Work

Thermodynamics-Informed Residual-Disentangled Network for Manufacturing Defect Diagnosis of Reciprocating Compressors Under Extreme Class Imbalance

Year: 2026
Type: article
Abstract: Manufacturing defect diagnosis of reciprocating compressors is limited by two coupled problems: extreme class imbalance and the physical opacity of data-driven models. A physics-informed residual-disentangled network (PIRD-Net) is proposed in this paper, which incorporates thermodynamic knowledge through four mechanisms: a differentiable thermodynamic residual layer that converts deviations from thermodynamic consistency relations into trainable diagnostic evidence; a dual-branch encoder with physics-gated fusion to arbitrate between physical and data-driven models at the sample level; label-free physics prior injection; and a composite physics-informed loss function. A physics-constrained SMOTE strategy restricts the synthetic minority samples to a range that does not violate thermodynamics. Validation of 921 end-of-line units with eight defect types and an imbalance ratio of 43.4:1 showed that PIRD-Net achieved 95.90% accuracy, 94.31% Macro-F1, and 92.99% Matthews correlation coefficient after 5 × 5 repeated stratified cross-validation, and performed better than the traditional classification methods and the label-free physics prior in imbalance-sensitive indicators. Accuracy is still 91.44% after adding 30% test noise to the natural process. Out-of-fold analysis shows that the errors are concentrated in mechanism-overlapping modes; the two rarest modes are correctly identified, and physics-dominated decisions are free from errors, so a low data-branch weight can be used as an internal reliability indicator.
Source: Processes
Language: English
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7236827632
FWCI:
Cites: 27
Cited by:
Related to: 10