Deep Ensemble Learning Based on Multi-Form Fusion in Gearbox Fault Recognition.

Area of Science:

• Engineering
• Computer Science
• Machine Learning

Background:

• Industrial equipment fault identification faces challenges due to single data sources, varying data sensitivity, and feature extraction limitations.
• These issues result in difficulties in information fusion, inadequate state representation, and low fault identification accuracy and robustness.

Purpose of the Study:

• To propose a novel multi-information fusion fault identification network model utilizing deep ensemble learning.
• To address the limitations of single-source data and improve the accuracy and robustness of equipment fault recognition.

Main Methods:

• A deep ensemble learning network comprising multiple sub-feature extraction units and feature fusion units was developed.
• Fault feature mapping information from each source was extracted and stored in distinct sub-models.
• Features from sub-models were fused using a dedicated feature fusion unit to obtain final fault recognition results.

Main Results:

• The proposed method was evaluated using two gearbox datasets.
• Achieved near 100% accuracy for all fault types, significantly outperforming simple stacking fusion and single-point methods.
• Demonstrated superior performance in gearbox fault recognition tasks.

Conclusions:

• The proposed multi-information fusion fault identification network based on deep ensemble learning is feasible and effective.
• The method offers a robust solution for accurate equipment fault recognition in industrial settings.
• Highlights the potential of deep ensemble learning for complex fault diagnosis applications.