution of the source domain to the target domain fault diagnosis. Finally, a validation using the PU dataset is performed and the experimental results show that the method is more advantageous for DWAMDA in cross-conditional tasks compared to other experiments.

Abnormal working conditions in coal mine production present critical risks to operational safety and miner welfare. Traditional methods based on time series anomaly detection suffer from insufficient integration of actual production processes, leading to detection outcomes with weak practical relevance. To meet the challenge, this paper proposes a novel AE-RNN-BiLSTM-GHMM framework (ARBG), which systematically integrates an AE-RNN-BiLSTM predictor (ARB) with a Gaussian Hidden Markov Model (GHMM)-based production process analyzer. The ARBG framework operates through a dual-modality architecture. First, the ARB component leverages historical electricity consumption patterns to forecast future energy usage profiles. Concurrently, the GHMM module dynamically models latent operational states within the coal mine production workflow. By synergistically integrating real-time predictions from ARB with GHMM-derived process state diagnostics, the framework achieves context-aware anomaly detection that aligns with physical production constraints. This fusion enables predictive identification of operational deviations through multidimensional correlation analysis between energy consumption forecasts and process state transitions. The experimental results show that the ARBG performs well in discriminating the load conditions and provides reliable technical support for improving the safety and operational efficiency of coal mine production.

Coal mine power load fluctuation serves as a critical precursor to safety accidents, with its clustering analysis holding significant value for safety production monitoring. To address the challenges of high-dimensional load data feature extraction, the separation between feature learning and clustering tasks in existing deep clustering methods, and the embedding space distortion caused by KL divergence loss, this paper proposes a Stacked Autoencoder with Combined Improved Loss for Coal Mine Daily Load Deep Embedded Clustering (SAE-CS-K-means). The methodology comprises three main components: Firstly, a layer-wise pretraining mechanism of stacked autoencoders extracts deep temporal features from load data, achieving effective mapping from high-dimensional data to low-dimensional embedding space. Secondly, a novel CS clustering loss function is designed to optimize intra-cluster compactness and inter-cluster differentiation. Finally, a joint optimization strategy synchronously updates network parameters and cluster centers, maintaining dynamic equilibrium between reconstruction loss and clustering loss, thereby ensuring embedding space stability and enhancing clustering performance. Experimental validation using actual power consumption data from a Yunnan coal mine demonstrates significant improvement in clustering quality through Silhouette Coefficient (SC), Davies-Bouldin Index (DBI), and Calinski-Harabasz Index (CHI) evaluations. The clustering results provide precise quantitative references for abnormal power consumption diagnosis and reliable technical support for coal mine safety monitoring.

Ovarian cancer is one of the most common and highly lethal gynecological malignancies, and its early diagnosis and accurate subtyping are of great clinical significance. Whole-slide images (WSI) contain rich tissue structural information required for pathological diagnosis and are often regarded as the "gold standard". However, due to their high resolution and high annotation cost, they typically only have patient-level weak labels. To address the insufficient feature representation under weak labeling and fully explore the potential histological correlations among instances, this paper proposes an ovarian cancer pathological image classification method based on Self-Supervised and multi-relational graph learning (SS-MRGL). First, self-supervised learning is employed to pre-train weakly labeled image instances, obtaining robust and discriminative compressed feature representations. Then, a multi-relational graph is constructed, and a multi-perspective joint modeling approach using Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) is introduced to capture the spatial topological and contextual semantic multi-dimensional information among instances, respectively. Finally, a gated attention mechanism is used to dynamically fuse graph-level features. Experimental results on the YN-OCPD dataset demonstrate that the proposed method can significantly improve the accuracy of ovarian cancer diagnosis under weak supervision.

Ovarian cancer is the disease with the highest mortality rate among malignant tumors of the female reproductive system. Its significant tissue heterogeneity leads to limitations in traditional treatment options. Therefore, accurate differentiation of ovarian cancer subtypes has become a key challenge to improve clinical efficacy. In recent years, histopathological image analysis methods based on multiple instance learning (MIL) mainly use complex representation networks to construct package-level feature representations, which requires a lot of computational cost and inefficient use of image information, or simply rely on data distribution modeling, which is easily affected by the noise of the data itself. To address the above problems, this paper innovatively proposes a prototype-guided deep embedding clustering multi-instance learning framework (Prototype-Guided Deep Embedding Clustering MIL, PGDEC-MIL), which constructs a binary discriminative prototype space by integrating tissue morphology prior knowledge, implements geometric structure constraints in a low-dimensional embedding space, and achieves efficient noise sample filtering, thereby generating a high-quality pseudo-label sample set with pathological interpretability for classifier training. This method not only achieves efficient use of pathological information and labels, but also has strong interpretability. Experiments on the YNC-OC ovarian cancer subtype dataset show that PGDEC-MIL is significantly better than the mainstream MIL model in terms of classification accuracy and F1 score. The visualization results further prove that our model has better tumor localization and subtype discrimination capabilities, providing a traceable decision-making basis for clinical diagnosis and treatment.