An interpretable Siamese-GRA-TOPSIS framework for early warning of quality risks in digital manufacturing

Quality development in digital manufacturing is shaped by intertwined technological, organizational, and environmental risks, which are often difficult to assess with conventional methods. This study proposes a hybrid early-warning framework, Siamese-GRA-TOPSIS, to support risk identification and prioritization in manufacturing settings. The model combines a Siamese neural network for discriminative feature learning, Grey Relational Analysis for objective indicator weighting under uncertainty, and TOPSIS for interpretable risk ranking. To describe quality-related risks in digital manufacturing, a multi-dimensional indicator system is constructed from technological, organizational, and environmental perspectives. Proxy-based experiments on four public datasets covering visual anomalies, logical constraint violations, distribution shifts, and time-series faults show that the proposed method consistently outperforms traditional MCDM approaches and several baseline models in AUROC, AUPR, F1-score, and Top-k Hit Rate. Ablation results further indicate that each component contributes to the overall performance, with the Siamese module providing the largest improvement. The framework also shows good robustness under distribution shifts and supports efficient inference, suggesting its practical value for manufacturing quality risk early warning.