Reliability‐Guided Recovery and Relation Learning for Fault Diagnosis in Coupled Industrial Processes

ABSTRACT Reliable fault diagnosis supports stable operation and product quality in complex industrial processes. Process coupling, operating variability and measurement disturbances influence both observed features and their interdependencies, motivating the integration of feature recovery with relation learning. This study proposes RGIFormer, a framework that couples adaptive multi‐scale recovery with global and sample‐adaptive relation learning. The recovery module combines local and global reconstruction features and uses a reconstruction‐agreement gate to regulate input correction. The corrected inputs then support dual‐relation graph learning, where a learned gate balances shared global associations and sample‐adaptive symmetric relations. These relations guide graph convolution and grouped‐descriptor attention, allowing recovered feature information to shape the diagnostic representation. Feature fusion and prototype‐supported scoring combine the resulting representations for classification. Experiments on Benchmark Simulation Model No. 2 (BSM2) snapshots and papermaking descriptors evaluate the framework across five random seeds. Under the original dataset‐specific splits, mean accuracies are 78.81% and 93.66%, with macro‐F1 scores of 77.76% and 93.78%. The framework connects measurement recovery with relation construction to provide an integrated diagnostic representation for coupled industrial processes.