Dynamic graph–based domain-adaptive fault diagnosis method for reciprocating compressors

Reciprocating compressors are widely used in critical industrial scenarios, such as energy and chemical engineering. Under variable-load conditions, their multi-sensor monitoring signals often show statistical distribution shifts and changes in sensor coupling relationships. This reduces the cross-condition generalization ability of traditional feature engineering methods and deep models based on the identical distribution assumption. To address this problem, this study proposes a fault diagnosis method that integrates a physical-prior–constrained adaptive sensor graph with cross-domain representation learning. First, a physical-prior graph is constructed according to sensor installation locations, subsystem mechanisms, and fault propagation relationships. A learnable residual term is then introduced to adjust the sensor connection strengths under different loads. Second, graph convolution and channel attention are combined to extract multi-sensor coupling features. Finally, source-domain classification, source/target-domain reconstruction, and maximum mean discrepancy–based distribution alignment are jointly optimized in a shared-encoder dual-decoder cross-domain stacked denoising autoencoder framework. This design enhances both the fault discriminability and domain invariance of the latent representations. Based on multi-load experimental data from a reciprocating compressor, single-source domain transfer tasks, multi-source domain transfer tasks, and open-set tests with unknown faults are constructed. A small reciprocating compressor dataset is also introduced for supplementary validation. The experimental results show that the proposed method achieves high diagnostic accuracy and stable performance in different cross-load tasks. It also outperforms typical domain adaptation methods and recent transfer diagnosis methods in overall performance. In addition, it maintains good rejection ability under unknown abnormal conditions. These results indicate that combining physical-prior–constrained relation-adaptive graph modeling with cross-domain representation learning can improve the robustness and reliability of variable-load health monitoring for reciprocating compressors.