Scanning viewpoint generation system and method for detecting defects of key parts of aerospace equipment

Abstract

The invention discloses a scanning viewpoint generation system and method for detecting defects of key parts of aerospace equipment, and belongs to the technical field of intelligent detection, and the system comprises an optimal next viewpoint generator module which is used for obtaining global point cloud data, carrying out spatial discretization and leading edge extraction, and generating an optimal next viewpoint; sampling in the neighborhood of the defect candidate region to generate a candidate viewpoint set, and selecting an optimal next viewpoint; the motion planner module is used for constructing a sampling area by taking the current viewpoint as a starting point and the optimal next viewpoint as an end point, generating an executable track and carrying out smoothing and time parameterization processing on a discrete path; and the adaptive scanning and termination module is used for dividing a three-dimensional space to be reconstructed into a plurality of sub-regions, distributing scanning resources according to region attention and information gain, judging a scanning termination moment and outputting a final scanning result and a viewpoint sequence. On the premise of ensuring scanning integrity and quality, global optimization and efficient operation of the scanning process are realized.

Description

Technical Field

[0001] This invention belongs to the field of intelligent detection technology, specifically relating to a scanning viewpoint generation system and method for key components of aerospace equipment. Background Technology

[0002] When critical components of aerospace equipment are subjected to extreme conditions such as high temperature, high speed, and high load over long periods, their surfaces are prone to damage including microcracks, fatigue spalling, localized pitting, and complex deformation. These defects are typically small in scale, concealed in shape, and distributed within complex curved surfaces. If they are not effectively identified in early detection, they will significantly affect structural performance and may even induce serious safety accidents. Therefore, achieving high coverage, high precision, and adaptive 3D scanning reconstruction and defect identification of critical components in a quality inspection and structural health monitoring system is of significant engineering importance.

[0003] Currently, 3D scanning technology has been widely used in equipment inspection, but its scanning methods still mainly rely on fixed paths or regular strategies, such as equidistant scanning, surround scanning, or spiral scanning. These methods are essentially passive scanning modes, unable to dynamically adjust the acquisition strategy according to the geometry of the object being measured, occlusion distribution, and changes in potential defect areas. This leads to problems such as large scanning redundancy, numerous blind spots, and low acquisition efficiency in practical applications. Especially for aerospace components with features such as deep cavities, abrupt curvature changes, and complex edges, fixed scanning paths easily cause the sensor to be in a suboptimal observation pose, resulting in insufficient acquisition of geometric information in local areas, or even significant data loss.

[0004] In existing related technologies, some techniques attempt to improve the adaptability and detection efficiency of scanning paths from different perspectives. For example, existing scanning path planning methods based on multi-view collaboration optimize the viewpoint layout by analyzing the geometric features of the components, but their real-time adjustment capability is limited when faced with dynamic occlusion and complex curved surfaces. Existing technologies have also introduced path decision-making mechanisms based on reinforcement learning, which can improve the autonomy of scanning to a certain extent, but a system framework that integrates geometric visibility, motion efficiency, and defect sensitivity has not yet been formed.

[0005] In summary, existing scanning path planning methods generally lack a comprehensive system that can uniformly consider geometric visibility, motion costs, defect attention, and real-time strategy adjustment capabilities. To achieve high coverage and high efficiency 3D point cloud scanning of complex curved surface components of aerospace equipment, there is an urgent need for a full-process active perception method that can dynamically generate optimal viewpoints, plan feasible trajectories in confined spaces, and possess adaptive termination capabilities during the scanning process. Summary of the Invention

[0006] The purpose of this invention is to provide a viewpoint generation scanning system and method for defect detection of key components of aerospace equipment, which solves the problems of point cloud scanning and robot scanning path planning.

[0007] To achieve the above objectives, this invention provides a viewpoint generation and scanning system for defect detection of key components in aerospace equipment. The system includes: an optimal next viewpoint generator module, used to acquire global point cloud data of the key components of the aerospace equipment, perform spatial discretization and front-edge extraction in the three-dimensional spatial model corresponding to the global point cloud data, sample and generate a set of candidate viewpoints within a pre-selected defect candidate region neighborhood, and select the optimal next viewpoint based on a multi-objective utility function; and a motion planner module, used to construct a sampling region based on a heuristic sampling motion planning algorithm, starting from the current viewpoint and ending at the optimal next viewpoint. Heuristic sampling search and collision detection are performed within the sampling area to generate an executable trajectory. The discrete path, which is formed by concatenating discrete pose points generated by the motion planning algorithm based on the sampling nodes, is smoothed and time-parameterized. The sampling nodes are obtained by sampling the sampling area. An adaptive scanning and termination module is used to divide the 3D space to be reconstructed into multiple sub-regions, allocate scanning resources according to region attention and information gain, and determine the scanning termination time based on the joint threshold condition of the information gain of the candidate viewpoints in the candidate viewpoint set and the attention gain of the defective region, and output the final scanning result and viewpoint sequence.

[0008] Preferably, the optimal next viewpoint generator module is further configured to: rasterize or voxelize the current 3D space, divide the 3D space into occupied voxels, free voxels, and unknown voxels, and define the common boundary between the free voxels and the unknown voxels as the front set: ;in, For a set of free voxels, For an unknown set of voxels, This is the adjacency threshold. Indicates belonging to the set of free voxels The spatial position of any voxel Represents the spatial location of any voxel belonging to the unknown voxel set. express and The Euclidean distance between the center points; sampling or clustering the front set to obtain a front point set; sampling the defect candidate region to obtain a defect candidate point set; jointly sampling the position and orientation within the neighborhood radius centered on the defect candidate point set and the front point set to construct the candidate viewpoint set; and evaluating the candidate viewpoints based on a multi-objective utility function, wherein the multi-objective utility function is: ;in, From the perspective The overall utility value, For information gain, For attention gain term, For the cost of movement, These are non-negative weight parameters.

[0009] Preferably, the motion planner module is further configured to employ a motion planning algorithm based on heuristic sampling, generating an executable trajectory that satisfies motion constraints by performing heuristic sampling search and collision detection in an ellipsoidal sampling region, and performing smoothing and time parameterization processing, including: using the current pose With the optimal next viewpoint pose As a path endpoint, when there exists an upper bound on the current optimal path cost. At that time, construct the ellipsoidal region As the sampling area, then in the ellipsoidal region Batch random sampling is performed to obtain a sample set. Samples that fall into obstacles or violate safety clearance constraints are removed to obtain a candidate node set. Based on the candidate node set Construct a search tree or search graph and perform a heuristic search to obtain a list of discrete path points for executable trajectories. ; and the discrete path point list As control points, B-splines or spline interpolation are used to generate smooth trajectory curves. So that the smooth trajectory curve The position and orientation are first- or second-order continuous, depending on the maximum speed of the robot platform. With maximum acceleration Constraints on the smooth trajectory curve Time reparameterization is performed to obtain an executable trajectory that satisfies the dynamic constraints.

[0010] Preferably, the motion planner module is further configured to perform heuristic sorting and expansion of nodes and edges using a lower bound function of cost, including: defining a set of candidate nodes from the starting point. Sampling points in The cumulative cost is From the set of candidate nodes Sampling points in The lower bound function of the heuristic cost to the destination is: The total assessment cost is: ;in, It can be obtained from sampling points The lower bound of the Euclidean distance or path length to the destination is given; the total evaluation cost is selected from the candidate viewpoint set. The tree structure is expanded using the smallest sampling point. Collision detection is performed on the newly generated edges. Edges that pass the collision detection are retained and the tree structure is updated. The new total evaluation cost is then used to expand the tree structure if it is better than the upper bound of the current optimal path cost. When the iteration converges or reaches the preset iteration time limit, the optimal path and ellipsoid parameters are updated until the iteration converges or reaches the preset iteration time limit. The ellipsoid parameters include the ellipsoid center, principal axis direction, major semi-axis length and minor semi-axis length.

[0011] Preferably, the adaptive scanning and termination module is further configured to: divide the three-dimensional space to be reconstructed into multiple sub-regions. j is the index of the sub-region, and M is a positive integer. For each sub-region... Maintain its access count Cumulative gain and average gain The information gain is the average gain. Priority indices are calculated for each sub-region using the upper confidence boundary criterion. According to priority indicators Sort the corresponding sub-regions by size, and select the first one from largest to smallest. Each sub-region is selected as a high-attention region. Each sub-region is designated as a low-attention region. Candidate viewpoints are generated and trajectories are planned within the high-attention region. For the high-attention region, a first scanning mode is selected, and for the low-attention region, a second scanning mode is selected to allocate scanning resources from global to local. The top 10% of all sub-regions The remaining 10% of all sub-regions, wherein the scanning speed of the first scanning mode is lower than that of the second scanning mode, and the resolution of the first scanning mode is higher than that of the second scanning mode; and defining the... The information gain of the optimal viewpoint in the next scan is: The defective attention gain is When the conditions are met simultaneously in multiple consecutive scans When the current reconstruction information is determined to be saturated, a scan termination signal is triggered, the update of the viewpoint sequence and path planning is stopped, and the final 3D scan result and viewpoint sequence are output. To preset the information gain threshold, The preset defect attention gain threshold is used.

[0012] Accordingly, the present invention also provides a viewpoint generation scanning method for defect detection of key components of aerospace equipment. The method includes: acquiring global point cloud data of the key components of the aerospace equipment, performing spatial discretization and front edge extraction in the three-dimensional spatial model corresponding to the global point cloud data, sampling to generate a candidate viewpoint set in the neighborhood of a pre-selected defect candidate region, and selecting the optimal next viewpoint based on a multi-objective utility function; constructing a sampling region based on a motion planning algorithm using heuristic sampling, with the current viewpoint as the starting point and the optimal next viewpoint as the ending point, performing heuristic sampling search and collision detection within the sampling region to generate an executable trajectory, and smoothing and time parameterizing the discrete path formed by concatenating discrete pose points generated by the motion planning algorithm based on sampling nodes, wherein the sampling nodes are obtained by sampling the sampling region; and dividing the three-dimensional space to be reconstructed into multiple sub-regions, allocating scanning resources according to region attention and information gain, and determining the scanning termination time based on the joint threshold condition of the information gain of the candidate viewpoints in the candidate viewpoint set and the attention gain of the defect region, and outputting the final scanning result and viewpoint sequence.

[0013] Preferably, the viewpoint generation scanning method for defect detection of key components of aerospace equipment provided by the present invention further includes: rasterizing or voxelizing the current three-dimensional space, dividing the three-dimensional space into occupied voxels, free voxels, and unknown voxels, and defining the common boundary between the free voxels and the unknown voxels as the front set: ;in, For a set of free voxels, For an unknown set of voxels, This is the adjacency threshold. Indicates belonging to the set of free voxels The spatial position of any voxel Represents the spatial location of any voxel belonging to the unknown voxel set. express and The Euclidean distance between the center points; sampling or clustering the front set to obtain a front point set; sampling the defect candidate region to obtain a defect candidate point set; jointly sampling the position and orientation within the neighborhood radius centered on the defect candidate point set and the front point set to construct the candidate viewpoint set; and evaluating the candidate viewpoints based on a multi-objective utility function, wherein the multi-objective utility function is: ;in, From the perspective The overall utility value, For information gain, For attention gain term, For the cost of movement, These are non-negative weight parameters.

[0014] Preferably, the viewpoint generation scanning method for defect detection of key components of aerospace equipment provided by the present invention further includes employing a motion planning algorithm based on heuristic sampling. This algorithm generates an executable trajectory that satisfies motion constraints by performing heuristic sampling search and collision detection in an ellipsoidal sampling region, and then performs smoothing and time parameterization processing, including: using the current pose... With the optimal next viewpoint pose As a path endpoint, when there exists an upper bound on the current optimal path cost. At that time, construct the ellipsoidal region As the sampling area, then in the ellipsoidal region Batch random sampling is performed to obtain a sample set. Samples that fall into obstacles or violate safety clearance constraints are removed to obtain a candidate node set. Based on the candidate node set Construct a search tree or search graph and perform a heuristic search to obtain a list of discrete path points for executable trajectories. ; and the discrete path point list As control points, B-splines or spline interpolation are used to generate smooth trajectory curves. So that the smooth trajectory curve The position and orientation are first- or second-order continuous, depending on the maximum speed of the robot platform. With maximum acceleration Constraints on the smooth trajectory curve Time reparameterization is performed to obtain an executable trajectory that satisfies the dynamic constraints.

[0015] Preferably, the viewpoint generation and scanning method for defect detection of key components of aerospace equipment provided by the present invention further includes heuristic sorting and expansion of nodes and edges through a lower bound function of cost, including: defining a set of candidate nodes from the starting point. Sampling points in The cumulative cost is From the set of candidate nodes Sampling points in The lower bound function of the heuristic cost to the destination is: The total assessment cost is: ;in, It can be obtained from sampling points The lower bound of the Euclidean distance or path length to the destination is given; the total evaluation cost is selected from the candidate viewpoint set. The tree structure is expanded using the smallest sampling point. Collision detection is performed on the newly generated edges. Edges that pass the collision detection are retained and the tree structure is updated. The new total evaluation cost is then used to expand the tree structure if it is better than the upper bound of the current optimal path cost. When the iteration converges or reaches the preset iteration time limit, the optimal path and ellipsoid parameters are updated until the iteration converges or reaches the preset iteration time limit. The ellipsoid parameters include the ellipsoid center, principal axis direction, major semi-axis length and minor semi-axis length.

[0016] Preferably, the viewpoint generation scanning method for defect detection of key components of aerospace equipment provided by the present invention further includes: dividing the three-dimensional space to be reconstructed into multiple sub-regions. j is the index of the sub-region, and M is a positive integer. For each sub-region... Maintain its access count Cumulative gain and average gain The information gain is the average gain. Priority indices are calculated for each sub-region using the upper confidence boundary criterion. According to priority indicators Sort the corresponding sub-regions by size, and select the first one from largest to smallest. Each sub-region is selected as a high-attention region. Each sub-region is designated as a low-attention region. Candidate viewpoints are generated and trajectories are planned within the high-attention region. For the high-attention region, a first scanning mode is selected, and for the low-attention region, a second scanning mode is selected to allocate scanning resources from global to local. The top 10% of all sub-regions The remaining 10% of all sub-regions, wherein the scanning speed of the first scanning mode is lower than that of the second scanning mode, and the resolution of the first scanning mode is higher than that of the second scanning mode; and defining the... The information gain of the optimal viewpoint in the next scan is: The defective attention gain is When the conditions are met simultaneously in multiple consecutive scans When the current reconstruction information is determined to be saturated, a scan termination signal is triggered, the update of the viewpoint sequence and path planning is stopped, and the final 3D scan result and viewpoint sequence are output. To preset the information gain threshold, The preset defect attention gain threshold is used.

[0017] The beneficial effects of this invention are as follows:

[0018] This invention addresses the problems of traditional 3D scanning, such as lack of targeted viewpoint selection, lack of continuity and motion constraints in path planning, significant scanning redundancy, and difficulty in eliminating occlusion blind spots. It proposes a scanning viewpoint generation technology for defect detection of key components in aerospace equipment. The technical solution provided by this invention constructs an active closed-loop system of "perception-planning-execution-feedback," organically combining optimal viewpoint generation, heuristic path planning, and adaptive scan termination strategies. Through viewpoint generation, an attention mechanism for key areas is constructed, achieving global optimization and efficient operation of the scanning process while ensuring scan integrity and quality. Specifically:

[0019] 1. By setting up a candidate viewpoint generation and multi-objective utility evaluation mechanism, the viewpoint selection can achieve a balance between information gain, defect attention and motion cost. Candidate viewpoints are sampled in the leading voxels and high-interest areas, making the viewpoint selection more targeted and globally guided. This effectively avoids the problems of traditional next-best viewpoint methods being prone to local optima and lacking motion cost constraints, and greatly improves the expansion speed of the visible area and the defect exposure capability during the scanning process.

[0020] 2. By setting up a motion planning algorithm based on heuristic sampling search, relying on ellipsoidal constraints, a constrained search space is constructed between the current viewpoint and the optimal next viewpoint. The path expansion is guided by the lower bound function of cost, making the path planning both efficient and feasible. Then, by combining trajectory smoothing and time parameterization techniques, the output trajectory is ensured to meet the actual execution requirements in terms of spatial continuity, dynamic reachability and obstacle avoidance safety, which significantly reduces the search overhead and improves the stability and real-time performance of path generation.

[0021] 3. By setting an adaptive scanning and termination mechanism, the scanning strategy is adjusted in real time according to the dynamic changes in environmental coverage and attention distribution during the scanning process. The Upper Confidence Bound (UCB) scheduling method based on regional attention allows for the allocation of more scanning resources in high-risk areas and rapid browsing in low-value areas, thereby achieving rational allocation of scanning resources and maximizing efficiency. Furthermore, by using information gain and attention gain as termination conditions, redundant acquisition is effectively avoided, improving overall detection efficiency. This mechanism enables the scanning process to maintain high coverage while possessing high adaptability and intelligent decision-making capabilities, achieving high-coverage and high-efficiency 3D point cloud scanning of key components of aerospace equipment in complex curved surface environments. Attached Figure Description

[0022] Figure 1 This is a block diagram illustrating the principle of scanning viewpoint generation for defect detection of key components in aerospace equipment, as provided by the present invention.

[0023] Figure 2This is a schematic diagram of the working principle of the motion planner module provided by the present invention.

[0024] Figure 3 The flowchart shows the scanning viewpoint generation method for defect detection of key components of aerospace equipment provided by this invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0026] This invention proposes an intelligent viewpoint generation and scanning method for defect detection of key components in aerospace equipment, comprising an optimal next viewpoint generator module, a motion planner module, and an adaptive scanning and termination module. The optimal next viewpoint generator module, motion planner module, and adaptive scanning and termination module have communication capabilities.

[0027] Figure 1 The present invention provides a principle block diagram for viewpoint generation scanning for defect detection of key components in aerospace equipment, as shown in the figure. Figure 1 As shown, the viewpoint generation and scanning system for defect detection of key components of aerospace equipment provided by the present invention includes an optimal next viewpoint generator module, a motion planner module, and an adaptive scanning and termination module.

[0028] The optimal next viewpoint generator module acquires global point cloud data of key components of aerospace equipment, performs spatial discretization and front-edge extraction on the corresponding 3D spatial model, samples and generates a set of candidate viewpoints within the neighborhood of a pre-selected defect candidate region, and selects the optimal next viewpoint based on a multi-objective utility function. The defect candidate region is a local area with a high probability of defect occurrence, which can be identified by those skilled in the art based on the actual situation.

[0029] The motion planner module is used for motion planning algorithms based on heuristic sampling. It constructs a sampling region with the current viewpoint as the starting point and the optimal next viewpoint as the ending point. Within the sampling region, it performs heuristic sampling search and collision detection to generate an executable trajectory. It also performs smoothing and time parameterization on the discrete path formed by concatenating discrete pose points generated by the motion planning algorithm based on sampling nodes. The sampling nodes are obtained by sampling the sampling region.

[0030] The adaptive scanning and termination module divides the 3D space to be reconstructed into multiple sub-regions, allocates scanning resources based on region attention and information gain, and determines the scanning termination time based on a joint threshold condition of the information gain of candidate viewpoints in the candidate viewpoint set and the attention gain of defective regions, outputting the final scanning result and viewpoint sequence. The determination of the scanning termination time is automatic.

[0031] This invention utilizes deep learning, utility evaluation, and neural network architecture to construct an attention mechanism for key regions through viewpoint generation. This provides high-density point clouds of key regions that exhibit a high probability of defects in terms of feature response, geometric deviation, or observation uncertainty, thereby achieving the goal of adaptively planning the 3D point cloud scanning path. This enables high-coverage and high-efficiency 3D point cloud scanning of key components of aerospace equipment in complex curved surface environments.

[0032] The key components of the aerospace equipment in this invention include turbine blades, integral bladed disks, and nozzles.

[0033] The defect types in this invention include surface cracks, surface unevenness damage, and processing deformation.

[0034] The optimal next viewpoint generator module is also used for:

[0035] The current 3D space (which may also be the reconstructed 3D space) is rasterized or voxelized, dividing the 3D space into occupied voxels, free voxels, and unknown voxels, and the common boundary between free voxels and unknown voxels is defined as the frontier set:

[0036] ;

[0037] in, For a set of free voxels, For an unknown set of voxels, This is the adjacency threshold. Indicates belonging to the set of free voxels The spatial position of any voxel Represents the spatial location of any voxel belonging to the unknown voxel set. express and The Euclidean distance between the center points.

[0038] The front set is sampled or clustered to obtain the front point set, and the defect candidate region is sampled to obtain the defect candidate point set. The radius of the neighborhood centered on the defect candidate point set and the front point set is... Within the site, joint sampling of position and orientation is performed to construct a candidate viewpoint set:

[0039] ;

[0040] in, For the first The position vectors of each candidate viewpoint in three-dimensional space In order to target the Camera or sensor pose rotation matrix for each candidate viewpoint For the first One candidate viewpoint, For three-dimensional real space, It is a three-dimensional special orthogonal group.

[0041] Candidate viewpoints are evaluated based on a multi-objective utility function, which is as follows:

[0042] ;

[0043] in, From the perspective The overall utility value, For information gain, For attention gain term, For the cost of movement, These are non-negative weighting parameters used to balance the contributions of the three factors.

[0044] The motion planner module is based on a heuristic sampling motion planning algorithm. It obtains a discrete path point list through iterative sampling, and then generates an executable trajectory through path smoothing processing. Specifically, it includes:

[0045] The motion planner module employs a heuristic sampling-based motion planning algorithm. By performing heuristic sampling search and collision detection within an ellipsoidal sampling region, it generates executable trajectories that satisfy motion constraints and performs smoothing and time parameterization processing, including:

[0046] With current pose With the optimal next viewpoint pose As a path endpoint, when there exists an upper bound on the current optimal path cost. At that time, construct the ellipsoidal region As the sampling area, then in the ellipsoidal region Batch random sampling is performed to obtain a sample set. Samples that fall into obstacles or violate safety clearance constraints are removed to obtain a candidate node set. Based on the candidate node set Construct a search tree or search graph and perform a heuristic search to obtain a list of discrete path points for executable trajectories. ;

[0047] Discrete path point list As control points, smooth trajectory curves are generated using B-splines or spline interpolation. To make the trajectory curve smooth The position and orientation are first- or second-order continuous, depending on the maximum speed of the robot platform. With maximum acceleration Constraints on smooth trajectory curves Time reparameterization is performed to obtain an executable trajectory that satisfies the dynamic constraints, and this trajectory is then sent to the actuator.

[0048] The adaptive scanning and termination module, based on an adaptive scanning and termination mechanism, divides the 3D space to be reconstructed into multiple sub-regions. It dynamically allocates scanning resources based on region attention and information gain indices, and determines the scanning termination time based on a joint threshold condition of the information gain of candidate viewpoints and the attention gain of defective regions. It then outputs the final scanning results and viewpoint sequence. The determination of the termination time can be completed automatically. The specific operation process of the adaptive scanning and termination module includes:

[0049] The three-dimensional space to be reconstructed is divided into multiple sub-regions. j is the index of the sub-region, and M is a positive integer. For each sub-region... Maintain its access count Cumulative gain and average gain Information gain is the average gain. Number of visits Cumulative gain With average gain The relational expression is as follows:

[0050] ;

[0051] in, To prevent division by zero of small constants.

[0052] Priority indices are calculated for each sub-region using the Upper Confidence Boundary Criterion (UCB criterion). According to priority indicators Sort the corresponding sub-regions by size, and select the first one from largest to smallest. Each sub-region is selected as a high-attention region. Each sub-region is designated as a low-attention region. Candidate viewpoints are generated and trajectories are planned within high-attention regions. For high-attention regions, a first scanning mode is selected, and for low-attention regions, a second scanning mode is selected to allocate scanning resources from global to local. The top 10% of all sub-regions This represents the last 10% of all sub-regions, where the scanning speed of the first scanning mode is lower than that of the second scanning mode, and the resolution of the first scanning mode is higher than that of the second scanning mode. Generally, those skilled in the art refer to the first scanning mode as a low-speed, high-resolution scanning mode, and the second scanning mode as a high-speed, low-resolution scanning mode. Specifically:

[0053] The UCB criterion is used to calculate priority indicators for each sub-region:

[0054] ;

[0055] in, This represents the total number of visits to all sub-regions. For the first The number of times each sub-region is visited, where M is the number of sub-regions. To adjust the hyperparameters that balance exploration and utilization, according to Sort regions by size, prioritizing those in the largest region. Generate candidate viewpoints and plan trajectories within a large area, according to Sort by size from largest to smallest and select the first size. Each sub-region serves as a high-attention area, for For larger high-attention areas, select a low-speed, high-resolution scanning mode. Smaller areas, by After sorting by size from largest to smallest Each sub-region is designated as a low-attention region, and a high-speed, low-resolution scanning mode is selected to achieve adaptive scanning resource allocation from global to local.

[0056] Definition of the first The information gain of the optimal viewpoint in the next scan is: The defective attention gain is When in a series of (e.g., Second-rate, (For consecutive number of times threshold) During scanning, simultaneously satisfying When the current reconstruction information is determined to be saturated, a scan termination signal is triggered (which may be automatically triggered), the update of the viewpoint sequence and path planning is stopped, and the final 3D scan result and viewpoint sequence are output. To preset the information gain threshold, The preset defect attention gain threshold is used.

[0057] Specifically, the scanning viewpoint generation technology for defect detection of key components in aerospace equipment relies on deep learning, utility evaluation, and neural network architecture. It utilizes an optimal next viewpoint generator module to construct a candidate viewpoint set using a candidate viewpoint generation method, and then selects the optimal next viewpoint based on a multi-objective utility model. A motion planner module, employing a heuristic sampling-based motion planning algorithm, generates an optimal scanning path that satisfies dynamic and obstacle avoidance constraints through cost lower bound sorting, collision detection, and trajectory updating. An adaptive scanning and termination module dynamically adjusts the scanning strategy based on changes in environmental coverage and information entropy, and automatically terminates the scan based on a dual-condition triggering mechanism. This constructs an active closed-loop system of "perception-planning-execution-feedback," building an attention mechanism for key areas through viewpoint generation to provide high-density point clouds of critical regions for defect detection. While ensuring scanning integrity and quality, it achieves high-coverage and high-efficiency 3D point cloud scanning of complex curved surface components, suitable for high-precision surface defect detection of key components in aerospace equipment.

[0058] Figure 2 The schematic diagram of the working principle of the motion planner module provided by the present invention is as follows: Figure 2 As shown, the motion planner module is based on a heuristic sampling motion planning algorithm. It generates an executable trajectory that satisfies motion constraints by performing heuristic sampling search and collision detection within an ellipsoidal sampling region, and then performs smoothing and time parameterization processing. (Based on the current pose...) With the optimal next viewpoint pose As a path endpoint, when there exists an upper bound on the current optimal path cost. At that time, construct the ellipsoidal region As the sampling area, then in the ellipsoidal region Batch random sampling is performed to obtain a sample set. Samples that fall into obstacles or violate safety clearance constraints are removed to obtain a candidate node set. Based on the candidate node set Construct a search tree or search graph and perform a heuristic search to obtain a list of discrete path points for executable trajectories. The specific algorithm includes:

[0059] If the upper bound of the current optimal path cost is Through the starting pose and the final pose The ellipsoidal region is constructed as follows:

[0060] ;

[0061] in, , This is a positive definite matrix determined based on path cost. This represents an n-dimensional real vector space.

[0062] Then, batch random sampling is performed within the ellipsoidal region to generate a sample set:

[0063] ;

[0064] in, Let m be the collection point in the sample set, where m is a positive integer.

[0065] Samples that overlap with obstacle voxels or violate safety gap constraints are removed to obtain a candidate node set. :

[0066] ;

[0067] in, This represents the sampling points in the sample set that meet the constraints. Represents the set of points that violate constraints.

[0068] The motion planner module is also used to heuristically sort and expand nodes and edges using a lower bound function of cost, specifically by defining a set of candidate nodes from the starting point. Sampling points in The cumulative cost is From the candidate node set Sampling points in The lower bound function of the heuristic cost to the destination is: The total assessment cost is:

[0069] ;

[0070] in, It can be obtained from sampling points The lower bound of the Euclidean distance or path length to the destination is given.

[0071] In each round of expansion, the total evaluation cost is selected from the candidate viewpoint set. The tree structure is expanded using the smallest sampling point. Collision detection is performed on the newly generated edges. Edges that pass the collision detection are retained and the tree structure is updated. When a new total evaluation cost is found to be better than the upper bound of the current optimal path cost, the tree structure is expanded. When updating the optimal path and ellipsoid parameters, the iteration converges or reaches the preset iteration time limit. The ellipsoid parameters include the ellipsoid center, principal axis direction, major semi-axis length, and minor semi-axis length.

[0072] Then the discrete path point list obtained from the search As control points, N is a positive integer, and a smooth trajectory curve is generated using the B-spline interpolation method. :

[0073] ;

[0074] in, Let be a B-spline basis function, and let it be non-zero over a finite interval. For points in a discrete path point list, Indicates time.

[0075] Maintain the trajectory as first- or second-order continuous in position and direction, and then base it on the maximum speed of the robot platform. With maximum acceleration Constraints on smooth trajectory curves Perform time reparameterization to obtain an executable trajectory that satisfies dynamic constraints, and then apply the dynamic constraints:

[0076] ;

[0077] in, The first derivative of the trajectory is the velocity. The second derivative of the trajectory, i.e., the potential acceleration, It represents its norm.

[0078] Finally, an executable trajectory is generated. The trajectory will then be sent to the implementing agency.

[0079] Furthermore, the motion planner module can leverage heuristic sampling-based motion planning algorithms, rely on ellipsoidal constraints, and further combine trajectory smoothing and time parameterization techniques to ensure that the output trajectory meets the actual execution requirements in terms of spatial continuity, dynamic reachability, and obstacle avoidance safety. This significantly reduces search overhead and improves the stability and real-time performance of path generation.

[0080] Figure 3 The flowchart of the scanning viewpoint generation method for defect detection of key components of aerospace equipment provided by the present invention is as follows: Figure 3 As shown, the method includes:

[0081] Step S301: Obtain global point cloud data of key components of aerospace equipment, and perform spatial discretization and front-end extraction in the three-dimensional spatial model corresponding to the global point cloud data. Sample and generate a set of candidate viewpoints in the neighborhood of the pre-selected defect candidate region, and select the optimal next viewpoint based on the multi-objective utility function.

[0082] Step S302: Based on the motion planning algorithm of heuristic sampling, a sampling region is constructed with the current viewpoint as the starting point and the optimal next viewpoint as the ending point. Heuristic sampling search and collision detection are performed within the sampling region to generate an executable trajectory. The discrete path formed by connecting discrete pose points generated by the motion planning algorithm based on sampling nodes is smoothed and time parameterized. The sampling nodes are obtained by sampling the sampling region.

[0083] Step S303: Divide the three-dimensional space to be reconstructed into multiple sub-regions, allocate scanning resources according to region attention and information gain, and determine the scanning termination time based on the joint threshold condition of information gain of candidate viewpoints and attention gain of defective regions in the candidate viewpoint set, and output the final scanning result and viewpoint sequence.

[0084] The viewpoint generation scanning method for defect detection of key components of aerospace equipment provided by this invention further includes: rasterizing or voxelizing the current three-dimensional space, dividing the three-dimensional space into occupied voxels, free voxels, and unknown voxels, and defining the common boundary between free voxels and unknown voxels as the front set. ;in, For a set of free voxels, For an unknown set of voxels, This is the adjacency threshold. Indicates belonging to the set of free voxels The spatial position of any voxel Represents the spatial location of any voxel belonging to the unknown voxel set. express and The Euclidean distance between the center points; sampling or clustering the front set to obtain the front point set, sampling the defect candidate region to obtain the defect candidate point set, jointly sampling the position and orientation within the neighborhood radius centered on the defect candidate point set and the front point set to construct the candidate viewpoint set, and evaluating the candidate viewpoints based on a multi-objective utility function, which is: ;in, From the perspective The overall utility value, For information gain, For attention gain term, For the cost of movement, These are non-negative weight parameters.

[0085] The viewpoint generation and scanning method for defect detection of key components of aerospace equipment provided by this invention also includes a motion planning algorithm based on heuristic sampling. This algorithm generates an executable trajectory that satisfies motion constraints by performing heuristic sampling search and collision detection in an ellipsoidal sampling region, and then performs smoothing and time parameterization processing, including: using the current pose... With the optimal next viewpoint pose As a path endpoint, when there exists an upper bound on the current optimal path cost. At that time, construct the ellipsoidal region As the sampling area, then in the ellipsoidal region Batch random sampling is performed to obtain a sample set. Samples that fall into obstacles or violate safety clearance constraints are removed to obtain a candidate node set. Based on the candidate node set Construct a search tree or search graph and perform a heuristic search to obtain a list of discrete path points for executable trajectories. ; to list discrete path points As control points, B-splines or spline interpolation are used to generate smooth trajectory curves. To make the trajectory curve smooth The position and orientation are first- or second-order continuous, depending on the maximum speed of the robot platform. With maximum acceleration Constraints on smooth trajectory curves Time reparameterization is performed to obtain an executable trajectory that satisfies the dynamic constraints.

[0086] The viewpoint generation and scanning method for defect detection of key components of aerospace equipment provided by this invention also includes heuristic sorting and expansion of nodes and edges through a cost lower bound function, including: defining a set of candidate nodes from the starting point. Sampling points in The cumulative cost is From the candidate node set Sampling points in The lower bound function of the heuristic cost to the destination is: The total assessment cost is: ;in, It can be obtained from sampling points The lower bound of the Euclidean distance or path length to the destination is given; the total evaluation cost is selected from the candidate viewpoint set. The tree structure is expanded using the smallest sampling point. Collision detection is performed on the newly generated edges. Edges that pass the collision detection are retained and the tree structure is updated. The new total evaluation cost is then used to expand the tree structure if it is better than the upper bound of the current optimal path cost. When updating the optimal path and ellipsoid parameters, the iteration converges or reaches the preset iteration time limit. The ellipsoid parameters include the ellipsoid center, principal axis direction, major semi-axis length, and minor semi-axis length.

[0087] The viewpoint generation scanning method for defect detection of key components of aerospace equipment provided by this invention further includes: dividing the three-dimensional space to be reconstructed into multiple sub-regions. j is the index of the sub-region, and M is a positive integer. For each sub-region... Maintain its access count Cumulative gain and average gain Information gain is the average gain. Priority indices are calculated for each sub-region using the upper confidence boundary criterion. According to priority indicators Sort the corresponding sub-regions by size, and select the first one from largest to smallest. Each sub-region is selected as a high-attention region. Each sub-region is designated as a low-attention region. Candidate viewpoints are generated and trajectories are planned within high-attention regions. For high-attention regions, a first scanning mode is selected, and for low-attention regions, a second scanning mode is selected to allocate scanning resources from global to local. The top 10% of all sub-regions The last 10% of all sub-regions, where the scanning speed of the first scanning mode is lower than that of the second scanning mode, and the resolution of the first scanning mode is higher than that of the second scanning mode; define the... The information gain of the optimal viewpoint in the next scan is: The defective attention gain is When the conditions are met simultaneously in multiple consecutive scans When the current reconstruction information is determined to be saturated, a scan termination signal is triggered, the update of the viewpoint sequence and path planning is stopped, and the final 3D scan result and viewpoint sequence are output. To preset the information gain threshold, A preset defect attention gain threshold is set. It should be noted that the specific details and benefits of the viewpoint generation scanning method for defect detection of key components of aerospace equipment provided by this invention are similar to those of the viewpoint generation scanning system for defect detection of key components of aerospace equipment provided by this invention, and will not be repeated here.

[0088] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above embodiments. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention.

[0089] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not describe the various possible combinations separately.

[0090] Furthermore, various different implementations of the present invention can be combined arbitrarily, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed in the present invention.

Claims

1. A viewpoint generation scanning system for defect detection of key components in aerospace equipment, characterized in that, The system includes: The optimal next viewpoint generator module is used to acquire global point cloud data of key components of the aerospace equipment, perform spatial discretization and front-end extraction in the three-dimensional spatial model corresponding to the global point cloud data, sample and generate a set of candidate viewpoints in the neighborhood of a pre-selected defect candidate region, and select the optimal next viewpoint based on a multi-objective utility function. A motion planner module is used for a motion planning algorithm based on heuristic sampling. Starting from the current viewpoint and ending at the optimal next viewpoint, a sampling region is constructed. Within this region, heuristic sampling search and collision detection are performed to generate an executable trajectory. The module also smooths and parameters the discrete path formed by concatenating discrete pose points generated by the motion planning algorithm based on sampling nodes, wherein the sampling nodes are obtained by sampling the sampling region. The adaptive scanning and termination module is used to divide the 3D space to be reconstructed into multiple sub-regions, allocate scanning resources according to region attention and information gain, and determine the scanning termination time based on the joint threshold condition of information gain of candidate viewpoints and attention gain of defective regions in the candidate viewpoint set, and output the final scanning result and viewpoint sequence.

2. The viewpoint generation and scanning system for defect detection of key components of aerospace equipment according to claim 1, characterized in that, The optimal next viewpoint generator module is also used for: The current 3D space is rasterized or voxelized, dividing the 3D space into occupied voxels, free voxels, and unknown voxels, and the common boundary between the free voxels and the unknown voxels is defined as the front set: ； in, For a set of free voxels, For an unknown set of voxels, This is the adjacency threshold. Indicates belonging to the set of free voxels The spatial position of any voxel. Represents the spatial location of any voxel belonging to the unknown voxel set. express and The Euclidean distance between the center points; The front set is sampled or clustered to obtain a front point set, and the defect candidate region is sampled to obtain a defect candidate point set. Joint sampling of position and orientation is performed within a neighborhood radius centered on the defect candidate point set and the front point set to construct the candidate viewpoint set. The candidate viewpoints are then evaluated based on a multi-objective utility function, which is: ； in, From the perspective The overall utility value, For information gain, For attention gain term, For the cost of movement, These are non-negative weight parameters.

3. The viewpoint generation and scanning system for defect detection of key components of aerospace equipment according to claim 2, characterized in that, The motion planner module is further configured to employ a heuristic sampling-based motion planning algorithm, generating an executable trajectory that satisfies motion constraints by performing heuristic sampling search and collision detection within an ellipsoidal sampling region, and then performing smoothing and time parameterization processing, including: With current pose With the optimal next viewpoint pose As a path endpoint, when there exists an upper bound on the current optimal path cost. At that time, construct the ellipsoidal region As the sampling area, then in the ellipsoidal region Batch random sampling is performed to obtain a sample set. Samples that fall into obstacles or violate safety clearance constraints are removed to obtain a candidate node set. Based on the candidate node set Construct a search tree or search graph and perform a heuristic search to obtain a list of discrete path points for executable trajectories. ;as well as: The discrete path point list As control points, B-splines or spline interpolation are used to generate smooth trajectory curves. So that the smooth trajectory curve The position and orientation are first- or second-order continuous, depending on the maximum speed of the robot platform. With maximum acceleration Constraints on the smooth trajectory curve Time reparameterization is performed to obtain an executable trajectory that satisfies the dynamic constraints.

4. The viewpoint generation and scanning system for defect detection of key components of aerospace equipment according to claim 3, characterized in that, The motion planner module is also used to heuristically sort and expand nodes and edges using a lower bound function of cost, including: Define the set of candidate nodes from the starting point. sampling points in The cumulative cost is From the set of candidate nodes sampling points in The lower bound function of the heuristic cost to the destination is: The total assessment cost is: ； in, It can be obtained from sampling points The lower bound of the Euclidean distance or path length to the destination is given; Select the total evaluation cost from the candidate viewpoint set. The tree structure is expanded using the smallest sampling point. Collision detection is performed on the newly generated edges. Edges that pass the collision detection are retained and the tree structure is updated. The new total evaluation cost is then used to expand the tree structure if it is better than the upper bound of the current optimal path cost. When the iteration converges or reaches the preset iteration time limit, the optimal path and ellipsoid parameters are updated until the iteration converges or reaches the preset iteration time limit. The ellipsoid parameters include the ellipsoid center, principal axis direction, major semi-axis length and minor semi-axis length.

5. The viewpoint generation and scanning system for defect detection of key components of aerospace equipment according to claim 4, characterized in that, The adaptive scanning and termination module is also used for: The three-dimensional space to be reconstructed is divided into multiple sub-regions. j is the index of the sub-region, and M is a positive integer. For each sub-region Maintain its access count Cumulative gain and average gain The information gain is the average gain. ; Priority indices are calculated for each sub-region using the upper confidence boundary criterion. According to priority indicators Sort the corresponding sub-regions by size, and select the first one from largest to smallest. Each sub-region is selected as a high-attention region. Each sub-region is designated as a low-attention region. Candidate viewpoints are generated and trajectories are planned within the high-attention region. For the high-attention region, a first scanning mode is selected, and for the low-attention region, a second scanning mode is selected to allocate scanning resources from global to local. The top 10% of all sub-regions For the last 10% of all sub-regions, wherein the scanning speed of the first scanning mode is lower than that of the second scanning mode, and the resolution of the first scanning mode is higher than that of the second scanning mode; and: Definition of the first The information gain of the optimal viewpoint in the next scan is: The defective attention gain is When the conditions are met simultaneously in multiple consecutive scans When the current reconstruction information is determined to be saturated, a scan termination signal is triggered, the update of the viewpoint sequence and path planning is stopped, and the final 3D scan result and viewpoint sequence are output. To preset the information gain threshold, The preset defect attention gain threshold is used.

6. A viewpoint generation and scanning method for defect detection of key components in aerospace equipment, characterized in that, The method includes: The global point cloud data of the key components of the aerospace equipment is acquired, and spatial discretization and front-end extraction are performed in the three-dimensional spatial model corresponding to the global point cloud data. A set of candidate viewpoints is generated by sampling in the neighborhood of the pre-selected defect candidate region, and the optimal next viewpoint is selected based on the multi-objective utility function. A motion planning algorithm based on heuristic sampling constructs a sampling region starting from the current viewpoint and ending at the optimal next viewpoint. Within this region, heuristic sampling search and collision detection are performed to generate an executable trajectory. The discrete path, formed by concatenating discrete pose points generated by the motion planning algorithm based on sampling nodes, is then smoothed and time-parameterized. These sampling nodes are obtained by sampling the sampling region. The three-dimensional space to be reconstructed is divided into multiple sub-regions. Scanning resources are allocated according to regional attention and information gain. Based on the joint threshold condition of the information gain of candidate viewpoints and the attention gain of defective regions in the candidate viewpoint set, the scanning termination time is determined, and the final scanning result and viewpoint sequence are output.

7. The viewpoint generation and scanning method for defect detection of key components of aerospace equipment according to claim 6, characterized in that, The method also includes: The current 3D space is rasterized or voxelized, dividing the 3D space into occupied voxels, free voxels, and unknown voxels, and the common boundary between the free voxels and the unknown voxels is defined as the front set: ； in, For a set of free voxels, For an unknown set of voxels, This is the adjacency threshold. Indicates belonging to the set of free voxels The spatial position of any voxel. Represents the spatial location of any voxel belonging to the unknown voxel set. express and The Euclidean distance between the center points; The front set is sampled or clustered to obtain a front point set, and the defect candidate region is sampled to obtain a defect candidate point set. Joint sampling of position and orientation is performed within a neighborhood radius centered on the defect candidate point set and the front point set to construct the candidate viewpoint set. The candidate viewpoints are then evaluated based on a multi-objective utility function, which is: ； in, From the perspective The overall utility value, For information gain, For attention gain term, For the cost of movement, These are non-negative weight parameters.

8. The viewpoint generation and scanning method for defect detection of key components of aerospace equipment according to claim 7, characterized in that, The method also includes a motion planning algorithm based on heuristic sampling, which generates an executable trajectory that satisfies motion constraints by performing heuristic sampling search and collision detection in an ellipsoidal sampling region, and performs smoothing and time parameterization processing, including: With current pose With the optimal next viewpoint pose As a path endpoint, when there exists an upper bound on the current optimal path cost. At that time, construct the ellipsoidal region As the sampling area, then in the ellipsoidal region Batch random sampling is performed to obtain a sample set. Samples that fall into obstacles or violate safety clearance constraints are removed to obtain a candidate node set. Based on the candidate node set Construct a search tree or search graph and perform a heuristic search to obtain a list of discrete path points for executable trajectories. ;as well as: The discrete path point list As control points, B-splines or spline interpolation are used to generate smooth trajectory curves. So that the smooth trajectory curve The position and orientation are first- or second-order continuous, depending on the maximum speed of the robot platform. With maximum acceleration Constraints on the smooth trajectory curve Time reparameterization is performed to obtain an executable trajectory that satisfies the dynamic constraints.

9. The viewpoint generation and scanning method for defect detection of key components of aerospace equipment according to claim 8, characterized in that, This method also includes heuristic sorting and expansion of nodes and edges using a lower bound function of cost, including: Define the set of candidate nodes from the starting point. sampling points in The cumulative cost is From the set of candidate nodes sampling points in The lower bound function of the heuristic cost to the destination is: The total assessment cost is: ； in, It can be obtained from sampling points The lower bound of the Euclidean distance or path length to the destination is given; Select the total evaluation cost from the candidate viewpoint set. The tree structure is expanded using the smallest sampling point. Collision detection is performed on the newly generated edges. Edges that pass the collision detection are retained and the tree structure is updated. The new total evaluation cost is then used to expand the tree structure if it is better than the upper bound of the current optimal path cost. When the iteration converges or reaches the preset iteration time limit, the optimal path and ellipsoid parameters are updated until the iteration converges or reaches the preset iteration time limit. The ellipsoid parameters include the ellipsoid center, principal axis direction, major semi-axis length and minor semi-axis length.

10. The viewpoint generation and scanning method for defect detection of key components of aerospace equipment according to claim 9, characterized in that, The method also includes: The three-dimensional space to be reconstructed is divided into multiple sub-regions. j is the index of the sub-region, and M is a positive integer. For each sub-region Maintain its access count Cumulative gain and average gain The information gain is the average gain. ; Priority indices are calculated for each sub-region using the upper confidence boundary criterion. According to priority indicators Sort the corresponding sub-regions by size, and select the first one from largest to smallest. Each sub-region is selected as a high-attention region. Each sub-region is designated as a low-attention region. Candidate viewpoints are generated and trajectories are planned within the high-attention region. For the high-attention region, a first scanning mode is selected, and for the low-attention region, a second scanning mode is selected to allocate scanning resources from global to local. The top 10% of all sub-regions For the last 10% of all sub-regions, wherein the scanning speed of the first scanning mode is lower than that of the second scanning mode, and the resolution of the first scanning mode is higher than that of the second scanning mode; and: Definition of the first The information gain of the optimal viewpoint in the next scan is: The defective attention gain is When the conditions are met simultaneously in multiple consecutive scans When the current reconstruction information is determined to be saturated, a scan termination signal is triggered, the update of the viewpoint sequence and path planning is stopped, and the final 3D scan result and viewpoint sequence are output. To preset the information gain threshold, The preset defect attention gain threshold is used.

Images