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

The identification and localization of water pipeline leakages based on ground penetrating radar (GPR) technology are gradually becoming a research hotspot. Current methods mostly focus on exploring the patterns of B-Scan images, heavily relying on the subjective experience of detection personnel, which can lead to misjudgments. Moreover, the large amount of data makes it difficult for manual processing. Therefore, a method based on wavelet transform (WT) and ResNet-50 is proposed to identify the time-frequency characteristics of GPR data, thereby achieving intelligent localization of pipeline leakages. The B-Scan images from GPR are transformed into time–frequency scale images using WT, and the features in both time and frequency domains are combined to enhance the representation of leakages. Subsequently, ResNet-50 is employed for feature extraction and leakage identification. Additionally, a deviation correction mechanism is proposed to improve the clarity of the prediction results. Experimental results demonstrate that ResNet-50 achieves an accuracy of 0.917 and a recall of 0.998 on the time-frequency dataset, almost detecting all leakages, with a recognition efficiency of 0.0165 s per data trace. The comprehensive method is validated in the field, indicating its capability to accurately identify and localize pipeline leakages.