Predictive Maintenance Using Bearing Fault Diagnosis

Overview: This project focuses on predictive maintenance using machine learning techniques for bearing fault diagnosis across multiple industrial datasets. The study combines feature engineering, exploratory data analysis (EDA), and classification models to identify bearing health conditions and evaluate cross-domain generalization.

The primary objective is to develop a scalable and reliable fault diagnosis pipeline capable of identifying bearing conditions using statistical vibration features extracted from sensor signals.

Problem Statement:

Industrial rotating machinery relies heavily on rolling element bearings for efficient operation. Unexpected bearing failures can result in: Production downtime Increased maintenance costs Equipment damage Safety risks

Traditional maintenance approaches such as reactive and preventive maintenance are often inefficient because they either respond after a failure occurs or rely on fixed maintenance schedules.

This project aims to develop an intelligent predictive maintenance framework capable of detecting bearing faults at an early stage through vibration signal analysis and machine learning.

Objectives: Perform vibration signal preprocessing and feature extraction Conduct exploratory data analysis (EDA) on bearing datasets Compare time-domain and frequency-domain vibration features Train machine learning models for bearing fault classification Evaluate cross-domain generalization performance Identify the most important features contributing to fault diagnosis

Research Questions:

RQ1: Can a machine learning model trained on one bearing dataset accurately predict faults on another dataset with different operating conditions?

RQ2: Which statistical vibration features contribute most to accurate bearing fault prediction across multiple datasets?

RQ3: Do frequency-domain features detect faults earlier than time-domain features?

RQ4: Do machine learning models outperform traditional threshold-based methods for fault detection in IMS data?

Datasets Used

IMS Bearing Dataset: The IMS dataset contains run-to-failure bearing vibration signals collected from rotating machinery under controlled operating conditions. XJTU-SY Bearing Dataset: A widely used benchmark dataset for intelligent fault diagnosis and predictive maintenance research.

Feature Engineering:

Time-Domain Features:

Mean Standard Deviation RMS (Root Mean Square) Variance Skewness Kurtosis Crest Factor Frequency-Domain Features:

Spectral Entropy Dominant Frequency FFT-Based Statistical Features

Project Workflow:

Experimental Results:

Model Output;

API Response Example

Project Structure:

predictive-maintenance/
├── data/
│   ├── ims_features.csv
│   ├── ims_features_crossdomain.csv
│   └── xjtu.csv
│
├── images/
│   ├── API.png
│   ├── OP.png
│   └── OP1.png
│
├── notebooks/
│   └── Cross_domain_Predictive_Maintenance_System_for_Industrial_AI_Fast_API.ipynb
│
├── README.md
└── requirements.txt

Key Findings:

Statistical vibration features are highly effective for bearing fault diagnosis. Time-domain features generally outperform frequency-domain features. Random Forest achieved near-perfect classification performance on the XJTU dataset. Cross-domain evaluation highlights challenges in model generalization across different operating conditions. Feature engineering significantly improves predictive performance.

Applications:

This project can be applied in: Manufacturing Industries Smart Factories Industrial IoT Systems Rotating Machinery Monitoring Aerospace Maintenance Automotive Predictive Maintenance Energy and Power Plants

Future Scope:

Potential future improvements include: Deep Learning-based fault diagnosis Transfer Learning for domain adaptation Real-time predictive maintenance systems Edge AI deployment for industrial monitoring Explainable AI (XAI) for fault interpretation Streaming sensor data integration

Technologies Used:

Python NumPy Pandas Scikit-learn SciPy Matplotlib FastAPI Jupyter Notebook

Author:

Yadnikee Bhole

Data Science | Machine Learning | Predictive Maintenance Research License