Convolutional Twin Point Network Blade Contour Stitching System Based on Multi-scale Feature Fusion

Abstract

The invention discloses a convolutional twin point network blade profile splicing system based on multi-scale feature fusion, including a data acquisition module, a convolutional twin point network and a data splicing module. The convolutional twin point network includes a network module that iterates several times, The network module includes a feature extraction module, a feature space matching module and a singular value decomposition module; the feature extraction module uses the edge convolution network structure of the improved pyramid structure to extract the high-dimensional spatial features in the source point cloud and the target point cloud, and then utilizes The high-dimensional spatial features calculate the feature space matching matrix, and use the feature space matching matrix to calculate the correspondence between the points in the two point cloud data (source point cloud and target point cloud), and finally obtain the two point cloud data ( The rigid body transformation between the source point cloud and the target point cloud), and the optimal rigid body transformation is solved according to multiple iterations. The experimental results show the feasibility and good practical application prospects of this method.

Description

technical field [0001] The invention relates to the field of blade profile detection, in particular to a convolution twin point network blade profile splicing system based on multi-scale feature fusion. Background technique [0002] Blades are known as the jewel in the crown of modern industry and are widely used in aero engines, steam turbines and wind turbines. To ensure perfect and stable aerodynamic performance at high speeds, blades require extremely high dimensional accuracy and surface integrity. Accurate measurement of blade profile is an important means to guide blade production. However, thin-walled, twisted and mirror-like free-form surfaces increase the difficulty of blade surface measurement. At present, the acquisition of blade profile is done by three-coordinate measurement, which is a high-precision and easy-to-implement method. However, the efficiency of three-coordinate measurement is low, which hinders the production efficiency of blades. The increased...

AI Technical Summary

Problems solved by technology

Because there must be mechanical errors in the four-axis detection system, there is a certain error between the rigid body transformation directly given by the system and the real rigid body transformation, which in turn causes errors between the spliced ​​blade outline and the actual blade

Existing point cloud registration algorithms include traditional splicing algorithm (ICP) and deep learning-based splicing algorithm (PointLK), but there are still the following problems: the thin wall of the blade, the distorted space free-form surface and the points under the two fields of view The small overlapping part of the cloud increases the difficulty of extracting features with rotation and translation invariance; and under different fields of view, the point cloud density in the overlapping part is inconsistent, and it is difficult to find point correspondences

Images