CNN-LSTM Electrical Fault Classification

A hybrid deep learning pipeline for classifying electrical faults in transmission lines using time-series voltage and current measurements.

📊 Project Overview

• Model: Hybrid CNN-LSTM Architecture
• Accuracy: 78.01% on test set
• Dataset: 7,861 electrical measurements
• Classes: 6 fault types (No Fault, LG, LL, LLG, LLL, LLLG)

📦 Data Source

This project uses the Electrical Fault Detection and Classification dataset from Kaggle:

• Dataset URL: https://www.kaggle.com/datasets/esathyaprakash/electrical-fault-detection-and-classification
• Features: Time-series voltage and current measurements from transmission lines
• Total Samples: 7,861 electrical measurements
• Format: CSV files with multiclass and binary classification variants

🔗 Reference Project

This implementation is built with reference to a previous Kaggle notebook:

• Reference Notebook: Electrical Faults Analysis & Classification
• Author: harshsingh2209
• Enhancements: Refactored into modular Python scripts, improved architecture, and comprehensive evaluation pipeline

📁 Project Structure

A Hybrid CNN - LSTM Model for fault detection in power distribution system/
├── src/                                    # Source code
│   ├── README.md                           # Source code documentation
│   ├── cnn_lstm_preprocessing.py           # Data preprocessing & sequence generation
│   ├── cnn_lstm_model.py                   # Model architecture definitions
│   ├── train_cnn_lstm.py                   # Training pipeline
│   └── evaluate_model.py                   # Evaluation & visualization
│
├── models/                                 # Trained models & artifacts
│   ├── best_cnn_lstm_model.h5              # Trained model weights (1.5 MB)
│   ├── scaler.pkl                          # Fitted StandardScaler
│   └── label_encoder.pkl                   # Label encoder
│
├── results/                                # Training results
│   ├── README.md                           # Results documentation
│   ├── visualizations/                     # Generated plots
│   │   ├── confusion_matrix.png
│   │   ├── training_history.png
│   │   ├── classwise_f1_scores.png
│   │   └── training_summary.png
│   └── metrics/                            # Performance metrics
│       ├── evaluation_metrics.txt
│       ├── training_history.csv
│       └── hyperparameters_log.json
│
├── data/                                   # Datasets
│   ├── classData.csv                       # Multiclass fault dataset (655 KB)
│   └── detect_dataset.csv                  # Binary fault detection dataset (952 KB)
│
├── docs/                                   # Documentation
│   └── README_CNN_LSTM.md                  # Detailed technical documentation
│
├── electrical-faults-analysis-classification.ipynb  # Jupyter notebook
├── electrical-faults-analysis-classification.py    # Converted Python script
├── model_architecture.txt                  # Model architecture summary
├── REFACTORING_SUMMARY.md                  # Project refactoring notes
└── README.md                               # This file

🚀 Quick Start

Prerequisites

pip install tensorflow numpy pandas scikit-learn matplotlib seaborn

Training

cd "c:\Users\HP\Downloads\A Hybrid CNN - LSTM Model for fault detection in power distribution system"
$env:PYTHONIOENCODING='utf-8'
python src/train_cnn_lstm.py

Evaluation

python src/evaluate_model.py

📈 Results

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%

🏗️ Architecture

Model Flow Diagram

flowchart TD
    A["Raw CSV Data"] --> B["Data Loading & Preparation"]
    B --> C["Feature Extraction"]
    C --> D["Train/Test Split"]
    D --> E["StandardScaler Normalization"]
    E --> F["Sequence Generation"]
    
    F --> G["CNN-LSTM Model"]
    
    G --> H["CNN Block 1"]
    H --> I["CNN Block 2"]
    I --> J["LSTM Layer"]
    J --> K["Dense Layer"]
    K --> L["Output Layer"]
    
    L --> M["Predictions"]
    M --> N["Evaluation Metrics"]
    
    N --> O["Confusion Matrix"]
    N --> P["ROC Curves"]
    N --> Q["PR Curves"]
    N --> R["F1 Scores"]
    
    style G fill:#e1f5ff
    style H fill:#fff3e0
    style I fill:#fff3e0
    style J fill:#f3e5f5
    style K fill:#e8f5e9
    style L fill:#e8f5e9

Loading

Layer Architecture

Input (10 timesteps, 6 features)
    ↓
Conv1D (64 filters) → BatchNorm → MaxPool → Dropout
    ↓
Conv1D (128 filters) → BatchNorm → MaxPool → Dropout
    ↓
LSTM (100 units, dropout=0.3)
    ↓
Dense (64, ReLU) → Dropout
    ↓
Dense (6, Softmax)

Parameters: 125,142 (488 KB)

Detailed Architecture: See model_architecture.txt for complete layer-by-layer breakdown

📚 Documentation

For detailed documentation, see docs/README_CNN_LSTM.md

🔬 Key Features

• ✅ Temporal Awareness: LSTM layer captures time-series patterns
• ✅ Automatic Feature Learning: CNN extracts relevant features
• ✅ Production Ready: Includes scaler and encoder for deployment
• ✅ Comprehensive Evaluation: Confusion matrix, F1 scores, training curves
• ✅ Reproducible: Fixed random seeds and hyperparameter logging

📊 Visualizations

All visualizations are automatically generated in results/visualizations/:

• Confusion Matrix - Prediction accuracy per class with heatmap
• Training History - Accuracy & loss curves over epochs
• Class-wise F1 Scores - Performance comparison across fault types
• ROC Curves - Receiver Operating Characteristic (One-vs-Rest) with AUC scores
• PR Curves - Precision-Recall curves with Average Precision scores

Run python src/evaluate_model.py to generate all visualizations.

🎓 Academic Use

This project was developed for a term paper on electrical fault classification. The implementation demonstrates:

• Deep learning for time-series classification
• Hybrid CNN-LSTM architecture
• Complete ML pipeline from preprocessing to evaluation

📝 License

Educational project for academic purposes.

👤 Author

Developed for academic purposes as part of a 400 level project on electrical fault detection and classification.