With the advancement of intelligent construction in coal mines, 5G communication technology has become a key support for underground core networks. To analyze the electromagnetic wave propagation characteristics of 5G signals in coal mine tunnels, an innovative method that integrates the Shooting and Bouncing Ray/Image Method (SBR/IM) with a Genetic Algorithm-Optimized Support Vector Machine (GA-SVM) is proposed. The SBR/IM method is used to construct a three-dimensional electromagnetic simulation model of the tunnel, simulating multi-path propagation and generating high-precision training data. The GA-SVM algorithm is then employed to intelligently correct and predict signal path loss, enhancing model accuracy and generalization capability. Experimental results under typical tunnel conditions show that the proposed fusion model achieves a maximum path loss of only 63 dB at 0.2 m wall thickness, which is 51 dB lower than that of the traditional Floating Interception (FI) model. Moreover, the signal-to-noise ratio remains above 0 dB within 225 m, significantly outperforming the FI model. At a wall thickness of 0.3 m, the path loss of the fusion model is 72 dB. Additionally, the fusion model exhibits a received signal strength index of -68 dBm, a coherent bandwidth of 25 MHz, a throughput of 320 Mbps, and a delay as low as 15ms, all of which are superior to those of comparative models. The proposed hybrid method effectively enhances the propagation performance of 5G signals in coal mine tunnels, reduces path loss, improves signal quality, and provides a valuable approach for studying electromagnetic wave propagation in such environments.
Existing cross-modal retrieval ignores key attribute features such as genre and author in art design, and independent encoding of images and texts leads to poor semantic alignment. Therefore, this study proposes an art design cross-modal retrieval model that combines attribute deep fusion and multi-task collaborative learning. This method first employs a prompt-template-based strategy to convert structured attribute information into natural-language sentences. From a causal inference perspective, separate encoding paths are designed to integrate attributes, text, and visual features in depth. In addition, the model introduces a multi-task collaborative learning framework to learn more robust shared representations by jointly optimizing cross-modal retrieval and artistic attribute classification. Experimental results on the Rijksmuseum data set showed that, in the text-retrieval image task, the average precision of the cross-modal retrieval model reached 62.7%, and the total recall rate reached 397.8%. In the image-retrieval text task, the average accuracy was 59.4%. The recall of the cross-modal retrieval model reached a maximum of 95.63%. Multi-task analysis showed that when using Genre as an auxiliary task, the Recall@50 (Number of relevant items in the top 50 results) of the image retrieval text task was 95.06%. The attribute deep fusion strategy based on prompt templates and the multi-task collaborative learning framework are the key to improving model performance. The combination of the two can effectively capture the specific semantics in the art field and achieve more accurate cross-modal alignment. The research results provide efficient and innovative solutions for solving cross-modal retrieval problems in art and design.
Deep learning-based skin disease recognition technology has been continuously developing, but existing methods still fall short in terms of detail representation and computational efficiency due to limitations such as limited image samples, small lesion areas, and significant fine-grained differences. Against this backdrop, this study proposes a skin disease recognition algorithm based on an attention mechanism and a Transformer architecture. While retaining the local texture-feature modeling capability of convolutional networks, a Transformer recognition framework that integrates convolutional features with an attention mechanism is constructed to jointly model both local fine-grained features and global semantic information. Experimental validation was conducted using the International Society for Skin Imaging Collaboration (ISIC) 2018 dataset. The proposed method achieved good classification performance on this dataset, with a Macro-F1 score of 0.90 and a Weighted-F1 score of 0.95, and demonstrated superior recognition performance in comparisons with various Transformer models. The results show that this method can effectively integrate local fine-grained features and global semantic information, which is significant in improving the accuracy and stability of automatic skin disease recognition and provides a feasible technical path for intelligent assisted diagnosis of skin diseases.
Currently, the development of the agricultural specialty industry in Southwest China faces challenges such as insufficient brand competitiveness and lagging processing technology, which restrict industrial development performance. Taking Sichuan Province in Southwest China as an example, the research employs the Analytic Hierarchy Process (AHP) combined with the Fuzzy Comprehensive Evaluation (FCE) method to evaluate performance. The former decomposes the problem into the target layer, criterion layer, and solution layer and uses an expert-scoring method to determine the weights of each indicator. The latter uses fuzzy mathematical theory to compute comprehensive evaluation results using membership functions and weight matrices. Based on the evaluation results, the factors affecting performance improvement are analyzed in depth, and industry optimization measures are developed. The overall performance level of the agricultural specialty industry in Sichuan Province in 2023 was moderate, with a comprehensive evaluation value of 3.10. The economic performance was poor, with a comprehensive evaluation value of only 2.95. The social and ecological performance was moderate, with comprehensive evaluation values of 3.26 and 3.38, respectively. The performance evaluation method can effectively quantify the multidimensional performance of the agricultural specialty industry in Sichuan Province, providing reliable technical support for the scientific diagnosis and industrial optimization of the agricultural specialty industry in Southwest China.
Due to the problems of dispersed business data, inconsistent regulatory interpretations and limited cross-domain modeling at the enterprise level, this paper puts forward a data governance model based on federated transfer learning and security auditing. The preprocessing stage adds salted pseudonymization, entity masking, rare-category pooling, 30-minute window aggregation, and lineage/access-log feature construction to ensure reproducibility without revealing raw records. The method builds metadata, lineage, access behavior and rule-hitting features locally in each business domain and achieves cross-domain knowledge transfer via transfer gating and alignment loss. At the same time, it records summary hashing, responsible parties, policy consistency and lineage continuity in audit quality scores, adds a risk-suppression aggregation mechanism, and synchronously links the model update process to the governance traceability process. Experiments are based on financial, procurement, logistics, sales and compliance data from a large local group from January 2024 to December 2025. According to the results above, the proposed method achieves an AUC of 0.957, a macro-F1 score of 0.912, a recall of 0.926, and an audit trace completeness of 0.987. It is 2.1 and 2.3 percentage points higher than the best-performing comparative method. Extra imbalance, threshold-sensitivity and edge deployment tests show that the transfer gate reduces negative transfer in a 1:50 class imbalance, and audit-suppressed aggregation maintains real-time latency below 160ms under the tested edge configuration. Thus, cross-domain risk identification and strategy implementation can be achieved in a closed-loop manner without aggregating raw data, and a technical path for enterprise data governance that balances accuracy, compliance, and feasibility has been provided. A separate public HDFS log benchmark supplies a limited proxy stress test; it does not validate enterprise audit completeness or reproduce an accuracy advantage over FedAvg.
