Economic uncertainty and increasingly interconnected supply chains create challenges for identifying financial distress among advanced manufacturing enterprises, where risks may evolve over time and extend across firm relationships. Conventional early warning approaches often have difficulty capturing both the temporal development of financial conditions and the structural dependencies through which vulnerabilities can spread. This study proposes a temporal–graph fusion deep learning framework for financial distress early warning. A dual-channel analytical structure integrates financial reporting indicators and risk management information. A gated temporal convolution module models changes in firm-level financial characteristics, while a dynamic graph convolution module represents evolving inter-firm relationships and potential channels of risk transmission. Macroeconomic variables and cross-environment regularization are incorporated to address distribution shifts across economic conditions. SHAP-based analysis is used to interpret the contribution of predictive features and to examine nonlinear patterns and interactions associated with financial vulnerability. The reported empirical comparisons indicate that the proposed framework achieves favorable predictive discrimination, early warning performance, and stability relative to the evaluated baseline models. The study thus presents a unified approach for integrating longitudinal financial data, inter-firm network information, and interpretable machine learning into corporate financial distress assessment.
