Siamese CNN-BiMGRU integrating multihead attention mechanism and its application to multipart covers fault detection under small-scale samples condition

Abstract

The detection of multipart covers (MPCs) is critical for the reliability of urban roads. In real-world scenarios, labeled data for MPCs are scarce, which presents a significant obstacle for employing deep learning techniques in MPCs fault detection. To address this issue, this article introduces a novel network named Siamese convolutional neural network-bidirectional Mogrifier-gated recurrent unit with multihead attention mechanism (CNN-BiMGRU-MHAM), which leverages metric learning to measure the similarity between different samples and facilitates network training without generating additional samples, even with limited data. Specifically, a MGRU is developed by integrating a Mogrifier coupling module into the traditional GRU to enhance the interaction of contextual information and mitigate information loss. The Siamese CNN-BiMGRU-MHAM network is designed to extract both spatial and temporal features effectively. A novel similarity measurement strategy is proposed, which considers both the distance and directional discrepancies between vectors, avoiding errors associated with distance measurements of vectors of identical length but different orientations. Additionally, this article introduces a dynamic weighted loss function that adaptively adjusts weights based on the trends in similarity value changes, enhancing training accuracy and speed. The efficacy of the proposed approach is validated through several comparisons on MPCs, demonstrating superior accuracy, noise resilience, and generalization capabilities.