Abstract: To address the issue where early faults in coal mills were easily masked due to strong parameter coupling and large inertia under frequent load-changing conditions, and the lack of effective physical mechanism constraints in existing methods, a fault early warning and classification method was proposed that combines physics-constrained causal calibration with an adaptive dynamic directed graph convolutional network (ADA-DGCN). Firstly, a radial basis function (RBF) kernel mapping was introduced to improve the PCMCI algorithm, enhancing nonlinear causal identification capability; initial causal matrices were physically corrected using energy consistency constraints and dynamic mass balance constraints to suppress pseudo-causal relationships that violate physical laws. Secondly, an adaptive dynamic directed attention mechanism was designed to update causal weights online, while extracting spatiotemporal evolution features using multi-step dynamic graph networks. Finally, the global features extracted by the dynamic graph convolutional network were mapped to the probability space via a Sigmoid activation function to calculate warning probabilities, combined with adaptive thresholds for warning decision, and classified based on distance matching using a standard fault fingerprint database. Experimental results show that this method achieves a fault warning accuracy of 96.43% under varying load disturbances, effectively distinguishes typical states such as normal operation, coal blockage, coal break, wear, and load variation, and provides substantial early warning lead time for typical fault conditions, balancing engineering interpretability, timeliness, and accuracy, thereby offering a physically interpretable approach for early fault warning in coal mills under load-changing conditions.