Abstract: Traditional machine learning methods for roughness prediction require the collection of various modalities of machine tool signals, and different predictive models must be designed for different grinding stages, making the process cumbersome and resulting in low prediction accuracy. Therefore, a surface roughness prediction method based on Swin-IMSA is proposed for the grinding process of external cylindrical grinding machines. Firstly, a machine vision system is established to capture microscopic images of the ground surface, from which texture feature maps are derived using fast gray-level co-occurrence matrix (GLCM) calculations. Secondly, these feature maps are classified into distinct categories according to roughness levels and subjected to data augmentation techniques to enhance the model's generalization capabilities. Finally, an improved Swin-IMSA roughness prediction network extracts relevant features from the texture characteristics for model training. Experimental results indicate that this model achieves a prediction accuracy of 97.56% on the test set within a Ra value range of 0.1 μm to 1.2 μm, with an average prediction error of 0.035 μm in assessing ground roll samples, thereby validating the efficacy of this approach.

Key words: grinding surface, roughness, precision roller, deep learning, textural features