Predictive Maintenance MLOps Pipeline

An end-to-end ML pipeline for predicting engine maintenance needs in commercial diesel vehicle fleets.

Overview

This project implements a predictive maintenance system that analyzes engine sensor data to predict whether maintenance is required before a breakdown occurs. The system achieves 99.8% recall (catches virtually all failures) with 63% precision.

Key Features

Real-time predictions via Streamlit web interface
Automated CI/CD pipeline with GitHub Actions
Model versioning on HuggingFace Model Hub
Dataset versioning on HuggingFace Datasets

Pipeline Stages

Stage	Script	Description
1. Data Registration	`model_building/register.py`	Upload raw sensor data to HuggingFace
2. Data Preparation	`model_building/prep.py`	Feature engineering (22 features from 6 sensors)
3. Model Training	`model_building/train.py`	Train AdaBoost classifier with recall gate
4. Deployment	`hosting/deploy.py`	Sync to HuggingFace Spaces

Resources

Resource	URL
Live Demo	https://huggingface.co/spaces/jskswamy/predictive-maintenance
Dataset	https://huggingface.co/datasets/jskswamy/predictive-maintenance-data
Model	https://huggingface.co/jskswamy/predictive-maintenance-model

Local Development

# Clone and setup
git clone https://github.com/jskswamy/predictive-maintenance.git
cd predictive-maintenance
pip install -r requirements.txt
export HF_TOKEN=your_huggingface_token

# Run pipeline
python model_building/register.py
python model_building/prep.py
python model_building/train.py
streamlit run deployment/app.py

Model Performance

Metric	Value
Recall	99.78%
Precision	63.2%
F2 Score	0.917
ROC-AUC	0.70

Threshold: 0.316 (optimized for maximum recall)

Business Impact

For a 100-truck fleet: 17 breakdowns prevented annually, $85,000-$170,000 in cost savings.

License

MIT License

Capstone Project for PGP-AIML program

jskswamy/predictive-maintenance