Problem
Maintenance teams need to prioritize limited inspection capacity without missing most of the rare machine-failure cases.
selected case study
Smart Factory App
A predictive-maintenance analytics case study on 10,000 AI4I records, combining EDA, engineered sensor signals, evaluated failure-risk scoring, DuckDB SQL, and a ranked maintenance queue — plus a second, genuine RUL-regression case study on NASA C-MAPSS turbofan data.
Quick technical review
The case study in four points.
Architecture
The workflow combines AI4I EDA, cross-sensor feature engineering, benchmark models, tuned HistGradientBoosting evaluation, SHAP explainability, DuckDB SQL artifacts, and a Dockerized Flask review surface with CI-driven training, drift checks, and…
Read complete note
The workflow combines AI4I EDA, cross-sensor feature engineering, benchmark models, tuned HistGradientBoosting evaluation, SHAP explainability, DuckDB SQL artifacts, and a Dockerized Flask review surface with CI-driven training, drift checks, and Playwright E2E tests.
Results
10,000 AI4I records
Limitations & scope
The 94.1% queue capture is measured on a 2,000-row holdout that contains 68 failures; across five splits PR-AUC is 0.880 plus or minus 0.032, so the single-split 0.9019 sits at the high end. The AI4I rows are simulated, the C-MAPSS result covers the…
Read complete note
The 94.1% queue capture is measured on a 2,000-row holdout that contains 68 failures; across five splits PR-AUC is 0.880 plus or minus 0.032, so the single-split 0.9019 sits at the high end. The AI4I rows are simulated, the C-MAPSS result covers the FD001 subset only, the live dashboard’s RUL tile is a simple heuristic derived from the AI4I risk score, and the drift check compares training rows against a synthetic shifted batch rather than a production sensor stream.
Summaries reuse the project’s documented evidence. Review the complete case study for evaluation boundaries.
Evidence-led visual review
Explore the actual work.
Actual project screenshots and repository figures are shown with source provenance.
Dated demo screenshot
Public demonstration screenshot dated 2026-10-14; does not establish that the demo is currently online.
Open published source10,000 AI4I records
0.9874 ROC-AUC / 0.9019 PR-AUC
0.8293 F1 · C-MAPSS RUL RMSE 18.05 cycles
Top 10% queue captures 94.1% of failures at 9.4x lift
Project proof
Verified external evidence
Last verified 2026-10-05
Problem and approach
Problem
Maintenance teams need to prioritize limited inspection capacity without missing most of the rare machine-failure cases.
Architecture
The workflow combines AI4I EDA, cross-sensor feature engineering, benchmark models, tuned HistGradientBoosting evaluation, SHAP explainability, DuckDB SQL artifacts, and a Dockerized Flask review surface with CI-driven training, drift checks, and Playwright E2E tests.
Workflow logic
Failure probabilities rank assets for review, allowing the project to compare inspection budget, failure capture, and lift over random selection.
What is not claimed
The 94.1% queue capture is measured on a 2,000-row holdout that contains 68 failures; across five splits PR-AUC is 0.880 plus or minus 0.032, so the single-split 0.9019 sits at the high end. The AI4I rows are simulated, the C-MAPSS result covers the FD001 subset only, the live dashboard’s RUL tile is a simple heuristic derived from the AI4I risk score, and the drift check compares training rows against a synthetic shifted batch rather than a production sensor stream.
Why it matters
It demonstrates how industrial context, imbalanced-classification evaluation, and operational queue design can work in one applied data-science case study — and pairs it with a second case study (C-MAPSS) that shows the same rigor applied where grouped, time-aware validation is actually possible.
Architecture and analysis
- Benchmarks failure-risk models on 10,000 AI4I records with a 3.39% failure rate.
- Engineers mechanical-load, thermal-stress, and tool-wear-to-load signals across sensor inputs.
- Connects model scores to a risk-ranked maintenance queue and inspection budget.
- Adds a genuine RUL-regression case study on NASA C-MAPSS turbofan data, GroupKFold-validated and scored on NASA’s own asymmetric PHM08 metric.
Testing and validation
- EDA, model-comparison, feature-importance, review-queue, and drift-report artifacts
- DuckDB SQL analysis over versioned model outputs
- Automated route, artifact-contract, metric-threshold, and asset validation, plus Playwright E2E browser tests
Operational value
- Achieves 0.9874 ROC-AUC, 0.9019 PR-AUC, and 0.8293 F1 with a tuned HistGradientBoosting model.
- The top 10% review queue captures 94.1% of failures at 9.4x lift over random selection.
- The companion C-MAPSS RUL model cuts RMSE 63.8% and PHM08 score 99.5% versus a naive median-RUL baseline, reaching 18.05 cycles RMSE.
- Packages EDA, DuckDB SQL evidence, model evaluation, drift detection, and maintenance-priority logic into one reviewable, live-deployed case study.
How to inspect
- Start with the README, model comparison, and review-queue artifacts.
- Inspect the engineered sensor features and DuckDB SQL outputs before reviewing the Flask surface.
- Check the RUL case study for the same rigor applied where grouped/time-aware validation is actually possible.
- Treat the public AI4I and C-MAPSS results as benchmark evidence, not a live production deployment claim.