Using deep convolutional networks combined with signal processing techniques for accurate prediction of surface...

Area of Science:

• Manufacturing Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Surface roughness is a critical quality indicator in milling operations.
• Accurate real-time monitoring of surface roughness is essential for process optimization.
• Traditional methods for surface roughness assessment can be time-consuming and labor-intensive.

Purpose of the Study:

• To develop and evaluate a deep learning framework for predicting and classifying surface roughness in milling parts.
• To compare the effectiveness of different signal processing techniques for converting acoustic emission (AE) signals into images.
• To identify optimal deep learning architectures for real-time surface roughness monitoring.

Main Methods:

• Acoustic emission (AE) signals from milling experiments were transformed into 2D images using four encoding methods: SSPC, SSSC, SSSC*, and RP.
• Convolutional neural networks (VGG16, ResNet18, ShuffleNet, CNN-LSTM) were employed to predict surface roughness categories (average surface roughness, Ra).
• The framework's robustness was tested against Gaussian noise, and machining parameters were incorporated as additional inputs.

Main Results:

• The SSPC signal processing technique achieved over 98% accuracy across most models.
• ShuffleNet offered a balance of high accuracy (96-98%) and low computational cost.
• SSPC and SSSC demonstrated superior noise resistance, maintaining over 90% accuracy under high noise conditions.
• Incorporating machining parameters improved model accuracy and convergence, especially with noisy data.

Conclusions:

• Deep convolutional networks combined with innovative signal encoding techniques can accurately predict surface roughness.
• ShuffleNet is identified as an optimal architecture for real-time monitoring due to its performance and efficiency.
• The proposed data-driven framework enables real-time monitoring and optimization of machining processes based on process signatures.