Bearing Life Prediction via HMM and Transfer Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current bearing life prediction methods often ignore the failure occurrence time (FOT) and struggle with distribution differences between source and target domains due to varying operating conditions, leading to reduced model generalization and accuracy.
Innovation Solution
A method using a hidden Markov model (HMM) to automatically detect FOT and multilayer perceptron (MLP)-based transfer learning to align feature sets from different domains, improving prediction accuracy and efficiency by extracting domain-invariant features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If conventional life prediction methods ignore or empirically determine failure occurrence time, then the prediction process is simplified, but the accuracy of life prediction deteriorates due to loss of important degradation information
Solution Approach 1:
The HMM automatically detects FOT by itself using the monitoring data without external intervention. The model self-learns the failure occurrence time point through unsupervised training on the degradation signals, eliminating the need for manual empirical determination while preserving critical degradation information.
Solution Approach 2:
The patent replaces the empirical/manual determination method with a computational model (HMM). Instead of relying on expert experience or simple threshold-based methods, the system uses probabilistic modeling to automatically identify FOT, substituting mechanical/empirical processes with intelligent algorithms.
2Ease of manufacture
If samples in source domains and target domains conform to the same data distribution, then the model training is simplified, but the adaptability to different operating conditions deteriorates
Solution Approach 1:
The patent changes the distribution parameters of the source domain data to match the target domain distribution. Through domain adaptation techniques, the model learns to transform or reweight features so that data from different operating conditions follow the same distribution, enabling the model to generalize across varying conditions while maintaining training effectiveness.
Solution Approach 2:
The patent introduces domain adaptation as an intermediary layer between source and target domains. This intermediary process aligns the feature distributions from different operating conditions, allowing the model to learn from diverse sources without being constrained by distribution mismatches, thereby improving generalization while keeping training feasible.
3Reliability
If domain adaptation is performed to reduce distribution differences, then the model generalization capability is improved, but the computational complexity increases
Solution Approach 1:
The patent applies partial domain adaptation by focusing on the most critical features that exhibit distribution shifts. Instead of attempting to align all features equally, the method identifies and adapts only the necessary subsets of features, reducing computational overhead while still achieving sufficient generalization improvement for practical applications.
Data Source
AI summary
The present invention discloses a method for predicting bearing life based on a hidden Markov model (HMM) and transfer learning, including the following steps: (1) acquiring an original signal of full life of a rolling bearing; and extracting a feature set including a time domain feature, a time-frequency domain feature, and a trigonometric function feature; (2) inputting the feature set into an HMM to predict a hidden state, to obtain a failure occurrence time (FOT); (3) constructing a multilayer perceptron (MLP) model, obtaining a domain invariant feature and an optimal model parameter, and obtaining a neural network life prediction model; and (4) inputting the remaining target domain feature sets into the neural network life prediction model, and predicting the remaining life of the bearing. In the present invention, MLP-based transfer learning is used to resolve distribution differences in a source domain and a target domain caused by different operating conditions.