Smart Water Meter Based on Deep Neural Network and Undersampling for PWNC Detection

Having become aware of how limited all the natural resources are, the water leakage problem in piping systems has become a stimulating topic. This problem increased over the past few years even though innovative tools and techniques appeared in the literature and in the consumer market. Identifying water leaks at the nearest point, the household level, is still an unsolved problem because most water meters are mechanical and, therefore, cannot detect leaks. While the issue is not important for water service providers since consumption is charged to the user, the resolution is crucial due to the increasingly relevant concern of saving natural resources. The detection of small but continuous leaks of drinking water in domestic systems is addressed in this work. Machine learning approaches enabled image processing techniques also in uncontrolled environments, overcoming the classical methods but introducing new challenges such as power consumption. Using a combination of convolutional neural networks (CNNs) and long short-term memory (LSTM) networks within an adaptive undersampling strategy, it is possible to process the images captured from the mechanical water meter dial and identify the period with null consumption (PWNC) or the consumption class. The presented solution can classify the water flow into four different classes, and, in the case of absence or small flow, its function becomes to detect leakages. Analyzing images from a mechanical water meter quadrant, it has been possible to identify PWNC and detect small water leakages in the domestic environment under common consumer flow profiles. In addition to the confusion matrices, the synthetic parameters of Sørensen–Dice coefficient (DSC) and Jaccard Index have been used and presented to quantify the performance of the proposed deep neural network (DNN). The conducted experiments on static and dynamic water flow demonstrated the applicability of this approach and the possibility of an increase in PWNC identification, thanks to the adaptative increase in the sampling time. Moreover, the reduction in sampling time allows for the reduction in computational load and power consumption in embedded scenarios where limited energy is available.