Abstract
The failure accidents of prestressed concrete cylinder pipe (PCCP) seriously affect the economic feasibility of the construction site. The traditional method of needing to stop construction for pipe inspection is time-consuming and laborious. This paper studies the PCCP broken wire identification algorithm based on deep learning. A PCCP wire-breaking test platform was built; the Distributed Fiber Acoustic Sensing Monitoring System (DAS) monitors wire-breakage events in DN4000mm PCCPs buried underground. The collected broken wire signal creates a time-frequency spectrum diagram dataset of the simulated broken wire signal through continuous wavelet transform (CWT). Considering the location of equipment limitations, based on the YOLOv5 algorithm, a lightweight algorithm, YOLOv5-Break is proposed for broken wire monitoring. Firstly, MobileNetV3 is used to replace the YOLOv5 network backbone, and Dynamic Conv is used to replace Conv in C3 to reduce redundant computation and memory access; the coordinate attention mechanism is integrated into the C3 module to make the algorithm pay more attention to location information; at the same time, CIOU is replaced by Focal_EIoU to make the algorithm pay more attention to high-quality samples and balance the uneven problem of complex and easy examples. The YOLOv5-Break algorithm achieves a mAP of 97.72% on the self-built broken wire dataset, outperforming YOLOv8, YOLOv9, and YOLOv10. Notably, YOLOv5-Break reduces the model weight to 7.74 MB, 46.25% smaller than YOLOv5 and significantly lighter than YOLOv8s and YOLOv9s. With a computational cost of 8.3 GFLOPs, YOLOv5-Break is 71.0% and 78.5% more efficient than YOLOv8s and YOLOv9s. It can be seen that the lightweight algorithm YOLOv5-Break proposed in this article simplifies the algorithm without losing accuracy. Moreover, the lightweight algorithm does not require high hardware computing power and can be better arranged in the PCCP broken wire monitoring system.