model_architecture.txt

================================================================================
CNN-LSTM MODEL ARCHITECTURE
Electrical Fault Classification System
================================================================================

MODEL OVERVIEW
--------------
Name: CNN_LSTM_FaultClassifier
Type: Hybrid Deep Learning Model (CNN + LSTM)
Task: Multiclass Time-Series Classification
Classes: 6 (No Fault, LG, LL, LLG, LLL, LLLG)
Input: (10 timesteps, 6 features)
Output: 6-class probability distribution

================================================================================
LAYER-BY-LAYER ARCHITECTURE
================================================================================

Layer 1: Conv1D (conv1d_1)
--------------------------
Type: 1D Convolutional Layer
Filters: 64
Kernel Size: 3
Activation: ReLU
Padding: Same
Input Shape: (10, 6)
Output Shape: (10, 64)
Parameters: 1,216

Layer 2: BatchNormalization (batch_norm_1)
-------------------------------------------
Type: Batch Normalization
Output Shape: (10, 64)
Parameters: 256

Layer 3: MaxPooling1D (maxpool_1)
----------------------------------
Type: Max Pooling
Pool Size: 2
Output Shape: (5, 64)
Parameters: 0

Layer 4: Dropout (dropout_1)
-----------------------------
Type: Dropout Regularization
Rate: 0.3 (30%)
Output Shape: (5, 64)
Parameters: 0

Layer 5: Conv1D (conv1d_2)
--------------------------
Type: 1D Convolutional Layer
Filters: 128
Kernel Size: 3
Activation: ReLU
Padding: Same
Input Shape: (5, 64)
Output Shape: (5, 128)
Parameters: 24,704

Layer 6: BatchNormalization (batch_norm_2)
-------------------------------------------
Type: Batch Normalization
Output Shape: (5, 128)
Parameters: 512

Layer 7: MaxPooling1D (maxpool_2)
----------------------------------
Type: Max Pooling
Pool Size: 2
Output Shape: (2, 128)
Parameters: 0

Layer 8: Dropout (dropout_2)
-----------------------------
Type: Dropout Regularization
Rate: 0.3 (30%)
Output Shape: (2, 128)
Parameters: 0

Layer 9: LSTM (lstm)
--------------------
Type: Long Short-Term Memory
Units: 100
Return Sequences: False
Dropout: 0.3
Recurrent Dropout: 0.2
Input Shape: (2, 128)
Output Shape: (100,)
Parameters: 91,600

Layer 10: Dense (dense_1)
-------------------------
Type: Fully Connected Layer
Units: 64
Activation: ReLU
Input Shape: (100,)
Output Shape: (64,)
Parameters: 6,464

Layer 11: Dropout (dropout_3)
------------------------------
Type: Dropout Regularization
Rate: 0.4 (40%)
Output Shape: (64,)
Parameters: 0

Layer 12: Dense (output)
------------------------
Type: Output Layer (Fully Connected)
Units: 6
Activation: Softmax
Input Shape: (64,)
Output Shape: (6,)
Parameters: 390

================================================================================
MODEL SUMMARY
================================================================================

Total Parameters: 125,142
Trainable Parameters: 125,142
Non-trainable Parameters: 0
Model Size: ~488 KB (saved as .h5)

================================================================================
HYPERPARAMETERS
================================================================================

Architecture Hyperparameters:
-----------------------------
- CNN Filters (Layer 1): 64
- CNN Filters (Layer 2): 128
- CNN Kernel Size: 3
- LSTM Units: 100
- Dense Layer Units: 64
- Dropout Rate (CNN): 0.3
- Dropout Rate (Dense): 0.4
- Recurrent Dropout: 0.2

Training Hyperparameters:
-------------------------
- Optimizer: Adam
- Learning Rate: 0.001
- Loss Function: Sparse Categorical Crossentropy
- Batch Size: 32
- Epochs: 50 (with early stopping)
- Validation Split: 0.2

Preprocessing:
--------------
- Normalization: StandardScaler (z-score normalization)
- Sequence Length: 10 timesteps
- Features: 6 (Ia, Ib, Ic, Va, Vb, Vc)

================================================================================
ARCHITECTURE FLOW DIAGRAM
================================================================================

Input Data (Raw CSV)
        |
        v
[Preprocessing & Normalization]
        |
        v
Time-Series Sequences (10 timesteps × 6 features)
        |
        v
┌───────────────────────────────────────┐
│     CNN FEATURE EXTRACTION BLOCK      │
├───────────────────────────────────────┤
│  Conv1D(64) → BatchNorm → MaxPool     │
│           ↓                           │
│       Dropout(0.3)                    │
│           ↓                           │
│  Conv1D(128) → BatchNorm → MaxPool    │
│           ↓                           │
│       Dropout(0.3)                    │
└───────────────────────────────────────┘
        |
        v
┌───────────────────────────────────────┐
│    LSTM TEMPORAL LEARNING BLOCK       │
├───────────────────────────────────────┤
│     LSTM(100 units)                   │
│  dropout=0.3, recurrent_dropout=0.2   │
└───────────────────────────────────────┘
        |
        v
┌───────────────────────────────────────┐
│    DENSE CLASSIFICATION BLOCK         │
├───────────────────────────────────────┤
│    Dense(64, ReLU)                    │
│           ↓                           │
│       Dropout(0.4)                    │
│           ↓                           │
│    Dense(6, Softmax)                  │
└───────────────────────────────────────┘
        |
        v
Output: Fault Class Probabilities [6 classes]
        |
        v
Predicted Fault Type

================================================================================
PERFORMANCE METRICS
================================================================================

Overall Performance:
-------------------
- Accuracy: 78.01%
- Precision: 77.53%
- Recall: 78.01%
- F1-Score: 77.47%

Class-wise Performance:
----------------------
| Fault Type      | Accuracy | F1-Score |
|-----------------|----------|----------|
| No Fault (0000) | 97.01%   | 95.18%   |
| LG Fault (1001) | 93.33%   | 88.24%   |
| LLG Fault (1011)| 88.50%   | 88.11%   |
| LL Fault (0110) | 80.50%   | 86.10%   |
| LLL Fault (0111)| 55.05%   | 50.21%   |
| LLLG Fault (1111)| 33.04%  | 38.27%   |

================================================================================
KEY FEATURES
================================================================================

1. Temporal Awareness
   - LSTM layer captures time-series dependencies
   - Learns patterns across 10 timesteps

2. Automatic Feature Extraction
   - CNN layers extract relevant spatial features
   - Hierarchical feature learning (64 → 128 filters)

3. Regularization
   - Batch Normalization for stable training
   - Dropout layers prevent overfitting
   - Early stopping with patience=10

4. Production Ready
   - Saved model weights (.h5)
   - Preprocessing artifacts (scaler, label encoder)
   - Comprehensive evaluation metrics

================================================================================
END OF ARCHITECTURE DOCUMENT
================================================================================