Multi-Task Deep Transfer Learning Method for Guided Wave-Based Integrated Health Monitoring Using Piezoelectric Transducers

Abstract

Deep learning networks provide an end-to-end monitoring method for guided wave based structural health monitoring when the model is deep enough and the training samples are sufficient. However, it is still a great challenge to conveniently transfer one monitoring task to another task to establish a multi-task monitoring model. This paper proposes a multi-task integrated health monitoring method based on deep transfer learning to realize the monitoring task transference in the plate structure. First, the guided wave-convolutional neural network is used as a general feature extraction model to extract the high-level features, and shallow regression network is trained to assess the damage level. Then multi-task feature sharing mechanism is applied to catch the shared features and to ensure the model not specifically fit into a single task. Finally, a deep network with branches is designed to output multiple monitoring labels to realize the intelligent recognition of multi-task monitoring. The optimization of the network hyperparameters and the influence of different transfer mechanisms are further discussed, and the accuracy of the proposed method with transferred feature is considerably higher than direct training. The experimental results illustrate that the proposed method can effectively transfer the damage level monitoring model to damage location monitoring model, and the location detection accuracy reached 98.14% with 15.24% improvement compare with direct training. The proposed method also presents a better detection performance compared with many other deep learning methods.