# Signature-Guided Data Augmentation: Neural Network for Diagnosing Faults in Induction Motors
The Signature-Guided Data Augmentation method, developed by specialists at HSE University, enables neural networks to detect faults in three-phase induction motors with 99% accuracy. Fault type classification reaches 86%. The approach generates synthetic data based on current signals from healthy motors, simulating real defects without physically damaging equipment. This solves the shortage of real fault datasets in industry.
Three-phase induction motors are a key component of production systems: from assembly lines at car factories to pumps in water supply networks. Faults lead to downtime, losses, and safety risks. Traditional current spectrum analysis requires manual filter tuning, frequency extraction, and expert knowledge, which slows down the process.
Challenges of Traditional Methods and ML Approaches
Classic MCSA (Motor Current Signature Analysis) relies on visual inspection of the current spectrum. Engineers look for characteristic harmonics associated with stator winding breaks, rotor eccentricity, or short circuits. The process is labor-intensive: it needs calibration for specific motors, accounting for load and operating conditions.
Machine learning promises automation but runs into a lack of data. Real faults are rare, and archives of fault signals aren't maintained. Standard augmentations (noise, shifts) don't account for defect physics, reducing model generalizability.
The HSE team—Artem Ryzhikov, Sara Ali, Alexander Khizhik, Stepan Svirin, Denis Derkach—introduced the Signature-Guided approach. The algorithm extracts 'signatures' of typical faults from physical models and injects them into clean signals.
How Signature-Guided Data Augmentation Works
The method relies on physical patterns:
- Stator turn breaks: adding sideband harmonics around the fundamental frequency.
- Rotor eccentricity: modulation at rotation frequency with a 2f_r coefficient.
- Short circuit: phase asymmetry with increased zero-sequence currents.
- Bearing defects: impulsive spikes at BPFO/BPFI frequencies.
Signals are generated by overlaying these signatures onto normal spectra. The neural network (CNN or Transformer-based) trains on a mixed dataset: 80% synthetic + 20% real (if available).
Process:
- Record current in normal mode (FFT spectrum).
- Extract the motor's baseline signature.
- Generate faults: inject harmonics based on models.
- Train classifier (binary: healthy/faulty; multiclass: fault type).
- Validate on hold-out set.
Denis Derkach emphasizes: the method doesn't require FEM simulations or real fault tests—it uses only analytical formulas from MCSA.
Experiment Results
Testing on datasets from two motors (4- and 6-pole, 50 Hz):
| Task | Accuracy | Recall | F1-score |
|--------|----------|--------|----------|
| Binary (healthy/faulty) | 99% | 98.5% | 98.7% |
| Multiclass (5 types) | 86% | 84% | 85% |
Sara Ali notes: the system starts with normal data, evolving into a detector without historical failures. Generalization to new motors is achieved by re-recording the baseline signature.
Advantages for mid/senior developers:
- Minimal data requirements: one healthy signal.
- Physical interpretability: traceable to harmonics.
- Scalability: applicable to any parameters (power, poles).
- Early detection: threshold on confidence score.
Key Takeaways
- 99% accuracy in binary classification without real fault data.
- Generation based on MCSA signatures ensures realism.
- Reduced downtime through predictive maintenance.
- Planned validation on industrial sites.
- Openness: method not tied to specific hardware.
The approach integrates into SCADA/IIoT systems: edge devices analyze current in real time, sending alerts. Future plans include expansion to vibration and temperature.
— Editorial Team
No comments yet.