Machine Learning for Sensorless Temperature Estimation of a BLDC Motor.

Area of Science:

• Electrical Engineering
• Machine Learning
• Thermal Management

Background:

• Accurate estimation of Brushless DC (BLDC) motor winding temperature is crucial for performance and longevity.
• Traditional methods often require direct temperature sensors, which can be intrusive or costly.
• Machine learning offers a promising avenue for developing advanced, non-invasive temperature estimation techniques.

Purpose of the Study:

• To develop and evaluate machine learning models for estimating BLDC motor winding temperatures.
• To assess the accuracy of sensorless temperature prediction and compare it with a model incorporating casing temperature data.
• To demonstrate the feasibility of using estimated winding temperatures for motor overheating protection and torque compensation.

Main Methods:

• Acquisition of over 160 hours of BLDC motor operation data.
• Preprocessing of collected motor operation data.
• Application of various machine learning algorithms including linear regression, ElasticNet, stochastic gradient descent regressor, support vector machines, decision trees, and AdaBoost for predictive modeling.
• Hyperparameter tuning using cross-validation to enhance model generalization.

Main Results:

• Model 1 (sensorless estimation) achieved a Mean Absolute Percentage Error (MAPE) below 4.5% and a coefficient of determination (R²) above 0.909.
• Model 2 (incorporating casing temperature) further improved accuracy, reducing error to approximately 1% and increasing R² to 0.990.
• Both models demonstrated high accuracy in predicting winding temperatures.

Conclusions:

• Machine learning models can effectively estimate BLDC motor winding temperatures with high accuracy.
• Sensorless temperature estimation (Model 1) is viable for motor overheating protection.
• Integrating casing temperature measurement (Model 2) significantly enhances estimation accuracy, suitable for torque compensation.