Non-contact Inspection of Electrically Discharged Materials Using Machine Learning
A machine learning approach for predicting surface roughness in EDM-machined materials using process parameters instead of contact-based inspection.
Devrajsinh Jhala, Nirmit Patel, Jemil Dharia, Jemin Butani, Devesh Patel, M. B. Kiran
Overview
This project predicts the surface roughness of metals machined by Electrical Discharge Machining using non-contact inspection methods and machine learning. Traditional inspection approaches are often contact-based, slow, and risky for precision surfaces.
The research introduces a data-driven alternative that uses regression algorithms on experimentally collected and augmented data to estimate surface roughness from process parameters.
What is EDM and Why It Matters
Electrical Discharge Machining is a non-conventional manufacturing process used for hard-to-machine materials. Surface roughness, represented as Ra, is a critical quality metric, but it is often measured with physical probes.
The aim was to remove the need for contact-based inspection by building a predictive model that estimates surface roughness from process parameters, enabling faster and safer evaluation.
Data Collection and Augmentation
Experimental data was collected using a real EDM machine setup at the university. Pulse on time, pulse off time, current, and voltage were recorded along with measured surface roughness.
- The initial dataset contained 31 experimental data points.
- Data augmentation used scaling, shifting, and controlled noise injection.
- The augmented dataset contained 9,300 samples for model training and evaluation.
Models Used
Three regression algorithms were trained and compared: K-Nearest Neighbors, Support Vector Regressor, and Random Forest Regressor. Each model used an 80:20 train-test split and was evaluated with R2 score and Mean Squared Error.
Key Findings
- K-Nearest Neighbors consistently outperformed the other models.
- KNN reached an R2 score of approximately 0.999.
- KNN produced an MSE of approximately 0.00157.
- Predicted-vs-actual plots and residual histograms confirmed tight, unbiased predictions.
Benchmarking
The work compared the results with approaches such as neural networks, W-ELM, SinGAN, Taguchi-ANN hybrids, and fuzzy logic systems. The simpler KNN-based approach matched or exceeded several reported results while remaining easier to deploy.
My Role
- Conducted EDM experiments.
- Designed the data augmentation strategy.
- Implemented all regression models.
- Analyzed and visualized model performance.
- Compared outcomes with state-of-the-art methods.
Future Scope
- Integrate the model with EDM machines for live surface monitoring.
- Use deep learning or transformers on surface image data for broader generalization.
- Explore 3D surface reconstruction with non-contact sensors.
Final Takeaway
This research shows how simple but well-prepared machine learning models can replace invasive inspection workflows, making manufacturing evaluation faster, safer, and easier to automate.