Automatic surface roughness recognition system under different manufacturing processes based on deep learning

Abstract

Automated high-precision online measurement of machined surface roughness for key parts holds significant importance in intelligent manufacturing processes. A deep learning-based system is developed for automatic surface roughness recognition, which is built upon a dual-source laser speckle imaging apparatus. The Relative Maximum Contrast Ratio (RMCR) method is proposed for the first time to determine the optimal measurement parameters, enabling the acquisition of high-quality speckle images. The precise recognition of surface roughness through datasets relies heavily on the performance of deep learning models. The Twins-SVT model excels in the field of image classification, leveraging a deep separable convolutional core architecture. To enhance the recognition precision of surface roughness, the Spatial and Channel Attention (SCA) module is integrated into the Twins-SVT model. This integration allows for comprehensive feature extraction by effectively combining spatial and channel information from speckle images. The outstanding recognition accuracy and generalization performance of the SCA-Twins-SVT model is validated using a substantial datasets of speckle images. Compared to mainstream deep learning models, the proposed SCA-Twins-SVT model demonstrates exceptional performance in recognizing surface roughness machined by horizontal and vertical milling.