Quantitative prediction of laser-cut slag adhesion by integrating image and frequency domain features

To achieve precise quantification of slag adhesion and process optimization in laser cutting, this study investigates a convolutional neural network (CNN)-based prediction method that integrates both image and frequency-domain features. A dataset of 2160 cross-sectional images of 1 mm thick 304 stainless steel was constructed. From these images, key dross characteristics-area, height, and perimeter were accurately extracted using a combination of image processing techniques including Gaussian blur, adaptive thresholding, and morphological closing operations. To evaluate the predictive potential of different input representations, both RGB images and binarized images transformed via wavelet packet decomposition (WPD) were used as model inputs. The regression performance of three CNN architectures-VGG16, ResNet50, and DenseNet121 was systematically compared. Experimental results demonstrate that VGG16 achieved the highest prediction accuracy for dross area and height using RGB images, with mean absolute errors (MAE) of 0.019 mm² and 0.044 mm, respectively. For predicting the perimeter, which better reflects dynamic process behavior, the WPD frequency-domain input path yielded a significantly improved MAE of 0.094 mm and a normalized MAE (nMAE) of 5.25%. The regression fit between predicted and actual values showed a slope of 0.83 and a coefficient of determination (R²) of 0.86, indicating a strong linear correlation. This study confirms the effectiveness of VGG16 in predicting dross-related features and demonstrates the capability of WPD-derived frequency-domain features in capturing transient process information during laser cutting. The proposed methodology offers a reliable quantitative tool for intelligent process evaluation and closed-loop optimization.