SRHS-Net: A Hybrid Attentive Architecture for Real-Time Steam Leak Detection in Power Plants

Steam pipeline leak detection is significant for identifying potential risks and improving the efficiency of inspections in power plants. However, due to various complex scenarios and densely distributed layouts, the current detection methods have difficulty obtaining high accuracy and efficiency. To address the problems, a hybrid attentive architecture for real-time steam leak detection model called SRHS-Net is proposed in this article. First, the SimAM attention mechanism is utilized to help the model effectively focus on key regions of the leakage target, improving detection accuracy and reducing interference from irrelevant background elements. Second, receptive field block and its variants are used to capture features at different scales, utilizing a multiscale receptive field strategy, which is essential for detecting small and dynamically changing steam leaks. Moreover, HorNet and space-to-depth are introduced to enhance the model's adaptability to complex scenarios. These modules process lateral information in feature maps and capture spatial details while maintaining computational efficiency. We experimentally demonstrate that the mean average precision of the proposed model is 83.2%, which is 9.4% higher than that of the YOLOv7. The proposed method demonstrates high performance on the self-constructed dataset and has been successfully tested in real-world industrial applications.