In mechanical systems, bearings play a vital role, yet their surfaces frequently develop defects during the production process. Hence, automated detection of such defects becomes indispensable for ensuring industrial quality control. To enhance the robustness of bearing surface defect detection, a dynamic sample matching detection network is proposed, which is guided by multi-level perceptual features. A lightweight adaptive feature extraction method is introduced to enhance defect representation while maintaining low parameter complexity and computational cost. To address diverse defect patterns, a multi-shape perception module with a dual-axis cross-weighting mechanism across multiple scales is designed to improve sensitivity to different defect types. Furthermore, a dynamic sample selection strategy with dual-label weighting is introduced to improve training sample quality. The proposed approach is evaluated on a self-developed bearing defect platform, where it outperforms mainstream detection models in complex environments and achieves superior performance and robustness.

Water supply networks, as critical urban infrastructure, play an essential role in ensuring stable city operations. Accurate flow forecasting is therefore of great significance for optimizing operational scheduling, reducing energy consumption, and maintaining system stability. With the strong capability of large language models (LLM) in sequence modeling and representation learning, their application to time-series forecasting has become an emerging research direction. However, a key challenge lies in the modality gap between numerical time-series data and the semantic embedding space of language models. To address this issue, this paper proposes a cross-modal alignment-based time-series foundation model for forecasting. The proposed method constructs a mapping between time-series features and the semantic embedding space, enabling effective projection of numerical sequences into a semantic domain. Furthermore, a cross-modal alignment mechanism is designed to enhance feature fusion, thereby improving the model’s ability to capture multi-scale periodic patterns, long-term temporal trends, and stochastic demand fluctuations commonly observed in water distribution systems. Experimental results demonstrate that the proposed method consistently outperforms baseline approaches across different prediction horizons in terms of mean absolute error (MAE) and mean squared erro (MSE), verifying the effectiveness and strong generalization capability of the cross-modal alignment strategy in water distribution network flow forecasting.