Point cloud registration method for aero-engine blade
By employing Tent chaotic initialization, nonlinear weighting factors, and an adaptive reverse learning mechanism, the initial position dependency and local optimum problems in point cloud registration of aero-engine blades are solved, achieving high-precision and fast point cloud registration, which is applicable to point cloud registration of objects with complex shapes.
Patent Information
- Application Number
- CN202511060713.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-30
- Publication Date
- 2025-11-25
AI Technical Summary
Existing technologies for point cloud registration of aero-engine blades suffer from problems such as initial position dependence, local optima, and slow convergence speed, making it difficult to meet the requirements of high-precision manufacturing.
A triple mechanism of Tent chaotic initialization, nonlinear weighting factor and adaptive back learning is adopted to generate a uniformly distributed initial population, dynamically adjust the search step size and learning intensity, and balance global exploration and local optimization.
It significantly improves point cloud registration accuracy and convergence speed, avoids local optima trapping, and is suitable for point cloud registration of complex-shaped objects, thus improving manufacturing accuracy and efficiency.
Smart Images
Figure CN121010631A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of machine vision, and particularly relates to a point cloud registration method for an aero-engine blade. BACKGROUND
[0002] The aero-engine blade is a key part of the aero-engine, and its dimensional accuracy directly determines the performance of the engine. Since the blade works in a harsh environment of high temperature and high pressure, its manufacturing quality is crucial. The manufacturing process of the blade mainly relies on mechanical machining, and machining allowance detection needs to be performed at each machining stage to ensure that the final dimensional accuracy meets the design requirements. The rough machining blade point cloud data is preliminarily registered with the ideal model, laying a foundation for subsequent machining allowance calculation.
[0003] Rough registration is of great significance in blade machining and manufacturing: first, rough registration can quickly align the scanned blade point cloud data with the ideal model, providing an initial position reference for subsequent fine registration and machining allowance calculation, significantly reducing the computational complexity. Secondly, the accuracy of rough registration directly affects the setting of subsequent machining parameters. If the rough registration error is large or falls into a local optimum, it may lead to the failure of subsequent allowance calculation, or even cause the blade to be scrapped. Therefore, rough registration is not only a key step in blade manufacturing, but also an important guarantee for ensuring the dimensional accuracy and performance of the blade.
[0004] With the development of three-dimensional scanning technology, line laser scanning has become an important means to obtain high-precision point cloud data of the blade. However, the classical three-dimensional point cloud registration method often needs hundreds of iterations to converge when dealing with complex blade point cloud data, not only slow convergence speed, but also running speed decline due to large amount of calculation, and easy to fall into local optimal solution, leading to machining allowance calculation error exceeding the allowed range, and then affecting the final dimensional accuracy of the blade and the performance of the engine.
[0005] In recent years, registration methods based on evolutionary algorithms have been widely used in industrial fields due to their adaptive search capabilities. Among them, the whale algorithm has shown good performance in many optimization problems due to its strong global search ability, simple parameter setting, etc. However, the traditional whale algorithm mostly uses random generation when generating the initial population, and the initial solution set cannot traverse the entire space, and the search span adjustment mechanism is not flexible enough, leading to the algorithm easily converging too early or falling into local optimum in complex point cloud registration tasks, making it difficult to meet the requirements of high-precision blade manufacturing. Based on this, the application realizes more accurate and stable point cloud registration results through the optimized evolutionary algorithm. SUMMARY
[0006] The technical problem to be solved is: To overcome the shortcomings of existing technologies, this invention provides a point cloud registration method for aero-engine blades. Through a triple mechanism of Tent chaotic initialization, nonlinear weighting factor, and adaptive back learning, it systematically solves the problems of initial position dependence, local optima, and convergence speed in point cloud registration. This method is simple to operate, easy to implement, and applicable to point cloud registration in the aero-engine blade manufacturing process.
[0007] The technical solution of this invention is: a point cloud registration method for aero-engine blades, comprising the following steps: Tent chaotic mapping is used to generate a uniformly distributed initial population as the initial solution set for the whale algorithm; The search step size is dynamically adjusted by nonlinear weighting factors: the weighting factors are adjusted by nonlinear functions according to the ratio of the current iteration number to the maximum iteration number, so that a large search step size is maintained in the early stage of the iteration to quickly locate the optimal solution region, and the step size is reduced in the later stage of the iteration to perform fine search. Introducing an adaptive reverse learning mechanism to generate reverse solutions :
[0008] in, For the first The intensity of reverse learning in the next iteration For the current solution, This is the current optimal pose transformation matrix solution; Reverse learning strength Based on fitness change rate Dynamic adjustment:
[0009] in, To adjust the step size, and These represent the maximum and minimum values of the reverse learning intensity, respectively. The current fitness change rate threshold, The rate of change of fitness; The preprocessed blade scan point cloud and the ideal model point cloud are compared using the above steps to improve the whale algorithm to optimize the pose transformation matrix and output the optimal registration result. A further technical solution of the present invention is: the Tent chaotic mapping formula is:
[0010] in, Indicates the first n Generations of populations, Indicates the first n +1 generation individuals in the population, To control parameters, in . A further technical solution of the present invention is: the nonlinear weighting factor The expression for dynamically adjusting the search step size is as follows:
[0011] in, For the first Weighting factor for the next iteration and These are the maximum and minimum values of the weighting factor, respectively. This represents the current iteration number. The maximum number of iterations, This is a non-linear adjustment coefficient.
[0012] A further technical solution of the present invention is: the fitness change rate The calculation method is as follows:
[0013] in, Indicates the first The fitness value of the next iteration.
[0014] A further technical solution of the present invention is: the fitness change rate threshold. The calculation method is as follows:
[0015] in, For the first The fitness change rate threshold for the next iteration and These are the maximum and minimum values of the threshold, respectively.
[0016] A further technical solution of the present invention is: the adaptive threshold The parameters are set as follows: =0.1, =0.01, so that the threshold decreases linearly from 0.1 to 0.01 with the number of iterations.
[0017] A further technical solution of the present invention is: the parameter of the reverse learning intensity is set as follows: =0.6, =0.9, =0.1, =0.1, ensuring the algorithm stops converging when it stalls. < Enhance exploration capabilities. A further technical solution of the present invention is: the fitness function is defined as:
[0018] in, The number of points, To scan points in a point cloud, For the corresponding point in the target point cloud, is the Euclidean distance between corresponding points. A further technical solution of the present invention is: the pose update method of the whale algorithm includes: Initialize the population size and the maximum number of iterations K; The initial pose solution set is generated using the Tent chaotic mapping. Calculate the fitness value for each pose solution; Update weighting factors and reverse learning strength ; Choose from encirclement, search, or spiral update mechanisms based on the random number r and the coefficient A; Generate a reverse solution and select the optimal solution; If the termination condition is met, output the optimal pose transformation matrix; otherwise, return to the step of calculating the fitness value for each pose solution.
[0019] A point cloud registration system for aero-engine blades, comprising: Data acquisition module: includes a line laser sensor and a rotary table. The line laser sensor is fixed to the end of the robotic arm and scans the rough-machined blade along the Z direction. The rotary table drives the blade to rotate around the Z axis at set angular intervals to acquire multi-view point cloud data. Preprocessing module: performs filtering, downsampling, and stitching operations on the acquired point cloud data, and outputs the preprocessed scanned point cloud; Improved whale algorithm processing module: including: Chaotic initialization unit: Generating a uniformly distributed initial pose solution set using Tent chaotic mapping; Dynamic weighting unit: using non-linear weighting factors Adjust the search step size; Adaptive Backward Learning Unit: Generates Backward Solutions And dynamically adjust the learning intensity ; Registration output module: Aligns the scanned point cloud with the ideal model point cloud based on the optimal pose transformation matrix, and outputs the registration result; Control module: Coordinates the execution timing of each module, controls the indexing rotation of the rotary table, and triggers the line laser sensor to collect data synchronously.
[0020] Beneficial effects The beneficial effects of this invention are as follows: 1. This invention generates an ergodic initial population through Tent chaotic mapping, which improves the global coverage of the search space and effectively avoids the algorithm getting trapped in local optima. The data in Table 1 show that the registration accuracy of this method is 71.3% higher than the traditional whale algorithm, 24.7% higher than the PCA registration algorithm, 64.6% higher than the ISS+FPFH registration algorithm, 59.6% higher than the gray wolf algorithm, and 79.8% higher than the particle swarm algorithm, significantly improving the registration accuracy.
[0021] Table 1
[0022] 2. This invention introduces a nonlinear weighting factor and a dynamic span adjustment strategy, balancing global exploration and local development capabilities. In the early stages of iteration, the algorithm rapidly explores the entire search space; in the later stages of iteration, it gradually narrows the search range for fine-tuning. Table 1 shows that this method's running time is 17.6% faster than the traditional whale algorithm, and 3.7 times and 14.3 times faster than the particle swarm optimization algorithm and the gray wolf algorithm, respectively, significantly improving the algorithm's convergence speed.
[0023] 3. This invention dynamically adjusts the search direction through a reverse learning strategy, further enhancing the algorithm's global search capability and avoiding registration failures caused by initial position dependence or data noise. Figure 11 The results of point cloud registration using this method are shown in several extreme cases, none of which have fallen into local optima.
[0024] 4. This invention is not only applicable to point cloud registration of aero-engine blades, but can also be widely applied to point cloud registration of other complex-shaped objects, such as turbine blades and wing structures. In practical applications, this method demonstrates high versatility and stability. Attached Figure Description Figure 1 This is a flowchart illustrating the registration process of a certain aero-engine blade in an embodiment of the present invention.
[0025] Figure 2 This is a flowchart of the improved whale algorithm registration in an embodiment of the present invention.
[0026] Figure 3 These are schematic diagrams of four types of chaotic mapping distributions.
[0027] Figure 4 This is a diagram of the registration iteration error of aero-engine blades using the improved whale algorithm.
[0028] Figure 5 These are comparison images of point cloud registration before and after using the improved whale algorithm on aero-engine blades.
[0029] Figure 6 These are comparison images of point cloud registration before and after using the whale algorithm on aero-engine blades.
[0030] Figure 7 These are comparison images of point cloud registration before and after using the PCA algorithm on aero-engine blades.
[0031] Figure 8 These are comparison images of point cloud registration before and after using the ISS+FPFH algorithm on aero-engine blades.
[0032] Figure 9 These are comparison images of point cloud registration before and after using the Grey Wolf algorithm on aero-engine blades.
[0033] Figure 10 These are comparison images of point cloud registration before and after using particle swarm optimization algorithm on aero-engine blades.
[0034] Figure 11 These are comparison images of point cloud registration of aero-engine blades under three poor initial conditions using the improved whale algorithm.
[0035] Figure 12 These are comparison images of point cloud registration before and after using the improved whale algorithm on three other models. Detailed Implementation The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the invention, and should not be construed as limiting the invention.
[0036] Traditional whale algorithms often use random generation for initial population formation, resulting in an initial solution set that cannot traverse the entire space. Furthermore, the search span adjustment mechanism is not flexible enough, leading to premature convergence or getting trapped in local optima in complex point cloud registration tasks. This makes it difficult to meet the requirements of high-precision blade manufacturing. This invention provides a coarse registration method for aero-engine blade point clouds based on an improved whale algorithm, comprising the following steps: Tent chaotic mapping is used to generate a uniformly distributed initial population as the initial solution set for the whale algorithm; The search step size is dynamically adjusted by nonlinear weighting factors: the weighting factors are adjusted by nonlinear functions according to the ratio of the current iteration number to the maximum iteration number, so that a large search step size is maintained in the early stage of the iteration to quickly locate the optimal solution region, and the step size is reduced in the later stage of the iteration to perform fine search. An adaptive reverse learning mechanism is introduced to generate a reverse solution; Reverse learning strength Based on fitness change rate Dynamic adjustment: Based on the comparison between the fitness change rate and the threshold, the back learning strength is dynamically adjusted. When the fitness change rate is less than the threshold, the back learning strength is increased; otherwise, the back learning strength is decreased, in order to balance global exploration and local optimization. The preprocessed blade scan point cloud and the ideal model point cloud are compared using the above steps to improve the whale algorithm to optimize the pose transformation matrix and output the optimal registration result.
[0037] This invention also proposes a point cloud registration system for aero-engine blades, comprising: Data acquisition module: includes a line laser sensor and a rotary table. The line laser sensor is fixed to the end of the robotic arm and scans the rough-machined blade along the Z direction. The rotary table drives the blade to rotate around the Z axis at set angular intervals to acquire multi-view point cloud data. Preprocessing module: performs filtering, downsampling, and stitching operations on the acquired point cloud data, and outputs the preprocessed scanned point cloud; Improved whale algorithm processing module: including: Chaotic initialization unit: Generating a uniformly distributed initial pose solution set using Tent chaotic mapping; Dynamic weighting unit: using non-linear weighting factors Adjust the search step size; Adaptive Backward Learning Unit: Generates Backward Solutions And dynamically adjust the learning intensity ; Registration output module: Aligns the scanned point cloud with the ideal model point cloud based on the optimal pose transformation matrix, and outputs the registration result; Control module: Coordinates the execution timing of each module, controls the indexing rotation of the rotary table, and triggers the line laser sensor to collect data synchronously.
[0038] The above technical solution will be further explained below with reference to the accompanying drawings and examples: In one embodiment, refer to Figure 1 As shown, the point cloud registration method based on the optimized whale algorithm in this embodiment mainly includes the following steps: Step 1. Acquire blade scanning data A robot equipped with a line laser sensor moves along the Z-axis to scan the blade. A turntable rotates the blade around the Z-axis, collecting a set of point cloud data every 30° of rotation, resulting in 12 perspectives of blade point clouds. Preprocessing techniques such as filtering and downsampling are used to reduce noise and accelerate computation. A calibration sphere is used to establish the transformation relationship between the robot's base coordinate system and the sensor's coordinate system, achieving initial alignment of the multi-view point clouds. Further feature extraction is then employed to achieve precise stitching.
[0039] The point cloud obtained by multi-view stitching is downsampled, and fitness thresholds, population size, and maximum number of iterations are set before proceeding. Figure 2 The point cloud initial registration shown is based on the improved whale algorithm.
[0040] The registration method involves solving the four-dimensional pose transformation matrix. accomplish
[0041] in For rotation matrix, It is a translation matrix.
[0042] Step 2. Generate an initial solution using the Tent chaotic map. In the initial stage of the optimization algorithm, the population generation strategy directly affects the final solution. Traditional whale optimization algorithms use random initialization to generate the pose solution set, resulting in non-uniform clustering in the search space, which limits global exploration capabilities. To improve this situation, this invention introduces a commonly used chaotic mapping to generate a population such as... Figure 3 The distribution map shown.
[0043] 1. Logistic mapping, its formula is as follows:
[0044] in This represents the population of generation n. This represents an individual in the (n+1)th generation of the population. To control the parameters, the values are taken in the range [0,4]. In this example, r=3.8.
[0045] 2. Tent mapping, its formula is as follows:
[0046] in As a control parameter, it takes a value in [0,1]. In this example, r=0.75.
[0047] 3. Chebyshev mapping, its formula is as follows:
[0048] in The parameter is set to a value in the range [1,2]. In this example, r=1.9.
[0049] 4. Kent mapping, its formula is as follows:
[0050] in To control the parameters, the values are taken in the range [0,4]. In this example, r=3.8.
[0051] Depend on Figure 3 The initial solutions generated by the Tent mapping have the most uniform distribution, so this experiment uses the Tent mapping to generate the initial population.
[0052] The method continuously optimizes the rotation parameters through individual operations in the whale algorithm, ultimately finding the optimal rotation matrix that minimizes the error between the scanned point cloud and the ideal point cloud. The position update formula during the whale encirclement phase is as follows:
[0053]
[0054]
[0055]
[0056] in, For the current solution, This represents the current optimal solution, where A and C are coefficients. It decreases linearly in the interval [0,2]. Let A be a random number within the interval [0, 1]. The range of A is... .
[0057] when Update the location during the search phase.
[0058]
[0059] in, The location of the whale is random.
[0060] when The position update formula for the spiral update stage is as follows:
[0061]
[0062] Let be a constant, and define the shape of the helix. It is a random number between [-1, 1].
[0063] The three optimization methods of the whale algorithm are as follows:
[0064] The formula for the nonlinear weighting factor is as follows:
[0065] in, For the first Weighting factor for the next iteration and These are the maximum and minimum values of the weighting factor, respectively. This represents the current iteration number. The maximum number of iterations, For non-linear adjustment coefficients, the maximum number of iterations in this example is... , This example achieves a balance between global search and local optimization by dynamically adjusting the weight factors.
[0066] The formula for generating the inverse solution is:
[0067] in For the first The strength of the reverse learning in each iteration.
[0068] The formula for back-learning strength update is:
[0069] in To adjust the step size, and These represent the maximum and minimum values of the reverse learning intensity, respectively. In this example, Set to 0.6, Set to 0.9, Set to 0.1, Set it to 0.1. The current fitness change rate threshold, This represents the rate of change in fitness.
[0070]
[0071] in Indicates the first Fitness value of the next iteration, fitness change rate threshold for:
[0072] in, For the first The fitness change rate threshold for the next iteration and These are the maximum and minimum values of the threshold, respectively. In this example, Set to 0.1, Set to 0.01.
[0073] The fitness function of the method Defined as:
[0074] in, The number of points, To scan points in a point cloud, For the corresponding point in the target point cloud, This represents the Euclidean distance between the corresponding points. In this example, the corresponding points are generated by finding the closest point in the target point cloud for each scanned point cloud based on a KD-tree.
[0075] Combination Figure 2 The improved whale algorithm optimization process used in this example is as follows: (1) Initialization parameters: Set the whale population size and the maximum number of iterations K; (2) Population initialization: The initial pose solution set is generated using the Tent chaotic mapping; (3) Fitness calculation: For each pose solution, calculate its fitness value; (4) Dynamic parameter update: Update the weight factor according to the number of iterations k. And adjust the reverse learning intensity ; (5) Location update: based on The value and the A value determine the update mechanism of the whale algorithm; (6) Generate inverse solution Re-compare the fitness values and select the optimal solution; (7) Termination judgment: When the fitness value is less than the threshold F or the number of iterations reaches K, output the optimal pose matrix and reverse the learning intensity. convergence to Otherwise, return to step (3).
[0076] The 3D measurement and preliminary registration process is as follows: (1) The robot is equipped with a line laser to scan the rough-machined blade along the Z direction to obtain a 0° angle scanning point cloud; (2) The turntable drives the blades to rotate around the Z-axis, and collects point cloud data every 30°. (3) Filter and combine the point clouds from each viewpoint; (4) Set the sampling voxel grid distance and downsample the scanned point cloud and the target point cloud; (5) Set the registration fitness threshold F to prepare for subsequent data analysis and processing; (6) Initial registration was performed using the improved whale algorithm.
[0077] To verify the effectiveness of the present invention, the following experiments were conducted: A line laser sensor was used to scan the blades of an aero-engine, acquiring point cloud data, which was then registered with an ideal model. The iteration error is as follows: Figure 4As shown, the algorithm achieves stable convergence within 50 iterations, without encountering any local optima. The iteration curves show that the initial larger step size of the nonlinear weight factor accelerates the convergence speed, while later fine-tuning is employed. The inclusion of the back-learning mechanism also prevents the algorithm from getting stuck in local optima. A comparison before and after registration is provided. Figure 5 As shown.
[0078] The algorithm also performed excellently in other point cloud tests: multiple registration simulation experiments and repeated experiments were conducted using three sets of point clouds with different shapes, poses, and sizes, and the comparison before and after registration was shown in the figure. Figure 12 As shown in the figure. Experimental results demonstrate that the algorithm achieves a 100% convergence success rate, showcasing its robustness.
[0079] The visual registration results clearly demonstrate that this method effectively balances global exploration and local development capabilities through the synergistic effect of chaotic initialization, dynamic weighting factors, and reverse learning, providing a stable and efficient solution for registration of complex industrial point clouds.
[0080] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.
Claims
1. A point cloud registration method for aero-engine blades, characterized in that... Includes the following steps: Tent chaotic mapping is used to generate a uniformly distributed initial population as the initial solution set for the whale algorithm; The search step size is dynamically adjusted by nonlinear weighting factors: the weighting factors are adjusted by nonlinear functions according to the ratio of the current iteration number to the maximum iteration number, so that a large search step size is maintained in the early stage of the iteration to quickly locate the optimal solution region, and the step size is reduced in the later stage of the iteration to perform fine search. Introducing an adaptive reverse learning mechanism to generate reverse solutions : in, For the first The intensity of reverse learning in the next iteration For the current solution, This is the current optimal pose transformation matrix solution; Reverse learning strength Based on fitness change rate Dynamic adjustment: in, To adjust the step size, and These represent the maximum and minimum values of the reverse learning intensity, respectively. The current fitness change rate threshold, The rate of change of fitness; The preprocessed blade scan point cloud and the ideal model point cloud are compared using the above steps to improve the whale algorithm to optimize the pose transformation matrix and output the optimal registration result.
2. The point cloud registration method for aero-engine blades according to claim 1, characterized in that: The Tent chaotic mapping formula is: in, Indicates the first n Generations of populations, Indicates the first n +1 generation individuals in the population, To control parameters, in .
3. The point cloud registration method for aero-engine blades according to claim 2, characterized in that: The nonlinear weighting factor The expression for dynamically adjusting the search step size is as follows: in, For the first Weighting factor for the next iteration and These are the maximum and minimum values of the weighting factor, respectively. This represents the current iteration number. The maximum number of iterations, This is a non-linear adjustment coefficient.
4. The point cloud registration method for aero-engine blades according to claim 3, characterized in that: The fitness change rate The calculation method is as follows: in, Indicates the first The fitness value of the next iteration.
5. The point cloud registration method for aero-engine blades according to claim 4, characterized in that: The fitness change rate threshold The calculation method is as follows: in, For the first The fitness change rate threshold for the next iteration and These are the maximum and minimum values of the threshold, respectively.
6. The point cloud registration method for aero-engine blades according to claim 5, characterized in that: The adaptive threshold The parameters are set as follows: =0.1, =0.01, so that the threshold decreases linearly from 0.1 to 0.01 with the number of iterations.
7. The point cloud registration method for aero-engine blades according to claim 6, characterized in that: The parameters for the reverse learning strength are set as follows: =0.6, =0.9, =0.1, =0.1, ensuring the algorithm stops converging when it stalls. < Enhance exploration capabilities.
8. The point cloud registration method for aero-engine blades according to claim 5, characterized in that: The fitness function is defined as follows: in, For the number of points, To scan points in a point cloud, For the corresponding point in the target point cloud, is the Euclidean distance between corresponding points.
9. A point cloud registration method for aero-engine blades according to any one of claims 1-8, characterized in that: The pose update method of the whale algorithm includes: Initialize population size and maximum number of iterations K ; The initial pose solution set is generated using the Tent chaotic mapping. Calculate the fitness value for each pose solution; Update weighting factors and reverse learning strength ; Choose from encirclement, search, or spiral update mechanisms based on the random number r and the coefficient A; Generate a reverse solution and select the optimal solution; If the termination condition is met, output the optimal pose transformation matrix; otherwise, return to the step of calculating the fitness value for each pose solution.
10. A point cloud registration system for aero-engine blades, used to implement the point cloud registration method for aero-engine blades as described in any one of claims 1-9; characterized in that, include: Data acquisition module: includes a line laser sensor and a rotary table. The line laser sensor is fixed to the end of the robotic arm and scans the rough-machined blade along the Z direction. The rotary table drives the blade to rotate around the Z axis at set angular intervals to acquire multi-view point cloud data. Preprocessing module: performs filtering, downsampling, and stitching operations on the acquired point cloud data, and outputs the preprocessed scanned point cloud; Improved whale algorithm processing module: including: Chaotic initialization unit: Generating a uniformly distributed initial pose solution set using Tent chaotic mapping; Dynamic weighting unit: using non-linear weighting factors Adjust the search step size; Adaptive Backward Learning Unit: Generates Backward Solutions And dynamically adjust the learning intensity ; Registration output module: Aligns the scanned point cloud with the ideal model point cloud based on the optimal pose transformation matrix, and outputs the registration result; Control module: Coordinates the execution timing of each module, controls the indexing rotation of the rotary table, and triggers the line laser sensor to collect data synchronously.