Predictive Maintenance: CNC Machine Failure Detection

Project Overview

This project focuses on Predictive Maintenance for CNC machines using a dataset of 10,000 records. The goal is to predict machine failures before they happen by analyzing sensor data (Temperature, Torque, Rotational Speed, etc.).

Moving from "Reactive Maintenance" to allows factories to reduce downtime and save costs.

---

Methodology & Techniques

1. Handling Multicollinearity:
   • Detected a high correlation (0.88) between Air_Temp and Process_Temp.
   • Removed Air_Temp to reduce redundancy and improve model stability.
2. Data Preprocessing:
   • Used StandardScaler to normalize features with different units (e.g., RPM vs. Torque).
3. Handling Imbalanced Data (Crucial Step):
   • The dataset was highly imbalanced (96% Normal vs. 4% Failure).
   • Initial models failed to detect failures (Recall was ~30%).
   • Applied SMOTE (Synthetic Minority Over-sampling Technique) to balance the classes. This significantly improved the model's ability to learn failure patterns.
4. Evaluation Metrics:
   • Evaluated models using Accuracy, Recall (Sensitivity), and F1-Score to ensure a balanced performance.

---

Model Results

The Random Forest Classifier achieved the best overall performance.

Model	Accuracy	Recall	F1-Score	Verdict
Logistic Regression (Base)	97.4%	~30%	Low	Missed most failures
Logistic Regression (SMOTE)	80.0%	79%	Moderate	Good detection, lower precision
Decision Tree	93.8%	High	Moderate	Prone to overfitting
Random Forest (Final)	94.2%	High & Stable	Balanced	Best Model

---

Visual Analysis

1. Confusion Matrix (Random Forest)

The model successfully identifies the majority of machine failures while keeping false alarms low.

2. Feature Importance

According to the Random Forest analysis, Torque and Rotational Speed are the most critical indicators of a potential failure.

3. Model Performance (ROC Curve)

The model shows excellent discrimination with an AUC score of 0.95.

---

---