Skip to Content
ProcessesProcesses
  • Article
  • Open Access

20 November 2026

31 Pages

Fault Diagnosis of Reciprocating Compressor Crosshead Nut Loosening Using Swin Transformer-Based Ordered Transfer Learning

,
,
,
and
1
College of New Energy, China University of Petroleum (East China), Qingdao 266580, China
2
PetroChina Tuha Oilfield Company, Hami 839009, China
3
Key Laboratory of Advanced Nuclear Energy Design and Safety, Ministry of Education, University of South China, Hengyang 421001, China
*
Author to whom correspondence should be addressed.

Abstract

The connecting nut between the piston rod and the crosshead is a key fastener of reciprocating compressors, and its loosening may lead to unplanned shutdowns and even safety accidents. Fault severities are commonly divided into normal, slight, moderate and severe states, but accurate data-driven diagnosis under variable working conditions remains difficult because labeled experimental fault data are scarce and the fault features of different severities are separated in scale. This paper proposes the Swin Transformer Deep Ordered Transfer Network with Multi-Scale Feature Fusion, CORAL Alignment and Ordinal Regression (Swin-MCO). The source domain is a rigid-flexible coupled multibody simulation validated against bench measurements, and the target domain is a test-bench dataset covering 25 speed and pressure conditions, of which only 20% are labeled for training. The model extracts hierarchical features from 224 × 224 angular-domain texture images, aligns second-order statistics across domains and encodes the physical fault progression in the classifier. The diagnosis accuracies reach 95.80%, 95.57% and 96.72% on the three unseen test sets, and misclassifications are confined to adjacent severity levels. The results indicate that the proposed model provides an effective solution for cross-domain diagnosis of connecting nut loosening. Future work will focus on industrial deployment, lightweight edge implementation and uncertainty quantification.

Article Metrics

Citations

Article Access Statistics

Article metric data becomes available approximately 24 hours after publication online.