Physics‐Informed Adaptive Feature Fusion for Cross‐Condition Fault Diagnosis of Pneumatic Control Valves

ABSTRACT

Pneumatic control valves are critical final control elements in industrial automation and process instrumentation systems. Their faults may degrade closed‐loop stability, product quality, and operational safety. Reliable diagnosis under variable operating conditions remains challenging because operating variations and fault‐induced responses are nonlinearly coupled through valve pressure–flow characteristics and actuator dynamics. To address this problem, this paper proposes a physics‐informed temporal and statistical adaptive fusion framework, termed PI‐TSAFNet, for cross‐condition fault diagnosis of pneumatic control valves. The proposed method first constructs physically interpretable residual representations from limited industrial measurements, including controller output, inlet and outlet pressures, temperature, valve stem displacement, and flow rate. Temporal features are then extracted from both measured response and residual streams, while compact statistical descriptors are generated to capture distribution‐level variations. A reliability‐aware adaptive fusion module integrates physics‐informed, temporal, and statistical evidence before fault classification. Experiments on the DAMADICS benchmark show that PI‐TSAFNet achieves 97.59% accuracy and 97.60% Macro‐F1 under the random‐sampling protocol. Under the more stringent leave‐one‐condition‐out (LOCO) protocol, it achieves 95.61% average accuracy and 95.62% average Macro‐F1. On the limited DAMADICS real‐data subset, the proposed method achieves 92.11% accuracy and 89.08% Macro‐F1 for binary group‐level fault detection. These results indicate that the proposed framework can transform limited industrial measurements into interpretable diagnostic evidence and retains useful binary fault‐detection capability on real industrial records. However, fine‐grained real‐data fault‐type recognition remains limited by sparse annotated events, class imbalance, and cross‐actuator domain shift.