A lightweight and real-time deep learning approach for pipeline bolt loosening detection using active guided waves

Abstract

Loosening of bolt connections in industrial pipeline systems poses significant risks to structural integrity and operational safety. However, conventional detection methods often suffer from low efficiency and poor robustness in complex environments. To address these challenges, this study proposes a lightweight and real-time bolt loosening detection framework based on active guided waves and multi-channel piezoelectric sensing, enhanced by advanced deep learning techniques. Specifically, an eight-channel piezoelectric sensor array is used to capture guided wave responses, which are transformed into two-dimensional (2D) representations via multi-scale feature fusion and local enhancement to facilitate deep learning. A total of 34 complex loosening scenarios—including single, adjacent, diagonal, and multi-bolt combinations—are experimentally simulated under diverse noise conditions to emulate real-world disturbances. An improved ResNet18 architecture is developed by integrating a multi-head attention mechanism for capturing long-range dependencies, along with a Squeeze-and-Excitation (SE) module for adaptive channel recalibration. Experimental results show that the proposed model achieves 99% detection accuracy under noisy conditions, with inference latency ranging from 7.2 to 10.1 ms and a throughput of 114–163 FPS (frames per second), fulfilling real-time requirements. Ablation studies confirm the effectiveness of the attention and SE components. Compared with deeper models such as ResNet50 and VGG16, the proposed method significantly reduces parameter count (13.02M) while maintaining competitive performance, enabling efficient edge deployment. Furthermore, few-shot learning experiments demonstrate that over 90% accuracy can be achieved with only five training samples in previously unseen working conditions. This research provides a robust, efficient, and scalable solution for intelligent structural health monitoring of pipeline bolt assemblies and offers valuable insights for fault diagnosis under complex industrial noise environments.