Research on the Bearing Remaining Useful Life Prediction Method Based on Optimized BiLSTM.

Area of Science:

• Mechanical Engineering
• Prognostics and Health Management (PHM)
• Artificial Intelligence in Engineering

Background:

• Accurate Remaining Useful Life (RUL) prediction for bearings is critical for machinery maintenance.
• Early degradation detection significantly impacts the performance of RUL estimation models.
• Existing methods for setting bearing degradation thresholds can limit RUL prediction accuracy.

Purpose of the Study:

• To propose a novel bearing RUL prediction method incorporating early degradation detection.
• To enhance the accuracy and reliability of RUL estimation models for bearings.
• To address the limitations of current bearing degradation threshold setting methods.

Main Methods:

• Optimized Variational Mode Decomposition (VMD) with Pearson correlation coefficient for signal denoising.
• Multi-domain feature extraction, evaluation, and optimal feature selection.
• Kernel Principal Component Analysis (KPCA) for feature integration and Health Indicator (HI) establishment.
• Sliding window method with 3σ (3-sigma) criterion and quartile method for early degradation detection.
• Bi-directional Long Short-Term Memory (BiLSTM) network with attention mechanism, optimized by Sine-Sine Algorithm (SSA) for RUL prediction.

Main Results:

• The proposed denoising method demonstrates high performance in processing bearing signals.
• The dynamic 3σ-threshold setting method effectively detects early bearing degradation points.
• The optimized BiLSTM-based RUL prediction model exhibits superior performance, fitting capacity, and generalization ability.

Conclusions:

• The integrated approach enables accurate early detection of bearing degradation.
• The developed RUL prediction model significantly improves prediction accuracy and applicability.
• This method offers a robust solution for early Remaining Useful Life prediction in bearings.