Abstract:The varying speed can cause the significant data distribution shift of bearings, making it difficult for deep learning-based bearing fault diagnosis models to ensure good generalization. Domain adaptation methods have been developed to address the domain shifts, while they struggle with the class-invariant features extraction under variable speed. Accordingly, a time–frequency supervised contrastive learning framework (TF-SupCon) is proposed for unsupervised cross-speed fault diagnosis of bearing. TF-SupCon adopts a pre-training-downstream task framework that aims to extract speed immune class-invariant features. During the pre-training phase, the physical consistency of the time-domain information and the frequency-domain information of bearing signal, a general feature applicable to different speed conditions, is learned in a targeted manner through supervised contrastive learning. Additionally, a K-Nearest Neighbor (KNN) algorithm based on cosine distance is designed to assign pseudo-labels to unlabeled target domain data , enabling cross-domain supervised contrastive pre-training. In the downstream task, unsupervised cross-domain fault diagnosis is performed using a KNN classifier and the speed immune time–frequency features by the trained encoders. It is worth noting that the same metric is maintained throughout both the pre-training phase and the downstream task, ensuring organic connection between the two stages. The superiority of TF-SupCon was demonstrated through a variety of fault diagnosis experiments conducted on both public and self-collected datasets. Lastly, the distances between time-domain and frequency-domain features of different categories were studied to verify the physical consistency between the features.

Dual Adversarial Contrast for Train Bearing Fault Diagnosis under Extremely Low Label Rate

Dual Contrastive Learning for Semi-Supervised Fault Diagnosis Under Extremely Low Label Rate

Self-supervised Knowledge Mining from Unlabeled Data for Bearing Fault Diagnosis under Limited Annotations

A self-attention based contrastive learning method for bearing fault diagnosis

A novel unsupervised deep transfer learning method based on contrast pre-training for fault diagnosis

A Semi-Supervised Intelligent Fault Diagnosis Method for Bearings Under Low Labeled Rates

Dual Classifier-Discriminator Adversarial Networks for Open Set Fault Diagnosis of Train Bearings

Federated temporal-context contrastive learning for fault diagnosis using multiple datasets with insufficient labels

A Fault Diagnosis Method for Rolling Bearing Based on Deep Adversarial Transfer Learning with Transferability Measurement

A contrastive learning-based fault diagnosis method for rotating machinery with limited and imbalanced labels

A Novel Multiscale Lightweight Fault Diagnosis Model Based on the Idea of Adversarial Learning

Dual prototypical contrastive network: a novel self-supervised method for cross-domain few-shot fault diagnosis

Differential contrast guidance for aeroengine fault diagnosis with limited data

Semi-Supervised Fault Diagnosis Method Based on Dual-Stream Attention Cyclic Auto-Encoder for Rotating Components With Low Label Rate

Triplet Loss Guided Adversarial Domain Adaptation for Bearing Fault Diagnosis.

Contrastive learning-enabled digital twin framework for fault diagnosis of rolling bearing

Time-frequency supervised contrastive learning via pseudo-labeling: An unsupervised domain adaptation network for rolling bearing fault diagnosis under time-varying speeds

Few-Shot Learning for Fault Diagnosis: Semi-Supervised Prototypical Network with Pseudo-Labels

Robust Mechanical Fault Diagnosis with Noisy Label Based on Multistage True Label Distribution Learning.

An unsupervised transfer learning approach for rolling bearing fault diagnosis based on dual pseudo-label screening

Joint Discriminative Adversarial Domain Adaptation for Cross-Domain Fault Diagnosis