A countersunk hole provides the main connection between the components and skin of an aircraft. Measuring the sizes and depths of countersunk holes to meet quality requirements is essential for ensuring connection accuracy and structural strength. The manual-contact method currently used in China requires only a low degree of intelligence and provides low measurement accuracy, while measurement methods using monocular vision require the camera to remain perpendicular to the measured object. Both of these methods directly influence the accuracy of perspective projections of the depth of the dimple. This study addresses these issues by proposing an online detection method for evaluating the quality of drilled countersink holes using binocular vision. A FANUC robot is adopted as the motion carrier for performing real-time detection of the hole-shape parameters using a visual-inspection system mounted on the end of the actuator.

In the round-hole detection algorithm based on binocular vision, the basic requirement is to accurately match the subpixel edge points. In the present study, a method for matching crossed laser lines was adopted that enables high-accuracy detection. In this method, a projection-mapping model of the boundary points is first constructed based on the intersection of the crossed laser lines and the countersunk hole. Using iterative projection transformations, the RANSAC algorithm is then applied to optimize the resulting single matrix. Second, to consider the influence of distortion in the perspective projection, a multilayer perceptron model based on the simulated-annealing algorithm (SA-MLP) is used to perform secondary corrective optimization of boundary points that were outside this mapping relationship. This model solved the problem of mismatch caused by small differences between light and dark regions of the boundary area and ensured matching accuracy of the boundary points. Finally, the spatial curvature of the three-dimensional reconstructed-contour point cloud data for the drilled countersink hole was employed to establish a mathematical model. The pore size and the depth of the reticle were calculated from the geometric relationship of the fitted cone model.

A preliminary mapping relationship between pairs of boundary points was established after implementing the perspective transformation and the RANSAC algorithm (Fig. 5). The boundary points on the left side of the countersunk hole were mapped onto the boundary contour on the right after a projection transformation. The parallel polar lines also indicated a linear relationship for the projection-mapping model based on the boundary points generated in this study. After error compensation by the SA-MLP network (Fig. 7), the boundary points from the left side of the hole were mapped precisely onto the right contour boundary (Fig. 9). To intuitively reflect the method of matching the laser cross, CloudCompare software was introduced to calculate the flatness of the point cloud generated by this method and the local feature matching method. The simulation results (Table 2, Fig. 11) showed that the point cloud generated by the proposed algorithm was increased in flatness by 77.7% and that the error volatility was reduced by 88.6%. Thus, the overall matching accuracy was considerably improved. Finally, the shape parameters of the drilled countersink hole were obtained using a space-cone model (Fig. 13) and were compared with the results from monocular visual inspection. The experimental results (Table 3, Fig. 18) showed that although the use of a telecentric lens improved the detection accuracy of monocular vision for the inner aperture (the directly measured index), the error generated by the projection process—for example, in the angle and depth of the dimple—were ignored by the overly idealized parallel-projection scheme. The measurement accuracy of the monocular method was considerably lower than that of the binocular results and cannot meet the requirements of the actual project. The measurement method based on binocular vision had a maximum error of 0.031 mm in dimple depth, and the error fluctuation was relatively stable, while the maximum errors in the aperture and the angle of the socket were 0.031 mm and 0.152°, respectively, which met the detection indices.

To improve the accuracy and robustness of the drilling-and-riveting detection process, this project constructs a visual-inspection system using a FANUC robot as the motion carrier and proposes a binocular-vision detection algorithm based on crossed laser lines. First, the algorithm uses the center of the laser cross as the matching benchmark and uses the RANSAC algorithm to optimize the projection matrix. The SA-MLP network model is used to compensate for the error of deviation point to obtain an accurate boundary-point matching relationship. Second, the generated boundary-point-cloud data is used to fit the inner and outer contour curves. On this basis, an ideal cone model is constructed for the drilled countersink hole. The dimple angle and dimple depth of the countersink hole can be derived from this model. Experimental results show that the point cloud data generated for round holes by the matching algorithm proposed in this paper increase the planarity of the point cloud by 77.7% and reduce the error volatility by 88.6% compared with the traditional algorithm. The final detection error in the aperture of the drilled countersink hole was less than 0.031 mm, dimple angle was less than 0.152°, dimple depth was within 0.04 mm, and the proposed algorithm had better stability and feasibility than methods in current use.