Ultrasonic Guided Wave-Based Real-Time Damage Detection via Lightweight Multitask Capsnet Under Imbalanced Data in Composite Structures

Composite materials are widely used in various engineering fields, and ensuring their safety and integrity is crucial. To monitor the structural health of composites, ultrasonic guided wave (UGW)-based methods are commonly employed. However, existing machine learning-based UGW detection methods face significant challenges, including imbalanced data distribution and limited computational resources in real-world applications. To address these challenges, this paper proposes a novel lightweight multi-task capsule (LMTC) model for real-time damage localization and quantification in composite structures. In the proposed approach, the Piecewise Aggregate Approximation (PAA) algorithm is first used to compress multi-path guided wave signals, significantly enhancing computational efficiency. The compressed ID UGW signals are then transformed into 2D Markov Transition Field (MTF) images, enabling the exploration of relationships between different amplitude components. Additionally, the LMTC model incorporates a lightweight multi-scale group convolution block and a batch capsule attention mechanism, reducing computational complexity while maintaining high accuracy, even under conditions of imbalanced data. The proposed method is evaluated on real-world datasets and achieves approximately 99% accuracy in damage identification, demonstrating its superior performance. It significantly outperforms other advanced damage detection methods, which achieve a maximum accuracy of only 89%, and exhibits excellent lightweight and generalization capabilities.