Systematical vibration data recovery based on novel convolutional self-attention networks

In the field of structural health monitoring (SHM) for civil infrastructures, data loss unavoidably occurs in measured data due to harsh operational conditions, significantly impacting the effectiveness and reliability of structural condition assessment. Prompt and effective recovery of lost data is, therefore, crucial. However, existing methods have limited generalization ability and struggle to handle complex data loss patterns. This study designs a convolutional cross self-attention (CCSA) module and proposes a CCSA based convolutional network (CCSA-ConvNet) for systematical data recovery. Its architecture facilitates automatic noise reduction through convolutional operations and highlights stably express and highly correlated elements among the convolutional feature maps by self-attention, enabling the extraction of robust and representative features from severely incomplete data. Experimental studies conducted with SHM data of Canton Tower validate the superiority of the CCSA-ConvNet architecture in data recovery and noise resistance, indicating its strong feature extraction and generalization capabilities. The identified modal parameters from recovered data consistently align with real values, demonstrating the applicability of recovered data. This systematic data recovery approach is user-friendly and highly promising for practical applications in civil engineering.