Predictive Maintenance (AI4I 2020)

Problem
This project explores how sensor data can be used to predict machine failures before they happen.
The idea is to help reduce unexpected downtime by giving maintenance teams enough time to plan repairs in advance.

• Why it matters: In maintenance, high precision reduces false alarms and avoids wasting technician time; recall at that precision shows how many real failures we can catch.

Data
AI4I 2020 Predictive Maintenance (UCI ML Repository). Synthetic but modeled after real industrial data.

• Download CSV: ai4i2020.csv from the UCI page and place it in data/ folder.
• Contains ~10,000 rows, 14 columns (temperature, rotational speed, torque, tool wear, and failure flags).
• Target: Machine failure (binary).

Failure Type Definitions

Code	Failure Mode	Description
TWF	Tool Wear Failure	Machine failed due to excessive tool wear
HDF	Heat Dissipation Failure	Failure caused by overheating or poor cooling
PWF	Power Failure	Electrical or power system failure
OSF	Overstrain Failure	Mechanical stress or overload failure
RNF	Random Failure	Unclassified or random event leading to failure

Data quality note:

• Verified that the overall Machine failure flag aligns with these five subtype failures (TWF–RNF).
• A small number of rows show minor inconsistencies; treated as negligible label noise.
• Leakage control: The five subtype flags (TWF, HDF, PWF, OSF, RNF) are dropped before modeling (used only to validate the target).

METHOD (plan)

1. EDA: distributions, correlations, label balance, leakage checks.
2. Target framing: failure in next 7 days (binary) or failure_now as provided.
3. Baselines: majority class; logistic regression with class weights.
4. Models: tree-based (RandomForest, GradientBoosting) and calibrated logistic.
5. Validation: stratified CV and temporal split sensitivity check.
6. Metrics: PR-AUC, ROC-AUC, Recall at fixed Precision (operating point).
7. Explainability: SHAP/feature importances.
8. Business impact: translate operating point/results to avoid downtime.

Method (actual)

1. EDA: Distributions, class balance, leakage check (drop TWF/HDF/PWF/OSF/RNF).
2. Feature engineering: Extract Product_Category from Product ID; remove Product ID.
3. Split & preprocessing: Stratified split; version-safe OneHotEncoder;
   • Logistic pipeline: scale numerics + OHE categoricals
   • Tree pipeline: passthrough numerics + OHE categoricals
4. Models:
   • Calibrated Logistic Regression (class_weight="balanced" + CalibratedClassifierCV)
   • Random Forest baseline (n_estimators=400, class_weight="balanced")
5. Evaluation: PR-AUC, ROC-AUC, and Recall @ 90% Precision via a custom utility.
6. Interpretation: Top-20 RF feature importances (Torque, Rotational speed, Tool wear typically dominate).
7. Business framing: Choose operating points that minimize false alarms while catching meaningful failures.

Results

Model	PR-AUC	Recall @ 90% Precision	Notes
Random Forest	see assets/metrics_summary.csv	see assets/metrics_summary.csv	Strong non-linear baseline
Calibrated Logistic	see assets/metrics_summary.csv	see assets/metrics_summary.csv	Interpretable, calibrated scores

Detailed values are in assets/metrics_summary.csv and the comparison (uplift) is in assets/operating_point_uplift.csv.

Key visuals

Business Impact
At 90% precision, the model detects 35.3% of true failures, versus 12.9% for the baseline. This is an improvement of 22.4 percentage points, yielding fewer false alarms and more predictable maintenance scheduling.

How to run

# (optional) python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# Add data
mkdir -p data
# Place ai4i2020.csv into data/

# Open and run the notebook
jupyter lab  # or: jupyter notebook
# Run: notebook/predictive_maintenance.ipynb end-to-end

The notebook saves curated outputs directly to assets/:

• pr_curve_comparison.png
• feature_importances_rf.png
• metrics_summary.csv
• operating_point_uplift.csv

---

Data source: UCI Machine Learning Repository – AI4I 2020 Predictive Maintenance Dataset