Deep learning-based time series classification for damage assessment in multi-layer composite structures

Purpose The purpose of this research is to address existing limitations in structural health monitoring (SHM) by proposing a novel deep learning model specifically designed for accurate time-series data classification. By integrating advanced neural network architectures – including 1-Dimensional Convolutional Neural Networks (1DCNN), Residual Networks (ResNet), Bidirectional Gated Recurrent Units (BiGRU), and Attention Mechanism (AM) – the proposed model effectively captures spatial relationships, highlights critical input sequences, maintains long-term dependencies, and mitigates vanishing gradient issues. The objective is to significantly enhance damage detection accuracy and reliability in multi-layer composite structures, improving upon existing ML methods by achieving superior precision and lower loss values in various structural damage scenarios. Design/methodology/approach The proposed methodology combines advanced deep learning techniques for classifying structural damage from time-series data. Initially, spatial features are extracted using a 1-Dimensional Convolutional Neural Network (1DCNN). A Bidirectional Gated Recurrent Unit (BiGRU) captures temporal dependencies in both past and future contexts. Residual Networks (ResNet) with double skip connections are employed to address vanishing gradient issues by merging initial inputs with intermediate features. An Attention Mechanism (AM) is integrated to emphasize critical segments within the input sequences. The model is validated experimentally on multi-layer composite structures under various damage scenarios, assessing performance through metrics such as precision and loss. Findings The study reveals that the proposed 1DCNN-BiGRU-ResNet-AM model significantly outperforms alternative methods, including standalone 1DCNN, GRU, BiGRU, combined 1DCNN-BiGRU, and 1DCNN-BiGRU-ResNet models. Specifically, it achieves superior damage detection accuracy with a precision of 0.902 and a loss value of 0.416. These findings demonstrate that integrating spatial feature extraction, bidirectional temporal analysis, residual connections, and attention mechanisms effectively enhances the identification and classification of structural damage. The results validate the proposed model's ability to overcome limitations related to spatial connectivity, gradient vanishing, and temporal dependencies, proving its practical effectiveness for structural health monitoring tasks involving multi-layer composite structures. Research limitations/implications This study focuses primarily on controlled laboratory scenarios using multi-layer composite structures, which may limit direct generalization to complex real-world structural health monitoring tasks. Future research should validate the proposed model using diverse field data, considering environmental noise and operational variability. Additionally, while the model addresses vanishing gradients and captures spatial-temporal dependencies effectively, its computational complexity and resource demands might be high, potentially constraining deployment in resource-limited settings. Exploring model optimization, efficiency enhancements, and assessing scalability for practical structural monitoring systems would extend its applicability and robustness. Originality/value The study's originality lies in developing a novel integrated 1DCNN-BiGRU-ResNet-AM model, uniquely combining advanced deep learning techniques to significantly enhance structural health monitoring accuracy. By leveraging 1DCNN for spatial feature extraction, BiGRU for effective temporal modeling, ResNet's dual skip connections to alleviate vanishing gradients, and Attention Mechanism (AM) for emphasizing critical data segments, the model uniquely addresses existing limitations. The approach also incorporates extensive feature integration, optimized network parameters, and sophisticated data augmentation methods. This comprehensive integration demonstrates notable advancements over traditional methods, providing robust, accurate, and reliable damage prediction in composite structures, substantially contributing to the field of intelligent structural health monitoring.