Maintenance-cost-oriented fault diagnosis framework: case study on Tennessee Eastman process

Within Prognostic and Health Management (PHM), fault diagnosis and maintenance scheduling are intrinsically interconnected. Fault diagnosis reduces maintenance costs indirectly by enhancing accuracy, making diagnostic accuracy ultimately a means to reducing maintenance expenditures. Current research prioritizes maximizing diagnostic accuracy—approaching perfection—by increasing model complexity and augmenting data, often overlooking maintenance-oriented strategies. This raises critical questions: Does the resource-intensive pursuit of marginal accuracy gains yield proportional maintenance savings? Are there efficient, low-resource alternatives? To address these questions, we propose a maintenance-cost-oriented fault diagnosis framework, implementing a production-maintenance cost model on the Tennessee Eastman Process (TEP) benchmark. This approach enhances Graph Isomorphism Network (GIN)-based diagnostics by integrating cost constraints in accuracy optimization. Maintenance expenditures are quantified comprehensively by evaluating the implications of preventive or corrective actions triggered by correct classifications, missed alarms, and false alarms—considering their effects on continuous production, including manufacturing, inventory, and maintenance-related costs. This framework dynamically balances accuracy objectives with misdiagnosis risks. Comparative analysis against data-enhanced GIN variants shows equivalent maintenance cost reduction performance, validating its superiority in targeted optimization and resource allocation efficiency for industrial PHM.