flight_dynamics_ML

flight_state estimator Using ML

Bearing Fault Classification: Machine Learning & Deep Learning

A comprehensive machine learning project for bearing condition monitoring using both Classical ML and Deep Learning approaches on the industry-standard CWRU Bearing Dataset.

---

🎯 Project Overview

This project implements dual approaches for bearing fault detection:

Classical Machine Learning:

• ✅ Manual feature extraction (time + frequency domain)
• ✅ Traditional ML models (Random Forest, SVM, Gradient Boosting)
• ✅ Fast training (~30 seconds)
• ✅ 92-98% accuracy

Deep Learning (1D CNN):

• ✅ Automatic feature learning from raw signals
• ✅ End-to-end learning pipeline
• ✅ Training time (~3-5 minutes)
• ✅ 95-99% accuracy

Real-World Application: Predictive maintenance, condition monitoring, industrial fault detection

Bearing Conditions:

• 🟢 Normal - Healthy bearing
• 🔴 Inner Race Fault - Inner race damage
• 🟠 Outer Race Fault - Outer race damage

---

📁 Project Structure

bearing-fault-ml/
│
├── main.py                      # Main entry (supports both approaches)
├── data_loader.py               # CWRU dataset loader
├── feature_extractor.py         # Feature engineering (for ML)
├── models.py                    # Classical ML models
├── visualizer.py                # Results visualization
├── deep_learning_model.py       # 1D CNN implementation
├── requirements.txt             # Dependencies
├── README.md                    # This file
│
├── data/
│   └── cwru/                    # CWRU dataset
│       ├── normal/
│       ├── inner_fault/
│       └── outer_fault/
│
└── results/                     # Auto-generated
    └── YYYY-MM-DD_HH-MM-SS/
        ├── results.png                  # ML comparison
        ├── confusion_matrix.png         # Confusion matrices
        ├── cnn_training_history.png     # CNN training curves
        ├── metrics.csv                  # Metrics
        ├── config.txt                   # Configuration
        └── cnn_model.h5                 # Saved CNN model

---

🚀 Quick Start Guide

Step 1: Clone and Install

# Clone repository
git clone https://github.com/ak5hay19/bearing-fault-ml.git
cd bearing-fault-ml

# Install dependencies
pip install -r requirements.txt

# For Deep Learning (optional)
pip install tensorflow

Step 2: Download CWRU Dataset

Visit: https://engineering.case.edu/bearingdatacenter

Download 12 files (right-click → Save Link As):

Normal Baseline Data (4 files):

File	Motor Load	Location
97.mat	0 HP	Normal Baseline → 12k Drive End
98.mat	1 HP	Normal Baseline → 12k Drive End
99.mat	2 HP	Normal Baseline → 12k Drive End
100.mat	3 HP	Normal Baseline → 12k Drive End

Inner Race Fault - 0.007" (4 files):

File	Motor Load	Location
105.mat	0 HP	12k Drive End → Inner Race → 0.007"
106.mat	1 HP	12k Drive End → Inner Race → 0.007"
107.mat	2 HP	12k Drive End → Inner Race → 0.007"
108.mat	3 HP	12k Drive End → Inner Race → 0.007"

Outer Race Fault - 0.007" @6:00 (4 files):

File	Motor Load	Location
130.mat	0 HP	12k Drive End → Outer Race → 0.007" → @6:00
131.mat	1 HP	12k Drive End → Outer Race → 0.007" → @6:00
132.mat	2 HP	12k Drive End → Outer Race → 0.007" → @6:00
133.mat	3 HP	12k Drive End → Outer Race → 0.007" → @6:00

Step 3: Organize Dataset

# Create folder structure
mkdir -p data/cwru/normal
mkdir -p data/cwru/inner_fault
mkdir -p data/cwru/outer_fault

# Place downloaded files:
# data/cwru/normal/       → 97-100.mat
# data/cwru/inner_fault/  → 105-108.mat
# data/cwru/outer_fault/  → 130-133.mat

Step 4: Choose Your Approach

Edit main.py (lines 17-18):

# For Classical ML only (fast)
USE_DEEP_LEARNING = False
USE_CLASSICAL_ML = True

# For Deep Learning only
USE_DEEP_LEARNING = True
USE_CLASSICAL_ML = False

# For comparison (both)
USE_DEEP_LEARNING = True
USE_CLASSICAL_ML = True

Step 5: Run

python main.py

---

🔬 Methodology

Approach 1: Classical Machine Learning

Pipeline:

Raw Signal (2048 points)
  ↓
Feature Extraction (11 features)
  • Time: Mean, Std, RMS, Peak, Kurtosis, Skewness
  • Frequency: Dominant freq, Energy bands, Spectral centroid
  ↓
Feature Scaling (StandardScaler)
  ↓
ML Models (RF, SVM, GB)
  ↓
Predictions

Models:

• Random Forest: 100 trees, robust ensemble
• SVM: RBF kernel, high-dimensional classification
• Gradient Boosting: Sequential learning, highest accuracy

Performance:

• Training time: 5-10 seconds
• Accuracy: 92-98%
• Interpretability: High ✅

---

Approach 2: Deep Learning (1D CNN)

Pipeline:

Raw Signal (2048 points)
  ↓
1D CNN Architecture:
  • Conv1D(64) → BatchNorm → MaxPool → Dropout
  • Conv1D(128) → BatchNorm → MaxPool → Dropout
  • Conv1D(256) → BatchNorm → MaxPool → Dropout
  • Conv1D(128) → BatchNorm → GlobalAvgPool
  • Dense(128) → Dropout
  • Dense(64) → Dropout
  • Dense(3, softmax)
  ↓
Predictions

CNN Details:

• Input: Raw vibration signals (no preprocessing)
• Total Parameters: ~487,000
• Optimizer: Adam (lr=0.001)
• Loss: Categorical crossentropy
• Callbacks: Early stopping, learning rate reduction

Performance:

• Training time: 2-5 minutes
• Accuracy: 95-99%
• Automatic feature learning ✅

---

📊 Performance Comparison

Metric	Classical ML	Deep Learning
Accuracy	92-98%	95-99%
Training Time	5-10 sec ⚡	2-5 min 🐢
Feature Engineering	Manual	Automatic
Interpretability	High	Low
Model Size	<1 MB	5-10 MB
GPU Required	No	Recommended
Data Required	Low	Medium
Best For	Quick results	Max accuracy

---

📈 Sample Output

Classical ML:

[4/5] Training models...
  Training Random Forest... Done (2.34s)
  Training SVM... Done (1.87s)
  Training Gradient Boosting... Done (3.45s)

[5/5] Evaluating...
Random Forest        : Accuracy = 0.9456 (94.56%)
SVM                  : Accuracy = 0.9523 (95.23%)
Gradient Boosting    : Accuracy = 0.9612 (96.12%)

✓ Best Classical Model: Gradient Boosting (96.12%)

Deep Learning:

[2/4] Building and training 1D CNN...
  Input shape: (2048, 1)
  Number of classes: 3

Model: "sequential"
Total params: 487,235
Trainable params: 486,083

Epoch 1/100
58/58 [======] - 2s - loss: 0.8234 - accuracy: 0.6543 - val_accuracy: 0.7012
...
Epoch 35/100
58/58 [======] - 1s - loss: 0.0234 - accuracy: 0.9921 - val_accuracy: 0.9812

[3/4] Evaluating CNN...
✓ CNN Test Accuracy: 0.9823 (98.23%)

Combined Results:

======================================================================
✓ RESULTS SUMMARY
======================================================================

📊 Classical ML:
   Best Model: Gradient Boosting
   Accuracy: 0.9612 (96.12%)

🧠 Deep Learning:
   CNN Accuracy: 0.9823 (98.23%)

📁 Output: results/2025-01-10_14-30-45/
======================================================================

---

📁 Output Files

1. results.png

Bar charts comparing:

• Model accuracies
• Training times
• Percentage view

2. confusion_matrix.png

Confusion matrices for all ML models showing classification errors

3. cnn_training_history.png

CNN training curves:

• Training vs validation accuracy
• Training vs validation loss
• Early stopping point

4. metrics.csv

Model,Accuracy,Accuracy_Percentage,Training_Time_seconds
Random Forest,0.9456,94.56,2.34
SVM,0.9523,95.23,1.87
Gradient Boosting,0.9612,96.12,3.45

5. config.txt

Complete configuration and dataset statistics