چكيده به لاتين
Surface roughness of machined parts is a crucial parameter in assessing surface quality,
influenced by a wide range of factors. This report presents a framework for predicting and
classifying surface roughness in milling operations using deep learning techniques applied to
data. A model is proposed to predict and classify surface roughness values, utilizing data from
162 experiments conducted on three aluminum alloys: 7075, 6061, and 2024. The data include
cutting speed, depth of cut, tool coating type, feed rate, and surface roughness output. In this
study, acoustic emission signals recorded during milling tests were converted into 2D images
and fed into convolutional neural networks (CNNs) such as ResNet18, ShuffleNet, MobileNet,
and CNN-LSTM. Four methods were used to convert the time-series acoustic emission signals:
Stacked Sample Permutation Channels (SSPC), Stacked Sampled Channels (SSSC/SSSC*) ,
which has two subsets, and Recurrence Plots (RP). Among these encoding techniques, SSPC
achieved the highest accuracy, over 98%, in most models, attributed to the minimal
preprocessing of the signals. MobileNet demonstrated a strong combination of accuracy (96-
98%) and low computational cost. The performance of these methods was also evaluated under
two noise levels, 40% and 80%, with both zero-mean and non-zero-mean noise conditions.
SSPC and SSSC were the most robust against noise, maintaining test accuracy above 90% even
in high-noise conditions. Adding acoustic emission data alongside process parameters (cutting
speed, depth of cut, feed rate, tool type) as additional inputs can enhance model accuracy and
convergence speed, particularly for noisy data. Finally, ShuffleNet and MobileNet were
identified as suitable architectures for real-time monitoring due to their high accuracy, noise
resistance, 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 under varying conditions. Based on process parameters,
this framework offers a data-driven approach for real-time monitoring and optimization of
machining processes.
