CNN-LSTM Neural Network Applied for Thermal Infrared Underground Water Leakage

In this paper is proposed a methodological approach for the detection of leaks in water pipelines. The approach is based on the use of Infrared Thermography (IRT) for the real time monitoring, and of Artificial Neural Networks (ANNs) for the identification of potential leaks not easy visible by IRT. The input data source consists of radiometric data processed by Convolutional Neural Networks (CNNs) providing as output information about the presence or absence of the water leakages. A preliminary study has been carried out about a fixed monitoring station created to identify leaks in underground pipelines. In addition, has been implemented a platform to remotely acquire the thermograms and to analyze them by CNN networks detecting leakages, combined with Long Short-Term Memory (LSTM) neural networks, and image filtering algorithms, such as image segmentation and active contour snake approach. The LSTM network allows the prediction and calculation of the propagation trend of the leak plume. Finally, image filtering improves the visualization of leaks as it allows to draw the contours of the pixel clusters representing the leakages areas in the thermograms. The work was developed within the research framework of an industrial project. The proposed approach is suitable also for oil spill and gas leakages detections.