🏭 Predictive Maintenance for Industrial Equipment

From Reactive Repairs to Smart Decisions — Aligned with Industry 5.0

"During my time at Rockwell Automation, I saw firsthand how equipment downtime at plants like IFFCO doesn't just impact revenue — it disrupts entire supply chains. That's when predictive maintenance stopped being a data science problem for me, and became a responsibility."

---

🎯 The Problem This Solves

In a 24×7 continuous manufacturing environment, a 3.4% failure rate isn't "low" — it's a recurring operational leak. Each unplanned breakdown can cost ₹10–25 Lakhs in lost production and emergency repairs.

This project builds an industrial-grade AI system that converts catastrophic ₹25 Lakh breakdowns into ₹2 Lakh planned maintenance windows — by catching failures hours before they happen.

---

🗺️ Project Architecture at a Glance

Raw Sensor Data ──► Physics-Driven Features ──► Multi-Model ML ──► SHAP Explainability
                                                        │
                                                        ▼
              Deployment-Ready Model ◄── Risk Tier Score ◄── Threshold Tuning
                        │
                        ▼
        [Human-Centric Output] + [Drift Monitoring] + [Sustainability Analysis]

Step	What Happens	Why It Matters
0–1	Setup & Data Loading	UCI AI4I 2020 dataset — 10,000 real machine records
2	Exploratory Data Analysis	8 visualization deep-dives with dual technical + business insights
3	Physics-Driven Feature Engineering	Power curves, thermal gradients, overstrain index — not just statistics
4	Multi-Model ML Training	LR · Random Forest · Gradient Boost · XGBoost · LightGBM (winner)
5	SHAP Explainability	Black box → trusted tool for the shop floor
6	Threshold Tuning	Recall-first strategy; tuned to minimize missed failures
7	Failure Mode-Specific Models	Predict which failure (TWF / HDF / PWF / OSF), not just if
8	Business Impact Analysis	ROI quantified in ₹ Crores, not just AUC scores
9	Maintenance Priority Score	Color-coded risk tiers a plant engineer can act on immediately
10	Industry 5.0 — Human-Centric AI	Plain-language operator explainer + uncertainty flagging
11	Sustainability & Energy Efficiency	CO₂ footprint estimation from inefficient operation
12	Model Drift Monitoring (PSI)	Resilience-first — knows when it needs retraining
13	Model Persistence & Deployment	Production-ready, joblib-serialized, inference-ready
14	Final Summary	Lessons from the shop floor

---

⚡ Key Technical Highlights

🔬 Physics-First Feature Engineering

Rather than blindly feeding raw sensor data into a model, every engineered feature has a physical justification:

Feature	Formula	Physical Meaning
Temp_Delta	Process_Temp − Air_Temp	Thermal gap; drops below 8.6K → HDF imminent
Power_W	Torque × (RPM × 2π/60)	Real mechanical output; outside 3500–9000W = danger
Strain_Index	Tool_Wear × Torque	Cumulative mechanical fatigue — catches OSF early
Power_Out_of_Range	Binary flag	Immediate high-signal guardrail for the classifier
Low_Temp_Delta	ΔT < 8.6K flag	Deterministic HDF trigger (captures 95%+ of HDF events)

🤖 Model Performance

• Recall: 0.85 — Catches 85% of all failures before they happen
• F1-Score: 0.78 — Robust despite 3.4% class imbalance (handled via SMOTE)
• LightGBM chosen for production: fastest inference, lowest memory, native categorical support — ideal for edge deployment on PLCs

🧠 SHAP Explainability

Discovered that engineered features (Power_Out_of_Range, Strain_Index) outrank raw sensors in the model's decision hierarchy — validating that leading indicators (strain/power) matter more than lagging indicators (temperature) for early warning.

🎛️ Threshold Strategy

Threshold shifted from 0.5 → 0.3 — because mathematically, you can afford 200–400 false alarms and still break even on a single prevented failure.

---

🏭 Industry 5.0 Alignment

This notebook goes beyond standard ML. It's designed around all three pillars of Industry 5.0:

Pillar	What's Built
🧑‍🔧 Human-Centric	Plain-language operator explainer · SHAP decision support · Human review queue for uncertain predictions
🛡️ Resilience	Ensemble uncertainty quantification · PSI drift monitoring · Recall-first threshold tuning
🌱 Sustainability	Energy efficiency analysis · CO₂ footprint estimation · Maintenance-as-emissions-reduction

---

🚀 Getting Started

Run in Google Colab (Recommended) — No setup needed:

Or run locally:

# Clone the repo
git clone https://github.com/CodeWithSrish/predictive-maintenance.git
cd predictive-maintenance

# Install dependencies
pip install ucimlrepo shap imbalanced-learn xgboost lightgbm joblib scikit-learn pandas matplotlib seaborn

# Launch notebook
jupyter notebook Predictive_Maintenance.ipynb

⏱️ Runtime: ~3–5 minutes on Google Colab (CPU). Dataset auto-downloads via ucimlrepo.

---

📊 Dataset

AI4I 2020 Predictive Maintenance Dataset — UCI Machine Learning Repository
10,000 records · 14 features · 5 failure modes · 3.4% failure rate

Features: Air Temp (K) · Process Temp (K) · RPM · Torque (Nm) · Tool Wear (min) · Product Type
Targets:  Machine Failure · TWF · HDF · PWF · OSF · RNF

---

🗂️ Repository Structure

📦 predictive-maintenance
 ┣ 📓 Predictive_Maintenance.ipynb   # Main notebook — all 14 steps
 ┣ 📄 README.md                      # You are here
 ┗ 📁 models/                        # Saved LightGBM model (generated on run)
     ┗ lgbm_maintenance_model.pkl

---

🧰 Tech Stack

Python 3.8+ · scikit-learn · XGBoost · LightGBM · SHAP · imbalanced-learn (SMOTE) · pandas · numpy · matplotlib · seaborn · joblib · ucimlrepo

---

👩‍💻 Author

Srishti Rajput
Data Scientist
Inspired by real-world experience at Rockwell Automation and exposure to large-scale process operations at IFFCO

---

💡 Five Lessons from the Shop Floor

1. Physics first, then ML — Engineer features from process knowledge, not just statistics
2. Recall over accuracy — In maintenance, a missed failure is never acceptable
3. Explainability = adoption — A model the engineer doesn't trust never gets deployed
4. Quantify in ₹, not AUC — ROI converts stakeholders faster than metrics
5. Build for the edge — Production models must be fast, light, and interpretable

---

This project is not just about predicting failures.
It's about building systems that manufacturing engineers can trust, operators can understand, and plant managers can justify to the board.

⭐ Star this repo if you found it useful!