# Model Training

This document explains how the TinyML model was trained using **Edge Impulse**, including dataset preparation, feature engineering, model architecture, optimizations, and export steps. This is a GitHub-friendly guide—not a research paper version.

---

## 📁 1. Dataset Preparation
All raw datasets are located in the project:
```

Data/
├── Data Final.zip
└── Motor LAB 2.0 (with load using dynamometer).zip

```
Each dataset contains:
- Three-phase voltage signals
- Three-phase current signals
- Temperature readings
- Vibration sensor data
- Labeled fault states

### **Classes Used in Training**
- **Healthy**
- **Single Phase Fault**
- **Bearing Fault**
- **Overcurrent / Overload**

### **Sampling Configuration**
- **Sampling rate:** 200 Hz
- **Window size:** 2 seconds (400 samples)
- **7 total channels:**
  - Voltage A, Voltage B, Voltage C
  - Current A, Current B, Current C
  - Vibration

---

## 🔍 2. Importing Data Into Edge Impulse
### Steps:
1. Create a new Edge Impulse project
2. Go to **Data Acquisition → Upload Data**
3. Upload all CSV/JSON/ZIP dataset files
4. Assign correct labels (Healthy, Phase Fault, etc.)
5. Verify data plots for any anomalies

Your project dashboard should now show all classes with balanced sample counts.

---

## 📊 3. Creating the Impulse
Go to **Create Impulse**.

### Impulse Structure
```

[Time Series Data] → [Spectral Analysis] → [Neural Network]

```

### Settings
- **Window size:** 2000 ms
- **Window increase:** 2000 ms
- **Frequency:** 200 Hz

---

## 🧮 4. Feature Engineering
Edge Impulse automatically extracts:

### **Time-Domain Features**
- RMS
- Mean
- Peak
- Standard deviation

### **Frequency-Domain (FFT) Features**
- Magnitude spectrum
- Energy distribution
- Harmonic signatures
- Important for:
  - Bearing faults (200–800 Hz range)
  - Phase faults (low-frequency distortion)

All spectral parameters were tuned using Edge Impulse’s visual FFT preview.

---

## 🤖 5. Neural Network Model
A compact neural network was chosen to fit within ESP32-S3 memory constraints.

### Architecture
```

Input (FFT + Time features)
↓
Dense 32
↓
Dense 16
↓
Dense 8
↓
Dropout 0.2
↓
Softmax Output (4 classes)

```

### Training Config
- **Epochs:** 50
- **Batch size:** 32
- **Learning rate:** 0.0005
- **Loss:** Categorical Crossentropy
- **Optimizer:** Adam

### Achieved Metrics
| Metric | Value |
|--------|-------|
| Accuracy | **94.3%** |
| Precision | 0.96 |
| Recall | 0.96 |
| F1-score | 0.96 |

---

## 🔧 6. Model Optimization
### Quantization Method
The model was converted to:  
**TFLite Micro (int8 quantized)**.

### Benefits
| Feature | Float32 Model | int8 Quantized Model |
|---------|----------------|------------------------|
| Classifier Latency | 24 ms | **4 ms** |
| Flash Size | 551 KB | **164 KB** |
| RAM Usage | 22.7 KB | **8.2 KB** |

Quantized model chosen for deployment due to faster inference and lower memory footprint.

---

## 📦 7. Exporting the Model
### Steps
1. Go to **Deployment** in Edge Impulse
2. Select **Arduino Library**
3. Choose **Quantized (int8)** model variant
4. Download `.zip`
5. Extract to your project:
```

lib/ei-model/

```

All generated files will be inside:
```

lib/ei-model/src/

```
These files are used directly by the firmware.

---

## 🧪 8. Testing the Model
Using Edge Impulse **Model Testing**, run the test dataset.

Expected output:
- Accuracy ~94%
- Confusion matrix showing strong separation among 4 classes
- Minor confusion only in early fault stages

---

## 🚀 9. On-Device Inference
Model runs on ESP32-S3 inside the FreeRTOS **TinyML Task**.

### Inference Cycle
1. Collect 400 samples
2. Run feature extraction (FFT + time-domain)
3. Execute TFLM forward pass
4. Display classification result on TFT + Web UI

### Average Inference Latency
```

287 ms

```
Fits easily inside 2-second sampling window.

---

## 📈 10. Prognostic Trend Analysis
After each inference, Softmax probabilities are passed through:
- **Exponential Moving Average (EMA)**
- **Trend slope detection**

Predicts faults **before threshold crossing**.

---

# ✅ Model Training Complete
This model is ready for deployment via the ESP32-S3 TinyML runtime.