Predictive Maintenance Pipeline

Problem

Sensor streams for maintenance are uneven, and the failure class is small. I needed a comparison that would not hide the rare cases.

What I built

I derived 22 features from 8 sensor streams and added automated data quality checks. I benchmarked 5 models with stratified cross-validation: logistic regression, random forest, gradient boosting, XGBoost, and SVM. The choice used F1 under class imbalance.

Architecture

Training is deterministic. Evaluation reports are generated automatically. pytest covers 14 tests. Docker starts the pipeline with one command.

Results

• 22 derived features from 8 sensor streams
• 5 models compared with stratified cross-validation
• Selection by F1 under class imbalance
• 14 pytest tests and automated evaluation reports

What I learned

I picked the model by F1 because the classes were imbalanced.

Stack

The benchmark included XGBoost. Docker packages the run. pytest locks the checks.