A Continuous Remaining Useful Life Prediction Method With Multistage Attention Convolutional Neural Network and...

Area of Science:

• Mechanical Engineering
• Artificial Intelligence

Background:

• Continuous monitoring of rotating machinery is crucial, but data storage and processing limitations challenge remaining useful life (RUL) prediction.
• Continuous learning (CL) offers a solution for dynamic RUL prediction models by enabling knowledge accumulation and updates.

Purpose of the Study:

• To develop an advanced RUL prediction methodology that overcomes data constraints and improves prediction accuracy.
• To introduce a novel approach combining a multistage attention convolutional neural network (MSACNN) with a knowledge weight constraint (KWC) for enhanced continuous RUL prediction.

Main Methods:

• A multistage attention convolutional neural network (MSACNN) was designed, incorporating an improved multihead full-channel sight self-attention (MFCSSA) mechanism.
• The MSACNN integrates MFCSSA, squeeze-and-excitation (SE), and convolutional block attention module (CBAM) for progressive feature refinement and global degradation information capture.
• A knowledge weight constraint (KWC) mechanism was developed, utilizing weight parameter importance and gradient information to mitigate catastrophic forgetting in CL.

Main Results:

• The proposed MSACNN demonstrated higher prediction accuracy compared to existing methods on bearing and gear datasets.
• The KWC mechanism effectively retained previously learned knowledge while acquiring new task knowledge, outperforming typical CL methods.
• The integrated MSACNN and KWC methodology showed superior performance in continuous RUL prediction tasks.

Conclusions:

• The developed MSACNN with KWC provides a robust and accurate solution for continuous remaining useful life prediction in rotating machinery.
• This methodology effectively addresses the challenges posed by limited data storage and processing capabilities.
• The approach offers significant advantages over existing continuous learning methods for machinery prognostics.