Long Short-Term Memory Neural Network with Transfer Learning and Ensemble Learning for Remaining Useful Life...

Area of Science:

• Machine Learning
• Predictive Maintenance
• Industrial IoT

Background:

• Accurate remaining useful life (RUL) prediction is crucial for manufacturing equipment safety and reliability.
• Existing RUL models struggle with accuracy when trained on limited fault data.
• Developing robust health indicators (HI) from raw sensor data is challenging.

Purpose of the Study:

• To propose an advanced RUL prediction method overcoming data limitations.
• To enhance the performance of RUL prediction models using transfer and ensemble learning.
• To develop an unsupervised method for constructing effective health indicators.

Main Methods:

• Utilized deep belief networks and self-organizing map networks for unsupervised health indicator (HI) construction from raw sensor data.
• Implemented a long short-term memory (LSTM) neural network for RUL prediction.
• Integrated transfer learning and ensemble learning to improve model generalization and accuracy.

Main Results:

• The proposed method demonstrated superior performance in RUL prediction compared to existing approaches.
• Validation on two experimental bearing datasets confirmed the effectiveness of the combined approach.
• The unsupervised HI construction method successfully translated sensor data into meaningful health status indicators.

Conclusions:

• The developed LSTM-based RUL prediction model, enhanced with transfer and ensemble learning and an unsupervised HI construction method, significantly improves prediction accuracy.
• This approach offers a viable solution for RUL prediction in industrial settings with limited fault data.
• The study highlights the potential of combining deep learning, transfer learning, and unsupervised feature engineering for enhanced predictive maintenance.