A novel temporal classification prototype network for few-shot bearing fault detection

Area of Science:

• Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Industrial bearing fault detection faces challenges due to limited fault data samples, hindering deep learning model training and generalization.
• Scarcity of massive datasets in industrial settings leads to underfitting and poor generalization in neural networks for fault detection.

Purpose of the Study:

• To propose an improved and efficient few-shot supervised learning method, the Temporal Classification Prototype Network (TCPN), to address data scarcity in bearing fault detection.
• To enhance both training efficacy and generalization capabilities of models under limited data conditions.

Main Methods:

• Fourier transform is used to highlight frequency domain characteristics of bearing signals for improved fault detection.
• An Enhanced Temporal Convolutional Network (ETCN) transforms discrete data points into a feature space.
• A ContractSim Classifier (CSC) utilizes support set features as anchors and similarity measures for classification, refining the model through query set data.

Main Results:

• The TCPN model demonstrated proficiency in few-shot learning across four standard bearing datasets through k-shot experiments.
• Comparative experiments showed TCPN outperforming baseline models in bearing fault detection.
• Ablation studies confirmed the robustness and effectiveness of the integrated modules within TCPN.

Conclusions:

• The proposed TCPN model offers an effective solution for bearing fault detection in data-scarce industrial environments.
• TCPN exhibits strong performance and generalization capabilities, outperforming existing methods in few-shot learning scenarios.
• The integration of Fourier transform, ETCN, and CSC contributes to a robust and rational fault detection system.