End-to-end predictive maintenance and operational analytics platform for industrial mining equipment. Demonstrates a production-style ML workflow: telemetry → daily KPIs → downtime-risk scoring (scikit-learn) → executive / operational / ML / data-quality dashboards → interactive what-if scenarios → cloud deployment.
Built on the NASA C-MAPSS turbofan benchmark recast into mining ops framing (haul trucks, excavators, drills).
🚀 Live demo: mining-ops-analytics.streamlit.app
Streamlit Cloud cold-starts on clean containers. If a page shows "missing artifacts," click Generate demo data now on the Home page to bootstrap the pipeline inside the container.
📦 Zero external downloads: the NASA C-MAPSS benchmark data is bundled under
data/raw/nasa_cmapss/CMAPSSData/— clone, install, and the pipeline runs end-to-end.
| Page | What it shows |
|---|---|
| Home | Project overview + one-click demo data bootstrap for cloud cold-starts |
| Executive Dashboard | 5 tabs: Overview · OEE & Reliability (with 85% world-class benchmark) · MTBF/MTTR · Downtime Pareto (80/20) · Cost Impact (configurable $/hr × catch-rate sliders) |
| Operations Drilldown | Risk heatmap (site × equipment-type matrix, mean predicted risk) + utilization/downtime trends + site comparison + filterable asset tables |
| Model Performance | 4 tabs: ROC & Threshold tuning · Calibration (reliability diagram + Brier score) · Lift & Gains (cumulative gains, lift curve, decile insights) · SHAP Explainability (global importance + per-asset waterfall) |
| Data Quality | Kolmogorov-Smirnov drift detection (configurable date split, per-feature KS stat / p-value / mean shift, distribution overlay) + missingness + row volume + range checks |
| What-If Simulator | Interactive sliders per model feature → live recomputed risk score + fleet percentile + on-demand SHAP local explanation. Sensitivity testing and intervention modeling in one tool. |
Cost Impact analysis (Executive Dashboard → Cost Impact tab): $450M annualized cost → 
Downtime Pareto (Executive Dashboard → Downtime Pareto tab): classic 80/20 — 74 of 100 assets account for 80% of downtime, with cumulative threshold line.
Per-point SHAP explanation (What-If Simulator): for any slider configuration, on-demand local explanation showing exactly which feature values push risk up (red) or down (blue) from baseline.
This is a portfolio project, not a production system. Disclosures:
- Synthetic data. Uses NASA C-MAPSS turbofan benchmark recast into mining ops framing. Physics is realistic; data is not from real mining operations.
- Model performance. ROC AUC 0.639 on the evaluation set. Realistic for raw C-MAPSS aggregates without deep feature engineering — and leakage-free: the training pipeline auto-detected and removed a leaky
high_sevfeature that was inflating scores to artificial perfection. - Drift is expected behavior. All seven features flag KS-significant drift between reference and current windows. This is correct — C-MAPSS simulates engine degradation, so distributions genuinely shift over time as components wear. The detector is catching the wear signal, not a pipeline bug.
- Cost impact is illustrative. The $/hour and model catch-rate sliders on the Executive Dashboard are configurable assumptions. The annualized cost-avoidance figure shows the calculation pattern, not a defensible business case for a real operator.
- No real-time ingestion. The pipeline is batch — telemetry generated, KPIs computed, model trained all upfront. Production would need streaming ingestion, incremental retraining, alert routing, ops integration.
flowchart LR
A[Telemetry + Maintenance Data] --> B[ETL + Feature Engineering]
B --> C[Daily KPI Table - Parquet]
B --> D[Model Training - scikit-learn]
C --> E[KPI Snapshot CSV]
D --> F[Downtime Risk Model - joblib]
C --> G[Streamlit Dashboards]
E --> G
F --> G
G --> H[Streamlit Cloud]
Streamlit Cloud runs on clean containers. The app includes a bootstrap mechanism that:
- Detects missing pipeline artifacts on cold start
- Generates demo artifacts on demand via a single button
- Prevents broken pages after redeploy
- Python 3.11 (pinned via
runtime.txt) - Streamlit — multi-page UI framework
- scikit-learn 1.4.2 — model training (Pipeline + classifier with leakage detection)
- Classifier — scikit-learn binary classifier wrapped in a Pipeline (see
src/miningops/train.pyfor algorithm choice and hyperparameters) - SHAP 0.46.0 — explainability (Kernel + Tree explainers)
- scipy.stats — Kolmogorov-Smirnov 2-sample test for drift detection
- Plotly — interactive charts (heatmaps, distributions, waterfalls)
- Pandas / NumPy / PyArrow — data layer (Parquet)
- Joblib — model serialization
mining-operations-analytics-platform/
├── streamlit_app/
│ ├── Home.py # Entry + cloud bootstrap
│ ├── bootstrap.py # Demo data regeneration logic
│ └── pages/
│ ├── 1_Executive_Dashboard.py # OEE, Pareto, cost impact
│ ├── 2_Operations_Drilldown.py # Risk heatmap + drilldowns
│ ├── 3_Model_Performance.py # ROC, calibration, lift, SHAP
│ ├── 4_Data_Quality.py # Drift detection + DQ checks
│ └── 5_What_If_Simulator.py # Interactive sensitivity tool
├── src/miningops/
│ ├── generate_data.py # Synthetic telemetry from C-MAPSS
│ ├── kpis.py # Daily aggregation + features
│ └── train.py # Model training + leakage detection
├── data/
│ ├── raw/nasa_cmapss/CMAPSSData/ # Bundled NASA benchmark dataset
│ └── processed/ # Pipeline outputs (gitignored)
├── docs/screenshots/ # Dashboard captures
├── reports/ # Published analysis reports + figures
├── .github/workflows/ # CI smoke tests
├── tests/ # Unit + integration tests
├── requirements.txt # Pinned dependencies
├── runtime.txt # Python 3.11
├── .env.example # Environment variables template
├── LICENSE # MIT
└── README.md
git clone https://github.com/fahadamjad009/mining-operations-analytics-platform.git
cd mining-operations-analytics-platform
pip install -r requirements.txt
# Generate demo data + train model
python -m src.miningops.generate_data
python -m src.miningops.kpis
python -m src.miningops.train
# Launch
streamlit run streamlit_app/Home.pyApp opens at http://localhost:8501.
- Tier 1 — Executive Dashboard (OEE, MTBF/MTTR, Pareto, cost impact)
- Tier 2 — Model Performance (ROC, calibration, lift & gains, SHAP)
- Tier 3 — Risk heatmap (Operations Drilldown)
- Tier 3 — KS-test drift detection (Data Quality)
- Tier 3 — What-If Simulator with live SHAP local explanations
- Refresh screenshots to reflect Tier 2/3 features
- Real-time telemetry ingestion (Kafka or polling)
- Drift-triggered automated retraining loop
- Alert routing (email / Slack / PagerDuty)
- MLflow experiment tracking integration
Pre-Tier-1 baseline preserved as a known-working snapshot:
- Tag:
savepoint-streamlit-stable-2026-02-22
MIT — see LICENSE.