Overview • Key results • Architecture • Quickstart • Dataset • Methodology

---

Overview

A predictive maintenance system that forecasts industrial equipment failures from sensor data using gradient boosting, survival analysis for remaining useful life estimation, and SHAP-based explanations.

Unplanned equipment downtime costs industrial operations thousands of dollars per hour. This project builds a failure prediction pipeline that ingests sensor readings (temperature, vibration, pressure, RPM, power consumption) from 50 machines and predicts whether a failure will occur within the next 7 days. Four classifiers are trained, a Weibull survival model estimates remaining useful life, and a cost-based threshold optimizer balances unplanned downtime ($15K) against preventive maintenance ($1.5K).

Problem   →  Predicting equipment failures before they happen using sensor data
Solution  →  XGBoost with survival analysis, SHAP explanations, and cost-based threshold tuning
Impact    →  AUC 0.94, catches 91% of failures with optimized maintenance scheduling

---

Key results

Metric	Value
AUC-ROC	0.94
Recall (failures caught)	91%
Precision	78%
PR-AUC	0.76
Best model	XGBoost

---

Architecture

┌──────────────────┐    ┌──────────────────┐    ┌──────────────────┐
│  Sensor data     │───▶│  Feature          │───▶│  Rolling         │
│  generation      │    │  extraction       │    │  aggregations    │
└──────────────────┘    └──────────────────┘    └────────┬─────────┘
                                                         │
                          ┌──────────────────────────────┘
                          ▼
              ┌──────────────────────┐    ┌──────────────────────┐
              │  Model training      │───▶│  Survival analysis   │
              │  (4 classifiers)     │    │  (Weibull AFT / RUL) │
              └──────────────────────┘    └──────────┬───────────┘
                                                     │
                          ┌──────────────────────────┘
                          ▼
              ┌──────────────────────┐    ┌──────────────────────┐
              │  SHAP explanations   │───▶│  Maintenance         │
              │  + threshold tuning  │    │  dashboard           │
              └──────────────────────┘    └──────────────────────┘

Project structure

project_21_predictive_maintenance/
├── data/
│   ├── sensor_readings.csv              # Sensor dataset
│   └── generate_data.py                 # Synthetic data generator
├── src/
│   ├── __init__.py
│   ├── data_loader.py                   # Data generation and loading
│   └── model.py                         # Training, evaluation, SHAP, RUL
├── notebooks/
│   ├── 01_eda.ipynb                     # Exploratory data analysis
│   ├── 02_feature_engineering.ipynb     # Rolling features, interactions
│   ├── 03_modeling.ipynb                # Model training and CV
│   └── 04_evaluation.ipynb             # ROC, SHAP, cost analysis
├── app.py                               # Streamlit dashboard
├── requirements.txt
└── README.md

---

Quickstart

# Clone and navigate
git clone https://github.com/guydev42/calgary-data-portfolio.git
cd calgary-data-portfolio/project_21_predictive_maintenance

# Install dependencies
pip install -r requirements.txt

# Generate sensor data
python data/generate_data.py

# Launch dashboard
streamlit run app.py

---

Dataset

Property	Details
Source	Synthetic industrial sensor data
Readings	15,000
Machines	50
Failure rate	~8% (1,200 pre-failure readings)
Features	11 (temperature, vibration, pressure, rpm, power, rolling stats)
Target	failure_within_7days (binary)

---

Tech stack

---

Methodology

Sensor feature engineering

• Rolling 24h mean temperature and standard deviation of vibration
• Temperature-pressure ratio as an interaction feature
• Machine-level attributes: age, operating hours, maintenance history

Model training

• Four classifiers: Logistic Regression, Random Forest, XGBoost, Gradient Boosting
• 5-fold StratifiedKFold cross-validation
• Class imbalance handled via class_weight and scale_pos_weight
• Metrics: AUC-ROC, precision, recall, F1, PR-AUC

Survival analysis

• Weibull Accelerated Failure Time (AFT) model from lifelines
• Covariates: machine age, mean temperature, mean vibration
• Outputs remaining useful life (RUL) estimates per machine

SHAP explainability

• TreeExplainer for gradient boosting models
• Global feature importance via mean absolute SHAP values
• Waterfall plots for individual sensor reading explanations

Cost-optimized threshold

• Business cost model: FN cost ($15,000 unplanned downtime) vs FP cost ($1,500 preventive maintenance)
• Sweep thresholds from 0.05 to 0.95 to minimize total cost
• Achieves 91% recall with optimized maintenance scheduling

---

Acknowledgements

Built as part of the Calgary Data Portfolio.

---

Ola K.