Lacofd: A Leakage-Aware Contrastive Few-Shot Intelligent Diagnosis Framework for Cross-Condition Rotating Machinery

Abstract

Intelligent industrial systems require reliable condition monitoring when operating conditions change, yet labeled fault data from a new condition are often scarce. This paper presents LACoFD, a leakage-aware contrastive few-shot intelligent diagnosis framework for cross-condition rotating machinery. The framework learns transferable diagnostic representations from abundant unlabeled source-condition vibration streams through momentum-based contrastive pretraining and then adapts to a target operating condition using only a small labeled support set. To make performance assessment reliable for intelligent condition-monitoring systems, LACoFD uses a blocked support-gap-query protocol that reduces the optimistic bias caused by overlapping vibration windows. Experiments on the CWRU bearing dataset over six load-transfer tasks show that contrastive pretraining provides an effective initialization for target adaptation, particularly in the 10 -shot and 20 -shot settings, whereas frozen linear probing remains insufficient under load shifts. The results indicate that leakage-aware representation learning can improve the reliability of data-driven intelligent diagnosis in changing industrial operating environments.