Predictive Maintenance — Machine Failure Prediction

End-to-end machine-learning project that predicts machine failures from sensor readings, using the AI4I 2020 Predictive Maintenance dataset. The focus is a correct, leakage-free pipeline and honest evaluation on imbalanced data.

Overview

Given sensor measurements (temperatures, rotational speed, torque, tool wear) and the product quality type, the goal is to predict whether a machine will fail. Because failures are rare (~3.4% of records), the project treats this as an imbalanced binary classification problem and evaluates models with precision / recall, not accuracy.

Dataset

• AI4I 2020 Predictive Maintenance Dataset — UCI Machine Learning Repository (10,000 rows, 14 columns).
• Target: Machine failure (0 = healthy, 1 = failure). Failure rate ≈ 3.4%.
• Link: https://archive.ics.uci.edu/dataset/601/ai4i+2020+predictive+maintenance+dataset

Pipeline

1. Load & inspect — no missing values; checked class balance.
2. Clean — dropped identifier columns (UDI, Product ID) and encoded the categorical Type (L/M/H) with one-hot encoding.
3. Prevent data leakage — dropped the individual failure-mode columns (TWF, HDF, PWF, OSF, RNF). The target is derived from these, so using them as features would leak the answer. Only true sensor readings, known before a failure, are kept as features.
4. Split — stratified train / validation / test split (60 / 20 / 20) to keep the ~3.4% failure ratio in every partition.
5. Scale — StandardScaler fit on the training set only, then applied to validation and test.
6. Handle imbalance — RandomOverSampler applied to the training set only, so validation and test still reflect the real-world failure rate.
7. Model & compare — Logistic Regression (baseline), KNN, and a small neural network (Keras). Model selection on the validation set.
8. Final evaluation — the chosen model evaluated once on the held-out test set.

Models & Results (failure class, recall-focused)

Validation set:

Model	Precision	Recall	F1
Logistic Regression	0.15	0.82	0.26
KNN	0.36	0.62	0.45
Neural Network	0.31	0.85	0.46

Held-out test set (Neural Network):

Metric (failure class)	Value
Recall	0.88 (caught 60 of 68 failures)
Precision	0.34
F1	0.49

Test performance closely matches validation, indicating the model generalizes rather than overfits.

Key decision

This is predictive maintenance, where a missed failure is far more costly than a false alarm. Recall was therefore prioritized, and the neural network was chosen for the highest failure recall (0.88 on test, missing only 8 of 68 failures). If false alarms were the dominant cost, the higher-precision KNN would be preferable — the right model depends on the cost trade-off.

Tech stack

Python, pandas, NumPy, scikit-learn, imbalanced-learn, TensorFlow/Keras, Matplotlib, Seaborn.

How to run

1. Download ai4i2020.csv from the UCI link above.
2. Open the notebook (Google Colab or Jupyter) and upload the CSV.
3. Run the cells top to bottom.

Limitations & next steps

• The neural network still produces many false alarms (low precision); tuning the decision threshold or trying class-weighted loss could improve the balance.
• Only three models were compared; tree-based models (Random Forest, XGBoost) are common strong baselines for tabular data and would be a natural next step.
• Hyperparameters were kept simple; a systematic search could improve results.