Bearing RUL Prediction With Bayesian GRNN Uncertainty Intervals
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional gated recurrent neural networks for bearing life prediction suffer from low accuracy and inability to handle prediction uncertainty due to sensor noise and unknown failure modes, and adding traditional attention mechanisms increases model complexity.
Innovation Solution
A method using a gated recurrent neural network with an attention mechanism and a Bayesian layer to directly calculate weights, integrating a stacked structure and dynamic time warping for improved information extraction, and converting point predictions to interval predictions to account for uncertainty.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional attention mechanism is introduced to improve prediction accuracy, then prediction accuracy is improved, but model complexity increases substantially
Solution Approach 1:
The patent extracts only the essential function of the attention mechanism (weight calculation) and implements it through a simplified direct calculation method rather than incorporating the full traditional attention mechanism with its additional network layers. This extracts the useful functionality while removing the complexity burden.
Solution Approach 2:
The patent applies attention weights locally to specific time steps in the temporal sequence rather than uniformly across the entire sequence. The attention mechanism focuses computational resources on critical time points that contribute most to the prediction, improving accuracy without requiring complex global processing.
2Device complexity
If conventional gated recurrent neural network is used for point prediction, then the model structure is simple, but it cannot handle prediction uncertainty caused by sensor noise and unknown failure modes
Solution Approach 1:
The patent transforms the static point prediction output into a dynamic interval prediction that adapts to uncertainty. The Bayesian layer dynamically adjusts the prediction intervals based on the learned uncertainty from the Gated Recurrent Neural Network, allowing the model to express confidence levels rather than fixed point estimates.
Solution Approach 2:
The patent creates a composite model structure by combining the Gated Recurrent Neural Network (for temporal feature extraction) with the Bayesian layer (for uncertainty quantification). This composite architecture integrates the strengths of both components to achieve both simple structure and reliable uncertainty handling.
3Measurement precision
If additional network layers are added to incorporate attention mechanism, then prediction accuracy is improved, but computation time increases substantially
Solution Approach 1:
The patent extracts only the essential weight calculation function from the traditional attention mechanism and implements it through direct calculation without the additional network layers. This extraction maintains the accuracy-improving functionality while eliminating the computational overhead of extra layers.
Solution Approach 2:
The patent replaces the mechanical neural network layer structure with a direct mathematical calculation approach for attention weights. Instead of passing data through additional neural network transformations, the system directly computes weights using a simplified formula, substituting computational mechanics with mathematical operations.
Data Source
AI summary
The present invention provides a method for predicting remaining useful life of bearings based on a gated recurrent neural network, comprising the following steps: S1. obtaining full life cycle vibration signals of bearings, extracting the vibration distribution features and creating a training set of gated recurrent neural network; S2. constructing a gated recurrent neural network model, and introducing an attention mechanism that directly calculates weights to improve the integrity of extracting temporal information; S3. adding a Bayesian layer to construct a nonlinear mapping relationship between temporal information and remaining useful life; S4. taking vibration signals of a test bearing as input, the output result of the gated recurrent neural network model is the remaining useful life of the test bearing at the current time. The present invention does not need to add additional neural network layers, which avoids the problem of increasing the complexity of the model. The integrity of extracted information is improved through weighted fusion of temporal information extracted at different time. Moreover, by adding a Bayesian layer, the traditional point prediction results are converted into interval predictions, to consider the prediction uncertainty of remaining useful life of bearings.