Conv-Transformer based few-shot learning for highly accurate multi-task structural health monitoring via piezoelectric impedance

Impedance signals for structural health monitoring are often sparse and difficult to acquire in damaged conditions. Increasing the damage categories would significantly reduce accuracy. In this study, we propose a Conv-Transformer model that is capable of multi-task structural health monitoring, addressing the complexities of small sample datasets while handling multiple fault detection tasks, including mass loss and bolt loosening. The model enhances feature extraction by combining convolutional layers and multi-head attention within the Transformer encoder, focusing on the relative location of the peaks and the local feature of each peak in the impedance signals. These advantages enable highly accurate multi-task SHM with small samples of impedance signals. The proposed model is first trained on a large amount of data in mixed conditions and then fine-tuned with small sample data for an eight-class fault classification task. Experimental results show that the model demonstrates strong learning ability and cross-condition transferability, achieving an accuracy of 92.12% for multi-task damage identification, a 4.49% improvement over a conventional Transformer baseline. The proposed method can be applied to health conditions identification of buildings, bridges, and trusses.