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

This paper uses deep learning techniques to present a framework for predicting and classifying surface roughness in milling parts. The acoustic emission (AE) signals captured during milling experiments were converted into 2D images using four encoding Signal processing: Segmented Stacked Permuted Channels (SSPC), Segmented sampled Stacked Channels (SSSC), Segmented sampled Stacked Channels with linear downsampling (SSSC*), and Recurrence Plots (RP). These images were fed into convolutional neural networks, including VGG16, ResNet18, ShuffleNet and CNN-LSTM for predicting the category of surface roughness values. This work used the average surface roughness (Ra) as the main roughness attribute. Among the Signal processing techniques, SSPC could achieve the highest accuracy, above 98%, across most models, owing to minimal preprocessing of signals. ShuffleNet demonstrated a strong combination of accuracy (96-98%) and low computational cost. The robustness of networks was evaluated by introducing Gaussian noise at two levels. SSPC and SSSC were the most noise-resistant approaches, maintaining testing accuracy above 90% at high noise. Augmenting acoustic data with machining parameters (cutting speed, depth, feed rate, tool type) as additional inputs could improve the model's accuracy and convergence rate, especially for noisy data. Finally, ShuffleNet was identified as an optimal architecture for real-time monitoring due to its accuracy, noise resilience, and low computational cost. In summary, this study demonstrates the capability of deep convolutional networks combined with innovative signal encoding techniques to accurately predict surface roughness values and categories under various cutting conditions. Based on process signatures, the framework provides a data-driven approach to monitoring and optimizing machining processes in real-time.