Evaluation of turned and milled surfaces roughness using convolutional neural network

Abstract

• This study develops deep learning models for evaluation of surface roughness. • The prediction is based on the pattern of ridge-valley of the cutter marks. • Five loss functions are analyzed to find its suitability for the roughness data. • The results of Ra , Rz , and RzJ prediction and its discrepancy are investigated. Existing computer vision methods to measure surface roughness rely on feature extraction to quantify the surface morphology and build prediction models. However, the feature extraction is a complicated process requiring advanced image filtering and segmentation steps, resulting in long prediction time and complex setup. This study proposes the use of convolutional neural network to evaluate the surface roughness directly from the digital image of surface textures. This method avoids feature extraction since this step is integrated inside the network during the convolution process. Five loss functions for the prediction models are selected and analyzed based on their suitability and accuracy. The predicted values obtained are compared to the actual surface roughness values measured using a stylus-based profilometer. The performance of the proposed model is evaluated for the prediction of the surface roughness of typical machining operations, such as outside diameter turning, slot milling, and side milling, at various cutting conditions.