Induction Motor Eccentricity Severity Estimation Using Sparse Regression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for eccentricity detection in induction motors are inadequate for accurately estimating the severity of eccentricity, which is crucial for effective maintenance and control, due to uncertainties in measurement relationships and noise amplification.
Innovation Solution
A sparsity-driven regression model using machine learning with physics-informed constraints to estimate eccentricity severity by learning weights from sensor data, incorporating features like load torque, rotor speed, vibration, and current spectra, while reducing noise and uncertainty through sparse combinations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional methods like MCSA or spectrum analysis are used for eccentricity detection, then the presence of eccentricity can be detected, but the accuracy of severity estimation is insufficient due to noise amplification and unknown measurement relationships
Solution Approach 1:
The patent transforms the estimation problem by changing parameters from direct spectral amplitude ratios to a regression model based on multiple harmonic components. The model uses parameters such as sideband amplitudes at frequencies (1±k·ε)·ωr and their relationships to estimate severity, converting an unreliable single-ratio measurement into a multi-parameter regression analysis that reduces noise sensitivity
Solution Approach 2:
The patent combines multiple measurement signals into a composite estimation model. Instead of relying on a single spectral ratio, it integrates information from multiple harmonic components (k=0,1,2,...) and their sidebands, creating a composite regression model that synthesizes multiple noisy measurements into a more reliable severity estimate
2Loss of information
If a comprehensive model relating multiple harmonics measurements is used, then more information is available for estimation, but the complexity of the model increases and noise from multiple measurements accumulates
Solution Approach 1:
The patent segments the complex spectral analysis into distinct harmonic components (k=0,1,2,...) with specific frequency relationships. Each component (sideband at (1±k·ε)·ωr) is analyzed separately with its own amplitude measurement, and then these segmented measurements are combined through a regression model. This segmentation prevents noise accumulation by treating each component independently rather than as a single complex measurement
3Adaptability or versatility
If measurements are taken under different load conditions, then the model becomes more adaptable to various operating conditions, but the uncertainty in measurement relationships increases
Solution Approach 1:
The patent creates a dynamic regression model where the relationships between harmonic amplitudes and severity are not fixed but adapt to different operating conditions. The model uses measurements taken under varying load conditions to establish regression coefficients that capture the dynamic relationships between spectral components and eccentricity severity across different operational states
Solution Approach 2:
The regression model incorporates feedback from multiple harmonic measurements to continuously refine the severity estimation. By using the relationships between different sideband amplitudes and their known frequency relationships to the rotor frequency, the model feedback-corrects for uncertainties in individual measurements, using the consistent mathematical relationships that hold across different load conditions
Data Source
Figure 1A
Figure 1B
Figure 2A
AI summary
A fault detection system of eccentricity severity of an induction machine including a rotor and stator is provided. The fault detection system includes a sensor interface configured to acquire sensor signals from sensors arranged at predetermined positions of the induction machine, wherein the sensor signals are indicative of an eccentricity level of a rotor of the induction machine, a memory coupled with a processor. The memory stores training data sets and instructions implementing a learning-based fault detection method for the induction machine. The instructions include steps of generating an eccentricity feature matrix from the sensor signals, where in the sensor signals include load torque, rotor speed, vibration acceleration of the rotor, vibration speed of the rotor, and current spectral of the stator or the induction machine, determining an eccentricity level of the induction machine based on the eccentricity feature matrix using the learning-based fault detection method, wherein the learning-based fault detection method is configured to find the eccentricity level from learning-based eccentricity feature matrix data sets.