A Real-Time Fault Diagnosis Method for Multi-Source Heterogeneous Information Fusion Based on Two-Level Transfer...

Area of Science:

• Engineering
• Artificial Intelligence
• Data Science

Background:

• Convolutional Neural Networks (CNNs) excel at feature extraction but have high computational costs.
• High sampling frequencies in equipment limit the online application of traditional CNN-based fault diagnosis.

Purpose of the Study:

• To develop a real-time fault diagnosis method for high-frequency equipment.
• To address the limitations of CNNs in terms of time complexity and computational load.
• To effectively fuse multi-source heterogeneous information for enhanced diagnostic accuracy.

Main Methods:

• Proposes a two-level transfer learning approach for multi-source heterogeneous information fusion.
• Constructs a feature extraction network model using screenshots.
• Designs a transfer mechanism from screenshot features to a deep learning model using one-dimensional sequence signals, transitioning from CNN to Deep Neural Network (DNN).

Main Results:

• The developed fault diagnosis model avoids computationally intensive convolution operations.
• Achieves low time complexity, enabling real-time fault diagnosis.
• Effectively integrates features from both one-dimensional sequence signals and screenshots.
• Demonstrated effectiveness on gearbox and bearing datasets.

Conclusions:

• The proposed two-level transfer learning method offers an efficient solution for real-time fault diagnosis.
• Successfully overcomes the computational limitations of CNNs for high-frequency data.
• Provides a robust approach for fusing diverse data sources in equipment monitoring.