Burst Localisation in Water Pressurised Pipelines Combining Numerical Data Generation and ANN Transient Signal Processing ^†

Abstract

Transient test-based techniques have been widely identified as one of the best non-intrusive techniques that exploit the propagation of pressure waves along pressurised pipelines, allowing the check of the status of the distribution systems. Although several studies have demonstrated the suitability of this technique for identifying anomalies in transmission pipelines, including leaks, the potential for automatically analysing transient signals through deep learning procedures has only been superficially investigated. With this aim, this study proposes how a proper synthetic generation of transient signals based on numerical simulations can support the development of neural network-based methodologies for water leak detection and localisation.

1. Introduction

Sustainable management of water distribution systems is imperative in the face of increasing water demand driven by socio-economic factors and exacerbated by climate change [1,2]. Ageing infrastructure and poor maintenance are often the cause of large water losses along the distribution network, resulting in critical issues such as intermittent service, increased energy consumption and costs, poor water quality, and scarce security [3,4].

Among the different anomaly detection methodologies, Transient Test-Based Techniques (TTBT) are non-invasive methods that do not require physical access to the pipeline, reducing the need for disruptive excavation or shutdowns for inspection, resulting in cheap and feasible even in the case of remote and difficult-to-access pipelines [5]. Usually, experts interpret the analysis of the recorded transient signals to identify anomalies on the base of the pressure signature specific to each system [6]. Instead, the automatic processing of the pressure signal has been poorly investigated, mainly due to the absence of big datasets needed to train ML methods [7].

Given this gap, this study proposes a framework for developing ML tools suitable to support the challenging task of leak identification in pipelines through pressure transient analysis. Therefore, this approach is founded on the need to collect big data on transient signals, including a large variety of conditions, with the final scope of training ML algorithms that can automatically identify leaks. Specifically, we propose a robust and flexible numerical method capable of generating many hydraulic unsteady simulations of pressurised pipelines with inducted pressure waves and pipe faults. The generated dataset is then exploited to train two artificial neural network (ANN) methodologies seeking to detect and localise leaks, respectively.

2. Methods

2.1. Numerical Hydraulic Model

In this paper, we adopt a one-dimensional semi-implicit scheme based on the work presented in [8]. While the pipe is assumed elastic and described by a rigidity coefficient, the fluid density can change as a function of the pressure. We then modify the algorithm to allow local material failure and mass loss due to localised fractures. The convective terms are discretised explicitly, while the pressure forces are treated implicitly. This led to a stability condition that is only limited by local velocity instead of the celerity that, in this context, is in the order of kilometres per second. The algorithm results are unconditionally stable if a semi-Lagrangian approach is adopted or convective effects can be neglected. Substituting the discrete momentum in the discrete continuity equation results in a mildly nonlinear system for the only unknown pressure. This type of system can efficiently be solved by the class of nested-Newton schemes proposed in [9]. The algorithm involves the solution of a linear system for every Newton stage. The system is symmetric and positive definite, so it can be solved using a matrix-free implementation of the conjugate gradient method. The resulting algorithm is mass and momentum conservative, hence allowing the precise mass loss due to pipe bursts to be quantified.

Furthermore, the high flexibility in the choice of the time step allows the simulation of both the steady state and the transient signal to be very efficient, allowing fast and accurate simulations as needed in the dataset generation. Since the target is to generate a large dataset, a simple parallelisation is employed by splitting the parameter set and not the nonlinear solver. In this way, each thread results independently from the others, and the parallel architecture is fully exploited.

2.2. Machine Learning Methodology

A data-driven framework is used to analyse the data generated by the numerical hydraulic generator for the twofold aim of firstly detecting if there is any leak in the pipeline systems and then, if there is, localising it. This methodology takes transient pressure signals at the metering point and returns the possible occurrence of any anomaly as the first step and the leak position as the second one. Specifically, two different models are used for the two different steps, but both are based on a multilayer perceptron (MLP) comprising an input layer with 256 inputs, hidden layers, and an output layer with one outcome, i.e., leak occurrence and leak position, respectively. The number of hidden layers and nodes for each are defined by a grid search to identify the best configuration.

3. Material

The synthetic case study consists of a single long pipe of 1 km with a diameter of 150 mm. At the top, reservoir S guarantees the piezometric head with a pressure of 55 m, while at the end of the transmission, the pipeline is located at a pressure valve PV, preceded upstream by a pressure meters MP. The PV is responsible for the manoeuvre that generates the pressure transient measured in MP. Specifically, in the hydraulic simulations, the transient phenomenon is propagated by closing the PV from 100% open to 50% in 0.1 s. Regarding dataset generation through unsteady simulations with the proposed numerical model, 9000 runs were performed, and 1000 for testing. The probability of leak occurrence in a simulation has been set to 70%. The parameters that have been randomly perturbated to generate a heterogeneous dataset are the water flow rate, ranging between 30 L/s and 100 L/s, the water leak characteristics, i.e., location along the pipe (20 m–980 m), discharge coefficient of the leak (2 × 10⁻¹ m5/2/s–1 × 10⁻² m5/2/s), and the pipe degradation of the pipe elasticity (5–50%). Finally, the parameters of the PV closure have also been perturbed by means of a Gaussian distribution of the manoeuvre uncertainty using a standard deviation of 0.01 s for the starting time, a standard deviation of 0.005 s for the duration of the operation, and the degree of valve closure operation.

4. Results

The first outcome proposed is the generation of the transient signals collected at the MP during the unsteady simulations with the main parameters randomly perturbated. The generation processes are performed in parallel with 20 workers, and it takes 183 min. Figure 1 shows twenty simulated signals with leaks in different positions.

The second and third findings detectan anomaly in the MP pressure signal and its localisation if presented. The detection model consists of an MLP with 64, 16, and 4 nodes for each of the three hidden layers, and the localisation one instead is three hidden layers with 48, 24, and 8 nodes. The first MLP model can predict the presence or absence of any leak with 100% accuracy, while the localisation model presets a performance of 0.6 m. The leak localisation performance by the second MLP model is illustrated in Figure 2.

5. Conclusions

This study highlights the critical rule of reliable unsteady hydraulic simulations using a flexible finite volume scheme that allows to generate a large number of pressure transient signals, which properly emulate metered pressure measurements, including a large variety of changing conditions, such as the occurrence or not of leaks, different failure characteristics, e.g., size, position and pipe deterioration, and several hydraulic boundary conditions, e.g., flow rate and hydraulic head at the reservoirs. Finally, the proposed analysis shows promising results in both detection of anomalies and localisation of the leak, which are suitable for future in-field applications with the goal of continuously assessing pressurised water pipelines through data collection by smart devices and real-time feedback by ML applications.