Leak detection techniques based on Machine Learning (ML) models can assist or even replace manual work in leak detection operations in water distribution systems (WDSs). However, studies on leakage detection based on on-site leak signals are limited compared to studies on lab-scale leak detection. The on-site leak signals have stronger interference and randomness, while leak signals in the laboratory are relatively simpler. To better assist on-site leak detection operations, the present paper develops and compares three ML-based models. For this purpose, many on-site tests were carried out, and tens of thousands of sets of on-site leak detection signals were collected. More than 6000 sets of these signals were marked and the signal features were extracted and analyzed from a statistical point of view. It was found that features such as the main frequency, the spectral roll-off rate, the spectral flatness, and one-dimensional (1-D) Mel Frequency Cepstrum Coefficient (MFCC) could well distinguish the leakage signals from non-leakage signals. After training the decision tree model, the performances of the random forest and Adaboost models were thoroughly compared. It was found that the false positive rates of the three models were 9.80%, 8.27% and 7.35%, all lower than 10%. In particular, the Adaboost model had the lowest false positive rate of 7.35%. The recall rate of the random forest and Adaboost models were 100% and 99.52%.

The leakage detection of a water distribution system (WDS) needs the support of a large number of field data. This paper collected over 6800 leak detection signals from cast iron pipelines used in a WDS. We found that 3280 signals indicated leakage, and the remaining indicated no leakage. The characteristics of the signals were extracted and analyzed from three perspectives: the central frequency of the power spectrum, the spectral roll-off rate, and the spectral flatness. Significant statistical distributions were found. The central frequencies of the leakage signals followed the normal distribution, and their spectral roll-off rates demonstrated the Burr distribution; the Birnbaum–Saunders distribution could describe the spectral flatness of the signals. Based on these characteristics, the recognition rate of the ML model for leak detection was improved. The Random Forest model was used to classify the leakage detection signals. The recall rate was 100%, and the false positive rate was 8.27%.

The discharge coefficient of an orifice is important for the outflow through the orifice. While it cannot be quantified theoretically as the outflow through an orifice depends on a number of parameters such as the pipe pressure, the liquid velocity and the shape and the area of the orifice. In this study, experiments and computational fluid dynamics (CFD) simulations were performed to find out the factors influencing discharge coefficient and the corresponding mechanism. The CFD simulation is based on Navier–Stokes equations combined with RNG k − ε turbulence model. The results show that a negative exponential function could fit the relationship between orifice discharge coefficient, pipe pressure, and orifice area more accurately. The relationship between the discharge coefficient of the orifice and the velocity was linear. In general, the simulation results fit well with the experimental results, which indicates that CFD simulation could be used to study pipeline leakage.

There are two key issues in the safety assessment of the water distribution system (WDS). One is how to evaluate the safety levels of water supply for customers, while the other is how to describe the importance of a pipe for the global or local WDS. The water demand guarantee rate (DGR) and the water demand failure rate (DFR) are proposed. The mathematical expectations of the DGR and DFR describe the average customer’s water safety levels for the first issue. Moreover, the unit influence of pipe failure (UIPF) is put forward for the second issue. It describes the importance of the pipe for the global or local system. Several cases show how to calculate the above values with the pressure-driven model. It is also shown how to find key pipelines in the WDS. The results show that the method can provide an effective reference for real-life WDS management.