Fault Detection and Monitoring in Induction Machines Using Data-Driven Model Drift Detection.

Area of Science:

• Electrical Engineering
• Machine Learning
• Predictive Maintenance

Background:

• Stator short-circuit faults (SSCFs) are a major cause of induction motor failures.
• Early detection is difficult due to scarce labeled data and impractical sensor installation in industrial settings.
• Traditional methods like spectral analysis and residual energy have reliability limitations.

Purpose of the Study:

• To propose a novel, data-driven framework for early fault detection and diagnosis of SSCFs.
• To overcome the limitations of existing methods by using the Residual Information Value (RIV) principle.
• To develop a non-intrusive solution that integrates seamlessly with Variable Speed Drives (VSDs).

Main Methods:

• The framework redefines fault detection as a statistical test of independence between voltage inputs and current residuals.
• A healthy nominal model (Multilayer Perceptron) is trained using data from the VSD's self-commissioning routine.
• The Residual Information Value (RIV) principle is employed for robust fault identification.

Main Results:

• The proposed method achieves superior diagnostic performance compared to traditional baselines.
• It demonstrates higher statistical separability and a reduced false alarm rate.
• The system detects 1% incipient faults in approximately 61 ms and identifies the faulty phase accurately.

Conclusions:

• The RIV-based strategy offers a robust, non-intrusive, and industry-ready solution for predictive maintenance.
• It effectively balances high-speed detection with enhanced statistical reliability.
• The framework eliminates the need for manual data collection or complex physical parameter identification.