Intelligent manufacturing has been extensively applied in the mass production of forgings within industries such as automotive manufacturing. While defect detection has achieved a relatively high degree of automation and accuracy, the subsequent repair phase still lacks adequate intelligence. This gap primarily stems from insufficient research on accurate three-dimensional spatial localization of defects, which is critical for guiding automated repair equipment with the necessary precision. To address the issues, this paper proposes a method for spatially locating surface defects on forgings by registering complete and partial point clouds. First, a binocular vision system is employed to detect surface defects and perform 3D reconstruction of the forging, thereby acquiring its spatial coordinate system and high-resolution defect images. Subsequently, the Iterative Closest Point (ICP) registration algorithm is applied to align the coordinate systems of the complete forging point cloud with the locally reconstructed defect point cloud. Finally, the 3D spatial coordinates of the defects are determined through point cloud projection, enabling precise spatial localization of the surface defects. Experimental results demonstrate that point cloud registration achieved high-precision alignment, with RMSE values consistently maintained around 0.6 mm. The defect dimensions calculated from the resulting 3D coordinates show excellent agreement with physical measurements, with errors of merely 1.9 % in length and 3.4 % in width. This performance corroborates the feasibility and effectiveness of the proposed method for the spatial localization of surface defects in forged components.

Robust rigid point cloud registration is effective for accurate positioning and measurement of complex components. The existing registration algorithms, however, fail to overcome the matching distortion caused by structural deviation, unknown abnormal allowance, and various measurement inherent defects. Although the recently proposed VMM and WPMAVM algorithms can inhibit the matching distortion to some extent, they still fail in the presence of numerous abnormal points. In this study, we present a progressive and adaptive variance minimization (PAVM) algorithm to address these issues. A progressive de-pseudo weight is established to ensure the involvement of all point pairs in optimization at the initial registration stage. Then, an approximately truncated weight function is employed to mitigate the influence of abnormal points on registration results. Furthermore, a novel adaptive coordination distance function is established by improving the symmetric point-to-plane distance metric and combining the first-order approximate point-to-point distance metric, which enhances the algorithm speed and stability. The analysis investigates the anti-abnormal interference ability and quadratic convergence, validating the feasibility of the PAVM algorithm. Experiments are undertaken to illustrate the notable benefits of our algorithm in convergence stability, matching speed, and universality. These attributes render the algorithm well-suited for registration tasks involving diverse complex components.

Three-dimensional temperature field monitoring of automotive aluminum alloy component forging dies under high temperature conditions is critical for process optimization and service life prediction. Addressing the limitations of existing single-point temperature measurement or two-dimensional thermal imaging technologies, which cannot provide complete three-dimensional temperature field information, this paper proposes a three-dimensional temperature field reconstruction method based on multi-sensor thermal field fusion. First, a heterogeneous three-camera system comprising two visible light cameras and one infrared thermal imaging camera was established. Then, a specially designed calibration plate with black-and-white checkerboard temperature differences was used to jointly calibrate the two visible light cameras and the infrared thermal imaging camera. Next, the three-dimensional geometric model of the mold is obtained using multi-view point cloud reconstruction technology from the two visible light cameras. To improve matching accuracy in areas with missing complex textures, an improved deep learning-based stereo matching algorithm, RAFT-Stereo, is introduced to overcome thermal noise issues. Furthermore, the two-dimensional temperature information collected by the infrared thermal imager is mapped to the three-dimensional point cloud. Through infrared-visible light camera multi-view geometric transformation temperature field fusion technology, a complete three-dimensional temperature field model containing both spatial coordinates and temperature attributes is generated. Experimental results show that compared to traditional stereo matching methods SGBM and PMS, the geometric reconstruction accuracy of this method has been improved by an average of approximately 68.2% and 51.9%, respectively; after fusion, the average accuracy of the three-dimensional temperature field temperature mapping reaches ±0.3°C. This method provides high-precision three dimensional data support for the thermal behavior analysis and thermal management of forging dies.

Robotic grinding is utilized for machining large complex components due to its exceptional flexibility and expansive workspace. However, the existing research primarily focuses on the global machining and lacks solutions for repair and remanufacturing of components with local defects. In order to perform the local repair efficiently, this paper proposes a 3D vision-guided robotic grinding framework for repairing random defects, by taking the removal of welding slags on the automotive body as an example. Based on the measured point cloud, a robust function weighted variance minimization (RFWVM) registration algorithm is utilized to position the automotive body with high precision. Meanwhile, the datum plane is reconstructed to extract the defect point clouds. On this basis, the welding slags are divided into regions in accordance with their height characteristics. For these regions, a robotic grinding depth prediction model is then constructed to determine the mapping between welding slag height and process parameters, and the genetic algorithm is further improved to solve the shortest path decision-making problem in robotic grinding. The experimental results conducted in a typical region of an automotive body confirm the practicality and effectiveness of the proposed framework. This study provides a valuable reference for local defects repair of complex components.