Non-contact Inspection of Electrically Discharged Materials Using Machine Learning

Overview

This project focuses on predicting surface roughness of metals machined by Electrical Discharge Machining (EDM) using non-contact inspection methods and machine learning. Traditional inspection techniques are often contact-based, slow, and risk damaging the surface. Our research introduces an intelligent, data-driven alternative that leverages regression algorithms on an experimentally collected and augmented dataset.

What is EDM & Why This Matters

Electrical Discharge Machining (EDM) is a widely used non-conventional manufacturing process especially suited for hard-to-machine materials. However, surface roughness (Ra) — a critical quality metric — is traditionally measured using physical probes. These contact-based inspections are not only time-intensive but also pose a risk to the integrity of high-precision surfaces.

Our goal was to eliminate this need by building a predictive model that estimates surface roughness using only input process parameters — enabling instant, accurate, and safe evaluation.

Data Collection & Augmentation

We collected experimental data using a real EDM machine setup at our university. Key process parameters such as:

• Pulse On Time
• Pulse Off Time
• Current
• Voltage
  were recorded along with the corresponding measured surface roughness (Ra).

To overcome the limited sample size (31 data points), we applied extensive data augmentation using techniques like:

• Scaling and shifting
• Controlled noise injection
  This increased our dataset to 9,300 samples, making it suitable for robust machine learning.

Models Used

We explored and compared the performance of three regression algorithms:

• K-Nearest Neighbors (KNN)
• Support Vector Regressor (SVR)
• Random Forest Regressor (RFR)

Each model was trained on an 80:20 split of the dataset and evaluated using standard metrics like R² score and Mean Squared Error (MSE).

Key Findings

Among all models, K-Nearest Neighbors (KNN) consistently outperformed the others with:

• R² Score ≈ 0.999
• MSE ≈ 0.00157

This suggests that for this dataset, local similarity-based methods like KNN are highly effective in capturing the relationship between EDM parameters and surface roughness.

We also visualized:

• Predicted vs. Actual scatter plots
• Residual histograms
  which confirmed the tight, unbiased predictions made by KNN across the data range.

Benchmarking vs Literature

We reviewed several other approaches used in industry and academia — including neural networks, W-ELM, SinGAN, Taguchi-ANN hybrids, and fuzzy logic systems. Surprisingly, our simpler KNN-based approach either matched or exceeded their performance while being computationally efficient and easier to deploy.

My Role

I was actively involved in:

• Conducting EDM experiments
• Designing the data augmentation strategy
• Implementing all regression models
• Analyzing and visualizing results
• Comparing outcomes with other state-of-the-art methods

Future Scope

• Real-time integration of this model with EDM machines for live surface monitoring
• Using deep learning or transformers on surface image data for further generalization
• Exploring 3D surface reconstruction with non-contact sensors

Final Takeaway

This research showcases how simple but well-prepared machine learning models can effectively replace traditional, invasive inspection techniques — enabling smarter, faster, and safer manufacturing practices.