IMPROVING THE EFFICIENCY OF PIPELINE LEAK DETECTION SYSTEMS USING NEURAL NETWORKS

Pipeline systems for oil and gas transportation are complex distributed infrastructure facilities whose efficient and safe operation largely depends on the application of modern information and communication technologies. In the context of industrial digital transformation, intelligent monitoring systems capable of continuous acquisition, transmission, and analysis of telemetry data for the early detection of emergency conditions have become particularly relevant. One of the most critical challenges is the timely detection and accurate localization of leaks, which can result in significant economic losses, environmental damage, and threats to public safety. The objective of this study is to develop an approach for determining the coordinates of a pipeline leak based on intelligent processing of measurement data using machine learning methods. The proposed solution is intended for integration into information and communication systems for supervisory control and digital monitoring of pipeline transport. A two-layer multilayer perceptron implemented in the MATLAB environment is employed as the data analysis tool, enabling the development of a computationally efficient algorithm suitable for practical use in decision-support systems operating in near-real-time conditions. The neural network was trained on experimental datasets generated for various leak locations and flow rate values of the transported medium and was tested on independent datasets. The influence of the number of neurons in the hidden layer on leak localization accuracy was investigated. Maximum and root mean square localization errors were used as performance metrics. The results demonstrate that increasing model complexity by raising the number of neurons beyond 3–4 does not lead to a significant improvement in accuracy and may be accompanied by overfitting, thereby reducing the reliability of the algorithm when processing new data. It was found that the optimal neural network architecture comprises three neurons in the hidden layer, providing a root mean square error of approximately 2 km and a maximum error not exceeding 5.5 km. The obtained results confirm the effectiveness of neural network methods for intelligent analysis of telemetry information and demonstrate the feasibility of developing scalable information and communication systems for early leak detection. The practical significance of this work lies in improving the accuracy of accident localization and reducing pipeline operation risks through the implementation of intelligent data processing algorithms within digital industrial monitoring platforms.