Machine Fault Monitoring using Spectral Correlation and Deep Learning
This project presents a robust deep learning approach to diagnose bearing faults in rotating machinery using vibration signal processing and Spectral Correlation Density (SCD) images. Early fault detection in bearings is vital to ensure equipment reliability, reduce maintenance costs, and prevent catastrophic failures in industrial systems.
Bearing Fault by Crack Size 
📂 Dataset
We used the publicly available dataset:
- 🔗 Mendeley Dataset Link
- 📄 Dataset Paper:
“Vibration, Acoustic, Temperature, and Motor Current Dataset of Rotating Machine Under Varying Load Conditions for Fault Diagnosis”
DOI: 10.17632/ztmf3m7h5x.6
Published: 8 February 2023, Version 6
📊 Problem Overview
Traditional signal processing techniques like Fast Fourier Transform (FFT) often struggle with the non-stationary nature of vibration signals. This project addresses those limitations by leveraging:
- Cyclostationary properties of vibration signals
- SCD image generation from time-series data
- CNN-based image classification for fault detection
🔍 Methodology
We developed and tested three deep learning models to classify seven bearing fault conditions under three load settings (0 Nm, 2 Nm, 4 Nm) across two machine housings (A and B).
📌 Models Used
| Model | Description |
|---|---|
| Custom CNN | Lightweight 4-layer convolutional model |
| ResNet152V2 | Deep residual network with skip-connections |
| EfficientNetB0 | Lightweight model optimized for edge deployment |
📈 Performance Summary
🖼️ Results Snapshot
Figure: Accuracy comparison of all models on both housings under varying loads.
🧠 Key Contributions
- 🌀 Transformed vibration time-series into SCD-based spectral images for enhanced interpretability.
- 🔍 Demonstrated CNNs’ effectiveness in complex, noisy industrial datasets.
- 🧾 Delivered practical insights for deploying condition monitoring solutions on edge hardware.
⚒️ Technologies Used
- Python
- TensorFlow & Keras
- OpenCV, NumPy, Matplotlib
- SciPy Signal Processing
- Jupyter Notebooks
🏁 Conclusion
This project provides a highly accurate and scalable solution for bearing fault classification. The use of SCD images combined with CNN architectures delivers a powerful framework for real-time, edge-compatible monitoring systems in industrial environments.
The models’ robustness across different housings and load settings confirms their generalization capability and readiness for real-world deployment.
Developed by Dilshara Herath — AI for Industrial Reliability & Safety