Manufacturing · SteelData Scientist · Noodle.ai PQAINov 2019 – Apr 2023
Defect Diagnosis & Root Cause Rule Engine
End-to-end AI system diagnosing mill defects and extracting decision-tree rules to guide plant operators and production scheduling — replacing manual recipe tuning with deterministic, transparent guidance.
25–35%
Total Cost of Quality
Reduction achieved
>30%
Defect Rate
Direct drop across mills
>60%
RCA Time
Inspection & root-cause acceleration
// THE PROBLEM
Problem formulation
Defects in rolled steel are influenced by dozens of interacting chemical and thermodynamic variables. Operators historically relied on tribal knowledge and manual correlation to diagnose root causes, leading to slow, inconsistent resolution and costly scrap.
// DATA FOUNDATIONS
Data engineering & ingestion
- Ingested high-frequency IoT sensor telemetry from casting, hot strip mill, pickling, and cold rolling stages
- Cleaned and aligned time series from heterogeneous PLC and instrumentation protocols
- Built feature stores for chemical composition (Si, Mn, C), furnace zones, speeds, and pressures
- Label engineering from downstream inspection quality outcomes
// MODELING STRATEGY
Mathematical & algorithmic approach
- Random Forest ensemble for stage-wise defect prediction
- Neural Networks with Transfer Learning for mechanical property forecasting
- Decision-Tree rule induction with pruning to extract human-readable IF-THEN recipes
- Multi-stage mill classification to localize defects to the responsible process stage
// EXPLAINABILITY
Interpretability & governance
- SHAP Tree Explainer attributions ranking top defect drivers for each prediction
- Partial Dependence Plots (PDP) to quantify marginal effect of each process variable
- Optimal vs. alternate rule generation so operators understand trade-offs and constraints
// PRODUCTION
Deployment & integration
- Batch prediction pipeline feeding production scheduling systems
- Automated visual rule cards surfaced directly to shop-floor operator dashboards
- Scorecards tracking cost of quality, defect rate, and RCA velocity over time
// TECH STACK
PythonRandom ForestSHAPPDPTensorFlowScikit-learnPandasSQL