The presence of water is the distinguishing factor that allows Earth to harbor life. Recent studies show that global water usage is predicted to surge by 55% and almost a quarter of major cities worldwide are already grappling with some degree of water strain [1]. A lack of safe drinking water is one of the consequences of water shortages. Nearly 2.2 billion individuals globally are struggling to drink safe water [1].

In an attempt to deal with the alarming situation of water worldwide, different initiatives have been proposed and deployed. RWesearchers can cite, for example, water restrictions, re-purposing water for non-potable use, and awareness campaigns.

However, one of the major problems faced by water management institutions worldwide is the high amounts of drinkable water that are wasted because of leaking pipes in the distribution networks [2]. Water wastage is caused by various factors, those that are related to human behavior, and those that are linked to the state and the quality of the pipes in distribution networks. Indeed, pipe flaws can lead to substantial losses of quality water and wasted energy in purification processes [3]. Morocco, representing a typical example from the Global South countries struggling with water scarcity, faces significant water wastage primarily due to leakages. According to a 2019 report by the World Bank, Morocco’s water supply experiences a loss of approximately 40% due to leaks, inefficient use and aging infrastructure [4]. This is a significant amount of water wastage that highlights the need for more efficient water management practices. Water leakages in buildings’ pipelines are a common and costly problem that can cause a wide range of issues. Leaks can be caused by a variety of factors such as age, wear and tear, corrosion, damage from external factors (temperature and humidity), poor installation, or the pipeline’s quality. They can occur in any part of the pipeline system, from the main water line to individual fixtures, and can result in significant water waste, increased water bills, structural damage, mold growth, and other related problems. The early detection and repair of leaks is crucial to minimize damage and avoid costly repairs.

Among the solutions to mitigate water stress is to improve water efficiency by reducing the quantity of non-revenue water. To this end, various techniques for leakage detection in buildings’ pipelines have been developed, ranging from visual inspection to advanced technologies such as cameras, acoustic, pressure, acceleration, and flow sensors. In this context, it is critical to identify the most effective and efficient methods to detect leaks.

Numerous studies have tackled the use of machine learning or deep learning based CNN architectures to discern patterns and detect irregularities within water and wastewater pipeline systems. For example, Fang et al. ^[5][11] present a CNN-based method for identifying numerous leakage spots. The CNN model collects important features from past leakage data and applies them to real-time data to detect whether there is a leak. A simulated water distribution system experimental platform was constructed within the AnhuWe Province Key Laboratory of Intelligent Building and Building Energy Saving. The platform spans 200 m², with pipe sections measuring 400 m in length and featuring pipe diameters ranging from 30 cm to 50 cm. A total of 21 water pressure sensors were strategically placed throughout the platform. They were able to collect a dataset that contains non-leakage data and four types of pipe network pressure data under the conditions of single-point leakages, two-point leakages, and three-point leakages. The experimental findings show great detection accuracies, with 99.63%, 98.58%, and 95.25% accuracy attained for one, two, and three leakage spots employing 21 sensors, respectively. When the number of sensors is reduced to eight, the accuracies for one, two, and three leakage points drop to 96.43%, 94.88%, and 91.56%, respectively. In Kang et al. ^[6][12], the authors conducted non-invasive measurement of the leakage signals using piezoelectric accelerometers (PCB-393B31). These accelerometers possess the ability to objectively measure vibrations and translate them into acceleration levels. Furthermore, the authors introduced a local search technique based on graphs to locate leaks. Their approach, however, was not completely tuned for categorizing 1-D signals. It required feature extraction from the recorded signal data prior to applying the classification layers. Furthermore, the detection range was determined by the clarity and correlation of acoustic signals, and the issue of mistakes caused by signals with low correlation coefficients remained unsolved. In a different study by Cody et al. ^[7][13], the authors designed a water system experimental test bed that is made up of many components, including a full-scale hydrant and PVC pipes with bends. These pipes are made of grayscale schedule 80 PVC with a 152.4 mm inner diameter, which is extensively used in water distribution networks (WDNs) in Canada and the United States. The authors used a CNN architecture whose output is sent into a variational autoencoder (VAE), which tries to recover the original spectrogram image. The mean squared error (MSE) between the original image and its reconstructed counterpart is used to calculate the loss function. This proposed method achieved an accuracy of 97.2% for detecting a 0.25 L/s leak. Shukla et al. ^[8][14] conducted a study where a CNN model is built using modified layers of the pre-trained AlexNet network ^[9][15]. The purpose of the model is to classify images based on different scenarios using a dataset of 9000 scalograms images. This was achieved by considering 25 scenarios, with 12 accelerometers per scenario, 10 samples per scenario, and three images per sample. The model excels at categorizing images based on their corresponding leakage scenarios, accurately identifying healthy configurations as well as various leaky situations with a 95% accuracy rate. Moreover, the average recall, precision, and F1 score for both validation and testing data are 94% and 95%, respectively. The findings of the authors indicate that the CNN model effectively detects true positive labels with a high degree of accuracy. Coelho et al. ^[10][16] introduced an IoT system that has the ability to monitor water distribution systems and accurately detect and locate water leaks. The proposed solution uses flow sensors and affordable microcontrollers (ESP32) to collect and process real-time data. Five different classification algorithms were considered, namely, random forests, decision trees, neural networks, support vector machines, and XGBoost. A total of 12 tests were performed for each method in order to find the algorithm with the best accuracy for implementation of the system. The random forest algorithm consistently achieved the highest accuracy across various scenarios, making it the preferred choice with an accuracy of nearly 85%. Loukatos et al. ^[11][17] presented a system that solves the issues of traditional IoT systems. Using embedded ML on a Raspberry Pi Pico microcontroller board, the authors trained a neural network to recognize three characteristic kinds of water utilization profiles which are Normal Use (NU), Water Leak (WL), and Water Waste (WW). The neural network structure has an input layer with 200 features (window size), two hidden layers, with the first one to have 20 neurons and the second one 10 neurons, and an output layer with three classes. Upon evaluating the testing data, the system achieved an accuracy of 77.8% for the NU category, indicating that it correctly identified Normal Use instances. Similarly, it achieved a 100% success rate for both the WW and WL categories, accurately identifying Water Waste and Water Leak scenarios. These performance results led to an expected accuracy of 98.5% for the final model when tested using the quantized (int8) version of the dataset. The investigation of domestic water leak detection has also been examined through the analysis of flow data ^[12][18]. This research has focused on training a random forest and a CNN-based model in the cloud. The classification problem addressed in this researchtudy revolves around detecting leak events from non-leak events. They also detect the magnitude of leaks categorized as small (≤1 L/h), medium-sized (1 to 10 L/h), or large (≥10 L/h). The CNN model exhibited the best performance with an accuracy, precision, and recall ranging from 92% to 96%. Additionally, the area under the Precision-Recall (PR) and Receiver Operating Characteristic (ROC) curves consistently achieved high values, ranging between 97% and 99%. A summary of the salient features from the literature researchview is provided by Table 1. The existing literature in the area of water leakage detection systems reveals a noticeable gap in the integration of TinyML methodologies. While prior efforts have been dedicated to tackling the water leakage challenge, a significant portion of these solutions has not capitalized on the potential benefits of TinyML approaches. Furthermore, the few endeavors that do involve TinyML techniques often exhibit limitations, particularly in terms of the scope and the considered CNN models. ResWearchers make a contribution by experimenting with state-of-the-art CNN models, such as ResNets ^[13][19], MobileNet ^[14][20] and EfficientNet ^[15][21] architectures, for highly accurate and efficient water leakage detection. Moreover, a distinctive aspect of ouresearchers' work is the attention researchers we devote to the integration of these models into small-scale devices using TinyML. This two-fold approach not only broadens the spectrum of the considered ML models but also paves the way for the seamless incorporation of cutting-edge technology into an efficient real-time operational framework of water leakage detection.