🌍 PakIndustry 4.0 AI Suite

🚀 End-to-end AI solutions for Pakistan’s manufacturing industry — developed for the Uraan AI Techathon 1.0.

We built a modular Industry 4.0 platform tackling three of the biggest challenges in local manufacturing:

1. 🏭 Defect Detection → Reduce wastage with automated visual inspection
2. 🔧 Predictive Maintenance → Prevent equipment breakdowns before they happen
3. 📊 Supply Chain Forecasting → Optimize inventory and demand planning

Together, these modules form a practical, deployable AI suite that directly addresses production inefficiency, unplanned downtime, and supply chain uncertainty.

---

✨ Executive Summary

Pakistan’s manufacturing sector struggles with:

• High production defects → wasted raw material, loss of export contracts
• Unexpected machine failures → costly downtime and emergency repairs
• Poor demand forecasting → overstocking, stockouts, and inefficiency

👉 Our AI Suite provides:

• Computer Vision QC (casting defects)
• Deep Learning RUL estimation (engine maintenance)
• Time-series forecasting (warehouse demand)

⚠️ Note on Limitations:

• Defect Detection is highly accurate (99.6%) but trained on one dataset → retraining needed for new materials/lighting conditions.
• Supply Chain Forecasting performs exceptionally (1.97% sMAPE), but relies on historical demand patterns → disruptive events could reduce accuracy.
• Predictive Maintenance currently underfits (flat predictions ~21 cycles) → framework works, but model accuracy is limited.

---

🧩 Modules Overview

1. 🏭 Defect Detection

• Goal: Detect casting defects in industrial parts
• Model: EfficientNet-B0 (transfer learning)
• Performance: 99.6% accuracy, 100% precision, 99.7% F1
• Deployment: Streamlit dashboard for easy factory-floor use
• Limitations: Lighting-sensitive, optimized for cast parts only

➡️ Read full Defect Detection README

➡️ Watch the demo

2. 🔧 Predictive Maintenance

• Goal: Predict Remaining Useful Life (RUL) of turbofan engines
• Dataset: NASA C-MAPSS (FD001 subset)
• Model: LSTM/GRU sequence models
• Performance: MAE ~13, MAPE ~22%, conservative flat predictions (~21 cycles)
• Deployment: Inference pipeline + Streamlit app for uploading test data
• Limitations: Underfits complex degradation patterns, needs further training on larger dataset

➡️ Read full Predictive Maintenance README ➡️ Watch the demo

---

3. 📊 Supply Chain Forecasting

• Goal: Daily order forecasting across 7 European warehouses (Rohlik dataset)
• Model: LightGBM with engineered lag/rolling features + calendar integration
• Performance: RMSE 208, MAE 127, sMAPE 1.97%, NRMSE 3.2%
• Deployment: Train/inference/evaluation scripts with reproducibility
• Limitations: Reliant on historical continuity, not yet tested for extreme disruptions

➡️ Read full Forecasting README ➡️ Watch the demo

---

🏗️ Repository Structure

pakindustry-4.0/
├── docs/                       # Contains project documentation
├── defect-detection/           # Module 1: Vision-based defect detection
├── predictive-maintenance/     # Module 2: RUL estimation
├── forecast/                   # Module 3: Supply chain demand forecasting
├── test_samples/               # Contains sample datasets for testing
├── LICENSE                     # MIT license file
├── README.md                   # Root overview
└── requirements.txt            # Project dependencies

---

🔄 Reproducibility & Setup

We’ve prioritized clear, reproducible pipelines for all modules.

1. Clone Repository

git clone https://github.com/sfarrukhm/pakindustry-4.0.git
cd pakindustry-4.0

2. Create Environment

python -m venv .venv
source .venv/bin/activate       # Linux/Mac
.venv\Scripts\activate          # Windows

3. Install Dependencies

pip install -r requirements.txt

4. Run Individual Modules

• Defect Detection:

  cd defect-detection
  python train.py
  streamlit run app.py

• Predictive Maintenance:

  cd predictive-maintenance
  python train.py
  streamlit run app.py

• Forecasting:

  cd forecast
  python train.py
  python inference.py

---

📊 Results Snapshot

Module	Metric Highlights	Status
🏭 Defect Detection	Accuracy 99.6%, Precision 100%	✅ Ready for demo
🔧 Predictive Maint.	MAE 13, MAPE 22%, flat ~21-cycle outputs	⚠️ Underfitting
📊 Forecasting	sMAPE 1.97%, NRMSE 3.2%	✅ Exceeds target

---

🌟 Why This Matters

• For Judges: Demonstrates 3 distinct, deployable AI solutions under one cohesive suite.
• For Industry: Provides a foundation for Pakistan’s manufacturers to experiment with AI tools that reduce costs and boost competitiveness.
• For Developers: Modular, reproducible pipelines that can be extended with better models, data, or deployment strategies.

---

🚧 Future Roadmap

1. Defect Detection → Multi-class defect classification, ONNX edge deployment
2. Predictive Maintenance → Attention/Transformer models, uncertainty quantification
3. Forecasting → Warehouse-specific models, time-series CV, real-time dashboards

---

🎤 Closing Note

PakIndustry 4.0 AI Suite is not just a hackathon project — it’s a vision for accessible AI in Pakistani manufacturing.