Adaptive weight updating multi-task deep learning for Lamb wave-based bolt group loosening evaluation

Abstract

Accurate evaluation of the bolt connection state in steel structures is of great significance for ensuring structural safety. In bolt group loosening detection, the identification of loosening location and loosening degree are two critical tasks. However, most existing detection methods are restricted to dealing with a single task. To address this issue, this study proposed an adaptive weight updating multi-task deep learning (AWU-MTDL) model based on Lamb waves, aiming to achieve simultaneous high-precision identification of both bolt loosening location and degree. In the AWU-MTDL model, the task-specific feature extractors and the adaptive weight updating mechanism were introduced to enhance multi-task feature extraction capability and improve collaborative learning across tasks. The Lamb wave signals were collected under different bolt loosening conditions in the laboratory environment and employed for training and testing the AWU-MTDL model. Results show that the AWU-MTDL model achieves high accuracy in identifying both bolt loosening location and degree, while also exhibiting excellent noise resistance. Compared with single-task deep learning models, the hard-parameter-sharing MTDL model, and the GradNorm-based method, the AWU-MTDL model achieves superior identification accuracy with higher computational efficiency and better task balance for heterogeneous sub-tasks. Although trained on single-bolt loosening data, the AWU-MTDL model still accurately identifies multiple loosened bolts, demonstrating strong generalization to complex unknown conditions.