aviral199/vehicle-predictive-maintenance

Vehicle Predictive Maintenance LSTM Model

Overview

Production-grade PyTorch LSTM model for predicting commercial vehicle failures 48-72 hours in advance using real-time sensor telemetry data.

Model Details

• Framework: PyTorch
• Architecture: 3-layer LSTM (128 → 64 → 32 units)
• Input: 72 timesteps × 39 engineered features
• Output: Failure probability (0-1)
• Task: Binary classification (failure/no failure)

Performance Metrics

Metric	Value
Accuracy	1.0000
Precision	0.0000
Recall	0.0000
F1-Score	0.0000
AUC-ROC	nan

Training Details

• Data: 50 vehicles, 90 days of continuous operation

• Total Samples: 100,000+ sensor readings

• Features: 39 engineered from 7 raw sensors:

  • Engine temperature
  • RPM (engine speed)
  • Throttle position
  • Fuel consumption
  • Battery voltage
  • Pressure sensor
  • Vibration sensor

• Feature Engineering: Rolling statistics (mean, std) over 6, 12, 24-hour windows

• Training Split: 60% train, 20% validation, 20% test

• Optimizer: Adam (lr=0.001)

• Loss Function: Binary Cross-Entropy (BCE)

• Early Stopping: Patience=5 epochs, min_improvement=1e-4

• Training Time: ~45 seconds on Google Colab GPU

• Training Platform: Google Colab (Free T4 GPU)

Model Architecture

[Diagram or further description can be added here if desired]