🏭 Industrial Predictive Maintenance & Fault Diagnostic Engine

An end-to-end machine learning system that predicts imminent industrial equipment failures before they happen and diagnoses their failure type, based on live sensor readings — to prevent unplanned downtime and reduce maintenance costs.

---

🎯 Problem & Solution

Problem	Unexpected equipment failures stop production lines and cost companies millions
Solution	A predictive model that determines when, which machine, and what type of failure — before it occurs

🧩 Task Definition

• Binary Classification — Will the machine fail? (Machine failure = 0/1)
• Multi-class Classification — What failure type? (TWF / HDF / PWF / OSF / RNF + co-occurring failures)

📊 Dataset

The AI4I 2020 Predictive Maintenance Dataset — 10,000 records of an industrial milling machine with 5 sensors:

Sensor	Unit
Air temperature	K
Process temperature	K
Rotational speed	rpm
Torque	Nm
Tool wear	min

---

🛠️ Methodology (CRISP-DM)

[Problem Framing] → [EDA] → [Feature Engineering] → [Modeling & Tuning] → [SHAP Interpretation] → [Packaging]

🔬 Exploratory Data Analysis

Sensor Distributions

Healthy vs Failed

Correlation Matrix

Failure Rate by Product Quality

Key Findings

1. Documented internal inconsistency in the original data between failure-type flags and the Machine failure column — reconciled with a clear policy.
2. Severe class imbalance (3.39% failures) — handled with scale_pos_weight + stratified cross-validation.
3. Tool wear is the leading driver of failures; product quality Type clearly affects failure rate.
4. Physics-based feature engineering derived from the known failure-generation rules: Power, Temp Diff, Overstrain.

---

🧪 Modeling

Three models compared fairly via 5-fold stratified cross-validation, then tuned with GridSearchCV:

Model	Role
Random Forest	Baseline
XGBoost	Advanced (best)
LightGBM	Advanced alternative

Chosen metric: Recall — because missing a real failure is far more costly than a false alarm in predictive maintenance.

📈 Results (Tuned XGBoost)

Metric	Value
Recall (failure detection)	88.2%
Precision	61.9%
F1-Score	72.7%
Accuracy	97.75%
ROC-AUC	98.6%

💡 Recall was deliberately prioritized over Precision: the cost of a missed failure (production downtime) vastly outweighs the cost of a false alarm (routine check).

Confusion Matrix

ROC Curve

SHAP Feature Importance

SHAP confirms that Tool Wear, followed by rotational speed drop and power/overstrain, are the strongest signals the model captures — consistent with the real physics of the failures.

---

✨ Key Features

• ✅ Full CRISP-DM lifecycle (6 phases) in clean, documented Python.
• ✅ Physics-based feature engineering (Power, Temp Diff, Overstrain).
• ✅ Imbalanced-data handling (scale_pos_weight) + stratified validation.
• ✅ Hyperparameter tuning with GridSearchCV.
• ✅ Model interpretability with SHAP — engineer-friendly, not a black box.
• ✅ Production-ready packaging: modular src/, config.py, saved .joblib model.

🚀 Quick Start

git clone <repo-url>
cd predictive-maintenance-engine
pip install -r requirements.txt
python src/train.py       # train & save the model
python src/evaluate.py    # evaluate the saved model

📁 Repository Structure

predictive-maintenance-engine/
├── README.md
├── LICENSE
├── requirements.txt
├── config.py                    # central configuration
├── data/
│   └── ai4i2020.csv
├── assets/                      # analysis figures
├── notebooks/                   # five phase notebooks
├── models/                      # saved model (.joblib)
└── src/
    ├── data_loader.py
    ├── features.py
    ├── preprocessing.py
    ├── train.py
    └── evaluate.py

📄 License

MIT License — see LICENSE.