Deep Learning for Induction Motor Fault Detection (MCSA + Transfer Learning)

This project develops a deep-learning-based condition monitoring pipeline for induction motor fault detection and classification using Motor Current Signature Analysis (MCSA) from stator current time-domain signatures. A three-phase squirrel-cage induction motor is modeled in MATLAB/Simulink, faults are induced under different load conditions, and transfer learning is used to classify motor health states.

Fault classes (multiclass):

• Healthy
• Single-phasing
• Phase-to-phase fault

Models evaluated:

• InceptionV3
• ResNet152
• VGG19

---

Why this matters (industry relevance)

Induction motors dominate industrial actuation, and unplanned downtime is expensive. Early fault detection enables:

• Predictive maintenance scheduling
• Reduced catastrophic failures and secondary damage
• Lower downtime and better production continuity
• Improved energy efficiency by detecting abnormal operation patterns

---

Technical summary

1) Motor modeling + fault injection (simulation)

A 3-phase induction motor model is created in MATLAB/Simulink and simulated under:

• Multiple load levels (0–100%)
• Healthy and faulty conditions

Faults are introduced by disconnecting one or more stator phases (single-phasing and phase-to-phase). The resulting stator current signatures are used as the diagnostic signal.

2) Dataset generation

• 276 current time signatures collected from simulation
• Train/validation split reported as ~82% / 18%
• Label mapping used:
  • Healthy → 0
  • Single-phasing → 1
  • Phase-to-phase → 2

Note: Some parts of the report mention additional stator fault types in dataset tables. In this repo, keep the labels consistent with the final class definition used during training.

3) Deep learning approach (transfer learning)

Transfer learning is used to reduce data requirements by fine-tuning pre-trained CNN backbones (originally trained on ImageNet) with a smaller custom dataset. The workflow includes:

• Preprocessing and labeling (Google Colab workflow)
• Model adaptation (Flatten + Dense layers + Softmax output)
• Training + validation + confusion matrix evaluation

---

Results snapshot

Model performance (reported)

ResNet152 shows the strongest overall performance:

• Precision ≈ 97%
• Recall ≈ 96%
• F1-score ≈ 97%
• Training accuracy ≈ 96.49%
• Validation accuracy ≈ 97.92%

Per-class accuracy comparison (reported)

Class	InceptionV3	ResNet152	VGG19
Healthy	75%	100%	75%
Single-phasing	81%	94%	88%
Phase-to-phase	100%	100%	100%

---

How to Run

This project can be run in two stages: (A) simulation/data generation and (B) model training + evaluation.

A) Generate stator current data (MATLAB/Simulink)

1. Open the Simulink induction motor model
2. Configure the simulation cases:
   • Health states: Healthy, Single-phasing, Phase-to-phase
   • Load levels: e.g., 0%, 25%, 50%, 75%, 100%
3. Run simulations and log stator current signals (per phase or a selected phase current).
4. Export each run to CSV and store in:
   • data/

Recommended CSV fields (minimum):

• time
• current_a (or current_b, current_c)
• label (0=Healthy, 1=Single-phasing, 2=Phase-to-phase)
• load_pct

Tip: Keep naming consistent so your notebook preprocessing can batch-load runs easily.

---

B) Train and evaluate the deep learning models (Google Colab / Local Python)

1. Open one of the notebooks:
   • ResNet152.ipynb (recommended starting point)
   • InceptionV3.ipynb
   • VGG19.ipynb
2. Run preprocessing cells:
   • Load dataset from data/
   • Apply any reshaping/segmentation steps used in the notebook
   • Confirm label mapping and class balance
3. Train the model and evaluate:
   • Confusion matrix
   • Accuracy/loss curves
   • Precision, Recall, F1-score

---

C) Reproduce reported results

To reproduce the headline results, run:

• ResNet152.ipynb start-to-finish
  using the same dataset split strategy and hyperparameters documented below.

---

Hyperparameters (Reported)

Parameter	InceptionV3	ResNet152	VGG19
Learning rate	0.0001	0.001	0.0001
Batch size	16	28	32
Loss function	Categorical cross-entropy	Categorical cross-entropy	Categorical cross-entropy
Epochs	20	15	20

Note: If you change the dataset split, preprocessing, or augmentation, your metrics may shift. Keep a fixed random seed for reproducibility when possible.

---

Limitations

• Simulation-only data: Results are obtained from simulated stator current signatures; real-world measurements may include additional noise sources, sensor drift, and non-idealities.
• Dataset size: The dataset is relatively small for deep CNN fine-tuning; larger datasets typically improve generalization and robustness.
• Fault coverage: The current study focuses on stator-related faults (healthy, single-phasing, phase-to-phase). It does not yet include common industrial fault modes such as bearing faults, rotor bar faults, or eccentricity.
• Domain shift risk: Models trained on simulated signatures may not transfer directly to field data without domain adaptation or re-training with real measurements.

---

Acknowledgements

• Project supervisor(s) and department support (as listed in the project report).
• Contributors and collaborators involved in simulation setup, dataset preparation, and evaluation.
• Open-source libraries and research community enabling transfer learning workflows.