Water Leakage Pre-Localization in Drinking Water Networks via the Cosmic-Ray Neutron Sensing Technique ^†

Abstract

Water leaks in drinking water networks contribute significantly to water losses and pose challenges to infrastructure sustainability. This study introduces a novel approach using Cosmic-Ray Neutron Sensing (CRNS) for pre-localizing leaks. We present a CRNS-based roving system that can detect the neutrons and muons produced by cosmic rays, providing real-time, below-ground water content data while addressing local variations. These data are analyzed using a one-class Support Vector Machine trained in an unsupervised manner. Finally, a brief overview of a proof of concept of the method conducted in a water district in Northern Italy is shown, highlighting preliminary results alongside some limitations.

1. Introduction

Water leaks in drinking water networks pose significant challenges to water resource management, infrastructure sustainability, and the overall efficiency of urban water supply systems. Recent studies estimate that leaks account for up to 70% of total water losses [1]. Due to its importance, innovative monitoring technologies have been developed during the past decade to detect and locate leaks quickly. This work presents a new approach for pre-localizing water leakage in drinking water networks using the Cosmic Ray Neutron Sensing (CRNS) technique. The CRNS method relies on detecting above-ground epithermal neutrons, which primarily interact with hydrogen nuclei. By exploiting the inverse correlation between the hydrogen content and epithermal neutron counts above ground, the CRNS technology allows for the estimation of the water content of soil using an above-ground probe [2,3]. Due to its ability to measure soil moisture on a large scale, in depth (tens of centimeters) and in real time, this method is widely adopted in agricultural applications and for monitoring snow water equivalent in mountain areas. In recent years, the use of CRNS sensors installed onto dedicated vehicles has extended monitoring capabilities from the field scale to the regional scale [4]. In this work, we propose using this technique to identify soil moisture anomalies below ground (up to 50 cm) in an urban environment potentially linked to a leakage in the drinking water network.

2. Materials and Methods

2.1. Cosmic Ray Neutron Rover System

The soil moisture mapping is carried out using a dedicated vehicle equipped with a set of 10 Finapp-5 probes (www.finapptech.com, accessed on 20 August 2024), a scintillator-based sensor, which mainly detects epithermal neutrons (up to 1600 counts per hour, in standard conditions, defined as reference pressure at sea level; air humidity < 5 g/m³; soil gravimetric moisture = 5%; geomagnetic cutoff rigidity = 5 GV; reference incoming cosmic ray flux = 166 counts from the NMDB-JUNG observatory). This setup has demonstrated strong performance relative to other commercial CRNS probes [5]. Moreover, it was recently used for agricultural rover applications [6]. The probe can also detect muons, enabling the implementation of corrections accounting for local effects (such as building density in an urban environment).

2.2. Data Collection and Analysis

During the mobile survey, raw neutron counts sampled at 1 Hz and tagged with GNSS locations are spatially grouped into a 10 m pixel size grid. Each pixel represents neutron counts, adjusted for the total time spent within the pixel area, which varies with vehicle speed. The vehicle speed was limited to below 30 km/h to maintain an optimal signal-to-noise ratio. After this aggregation, the neutrons and muons count rate in each cell was smoothed, considering the probe footprint in the following way:

$$P\left(c_{ij}\right)={\displaystyle \frac{{\sum }_k{\sum }_l F\left(c_{ij},c_{kl}\right)P\left(c_{kl}\right)}{{\sum }_k{\sum }_{l}F\left(p_{ij},p_{kl}\right)}}$$

(1)

where $P\left(c_{ij}\right)$ is the particle count rate at the cell (i, j) (i.e., neutrons or muons counts), $F\left(c_{ij},c_{kl}\right)$ is a function that weights the contribution depending on the distance between the cell (i,j) and the cell (k,l) using the probe footprint [7].

These data serve as input for anomaly detection using One-Class Support Vector Machines (SVMs) [8]. One-class SVMs are particularly suitable for unsupervised anomaly detection tasks where only normal (i.e., non-anomalous) data are available for training. By learning the characteristics of normal data, One-Class SVMs can identify deviations from these patterns, highlight potential anomalies in the dataset, and represent leaks. For this purpose, normal data were acquired along a set of roads without leaks, as they were recently inspected using traditional electroacoustic methods.

3. Results

Case Study

In collaboration with Piave Servizi S.p.A., we conducted a comprehensive proof of concept (PoC) in the Conegliano Sud district (Treviso, Italy). This district, encompassing 43 km of drinking water network, provided an ideal testing ground with its extensive urban environment and limited rural areas. Data collection occurred throughout the district on a typical working day in late November 2023.

The final output of the data analysis pipeline consists of a GIS map showcasing pre-localized areas suspected of water leaks (see Figure 1). These areas are represented by segments along the water network pipes varying in length between 20 and 160 m, providing a targeted approach for traditional investigation. Indeed, the pre-localized areas identified through CRNS analysis were further investigated using electroacoustic instrumentation, including listening sticks, geophones, and noise correlators. About 9.5% of the global water network is highlighted on the pre-localization map, which is 4.1 km long.

Additionally, to further validate and provide a point of comparison, a systematic search using electroacoustic methods was conducted independently from the search performed through pre-localization mapping. Although this systematic search did not cover the entire district, it examined 10 km of the water network. These data help estimate, albeit partially, any false negatives produced by the pre-localization mapping. The results of the proof of concept are summarized in

Table 1. Overview and quantitative features of the experimental campaign.

Feature	Value
Length of the global water network	43.5 km
Length of pre-localized water network	4.1 km
Number of pre-localization spots	43
True positive	11
False negative 1	2

¹ with respect to 10 km of systematic electroacoustic searching.

as follows:

4. Discussion

The PoC’s outcomes showed promising results for water leaks pre-localization. However, it is essential to acknowledge that the technology still exhibits limitations due to high heterogeneity in real-world scenarios. For instance, factors that can interfere with measurements, such as hydrogen-rich sources like petrol pumps, pose challenges to accurate detection. Moreover, this method cannot detect all the leaks that have not raised the soil moisture of the first 50 cm of soil. These interferences may contribute to false positives or false negatives, impacting the reliability of the CRNS-based leak detection system. Addressing these limitations will require further research and development efforts to enhance the robustness and accuracy of the technology.