Comparative Study of Popular Deep Learning Models for Machining Roughness Classification Using Sound and Force...

Area of Science:

• Manufacturing Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Machining surface roughness classification is crucial for quality control.
• Traditional methods often lack efficiency and accuracy.
• Deep learning offers potential for automated, data-driven classification.

Purpose of the Study:

• To compare various deep learning (DL) architectures for classifying machining surface roughness.
• To evaluate the effectiveness of sound and force data for this classification task.
• To identify the optimal DL model and feature extraction technique.

Main Methods:

• Compared Multi-Layer Perceptron (MLP), Convolution Neural Network (CNN), Long Short-Term Memory (LSTM), and transformer models.
• Utilized Mel-Spectrogram and Mel Frequency Cepstral Coefficients (MFCCs) for audio feature extraction.
• Trained and validated models on sound and force data from aluminum machining experiments.

Main Results:

• Transformer models achieved the highest training (>96%) and validation (≈90%) accuracies.
• All models, except CNN with Mel-Spectrogram, provided satisfactory roughness classification.
• Transformer models demonstrated superior performance in classifying time-series sound and force data.

Conclusions:

• The transformer model is highly suitable and superior for classifying machining surface roughness using time-series sound and force data.
• Feature extraction techniques like Mel-Spectrogram and MFCCs are effective when used with appropriate DL architectures.
• Deep learning, particularly transformers, offers a promising avenue for advanced manufacturing quality assessment.