| In industrial environments,liquid leakage can cause safety accidents and economic losses.The liquid leakage detection method is mainly to use intrusive sensors to collect sound,pressure and other one-dimensional leakage signal,through the analysis of signal changes to detect leakage.Due to the limited sensitivity of the sensor,early liquid micro-leakage is usually not detectable.Visual information from surveillance can more intuitively reflect the micro-leakage state.Based on Tianjin Binhai heating station,with the help of the established video platform,we designed the framework of micro-leakage detection system based on dynamic detection and network feature extraction,and proposes two methods of micro-leakage detection.In view of the problem that micro-leakage is difficult to detect and recognition because its feature is not obvious,we proposed micro-leak detection method based on visual background extractor and deep neural network.The visual background extractor can sensitively detect potential leakage targets in the video stream,and then the deep neural network recognizes the real leakage in potential leakage targets.Experiments show that the accuracy of the model based on Dense Net121 in leakage recognition is99.274%.This method has strong anti-jamming ability,high sensitivity,high real-time and high accuracy,which meets the requirements of early micro-leakage detection in industry.Considering the robustness,generalization ability and deployment of microleakage detection system in the actual industry,we proposed a micro-leakage detection method based on pixel based adaptive segmenter and Efficient Net B0.Pixel based adaptive segmenter combines cybernetics and background complexity measurement,so that the update rate of the background model can be adaptively changed with the complexity of the background,which is more suitable for leakage detection in actual changeable industrial environments.The central loss function enhances the robustness of the leakage detection system in the actual industry by enhancing the compactness within the class.Experiments show that the accuracy of the model based on Efficient Net B0 is 99.495% and the model size is 11.58 M.This proposed method has higher accuracy,smaller model size and lower complexity,and is more suitable for deployment in actual industrial environments.Finally,based on Tianjin Binhai Heating Station,we verified the effectiveness of the proposed method in industry. |
