Intelligent identification method for pipeline leakage based on GPR time–frequency features and deep learning

Pipeline leakage has long been a challenge for water supply systems, and research on leakage identification has always been of interest to scholars. Currently, common leakage identification techniques mainly include model-based methods and equipment-based methods (Chen et al. 2023). Model-based methods use monitoring data such as flow meters and pressure meters to determine whether pipelines are leaking and preliminarily locate leakage areas in the pipe section between valves. Equipment-based methods (Hao et al. 2012; Juliano et al. 2013; Karthikeyan et al. 2014; Jinguuji & Yokota 2022) can achieve meter-level leakage localization, mainly including acoustic method, correlation instrument, tracer, infrared thermography, and resistivity method. The simple, low-investment, and flexible acoustic method (Shirajuddin et al. 2022) is the most commonly used method. Detection personnel analyze the leakage situation while advancing along the pipeline on the road surface relying on listening rods, which determines that the acoustic method is often interfered with by environmental noise. Therefore, researchers have begun to search for other methods that can replace manual labor in order to achieve intelligent leakage recognition.

The B-Scan image of GPR is actually composed of multiple ordered A-Scan signals, and spectral analysis can be performed on the A-Scan signals. Spectral analysis is commonly used in fields such as pollutant migration investigation, cave exploration, and pavement assessment (Marcak & Gołębiowski 2008; Szymczyk & Szymczyk 2015; Rodés et al. 2020), and can be used to distinguish different types of defects based on the sensitivity characteristics of materials to the spectrum. At present, there is limited research on the evaluation of pipeline leakages based on frequency domain analysis. Benedetto & Benedetto (2011) found that the degree to which radar waves of different frequencies are absorbed by water varies, and the frequency of scattered waves shifts toward lower frequencies with increasing moisture content. Therefore, the stratum moisture content can be evaluated based on the changes in spectral peaks, but it also requires the signal to have high resolution in spectral distribution, which limits the effectiveness of Fast Fourier Transform (FFT) (Benedetto & Tosti 2013). An effective alternative method is wavelet transform (WT) (Zhang et al. 2023), which has the properties of frequency analysis, highlighting frequency signals in A-Scan data, and reflecting the two-way-travel-time of target reflections. This allows for the extraction of changes in specified time and frequency from 1D signals simultaneously. By combining time domain B-Scan images with frequency domain A-Scan signals, leakage identification can be studied from a new perspective.

However, the vast amount of time-frequency data obtained from GPR requires the development of automated high-precision recognition methods capable of processing data in batches. Machine learning classification methods such as support vector machine (SVM) (El-Mahallawy & Hashim 2013) cannot extract detailed features due to their simple decision boundaries, and the classification accuracy depends on the manual selection of features. Recently, deep learning models represented by convolutional neural networks (CNNs) have been widely used in non-invasive underground targets or damage recognition based on GPR. According to the characteristics of the recognition object, it can be mainly divided into 1D signal and 2D B-Scan image recognition. Generally, the accuracy of signal-based recognition is higher than that of image (Tong et al. 2020; Li et al. 2022), but due to the strong interpretability of images, more research is conducted using B-Scan profiles as training and recognition targets. Li et al. (2016) evaluated the application of random Hough transform in root target recognition, locating underground tree roots based on hyperbolic diffraction waves. Zhang et al. (2020) intelligently identified abnormal reflections in B-Scan images using deep learning and incremental random sampling, thereby assessing damage in asphalt pavement. Qin et al. (2021) used ResNet-101 and feature pyramid networks (Lin et al. 2017) to extract features from B-Scan images, and then used Mask R-CNN (He et al. 2017) to detect defects in the steel ribs, voids, and initial linings of the tunnel. Regarding pipeline leakage, Liu et al. (2023) used CNNs and transfer learning to identify pipeline acoustic emission data, achieving optimal performance among multiple models. Choi & Im (2023) used CNNs to identify leakages in vibration data of water pipes, demonstrating a significant performance improvement compared to SVM. Xie et al. (2023) automatically detected leakages in infrared thermal imaging data using Faster R-CNN (Ren et al. 2015), achieving leakage identification in complex backgrounds. Typically, only 1D signals or 2D images are independently recognized in the study. Deep neural networks such as R-CNN and Faster R-CNN benefit from multiple hidden layers and non-linear structures, allowing them to uncover fine features in data and achieve intelligent classification of 1D–2D comprehensive time-frequency data. Compared to these networks, ResNet-50 (He et al. 2016) with a deeper network structure introduces residual connections to reduce the number of parameters and solve the problem of gradient vanishing or exploding. Furthermore, ResNet-50 is an end-to-end neural network that, compared to the two-stage models of the R-CNN series, has higher efficiency and is suitable for pipeline leakage recognition tasks. Therefore, this study chose ResNet-50 to intelligently identify the time-frequency scale images obtained through WT, thus achieving intelligent interpretation of the joint time-frequency domain features for water supply pipeline leakage identification.

It should be noted that the research scope of this paper is limited to pipelines with known information regarding construction location, diameter, and burial depth, a situation not uncommon in practical engineering and with considerable engineering demand. Additionally, the groundwater level should not inundate the pipeline, and areas with saturated soil medium are not within the scope of this study. The structure of this paper is organized as follows: Section 2 introduces the methodology, including the formation of GPR data, the principles of WT, and the structure of ResNet-50. Section 3 elaborates on the process of dataset establishment, the features of the data, and the preprocessing results. Section 4 explains the training and testing results of ResNet-50. Section 5 concludes and provides future prospects.

The raw data of GPR are divided into A-Scan and B-Scan according to the scanning dimension. A-Scan is a 1D profile produced by the propagation of electromagnetic waves along the longitudinal axis with the signal intensity varying with time is called A-Scan. It is essentially composed of multiple signed 16-bit binary numbers, with values ranging from −32,768 to 32,767. A set of A-Scans is arranged in an orderly manner with a certain trace spacing along the horizontal axis to form a B-Scan. B-Scan is usually transformed into a grayscale image for display according to a certain mapping method.

The temporal characteristics of radar data are reflected in the reflection wave group. When there is an electrical difference in the detected target medium, a reflection signal will be generated in the radar record. The amplitude, continuity, and phase coherence of the reflection waves are the focus of attention. A-Scan data contain waveform information of reflection waves, while B-Scan images provide a more intuitive comparison of the time domain signals of each trace data. Detection personnel can ultimately complete the geological interpretation based on processed GPR profile features such as amplitude, phase axis, and frequency.

In the B-Scan images before and after the occurrence of leakage, diffraction wave signals of the water supply pipeline will be displayed. Although the hyperbolic curves of the pipeline will shift downward after leakage, using this as a basis for discrimination may mistakenly identify deeply buried pipelines and underground unidentified buried objects in the inspection area as leakages. Therefore, only extracting the high-frequency details and low-frequency backgrounds of the B-Scan images without accurately perceiving the spectral changes in the images will not achieve accurate leakage identification. Leakage identification of pipelines must consider both the abnormal reflection waves in the time domain and the attenuation changes in the frequency domain.

The continuous updates and iterations of radar equipment have made data collection more and more efficient, but the interpretation of images still relies on experienced professional inspectors, leading to extremely high costs and very low efficiency. In recent years, various deep learning frameworks have been developed in the field of artificial intelligence, such as CNNs and transformers (Vaswani et al. 2017; Han et al. 2023), which can be used for the automatic interpretation of radar return images.

Compared with leakage identification methods based on GPR and deep learning networks, traditional model or data-based methods (Zhou et al. 2019) require obtaining parameters such as flow and pressure at multiple points, and training in hydraulic calculation models to determine the pipeline segment where the leak occurs through parameter changes. The localization is coarse and cannot achieve meter-level accuracy, resulting in considerable excavation work and a significant impact on the surrounding environment. The advantage of GPR lies in its independence from the variations of parameters within the pipeline system. It is applicable to pipes of any length and diameter, achieving meter-level precision in localization. In addition, current mainstream acoustic methods (Kang et al. 2018) are sensitive to environmental noise and require the processing of abnormal data. If the leak in the pipeline is small, resulting in weak vibration signals, sensors may fail to transmit data accurately, thus affecting the recognition accuracy. However, GPR combined with deep learning networks is not affected by the size of leakage points and environmental factors. Therefore, the method proposed in this study has significant advantages, and the relevant results will be presented in Section 4.

The collected data include both soil cover, pipeline, and pipeline leakage information, as well as various noises, which need to be weakened or eliminated to improve the signal-to-noise ratio. Since the pipeline is shallowly buried and the signal characteristics of the pipeline are already very clear, no enhancement was applied to avoid image distortion. After processing steps such as trace editing, removing the DC component, static correction, and average filtering, the contrast of the radar 2D profile was enhanced. In practical engineering, the burial depth of water supply pipelines is typically between 0.7 and 2.0 m, with soil types mainly consisting of gravel, cobble, sand, and plain fill. The radar antennas used in this study, ranging from 500 to 800 MHz, can meet the penetration depth requirements of the above-mentioned soil types, and this has been validated in subsequent field tests.

The B-Scan images in the X direction before and after the leakage can be seen in Figure 6(c) and 6(d), respectively. In Figure 6(c), on the measurement lines X16 and X19, there are bright black bands at the interface between the buried soil and the pipeline before the leakage occurs, appearing at 12 ns. This is because there is a significant difference in the dielectric constant between the buried soil and the pipeline boundary, leading to the reflection of electromagnetic waves at the interface. It can be observed from the results in Figure 6(d) that after the leakage, the previously continuous black bands are interrupted, and the two-way-travel-time of the reflection signal is prolonged, manifesting as an upward-opening hyperbola. This is because the previously uniform underground medium is disturbed by high-pressure water flow, resulting in a gradient change in soil moisture content at the same depth, causing the propagation rate of electromagnetic waves to vary.

Leakage defects are reflected differently in the time domain and frequency domain. For example, the phenomenon of the downward shift of hyperbolas in B-Scan images is intuitive, but it is prone to interference in areas where multiple pipelines are buried. On the other hand, comparing the main frequency distribution in the spectrum makes it easier to identify the presence of water, but it cannot determine whether there are pipelines in the area without excavation, which is not conducive to pipeline localization. Based on the advantages and disadvantages of the time domain and frequency domain mentioned earlier, this study uses the WT algorithm to construct time–frequency scale images, thereby conducting a time–frequency comprehensive analysis to identify pipelines and leakages.

The features perceptible to the naked eye are only a small part of the time–frequency scale image, and larger-scale frequency variations or finer features are difficult to capture manually. Furthermore, the vast amount of radar data does not support all tasks being completed manually. This study uses the ResNet-50 model to identify and classify pipeline leakages in GPR images.

The metrics achieved by ResNet-50 on the test set are shown in Table 1. The Pr of the normal area and the pipeline area reaches 0.990 and 0.968, respectively, indicating that the model can accurately determine the presence of underground pipelines. Due to significant changes in the moisture content at the leakage boundary, the Pr of the leakage area is lower, reaching only 0.812, indicating that data at the leakage boundary are prone to misdiagnosis. In terms of Re, it is the opposite, with the normal area and pipeline area reaching 0.882 and 0.880, respectively, while the leakage area is 0.998, indicating that almost all areas with leakages can be accurately identified by the model, which is crucial in practical leakage identification tasks. Therefore, based on the use of WT for image preprocessing, ResNet-50 can accurately predict time–frequency scale images, meeting the needs of engineering applications.

Efficiency experiments were conducted on the overall process of WT, ResNet-50 identification, and deviation correction. Experiments were conducted on 300 sets of data, and the final time spent was 4.95 s, with an average time of 0.0165 s per set, which fully meets the requirements of engineering inspection and can be used for real-time leakage diagnosis.

This study applies WT to convert GPR data into the time-frequency domain, then employs ResNet-50 to detect time-frequency scale images, and proposes a deviation correction mechanism to enhance result clarity at the visualization level. The combination of time domain B-Scan images and frequency domain A-Scan signals enables more accurate identification and localization of leakages, eliminating the uncertainty introduced by B-Scan images, and addressing the inability of A-Scan signals to locate leakage positions. The conclusions drawn from this study are summarized as follows:

• (1) WT applied to GPR data yields time-frequency scale images, capable of extracting temporal and spectral information from raw radar data with extremely high resolution, visually displaying differences between pipeline leakage areas and normal areas in the time-frequency domain, thus avoiding the need for manual adjustment of mapping ranges in B-Scan charts to enhance image contrast.

• (2) The trained ResNet-50 model achieved an accuracy of 0.917 and a recall of 0.998, indicating that the model can recognize almost all leakages in the dataset, while also demonstrating robustness to complex underground terrain, with minimal impact from irrelevant interfering targets on the model.

• (3) The proposed bias correction method improves the clarity of leakage prediction results at the visual level. Furthermore, the leakage intelligent recognition method proposed in this study is capable of processing radar data at a speed of 0.0165 s per trace, meeting the requirements for real-time identification.

This study also has some limitations. Water supply pipelines are just one type of underground target to be detected. Other targets such as inspection wells, communication systems, heating, and power supply systems can also affect the data characteristics of GPR. Future research should consider the detection of leakages or damages in multiple types of underground targets. On the other hand, GPR data are actually 3D data. C-Scan images can also reflect the characteristics of pipelines and leakages. In future research, the utilization of GPR data should be improved to enhance the accuracy of leakage recognition from a 3D feature perspective.

We gratefully acknowledge the support of the National Key Research and Development Program (Grant No. 2022YFF06069003-03) and the ZJU-ZCCC Institute of Collaborative Innovation (Grant No. ZDJG2021009).

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.