Deep learning-based three-dimensional topography measurement of additively manufactured components

To solve the problem that the dimensions of friction stir deposition components easily deviate from model slicing planning, which makes it difficult to deposit complex-shaped components, it is urgent to conduct high-precision surface topography measurement to provide data support for the hybrid additive and subtractive manufacturing of precision components. In this study, a laser stripe centerline extraction algorithm integrating multi-scale convolution and self-attention mechanism was proposed to solve the problem of poor accuracy of traditional algorithms under severe interference conditions by enhancing local feature representation and stripe continuity. The model was trained using a laser stripe dataset containing different types of noise, and its coefficient of determination (R²) reached 0.945. The measurement error of the three-dimensional topography measurement method based on convolutional neural networks for standard parts was 0.37% of the full scale. Subsequently, the proposed method was applied to the scanning of aluminum alloy friction stir deposition components, and the noise ratio was reduced by 13.49% compared with the gray centroid method. Finally, the width and height data of single-layer and multi-layer deposition additive components were measured, and high-precision topography reconstruction and feature extraction of highly reflective alloy friction stir deposition components were achieved.