⚙ Predictive Maintenance — A Failure Analysis Approach

Identifying Operational Thresholds for Predictive Maintenance

⚡ Why wait for machines to fail when you can see it coming?
This project applies failure analysis to spot early warning signs of mechanical failure in a milling process — helping prevent downtime, extend machine life, and reduce maintenance costs.

---

📌 Overview

This project demonstrates how data-driven analysis can reduce unplanned downtime by detecting operational thresholds that signal machine failure risks.
Using a synthetic dataset simulating a milling process, I performed Exploratory Data Analysis (EDA), correlation studies, and threshold identification to support predictive maintenance strategies.

---

📂 Dataset Description

• Source: Predictive Maintenance Dataset (AI4I 2020)
• Size: 10,000 rows × multiple operational and failure-related variables
• Key Variables:
  • Air temperature [K]
  • Process temperature [K]
  • Torque [Nm]
  • Tool wear [min]
  • Rotational speed [rpm]
  • Machine failure (binary)
  • Failure types: TWF, HDF, PWF, OSF, RNF
• Note: This is synthetic data, generated using simulation models to mimic real-world machine behavior.

---

🎯 Research Goal

Main Question:
How can operational thresholds be detected to predict and prevent machine failures effectively?

---

🛠 Steps Taken

1. Data Cleaning
   • Removed anomalies, renamed columns, encoded categories.
2. EDA
   • Explored variable distributions, relationships, and failure patterns.
3. Visualization
   • Boxplots for threshold detection
   • Histograms for distribution analysis
   • Heatmaps for correlation checks
   • Pairplots for variable interactions
4. SQL Analysis
   • Wrote queries to extract failure counts, success rates, and variable relationships directly from the database.

---

📊 Key Findings

• Tool Wear > 175 min significantly increases the probability of machine failure.
• Torque > 60 Nm combined with high tool wear is a strong risk factor.
• Process Temperature > 310 K often coincides with failures.
• Air temperature and process temperature are highly correlated, influencing operational stress levels.

---

🔄 Update 1.0 — First Machine Learning Model 🤖⚙️

Building on the exploratory and threshold analysis, I applied my first machine learning model — K-Nearest Neighbors (KNN) — to the predictive maintenance dataset.

---

⚙️ Approach

• ✨ Features (X): Torque, Tool Wear, Rotational Speed, Air Temperature, Process Temperature
• 🎯 Target (y): Machine Failure (binary: 0 = success, 1 = failure)
• 📏 Preprocessing: Applied StandardScaler for feature scaling + train/test split
• 🧮 Model: Implemented baseline KNN (k=5) for classification

---

📊 Results

• ⚖️ Class Imbalance: 96.6% machine success vs 3.4% machine failure
• ✅ Accuracy (baseline): ~96% across k = 1–25
• 🧾 Confusion Matrix (k=5):
  • 🟩 True Negatives (TN): 1914
  • 🟥 False Positives (FP): 11
  • 🟦 False Negatives (FN): 60
  • 🟨 True Positives (TP): 15
• 📉 Recall (failures): 0.20 → Only 20% of actual failures detected
• ⚡ Insight: High accuracy is misleading in imbalanced datasets; the model misses most failures

---

🧾 Key Takeaways

• 🎯 In predictive maintenance, recall matters more than accuracy
• 🚨 False negatives (missed failures) are riskier than false positives (extra checks)
• 📊 Accuracy alone ≠ success when dealing with imbalance

---

🔄 Update 2.0 — Handling Imbalanced Data & Model Optimization ⚖️🤖

The dataset used in this project was highly imbalanced: only ~3% of machines experienced failures.
This made it difficult for baseline models to detect breakdowns, despite high overall accuracy.

---

🛠 What I implemented

• Baseline Models: Logistic Regression, Random Forest, XGBoost
• Data Resampling: Applied SMOTE to balance failure vs. success cases
• Class Weighting: Used scale_pos_weight in XGBoost to address imbalance without oversampling
• Evaluation Metrics: Focused on Precision, Recall, and F1-score (catching failures is more important than accuracy).

---

📊 Key Results

• Logistic Regression → High accuracy but poor recall (missed most failures).
• Random Forest & XGBoost → Performed better but still missed ~35–40% of failures.
• SMOTE → Boosted recall (~80%) but lowered precision (more false alarms).
• Best Trade-Off: XGBoost with class weighting →
  • Recall: ~78%
  • Precision: ~64%
  • F1 Score: ~0.70
  • Best balance between catching failures and limiting false alarms.

---

📈 Visualization

To compare models, I built a Plotly leaderboard showing Precision, Recall, and F1 across all approaches.
This made it easy to visualize trade-offs and identify the best-performing models.

---

🚀 Why this matters

In predictive maintenance, missing a failure can be costly.
Class-weighted XGBoost provided the most reliable solution, detecting failures effectively while keeping false alarms manageable.

---

🚀 Next Steps

1. Model Explainability

   • Use SHAP or LIME to explain why models predict a machine failure.
   • This makes the results more trustworthy and actionable for engineers.

2. Time-Series & Temporal Patterns

   • Extend the analysis to capture time-based degradation trends (e.g., tool wear over cycles).
   • Try models like LSTM or Prophet for forecasting failure risk.

3. Feature Optimization

   • Investigate interaction features (e.g., torque × rotational speed).
   • Run feature importance analysis to refine the dataset further.

4. Advanced Imbalance Handling

   • Try ensemble resampling methods (SMOTE + Tomek Links, ADASYN).
   • Compare with cost-sensitive learning beyond just XGBoost’s scale_pos_weight.

5. Deployment Pipeline

   • Build a real-time prediction API with FastAPI/Flask.
   • Simulate how predictions could integrate into a monitoring dashboard for factory use.

---

💻 Reproduction Guide

Requirements:
pandas, matplotlib, seaborn, numpy, sqlite3

Run Instructions:

1. Clone this repository

   git clone https://github.com/sergie-o/Predictive-Maintenance-Project.git

2. Navigate to the project folder

    cd Predictive-Maintenance-Project

3. Open the Jupyter Notebook

• If you use Jupyter Notebook:

  jupyter notebook "Project_predictive_maintenance.ipynb"

• Or, open it in VSCode by double-clicking the file or using:

   code "Project_predictive_maintenance.ipynb"

4. Ensure the dataset is in the correct location

• The file ai4i2020.csv must be in the same directory as the notebook.

5. Run all cells

• Select Cell > Run All in Jupyter Notebook or VSCode to reproduce the analysis.

🚀 Next Steps

• Predict Failures Before They Happen – Using the identified operational thresholds, models can be trained to recognize early warning signs and flag machines before they reach critical failure points.
• Keep Machines in the Safe Zone – By continuously monitoring operational variables (e.g., torque, tool wear, process temperature), predictive systems can regulate performance to remain within safe operational ranges, reducing stress on components.
• Enable Automated Preventive Actions – Integrating these predictions with control s ystems could automatically trigger adjustments, slowdowns, or maintenance requests when a threshold is exceeded.

📁 Repository Structure

predictive-maintenance-failure-analysis/
│
├── data/                               # Raw and cleaned datasets
│   ├── cleaned_data.db                  # SQLite database with processed data
│   └── ai4i2020.csv                     # Original dataset
│
├── notebooks/                          # Jupyter notebooks for analysis
│   └── predictive_maintenance_analysis.ipynb
│
├── sql/                                # SQL queries for analysis
│   └── predictive_maintenance_queries.sql
│
├── visuals/                            # Generated plots and charts
│
├── README.md                           # Project documentation
└── requirements.txt                    # Dependencies list