A full end-to-end AI-assisted SPC system that detects out-of-spec batches, process drifts, and abnormal variance in batch manufacturing using classical statistics and machine learning.
Built a complete SPC pipeline for batch manufacturing that combines classical statistical process control with ML anomaly detection, running across 200 simulated batches with intentional drift and variance anomalies baked in.
| Layer | Tool |
|---|---|
| Data Generation | Python, NumPy, Pandas |
| Control Charts | Matplotlib, SPC custom classes |
| Capability Analysis | SciPy, NumPy |
| WE Rules Engine | Custom Python (Z-score based) |
| ML Anomaly Detection | Scikit-learn (IsolationForest, OneClassSVM) |
| Change-Point Detection | Ruptures (PELT algorithm) |
| Web App | Streamlit |
| Version Control | Git & GitHub |
- 200 batch records generated and processed
- Cpk: 0.844 — measurement_1 NOT CAPABLE (PPM: 17,482)
- Cpk: 0.975 — measurement_2 NOT CAPABLE (PPM: 2,523)
- 4 WE Rules triggered on measurement_1 including Critical R1
- 4 consensus ML anomalies confirmed by both IF and OC-SVM
- 6 phases fully automated — one command runs the full pipeline
- Simulated drift (batches 80–100) caught by CUSUM and Rule 2
- Variance spike (batches 140–160) caught by I-MR and Rule 1
- ML consensus reduced false positives from 22 (SVM) to 4 (consensus)
- Cp > Cpk gap on both measurements indicates mean offset from target
| Metric | Before SPC System | After SPC System |
|---|---|---|
| Defect detection | End-of-shift manual review | Real-time automated alerts |
| Analysis time | 2–4 hours per shift | 30 seconds per run |
| Defect rate (M1) | 17,482 PPM undetected | Flagged immediately |
| Defect rate (M2) | 2,523 PPM undetected | Flagged immediately |
| Process drift | Invisible until product fails | Caught by CUSUM in batches 80–100 |
| Variance spike | Found during quality audit | Caught by I-MR in batches 140–160 |
| False positive rate | N/A (no system) | Reduced 80% via ML consensus |
| Compliance evidence | Manual logbooks | Automated JSON audit trail |
| Cost Driver | Calculation | Annual Estimate |
|---|---|---|
| Defect reduction (M1) | 17,482 PPM × $50/defect × 4,380 batches/yr | ~$3.78M saved |
| Defect reduction (M2) | 2,523 PPM × $50/defect × 4,380 batches/yr | ~$552K saved |
| Labor saving | 3.5 hrs/shift × 3 shifts × $40/hr × 365 days | ~$153K saved |
| Recall prevention | 1 recall avoided × $10M average cost | ~$10M protected |
- ISO 9001 — Quality management system evidence
- GMP (Good Manufacturing Practice) — Full audit trail
- FDA 21 CFR Part 11 — Electronic records with timestamps
- IATF 16949 — Automotive SPC requirements met
spc_system/
├── core/
│ ├── schema.py → Data contracts
│ ├── ingestion.py → CSV loader + validator
│ ├── preprocessor.py → Cleaning + OOS flagging
│ ├── charts/ → I-MR, X̄-R, CUSUM, EWMA
│ ├── capability/ → Cp, Cpk, Pp, Ppk dashboard
│ ├── rules/ → 8 Western Electric rules
│ ├── anomaly/ → Drift, variance, ML detection
│ └── reporting/ → Alert engine + HTML report
├── sample_data/ → Generated batch CSV
├── tests/ → Phase test scripts
├── docs/ → Screenshots
└── requirements.txt
| Component | Type |
|---|---|
| Control Charts, Cpk, WE Rules | Automated classical statistics |
| Isolation Forest, One-Class SVM | Genuine unsupervised ML |
| PELT Change-point Detection | Adaptive algorithm |
| Full pipeline orchestration | Automated end-to-end |
This project was built and designed by beebzy-droid.
Licensed under the MIT License — you are free to use this code but must include attribution linking back to this repository.
© 2026 beebzy-droid 🔗 https://github.com/beebzy-droid/spc-system