Mutual Information Feature Selection for Rotating Machine Monitoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine monitoring techniques for rotating machinery, such as those with rolling element bearings, face limitations in feature extraction and dimensionality reduction, particularly in real-time fault detection and prediction, as they often require prior knowledge of defect frequencies and can discard relevant information or become overwhelmed by large feature sets.
Innovation Solution
The method employs mutual information-based feature selection on modulation spectral features to reduce dimensionality and select a subset of relevant features for real-time monitoring, allowing for the prediction of machine performance and time to failure without prior knowledge of defect locations or frequencies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If envelope analysis is used to extract defect frequency features, then defect detection capability is improved, but prior knowledge of spectral location is required and relevant information can be discarded
Solution Approach 1:
The system performs self-diagnosis by automatically identifying defect frequencies through mutual information analysis without requiring external expert knowledge or manual spectral inspection. The algorithm autonomously determines which frequency bands contain defect information by measuring the mutual information between frequency components and the overall signal envelope.
Solution Approach 2:
The patent replaces the manual mechanical process of spectral inspection and defect frequency identification with an automated computational approach using mutual information calculation. This substitution eliminates the need for manual determination of defect frequencies while preserving the ability to detect defects accurately.
2Adaptability or versatility
If a wide frequency band is analysed to incorporate all defect locations, then defect coverage is improved, but neighbouring noise swamps important information
Solution Approach 1:
The patent dynamically adjusts the effective frequency band analysis by calculating mutual information for different frequency components and selectively focusing on those with high mutual information values. This parameter-based selection approach adapts the analysis bandwidth to the actual defect characteristics present in the signal, rather than using a fixed wide or narrow band.
Solution Approach 2:
The system extracts only the relevant frequency components that contain defect information by identifying and isolating frequency bands with high mutual information. This extraction process separates the useful defect signals from the surrounding noise, achieving both wide defect coverage and high signal-to-noise ratio.
3Measurement precision
If STFT with fixed window width is used to maintain temporal location information, then temporal resolution is improved, but features with intervals longer than window width cannot be resolved
Solution Approach 1:
The patent introduces dynamic adaptability to the fixed-window STFT approach by using mutual information analysis to identify and adjust the effective analysis parameters based on the actual signal characteristics. This allows the system to handle varying feature intervals while maintaining temporal precision through the information-theoretic selection of relevant frequency-temporal components.
4Loss of information
If modulation spectral analysis is used to extract long term features, then feature information completeness is improved, but the number of features becomes extremely large
Solution Approach 1:
The system extracts only the essential modulation spectral features by applying mutual information analysis to identify and select the most informative frequency components. This selective extraction reduces the feature dimensionality from hundreds of thousands to a manageable subset while preserving the complete diagnostic information needed for defect detection and RUL prediction.
Solution Approach 2:
The patent transforms the high-dimensional modulation spectral feature space into a lower-dimensional representation by changing the parameter selection criteria from exhaustive feature capture to information-theoretic feature selection. This parameter change maintains information completeness while dramatically reducing computational complexity.
Data Source
Figure 1~3
Figure 2~4
Figure 5
AI summary
A method of machine monitoring for real-time detection and/or prediction of faults or performance variations is disclosed, comprising detecting noise emitted by the machine and obtaining a digital signal representative thereof; conducting a spectral analysis of the signal to obtain modulation spectral features; applying a mutual information feature selection technique to the modulation features to obtain a sub-set of features which provide useful information about performance and/or expected time to failure of the machine for monitoring.