Skip to content
Open access

Supervised Contrastive Learning for Cross-Load Bearing Fault Diagnosis with Limited Training Labels

Oct 2026 · Applied and Computational Engineering · 0 citations

Abstract

Vibration-based fault diagnosis requires annotated examples that may not be easy to come by in service. This paper compares the effectiveness of supervised contrastive pretraining to cross-entropy training using the same compact convolutional encoder and 40 recordings of bearings from Case Western Reserve University. Motor load data with 0 and 1 hp load provided the training set, 2 hp provided the validation data, and 3 hp the test set. Five random seeds were evaluated at three training budgets of 5, 10 and 20 examples per class. A separate fixed validation set with 944 labeled windows was used to select checkpoints and was excluded from the budgets above. Linear classification on contrastively pretrained embeddings resulted in higher mean test accuracy and macro-F1 score compared to cross-entropy training at all three budgets. The biggest improvement came at a budget of 10 windows/class: accuracy increased from 84.81% to 89.22% and macro-F1 score from 84.62% to 89.24%. Recall for ball faults was primarily responsible for the increase. These results support the effectiveness of complete contrastive training based on this cross-load test scheme but do not isolate the effect of the loss or establish performance on other machines.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.