标签引导多尺度自适应特征对比的消防管网跨域故障诊断OA

As an important part of urban infrastructure,the fire protection pipe network faces dual challenges of small samples and cross-domain in its fault diagnosis.Traditional fault diagnosis methods usually suffer from problems such as insufficient generalization performance and poor model adaptability when dealing with cross-domain tasks with large differences in data distribution.After constructing a pipeline network fault dataset covering different leakage degrees and leakage locations,a cross-domain fault diagnosis method based on label-guided characteristic analysis and multi-scale attention mechanisms(LCA-MSA)is proposed.In this method,a learning strategy that combines multi-task learning with label-guided feature contrast is adopted,and the multi-scale convolution and attention mechanisms are introduced simultaneously,which enhances the model's ability to extract multi-level fault features.The experimental results demonstrate that the LCA-MSA model exhibits significant advantages in small-sample cross-domain fault diagnosis tasks for fire pipeline networks,achieving a diagnostic accuracy of 95.16%on the target domain test set.In comparison with traditional transfer learning and contrastive learning methods,the proposed method can show superior performance and better adaptability in fire pipeline fault diagnosis scenarios.