Comparative ML study for industrial predictive maintenance — peer-reviewed research with reproducible notebooks and benchmark curves.
| Title | Change of Maintenance Strategy for Effective Maintenance |
| Authors | Pranav Singh Rathore, Priya Singh |
| Journal | Grenze International Journal of Engineering and Technology (GIJET) |
| Volume / Issue | Vol. 11, No. 1, January 2025 |
| Pages | 3702–3709 |
| Citation ID | 01.GIJET.11.1.239 |
Download PDF from Grenze (official publisher)
Journal issue listing: GIJET Vol. 11, Issue 1 (2025)
| Result | Detail |
|---|---|
| Best classifier | XGBoost — AUC = 0.98 on 10,000-point industrial sensor dataset |
| Baselines | Random Forest (AUC = 0.92), Decision Tree (AUC = 0.71) |
| Dataset | 6 sensor features: air temperature, process temperature, rotational speed, torque, tool wear, product quality |
| Strategies compared | Predictive · Condition-based · Preventive · Remedial maintenance |
| Operational impact | 58% reduction in maintenance hours by shifting from calendar-based to operating-time preventive maintenance |
Industrial equipment failure is costly: unplanned downtime, spare parts, and labour add up quickly. This study asks a practical question: which maintenance strategy works best when you have sensor data and ML models to guide decisions?
We benchmarked three classifiers (XGBoost, Random Forest, Decision Tree) on the AI4I 2020 predictive maintenance dataset to predict machine failure from multi-sensor readings. XGBoost consistently outperformed the baselines across classification thresholds.
Beyond model accuracy, we evaluated four maintenance strategies across cost, resource use, and machine availability. The key finding: moving from rigid calendar-based preventive maintenance to operating-time-based preventive maintenance cuts maintenance hours substantially while keeping reliability high — a result directly actionable for plant engineers.
This repository contains the Jupyter notebooks, dataset, and ROC curves used in the published paper.
| XGBoost (AUC ≈ 0.98) | Random Forest (AUC ≈ 0.92) | Decision Tree (AUC ≈ 0.71) |
|---|---|---|
 |
 |
 |
Maintenance-strategy/
├── README.md
└── predictive-analysis-main/
├── ai4i2020.csv # Industrial sensor dataset
├── predictive_xgb.ipynb # XGBoost classifier
├── predictive_rfc.ipynb # Random Forest classifier
├── predictive_decisionTree.ipynb # Decision Tree classifier
├── predictive_combind.ipynb # Combined analysis
└── curves/
├── combined.png
├── XGBCurve.png
├── RFCCurve.png
└── decisionTreeCurve.png
git clone https://github.com/pranav-singh-rathore/Maintenance-strategy.git
cd Maintenance-strategy/predictive-analysis-main
pip install pandas scikit-learn xgboost matplotlib jupyter
jupyter notebook predictive_xgb.ipynbOpen predictive_combind.ipynb for the full comparative analysis across all three models.
If you use this work, please cite:
@article{rathore2025maintenance,
title = {Change of Maintenance Strategy for Effective Maintenance},
author = {Rathore, Pranav Singh and Singh, Priya},
journal = {Grenze International Journal of Engineering and Technology},
volume = {11},
number = {1},
pages = {3702--3709},
year = {2025},
note = {Grenze ID: 01.GIJET.11.1.239}
}This research is extended in PredictIQ — edge-deployed LSTM anomaly detection with MQTT telemetry and Grafana dashboards.
Portfolio: pranavrathore.dev
Pranav Singh Rathore — GitHub · LinkedIn
Co-authored with Priya Singh — Professor, Delhi Technological University · LinkedIn