Employing Machine Learning and Deep Learning Techniques for Leak Detection Based on Infrared Images of Pipelines

Abstract

Leakage in pipelines conveying water or hydrocarbon fluids poses significant risks, including injuries, environmental disasters, and economic losses. To mitigate these risks, a preventive and non-destructive inspection method is crucial. This research proposes a hybrid model combining two-steps control: an artificial neural network Multi Layer Perceptron (MLPClassifier) model utilizing metadata and a Convolutional Neural Network (CNN) analyzing thermal images for leak detection. Moreover, the severity of leaks is assessed through leak propagation analysis using a time-based image dataset. The analysis of the infrared images dataset, along with the associated metadata, provides valuable information, such as leak detection, pipe failure conditions, and leak propagation assessment. Initially, the model analyzes the background information about the pipes, including factors such as pipe age, material, and installation quality, to identify potential leaking pipes. Subsequently, thermal images of the detected pipes, classified into the category of 'exhibiting pipe failure' based on a predefined threshold, are captured using a thermal camera-equipped drone, eliminating the need for open inspections. These thermal images are then input to the CNN model for potential leak detection. Experimental results demonstrate that the two-steps based hybrid model achieves 94% accuracy with the MLPClassifier model and 96.4% accuracy with the CNN model. This approach offers an effective and reliable solution for non-destructive leak detection in pipeline networks, combining the advantages of both metadata analysis and thermal image processing. © 2025 IEEE.