Predictive maintenance: why I ranked machines instead of just predicting failures
I came to data science from industrial automation, so for my predictive maintenance project I started from a maintenance team's question: which machines should we inspect first?
On the UCI AI4I 2020 dataset, a HistGradientBoosting classifier scores each machine's failure risk, and the scores become a ranked work queue in a small Flask dashboard.
A few things I learned:
- Failures are rare, so ROC-AUC looked great (0.987) while PR-AUC (0.902) was the more honest number.
- Across five different splits, PR-AUC averaged 0.880 ± 0.032, so one split overstates how stable the model is.
- Reviewing the top 10% of machines catches 94.1% of the failures in the holdout set. For a maintenance team, that queue is more useful than a raw probability.
I also added a separate remaining-useful-life study on NASA C-MAPSS FD001: RMSE 18.05 cycles against 49.82 for a naive baseline. The sensor data in the live dashboard is simulated, and the repo says so.
How do you turn risk scores into decisions in your work: a fixed top-k queue, a threshold, or a cost-based cutoff?
Demo: smart-factory-app.onrender.com
Code: github.com/senanurcetin/smart-factory-app






Comments
0 comments