A hybrid neural network approach for preventing accidents and leaks in pipelines and boiler rooms using video stream analysis

Abstract

A comprehensive approach to solving the problem of accidents and leaks in pipeline and boiler systems is considered, combining risk forecasting methods with modern computer vision algorithms for video stream analysis. The relevance of the study is due to the high accident rate of engineering networks, significant economic losses, and environmental risks associated with delayed leak detection. The aim of the work is to develop and experimentally evaluate a hybrid method for accident and leak prevention based on risk forecasting and real-time neural network analysis of video data. Within the framework of the study, convolutional neural network architectures used for video stream processing are analyzed, and the limitations of traditional monitoring methods based on point pressure and flow sensors are shown. Based on historical, operational, and telemetry data, a predictive risk table is formed determining the priority of monitoring potentially hazardous sections of engineering systems. For sections with elevated risk levels, a hybrid neural network model is applied, including a spatiotemporal anomaly detection module and a semantic segmentation module for identifying visual indicators of leaks. The results of the experimental study, conducted on a dataset of 120 hours of video recordings from 20 industrial facilities, demonstrate detection accuracy of up to 94.7%, an F1-score of 92.9%, and an average response time of approximately 22 seconds. In conclusion it is noted that the proposed solution allows a transition from a reactive to a proactive model of engineering systems, enhances industrial and environmental safety, and can be integrated with existing monitoring means to expand the monitoring zone.