Industrial AI: Predictive Maintenance for Motor Bearings

This project demonstrates an end-to-end Condition Monitoring and Fault Diagnosis pipeline. It uses signal processing and statistical feature extraction to detect early-stage motor failures using the CWRU Bearing Dataset.

1. Exploratory Data Analysis (EDA)

We analyze vibration signals in the time domain. By comparing a healthy motor with one having an inner race fault, we can visually identify high-amplitude impacts caused by the defect.

2. AI Feature Extraction & Comparison

The raw vibration data is converted into statistical "features" that a Machine Learning model can understand. We specifically focus on:

• RMS (Root Mean Square): Represents the overall energy of the vibration.
• Kurtosis: Measures the "peakiness" of the signal to detect sharp impacts.

The bar charts below clearly show that the Faulty Motor has significantly higher energy and impact levels compared to the Healthy Motor.

3. Diagnostic Results

Based on our AI-driven analysis, the system successfully differentiates between motor states:

Feature	Healthy Motor	Faulty Motor	Change
RMS	0.0738	0.2915	+295% Increase
Kurtosis	2.7642	5.3959	+95% Increase

4. Automated Decision Logic

The system includes a rule-based classifier:

• Threshold: 0.15 RMS
• Status: If RMS > Threshold -> [!!!] FAULT DETECTED
• This logic enables 24/7 automated monitoring without human intervention.

5. Hardware Implementation (C/Embedded)

To bridge the gap between simulation and real-world industrial application, the diagnostic algorithms have been ported to C code. This implementation is optimized for microcontrollers such as STM32, enabling real-time edge processing.

Embedded Module Breakdown:

• fault_detection.c: Contains the core mathematical engine for calculating RMS and Kurtosis from raw sensor buffers.
• main_diagnostic.c: Implements the high-level application logic. It integrates with the STM32 HAL (Hardware Abstraction Layer) to trigger physical alerts (e.g., toggling a Red LED on GPIO PA5) when an anomaly is detected.

Transition from Simulation to Reality

By translating MATLAB models into efficient C functions, this project demonstrates a complete workflow for Industrial AI:

1. Analyze (MATLAB) -> 2. Validate (Simulation) -> 3. Deploy (Embedded C).

6. Statistical Validation: Why AI?

To prove the necessity of AI, we analyzed the raw data distribution versus the extracted feature space.

Analysis Breakdown:

• Data Overlap (Left): Raw vibration signals show significant overlap between healthy and faulty states. Traditional amplitude monitoring would result in high false-alarm rates.
• Feature Space (Right): By mapping data into the RMS-Kurtosis plane, we achieve clear linear separability. The black dashed line represents our AI Decision Boundary.

Approach	Reliability	Efficiency
Raw Signal Monitoring	Low (High Overlap)	Manual / Inefficient
AI Feature Engineering	High (Clear Clusters)	Automated / Precise

7. Machine Learning & AI Reliability

To move beyond basic thresholds, we implemented a supervised learning workflow using the MATLAB Classification Learner app. This ensures the system can autonomously adapt to different motor conditions with high precision.

Feature Space & Methodology

As demonstrated in the Statistical Validation (Section 6), our extracted features (RMS and Kurtosis) provide perfect Linear Separability. This distinct clustering allowed the machine learning models to define a clear decision boundary, ensuring 100% reliable classification without manual tuning.

Data Preparation (generate_training_data.m)

The training dataset was generated using a custom script that segments the raw vibration signals into labeled observations. This process is crucial for "teaching" the AI the difference between motor states.

• Method: 120 samples (60 Healthy / 60 Faulty) were extracted.
• Script: Detailed logic is available in generate_training_data.m.

Model Performance & Validation

We trained multiple classifiers, including SVM (Support Vector Machines), KNN, and Decision Trees. Every model achieved 100% Validation Accuracy.

Confusion Matrix Analysis:

The following matrices confirm that our "AI Brain" makes zero mistakes:

Sample Counts	True Positive Rates	Precision / FDR

• Zero False Positives: The model never incorrectly flags a healthy motor as faulty.
• Zero False Negatives: Every single fault is detected instantly.

---

Authors

Mevlut Korkmaz Electrical and Electronics Engineering Student at Ege University

Ali Özdemir Electrical and Electronics Engineering Student at Ege University