Abstract: ​​ Thin plate welding constitutes a core process in shipbuilding. During this process, thin plates are prone to various defects, posing potential safety hazards during vessel service. This paper proposes a ​data-driven full-chain diagnostic solution​ for defects in marine thin plate welding. First, the ​Mamba Residual Graph Convolutional Network for Spatio-Temporal Reconstruction (MRGSTR)​ is introduced, which facilitates ​real-time unsupervised anomaly detection​ on multivariate time-series (MTS) data from welding sensors based on a reconstruction approach. Subsequently, a ​CNN-BiGRU-based defect detection method​ is designed, and ​Bayesian networks​ are employed to fuse time-series data with a ​process knowledge graph​ for ​root cause analysis. Finally, a ​multi-agent Retrieval-Augmented Generation (RAG) system​ for operation and maintenance of thin plate welding defect detection is constructed, when a comprehensive defect detection chain is established: “Sensor data anomaly detection — Defect identification — Root cause analysis — Repair recommendations”.