Abstract:Recently, supervised learning methods for intelligent mechanical fault diagnosis have achieved considerable advancements, while they heavily rely on labeled information and neglect massive unlabeled signals. Consequently, some self-supervised research has emerged and employs contrastive learning to mine general knowledge from unlabeled samples. However, these methods emphasize independent feature extraction and lack deep feature interaction, which limits better representation learning. Furthermore, it also restricts the performance improvement of fault diagnosis. To fill this gap, a novel self-supervised representation learning framework based on time-frequency alignment and interaction (TFAI) is proposed to improve the diagnosis reliability under limited labeled data. Comprising a dual encoder and a cross-modal encoder based on Transformers, the proposed TFAI model takes paired time-frequency data as input for self-supervised pretraining. Two self-supervised learning strategies, namely time-frequency alignment and time-frequency interaction, are proposed to enable the model to gain valuable knowledge from unlabeled data. The former leverages a time-frequency contrastive loss to drive the dual encoder to extract profound features from unlabeled data. The latter involves utilizing a time-frequency matching loss to guide the cross-modal encoder to perform deep feature fusion in an unsupervised manner. The well pretrained TFAI model can learn informative representations and can be fine-tuned for specific downstream fault diagnosis tasks to improve the diagnosis accuracy. By conducting three experimental cases on an axial flow pump system and a public gearbox platform system, the effectiveness, generalizability, and domain adaptability of the proposed method are validated.

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

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

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

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

Mining knowledge from unlabeled data for fault diagnosis: A multi-task self-supervised approach

Information-guided signal multi-granularity contrastive feature learning for fault diagnosis with few labeled data

A limited labeled data bearing fault diagnosis method based on self-supervised learning

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

Imbalanced Fault Diagnosis by Supervised Contrastive Learning

A Hybrid Supervised Approach for Fault Diagnosis Based on Temporal and Frequency Domains

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

A New Semi-Supervised Fault Diagnosis Method via Deep CORAL and Transfer Component Analysis

A Universal Feature Extractor Based on Self-Supervised Pre-Training for Fault Diagnosis of Rotating Machinery under Limited Data

Self-attention Metric Learning Based on Multi-Scale Feature Fusion for Few-shot Fault Diagnosis

Fault diagnosis of wind turbine gearbox under limited labeled data through temporal predictive and similarity contrast learning embedded with self-attention mechanism

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

A novel self-supervised representation learning framework based on time-frequency alignment and interaction for mechanical fault diagnosis

Fault Feature Extractor Based on Bootstrap Your Own Latent and Data Augmentation Algorithm for Unlabeled Vibration Signals

A Novel Imbalanced Data Classification Method Based on Weakly Supervised Learning for Fault Diagnosis

Fault Diagnosis of Rotating Machinery Based on Improved Self-Supervised Learning Method and Very Few Labeled Samples

A Few-Shot Machinery Fault Diagnosis Framework Based on Self-Supervised Signal Representation Learning