Damage identification using ultrasonic Lamb waves with multi-scale adaptive attention Transformer-based unsupervised domain adaptation

Abstract

In this paper, a novel unsupervised transfer learning method – the Multi-scale Adaptive Attention Transformer Domain Adaptation Network (MAAT-DAN) is proposed for structural damage identification in regression and classification tasks. MAAT-DAN comprises three main components: a feature extractor, a domain adapter, and a label predictor. The feature extractor, for the first time, introduces a novel dynamic adaptive weighting module that enhances multi-head attention by adjusting attention head weights based on the input. Additionally, a multi-scale feature extraction technique is employed to simultaneously extract damage features at three different scales from Lamb waves, a popular means in structural health monitoring. The label predictor is composed of multiple fully connected layers. MAAT-DAN’s domain adapter integrates a gradient reversal layer for managing high-dimensional data, maximum mean discrepancy for improved stability, and feature reconstruction techniques to support effective domain adaptation. MAAT-DAN’s performance is validated in both regression and classification tasks using two specially designed ultrasonic experiments: (1) for regression, an experiment on a reinforced aluminum plate using test beams of two different materials and Blu-Tack to simulate repeatable damage; (2) for classification, an experiment on a steel water tank deploying two pairs of sensors to collect data over two years. Its feature extractors and domain adapters are compared across customized metrics against five other methods using these datasets. Results show that MAAT-DAN achieved the highest prediction accuracy and consistent prediction stability in both tasks, demonstrating its effectiveness in unsupervised domain adaptation for damage identification.