Back to Projects
Predictive Fault Diagnosis System for Induction Motors

Predictive Fault Diagnosis System for Induction Motors

University of Moratuwa•2019 - 2020•Undergraduate Capstone Project

Technologies

pythonaws

Key Highlights

  • •Published in IESL Journal 'Engineer' (2021)
  • •Real-time fault detection using machine learning on embedded systems
  • •Raspberry Pi-based standalone system with Python/MATLAB backend
  • •Java GUI for monitoring and visualization
  • •Multi-class fault classification with high accuracy
Read Publication

Overview

Undergraduate capstone project developing an intelligent, standalone fault diagnosis system for three-phase induction motors. The system uses machine learning to detect and classify motor faults in real-time, enabling predictive maintenance and preventing catastrophic failures.

Publication

Title: "Predictive and Standalone Fault Diagnosis System for Induction Motors"
Journal: Engineer - Journal of the Institution of Engineers Sri Lanka (IESL)
Year: 2021
Authors: Imantha Ahangama et al.

Motivation

Induction motors are workhorses of industry, but unexpected failures cause:

  • Costly downtime and production losses
  • Safety hazards for workers
  • Damage to connected equipment
  • Expensive emergency repairs

Traditional monitoring requires expert analysis. Our system automates fault detection using ML, making predictive maintenance accessible even to smaller operations.

System Design

Hardware Architecture

Embedded Platform: Raspberry Pi 3B+

  • Sufficient processing for real-time ML inference
  • GPIO for sensor interfacing
  • Network connectivity for remote monitoring
  • Cost-effective for industrial deployment

Sensors:

  • Current sensors (3-phase)
  • Vibration sensors (accelerometer)
  • Temperature sensors
  • Motor speed encoder

Software Components

Data Acquisition:

  • Real-time sensor data collection
  • Signal conditioning and filtering
  • Sampling at appropriate frequencies
  • Circular buffer for continuous monitoring

Feature Extraction:

  • Time-domain features (RMS, peak, crest factor)
  • Frequency-domain analysis (FFT, harmonics)
  • Statistical features (mean, variance, skewness)
  • Motor-specific features (slip, power factor)

ML Classification:

  • Multi-class classifier trained on fault signatures
  • MATLAB for model development and validation
  • Python (scikit-learn) for embedded inference
  • Support for various fault types:
    • Bearing faults
    • Stator winding faults
    • Rotor bar faults
    • Misalignment
    • Unbalance

User Interface:

  • Java Swing desktop application
  • Real-time data visualization
  • Fault alerts and notifications
  • Historical trend analysis
  • System configuration and calibration

Machine Learning Approach

Dataset

  • Experimental data from motors with induced faults
  • Simulation data from motor models
  • Balanced dataset across fault classes
  • Multiple operating conditions (load, speed)

Models Evaluated

  • Support Vector Machines (SVM)
  • Random Forest
  • Neural Networks
  • K-Nearest Neighbors

Best Performance: SVM with RBF kernel achieving 95%+ accuracy

Model Deployment

  • Model serialization for embedded deployment
  • Optimization for Raspberry Pi constraints
  • Inference time < 100ms for real-time operation
  • Periodic retraining capability

Technical Challenges

Real-Time Processing

  • Signal processing on embedded hardware
  • Meeting latency requirements for safety-critical faults
  • Efficient feature computation

Robustness

  • Handling noisy industrial environments
  • Adapting to varying motor loads and speeds
  • False positive/negative trade-offs

Deployment

  • Standalone operation without cloud dependency
  • Reliability for 24/7 operation
  • Easy installation and commissioning

Results

Fault Detection Performance

  • Accuracy: 95%+ across all fault types
  • False Positive Rate: < 3%
  • Inference Time: ~80ms (real-time capable)
  • Detection Latency: Faults detected within seconds of onset

Practical Validation

  • Tested on laboratory test bench
  • Validation with multiple motor sizes (0.5-5 HP)
  • Successful detection of actual bearing failures
  • Comparison with expert diagnosis (high agreement)

Impact

  • Academic: Published in peer-reviewed IESL journal
  • Industrial Interest: Demonstrations to local manufacturers
  • Educational: Case study for ML in embedded systems
  • Personal Growth: Deep dive into signal processing, embedded systems, and ML deployment

Technology Stack

Hardware: Raspberry Pi, current sensors, accelerometers, temperature sensors
Embedded: Python 3, GPIO libraries, threading
ML: scikit-learn, NumPy, SciPy, MATLAB
Signal Processing: FFT, digital filters, feature extraction
UI: Java Swing, JFreeChart
Communication: MQTT for optional remote monitoring

Future Enhancements

  • Edge ML model retraining based on operational data
  • Integration with industrial IoT platforms
  • Mobile app for maintenance technicians
  • Support for additional motor types
  • Prognostics (remaining useful life estimation)