A laser processing adaptive path planning method and system based on probe data analysis

Abstract

The application relates to the technical field of intelligent control, in particular to a laser processing adaptive path planning method and system based on probe data analysis. The laser processing adaptive path planning system based on probe data analysis comprises a point cloud data acquisition module, a point cloud data labeling module, a point cloud data noise reduction module, a filter parameter vector adjustment module and a laser processing path determination module. The application carries out structure labeling and region division on point cloud data of a workpiece to be processed through a multi-dimensional feature vector, adaptively configures filter parameters according to region labels, introduces a self-feedback noise reduction strategy driven by fitting residual error, constructs a fitting surface and calculates point cloud residual error distribution, realizes intelligent identification and local complex filtering adjustment of high residual error aggregation regions and low error edge regions, so that the overall noise reduction process has convergence and local sensitivity control ability, and the problem that structure distortion is caused by insufficient noise reduction or excessive smoothing is avoided.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, specifically to a laser processing adaptive path planning method and system based on probe data analysis. Background Technology

[0002] With the development of intelligent manufacturing technology, laser cutting, welding, marking, and other processing methods are widely used in fields such as metals and composite materials. To improve processing accuracy and automation, a probe is commonly used to perform 3D scanning of the workpiece to obtain its point cloud data, which is then used to guide subsequent laser path planning. However, in actual industrial environments, the measured point cloud data is usually affected by various factors, resulting in significant errors and noise.

[0003] Specifically, highly reflective materials such as metals are prone to generating strong reflection points and laser drift during laser scanning, resulting in localized uneven density of point cloud data or the appearance of "floating points." Furthermore, vibrations at the processing site or insufficient probe clamping precision often cause scanning misalignment or point cloud ghosting, leading to inaccurate structural information. On the other hand, in high-temperature processing areas, such as near weld start and end points or power hotspots, hotspot interference is likely to occur, causing dense distortion or non-physical jumps in the point cloud in that area.

[0004] The above problems not only affect the structural consistency and fitting quality of point cloud data, but also seriously interfere with the subsequent processing path extraction process, which can easily cause trajectory deviation, structural misjudgment or processing damage. Summary of the Invention

[0005] This invention uses multi-dimensional feature vectors to perform structural annotation and region division on the point cloud data of the workpiece to be processed. It adaptively configures filtering parameters based on region labels and introduces a self-feedback noise reduction strategy driven by fitting residuals. By constructing a fitting surface and calculating the point cloud residual distribution, it achieves intelligent identification and local re-filtering adjustment of high residual clustering regions and low error edge regions. This enables the overall noise reduction process to have convergence and local sensitivity control capabilities, avoiding structural distortion caused by insufficient noise reduction or excessive smoothing. This effectively avoids error propagation problems caused by factors such as reflection, misalignment, and hot spots.

[0006] This invention provides an adaptive path planning method for laser processing based on probe data analysis, comprising:

[0007] Step S1: Obtain the point cloud dataset of the workpiece to be processed using the probe;

[0008] Step S2: Determine the multidimensional feature vector corresponding to each point cloud data based on the point cloud dataset, label the point cloud dataset based on all multidimensional feature vectors, and add corresponding region labels and noise probabilities to the point cloud data;

[0009] Step S3: Determine the filtering parameter vector based on the region label corresponding to the point cloud data. The filtering parameter vector includes the filtering kernel size and the filtering kernel function weight. Perform noise reduction operation on all point cloud data in the point cloud dataset based on the corresponding filtering parameter vector to obtain the reconstructed point cloud dataset.

[0010] Step S4: Perform surface reconstruction based on all point cloud data in the reconstructed point cloud dataset, then project the point cloud data in the dataset onto the simulated workpiece surface to obtain the corresponding simulated point cloud data. Calculate the distance between the point cloud data and the corresponding simulated point cloud data as the point cloud residual corresponding to the point cloud data. Calculate the root mean square error of the point cloud residuals corresponding to all point cloud data as the fluctuation amplitude value. If the standard deviation of the fluctuation amplitude value of the current iteration and the fluctuation amplitude value of the previous N-1 iterations is less than the fluctuation threshold, output the reconstructed point cloud dataset output by the previous iteration as the optimized point cloud dataset and proceed to step S6; otherwise, record the reconstructed point cloud dataset output by the previous iteration as the point cloud dataset and proceed to step S6.

[0011] Step S5: Based on the point cloud residuals corresponding to the point cloud data in the point cloud dataset, determine the high residual clustering region and the low error edge region. Adjust the filtering parameter vectors corresponding to the point cloud data in the high residual clustering region and the low error edge region, and perform noise reduction operation on the point cloud data in the high residual clustering region and the low error edge region. Update the point cloud dataset to obtain the reconstructed point cloud dataset. Record the reconstructed point cloud dataset before the noise reduction operation as the point cloud dataset, and return to step S4.

[0012] Step S6: Determine the laser processing path based on the optimized point cloud dataset.

[0013] Preferably, the point cloud data is subjected to noise reduction, which specifically includes the following operations:

[0014] The point cloud data is denoted as the target point cloud data. Based on the filter kernel size r in the filter parameter vector corresponding to the target point cloud data, all point cloud data whose distance to the target point cloud data is within the range of r are selected to form a neighborhood set.

[0015] For the i-th point cloud data in the neighborhood set, calculate the filtering weight W corresponding to the i-th point cloud data. i =f1(D i )·f2(P i )·(1-Z i ) γ Where f1 and f2 are both Gaussian kernel functions, specifically: f1(D i ) = exp(-(D i ) 2 / 2α 2 ), f2(P i=exp(-(1-(P)) i ) 2 ) / 2β 2 ), D i Let P be the distance between the i-th point cloud data in the neighborhood set and the target point cloud data. i Z is the dot product between the normal vector corresponding to the i-th point cloud data in the neighborhood set and the normal vector corresponding to the target point cloud data, where α is the distance weight, β is the normal consistency weight, and Z is the distance weight. i γ represents the noise probability corresponding to the target point cloud data, α, β and γ form the filter kernel function weights in the filter parameter vector;

[0016] The noise reduction point cloud data is obtained by weighting all point cloud data in the neighborhood set with their corresponding filter weights and then dividing by the sum of all filter weights. The noise reduction point cloud data is then used to replace the target point cloud data to complete the noise reduction operation.

[0017] Preferably, determining high residual clustering regions and low error edge regions based on the point cloud residuals corresponding to the point cloud data in the point cloud dataset specifically includes the following steps:

[0018] Point cloud data with residuals higher than the residual threshold are grouped into a high residual set. The residual threshold is set manually to reflect the degree of insufficient noise reduction. Connectivity component extraction is performed on the high residual set to construct several connected sets. Each connected set contains several point cloud data. The region corresponding to the connected set with a number of point cloud data exceeding the clustering threshold is selected as the high residual clustering region.

[0019] Select all point cloud data labeled as edge regions to form an edge point set. Calculate the root mean square error of the point cloud residuals in the edge point set as the edge fluctuation amplitude value. If the edge fluctuation amplitude value is lower than the fluctuation amplitude value at the current iteration, the region corresponding to the edge point set is recorded as a low-error edge region.

[0020] Preferably, the filter parameter vectors corresponding to the point cloud data in the high residual clustering region and the low error edge region are adjusted, specifically including the following steps:

[0021] For point cloud data within high residual clustering regions, a Hadamard product operation is performed between the filtering parameter vector used in the previous iteration of the point cloud data during the denoising operation and the high residual clustering adjustment vector. This completes the adjustment of the filtering parameter vector corresponding to the point cloud data within the high residual clustering regions, and the adjusted filtering parameter vector is used for the denoising operation in the current iteration.

[0022] For point cloud data in the low-error edge region, perform a Hadamard product operation between the filter parameter vector used in the previous iteration of the noise reduction operation and the low-error edge region adjustment vector to complete the adjustment of the filter parameter vector corresponding to the point cloud data in the low-error edge region. The adjusted filter parameter vector is used for the noise reduction operation in the current iteration.

[0023] Both the high residual clustering adjustment vector and the low error edge region adjustment vector were obtained by simulation using the sparrow search algorithm.

[0024] Preferably, the point cloud dataset is labeled based on all multidimensional feature vectors, and corresponding region labels and noise probabilities are added to the point cloud data. Specifically, this includes the following: the multidimensional feature vectors corresponding to the point cloud data are fed into the region labeling network and the noise analysis network for processing to obtain the corresponding region labels and noise probabilities. The region labels include flat regions, curvature regions and edge regions. The region labeling network and the noise analysis network are both based on BP neural networks and include an input layer, a hidden layer and an output layer.

[0025] Preferably, the filtering parameter vector is determined based on the region label corresponding to the point cloud data. Specifically, this includes the following: concatenating the point cloud data and the corresponding region label end to end to obtain the filtering parameter analysis vector, and then feeding the filtering parameter analysis vector into the filtering parameter analysis network for processing to obtain the filtering parameter vector. The filtering parameter analysis network is also based on a BP neural network and includes an input layer, a hidden layer, and an output layer.

[0026] This invention also provides a laser processing adaptive path planning system based on probe data analysis, comprising:

[0027] The point cloud data acquisition module is used to acquire the point cloud dataset of the workpiece to be processed through the probe;

[0028] The point cloud data annotation module is used to determine the multi-dimensional feature vector corresponding to each point cloud data based on the point cloud dataset, annotate the point cloud dataset based on all multi-dimensional feature vectors, and add corresponding region labels and noise probabilities to the point cloud data.

[0029] The point cloud data denoising module is used to perform denoising operations on all point cloud data in the point cloud dataset based on the corresponding filter parameter vector to obtain the reconstructed point cloud dataset.

[0030] The filter parameter vector adjustment module is used to determine high residual clustering regions and low error edge regions based on the point cloud residuals corresponding to the point cloud data in the point cloud dataset, and to adjust the filter parameter vectors corresponding to the point cloud data in the high residual clustering regions and low error edge regions.

[0031] The laser processing path determination module is used to determine the laser processing path based on an optimized point cloud dataset.

[0032] The present invention has the following advantages:

[0033] This invention uses multi-dimensional feature vectors to perform structural annotation and region division on the point cloud data of the workpiece to be processed. It adaptively configures filtering parameters based on region labels and introduces a self-feedback noise reduction strategy driven by fitting residuals. By constructing a fitting surface and calculating the point cloud residual distribution, it achieves intelligent identification and local re-filtering adjustment of high residual clustering regions and low error edge regions. This enables the overall noise reduction process to have convergence and local sensitivity control capabilities, avoiding structural distortion caused by insufficient noise reduction or excessive smoothing. This effectively avoids error propagation problems caused by factors such as reflection, misalignment, and hot spots. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the adaptive path planning system for laser processing based on probe data analysis used in an embodiment of the present invention. Detailed Implementation

[0035] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of this invention.

[0036] Example 1: A laser processing adaptive path planning method based on probe data analysis, comprising:

[0037] Step S1: Obtain the point cloud dataset of the workpiece to be processed through the probe. The point cloud dataset includes the point cloud data corresponding to several workpieces to be processed. The point cloud data includes the three-dimensional coordinates in the coordinate system constructed by the probe. The point cloud dataset reflects the size of the workpiece to be processed and also provides a data basis for the path planning of laser processing. It should be noted that this embodiment is for a five-axis CNC laser processing machine tool.

[0038] Step S2: Determine the multidimensional feature vector corresponding to each point cloud data based on the point cloud dataset. The multidimensional feature vector includes normal vector, curvature, normal uniformity, and point density. The multidimensional feature vector can reflect the feature information of the corresponding point in the point cloud data, and can be used to determine the region to which the corresponding point cloud belongs. The point cloud dataset is labeled based on all multidimensional feature vectors, and corresponding region labels and noise probabilities are added to the point cloud data. Here, labeling refers to the local structure recognition of the point cloud. For example, point clouds with low curvature, low normal uniformity, and uniform point density belong to flat regions. These flat regions are less affected by noise, and larger-scale filtering operations can be performed to reduce noise. The noise probability reflects the probability that the corresponding point belongs to noise. It should be noted that the determination of the multidimensional feature vector corresponding to each point cloud data based on the point cloud dataset can be achieved through the point cloud algorithm built into the PCL library.

[0039] This section describes the annotation of the point cloud dataset based on all multidimensional feature vectors, adding corresponding region labels and noise probabilities to the point cloud data. Specifically, this includes the following:

[0040] The multidimensional feature vectors corresponding to the point cloud data are fed into the region labeling network and the noise analysis network for processing to obtain the corresponding region labels and noise probabilities. The region labels include flat regions, curvature regions, and edge regions. Flat regions refer to areas in the point cloud with consistent local normal vectors, near-zero curvature, and uniform density. They typically correspond to smooth planes, weld bead areas, and the main body of the sheet metal of the workpiece to be processed. Flat regions can be filtered using large-scale kernels. Curvature regions refer to curved areas or areas near concave or convex structures with moderate curvature values ​​and obvious surface undulations but not boundary jumps. They are commonly found in the cavities, hole walls, and curved weld seams of the workpiece to be processed. Edge regions refer to areas at workpiece boundaries, surface discontinuities, areas with local normal jumps, severe curvature, and abrupt changes in point density. The region labels are numerical forms of word embeddings, for example, "100" for flat regions, "010" for curvature regions, and "001" for edge regions.

[0041] Both the region labeling network and the noise analysis network are based on BP neural networks, which include an input layer, a hidden layer, and an output layer.

[0042] Step S3: Determine the filtering parameter vector based on the region label corresponding to the point cloud data. The filtering parameter vector includes the filtering kernel size (i.e., neighborhood radius) and the filtering kernel function weight. The filtering kernel size represents the range of point clouds selected during the denoising operation, and the filtering kernel function weight represents the weight ratio of the influence of different filtering kernel functions and noise probability during the denoising operation. Based on the corresponding filtering parameter vector, perform denoising operation on all point cloud data in the point cloud dataset to obtain the reconstructed point cloud dataset.

[0043] The filtering parameter vector is determined based on the region labels corresponding to the point cloud data, specifically including the following:

[0044] After concatenating the point cloud data and the corresponding region labels, the filter parameter analysis vector is obtained. The filter parameter analysis vector is then fed into the filter parameter analysis network for processing to obtain the filter parameter vector.

[0045] The filter parameter analysis network is also based on a BP neural network, including an input layer, a hidden layer, and an output layer;

[0046] Noise reduction of point cloud data includes the following operations:

[0047] The point cloud data is denoted as the target point cloud data. Based on the filter kernel size r in the filter parameter vector corresponding to the target point cloud data, all point cloud data whose distance to the target point cloud data is within the range of r are selected to form a neighborhood set.

[0048] For the i-th point cloud data in the neighborhood set, calculate the filtering weight W corresponding to the i-th point cloud data. i =f1(D i )·f2(P i )·(1-Z i ) γ Where f1 and f2 are both Gaussian kernel functions, specifically: f1(D i ) = exp(-(D i ) 2 / 2α 2 ), f2(P i =exp(-(1-(P)) i ) 2 ) / 2β 2 ), D i Let P be the distance between the i-th point cloud data in the neighborhood set and the target point cloud data. i Z is the dot product between the normal vector corresponding to the i-th point cloud data in the neighborhood set and the normal vector corresponding to the target point cloud data, where α is the distance weight, β is the normal consistency weight, and Z is the distance weight. i γ represents the noise probability corresponding to the target point cloud data, α, β and γ form the filter kernel function weights in the filter parameter vector;

[0049] The noise reduction point cloud data is obtained by weighting and summing all point cloud data in the neighborhood set with their corresponding filter weights and then dividing by the sum of all filter weights. The noise reduction point cloud data is then used to replace the target point cloud data to complete the noise reduction operation.

[0050] Step S4: Perform surface reconstruction based on all point cloud data in the reconstructed point cloud dataset. The surface reconstruction method can be MLS or PCA. Construct a simulated workpiece surface, and then project the point cloud data from the point cloud dataset onto the simulated workpiece surface to obtain the corresponding simulated point cloud data. The projection method here is normal projection, that is, projecting the point cloud data onto the simulated workpiece surface according to the normal vector direction of the point cloud data. Find the point cloud data corresponding to the point closest to the point cloud data. Calculate the distance between the point cloud data and the corresponding simulated point cloud data as the point cloud residual corresponding to the point cloud data. Calculate the mean square of the point cloud residuals corresponding to all point cloud data. The root error is used as the fluctuation amplitude value, which represents the magnitude of the fitting error between the points in the entire point cloud dataset and the simulated workpiece surface. If the standard deviation of the fluctuation amplitude value of the current iteration and the fluctuation amplitude values ​​of the previous N-1 iterations is less than the fluctuation threshold, it indicates that after multiple rounds of iteration denoising, the denoising effect has converged to the optimal equilibrium state. The reconstructed point cloud dataset output by the previous iteration is output as the optimized point cloud dataset, and the process proceeds to step S6. Otherwise, the reconstructed point cloud dataset output by the previous iteration is recorded as the point cloud dataset, and the process proceeds to step S6. It should be noted that each denoising operation is considered as completing one iteration.

[0051] Step S5: Based on the point cloud residuals corresponding to the point cloud data in the point cloud dataset, determine the high residual clustering region and the low error edge region. The high residual clustering region refers to a region where multiple points have significantly larger residuals compared to the simulated workpiece surface after denoising, and these points are spatially close to each other, exhibiting a clustered state. The appearance of the high residual clustering region is clearly due to insufficient denoising. The low error edge region refers to edge points that should have geometric abrupt changes (such as abrupt changes in normal or curvature), but whose fitting residuals are very small after denoising, indicating that the structure has been "over-smoothed." For high residual... The filtering parameter vectors corresponding to the point cloud data in the high residual clustering region and the low error edge region are adjusted, and the point cloud data in the high residual clustering region and the low error edge region are denoised. The point cloud dataset is then updated to obtain the reconstructed point cloud dataset. The reconstructed point cloud dataset before the denoising operation is recorded as the point cloud dataset, and the process returns to step S4. It should be noted that when updating the point cloud dataset, the denoising operation is performed on the point cloud data located in the high residual clustering region and the low error edge region. The other point cloud data is not processed.

[0052] Based on the point cloud residuals corresponding to the point cloud data in the point cloud dataset, high residual clustering regions and low error edge regions are determined, specifically including the following steps:

[0053] Point cloud data with residuals higher than a residual threshold are grouped into a high residual set. The residual threshold is manually set to reflect the degree of insufficient noise reduction. Connected component extraction is performed on the high residual set to construct several connected sets, each containing a number of point cloud data. The region corresponding to the connected set with a number of point cloud data exceeding a clustering threshold is selected as the high residual clustering region. The clustering threshold is manually set to complement the residual threshold in reflecting the degree of insufficient noise reduction. Connected component extraction specifically includes: constructing a 3D voxel model based on all point cloud data in the high residual set; mapping all point cloud data in the high residual set to the 3D voxel model; starting from any voxel in the 3D voxel model, scanning its neighboring voxels (one circle of the cube); and forming connected sets from the point cloud data within the neighboring voxels. For each voxel containing point cloud data, the neighboring voxels are queried until no point cloud data is found, thus completing the construction of a connected set. Then, a voxel containing point cloud data not in a connected set is selected.

[0054] Select all point cloud data labeled as edge regions to form an edge point set. Calculate the root mean square error of the residuals of all point clouds in the edge point set as the edge fluctuation amplitude value. If the edge fluctuation amplitude value is lower than the fluctuation amplitude value at the current iteration, it indicates that the point cloud data in the edge point set is abnormally denoised, causing the point cloud data in the edge point set to be overly smoothed. This may be because points from non-similar structural regions (e.g., flattening from two surfaces to the middle) have been mixed into the neighborhood of the edge points during denoising. Record the region corresponding to the edge point set as a low-error edge region.

[0055] The filter parameter vectors for point cloud data in high residual clustering regions and low error edge regions are adjusted, specifically including the following steps:

[0056] For point cloud data within high residual clustering regions, a Hadamard product operation is performed between the filtering parameter vector used in the previous iteration of the point cloud data during the denoising operation and the high residual clustering adjustment vector. This completes the adjustment of the filtering parameter vector corresponding to the point cloud data within the high residual clustering regions, and the adjusted filtering parameter vector is used for the denoising operation in the current iteration.

[0057] For point cloud data in the low-error edge region, perform a Hadamard product operation between the filter parameter vector used in the previous iteration of the noise reduction operation and the low-error edge region adjustment vector to complete the adjustment of the filter parameter vector corresponding to the point cloud data in the low-error edge region. The adjusted filter parameter vector is used for the noise reduction operation in the current iteration.

[0058] Both the high residual clustering adjustment vector and the low error edge region adjustment vector were simulated using the sparrow search algorithm.

[0059] Step S6: Determine the laser processing path based on the optimized point cloud dataset;

[0060] In a five-axis CNC laser processing machine tool, taking laser cutting as an example, a simulated workpiece model to be processed is constructed based on an optimized point cloud dataset, and then matched with a pre-set target workpiece model to construct a cutting line trajectory. The laser processing path is then set based on the cutting line trajectory. This can be achieved by the program in the five-axis CNC laser processing machine tool, and will not be elaborated here.

[0061] This application uses multi-dimensional feature vectors to perform structural annotation and region division on the point cloud data of the workpiece to be processed. It adaptively configures filtering parameters based on region labels and introduces a self-feedback noise reduction strategy driven by fitting residuals. By constructing a fitting surface and calculating the point cloud residual distribution, it achieves intelligent identification and local re-filtering adjustment of high residual clustering regions and low error edge regions. This enables the overall noise reduction process to have convergence and local sensitivity control capabilities, avoiding structural distortion caused by insufficient noise reduction or excessive smoothing. This effectively avoids error diffusion problems caused by factors such as reflection, misalignment, and hot spots.

[0062] The training of the region labeling network and the noise analysis network includes the following steps:

[0063] Several region annotation training samples are obtained. These samples include multi-dimensional feature vectors corresponding to point cloud data. These multi-dimensional feature vectors are calculated from the point cloud data measured from the actual workpiece to be processed. The region annotation training samples are labeled with region labels, which are manually labeled based on the 3D model of the workpiece to be processed. All labeled region annotation training samples are combined into a region annotation training set. The region annotation network is trained using the region annotation training set. During training, the labeled region labels are used as the target to determine whether the accuracy of the region annotation network meets expectations. If the accuracy of the region annotation network meets expectations, the trained region annotation network is output. Otherwise, the region annotation network is trained again using the region annotation training set.

[0064] Several noise analysis training samples are obtained, including multi-dimensional feature vectors corresponding to point cloud data. These multi-dimensional feature vectors are calculated from the point cloud data measured from the actual workpiece to be processed. The noise analysis training samples are labeled with noise probabilities, which are manually labeled based on the 3D model of the workpiece. Generally, the noise probability of point cloud data corresponding to noise is directly recorded as 1, and the noise probability of point cloud data not corresponding to noise is recorded as 0. All labeled noise analysis training samples are combined into a noise analysis training set. The noise analysis network is trained using the noise analysis training set. During training, the labeled region labels are used as targets to determine whether the accuracy of the noise analysis network meets expectations. If the accuracy of the noise analysis network meets expectations, the trained noise analysis network is output; otherwise, the noise analysis network is trained again using the noise analysis training set.

[0065] Training the filter parameter analysis network involves the following steps:

[0066] Several training samples for filtering parameter analysis are obtained. These training samples include filtering parameter analysis vectors, which are constructed from point cloud data measured from the actual workpiece and labeled region tags. The training samples are labeled using these filtering parameter vectors, which are the optimal parameters manually selected when performing noise reduction on the point cloud data. All labeled training samples are combined into a filtering parameter analysis training set. The filtering parameter analysis network is trained using this training set. During training, the labeled region tags are used as targets to determine whether the accuracy of the filtering parameter analysis network meets expectations. If the accuracy meets expectations, the trained filtering parameter analysis network is output; otherwise, the filtering parameter analysis network is trained again using the training set.

[0067] The high residual clustering adjustment vector is obtained through simulation using the sparrow search algorithm, specifically including the following:

[0068] Construct several simulated individuals with the same storage format as the high residual clustering adjustment vector. The elements in the simulated individuals are all random values ​​between 0 and 1. Combine all the simulated individuals into a population set and set the maximum number of iterations.

[0069] Calculate the fitness of the simulated individual: The simulated individual is directly used as the high residual clustering adjustment vector. For any workpiece to be processed, steps S1-S6 are executed to perform noise reduction on the point cloud dataset corresponding to the workpiece to be processed obtained by the probe. The reciprocal of the total number of iterations in steps S1-S6 is used as the fitness of the simulated individual.

[0070] Based on the simulated fitness of individuals, the population set is updated using a sparrow search algorithm;

[0071] The process continues until the maximum number of iterations is reached, at which point the simulated individual with the highest fitness is output as the high residual clustering adjustment vector.

[0072] The method of simulating low-error edge region adjustment vectors using the sparrow search algorithm is consistent.

[0073] Example 2: An adaptive path planning system for laser processing based on probe data analysis, see [link to example]. Figure 1 ,include:

[0074] The point cloud data acquisition module is used to acquire the point cloud dataset of the workpiece to be processed through the probe. The point cloud dataset includes the point cloud data corresponding to several workpieces to be processed. The point cloud data includes the three-dimensional coordinates in the coordinate system constructed by the probe. The point cloud dataset reflects the size of the workpiece to be processed and also provides a data basis for the path planning of laser processing. It should be noted that this embodiment is for a five-axis CNC laser processing machine tool.

[0075] The point cloud data annotation module is used to determine the multidimensional feature vector corresponding to each point cloud data point based on the point cloud dataset. The multidimensional feature vector includes normal vector, curvature, normal uniformity, and point density, etc. The multidimensional feature vector can reflect the feature information of the corresponding point in the point cloud data, and can be used to determine the region to which the corresponding point cloud belongs. The point cloud dataset is annotated based on all multidimensional feature vectors, adding corresponding region labels and noise probabilities to the point cloud data. Here, annotation refers to the local structure recognition of the point cloud. For example, point clouds with low curvature, low normal uniformity, and uniform point density belong to flat regions. These flat regions are less affected by noise, and larger-scale filtering operations can be performed to reduce noise. The noise probability reflects the probability that the corresponding point belongs to noise. It should be noted that determining the multidimensional feature vector corresponding to each point cloud data point based on the point cloud dataset can be achieved through the point cloud algorithm built into the PCL library.

[0076] The point cloud data denoising module is used to perform denoising operations on all point cloud data in the point cloud dataset based on the corresponding filter parameter vector.

[0077] The filter parameter vector adjustment module is used to determine high residual clustering regions and low error edge regions based on the point cloud residuals corresponding to the point cloud data in the point cloud dataset. High residual clustering regions refer to multiple points within a region whose residuals after noise reduction are significantly larger than those on the simulated workpiece surface. These points are also spatially close to each other, exhibiting a clustered state. The appearance of high residual clustering regions is clearly due to insufficient noise reduction. Low error edge regions refer to edge points that should have geometric abrupt changes (such as abrupt changes in normal or curvature), but whose fitting residuals are very small after noise reduction, indicating that the structure has been "over-smoothed." The filtering parameter vectors corresponding to the point cloud data in the high residual clustering region and the low error edge region are adjusted, and denoising operation is performed on the point cloud data in the high residual clustering region and the low error edge region. The point cloud dataset is then updated to obtain the reconstructed point cloud dataset. The reconstructed point cloud dataset before the denoising operation is recorded as the point cloud dataset, and the process returns to step S4. It should be noted that when updating the point cloud dataset, the denoising operation is performed on the point cloud data located in the high residual clustering region and the low error edge region. The other point cloud data is not processed.

[0078] The laser processing path determination module is used to determine the laser processing path based on an optimized point cloud dataset.

[0079] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims. Parts not described in detail in this specification are prior art known to those skilled in the art.

Claims

1. A laser processing adaptive path planning method based on probe data analysis, characterized in that, The method comprises the following steps: Step S1: obtaining a point cloud data set of a workpiece to be processed by a measuring head; Step S2: determining a multi-dimensional feature vector corresponding to each point cloud data according to the point cloud data set, labeling the point cloud data set based on all the multi-dimensional feature vectors, and adding a corresponding region label and a noise probability to the point cloud data; Step S3: determining a filtering parameter vector according to the region label corresponding to the point cloud data, the filtering parameter vector comprising a filtering kernel size and a filtering kernel function weight, performing a noise reduction operation on all the point cloud data in the point cloud data set based on the corresponding filtering parameter vector, and obtaining a reconstructed point cloud data set; Step S4: performing surface reconstruction based on all the point cloud data in the reconstructed point cloud data set, projecting the point cloud data in the point cloud data set onto a simulated workpiece surface to obtain corresponding simulated point cloud data, calculating a distance between the point cloud data and the corresponding simulated point cloud data as a point cloud residual error corresponding to the point cloud data, calculating a root mean square error of the point cloud residual errors corresponding to all the point cloud data as a fluctuation amplitude value, and if a standard deviation of the fluctuation amplitude values of the previous N-1 iterations is less than a fluctuation threshold, outputting the reconstructed point cloud data set output in the last iteration as an optimized point cloud data set and entering step S6; otherwise, recording the reconstructed point cloud data set output in the last iteration as the point cloud data set and entering step S6; Step S5: determining a high residual aggregation region and a low error edge region based on the point cloud residual error corresponding to the point cloud data in the point cloud data set, adjusting the filtering parameter vector corresponding to the point cloud data in the high residual aggregation region and the low error edge region, performing a noise reduction operation on the point cloud data in the high residual aggregation region and the low error edge region, updating the point cloud data set, obtaining a reconstructed point cloud data set, and recording the reconstructed point cloud data set before the noise reduction operation as the point cloud data set, and returning to step S4; Step S6: determining a laser processing path based on the optimized point cloud data set.

2. The laser processing adaptive path planning method based on probe data analysis according to claim 1, characterized in that, The noise reduction operation on the point cloud data comprises the following operations: Recording the point cloud data as target point cloud data, selecting all the point cloud data within a distance of r from the target point cloud data as a neighborhood set according to the filtering kernel size r in the filtering parameter vector corresponding to the target point cloud data; For the i-th point cloud data in the neighborhood set, a filtering weight W corresponding to the i-th point cloud data is calculated i =f1(D i )·f2(P i )·(1-Z i ) γ , wherein f1 and f2 are Gaussian kernel functions, specifically: f1(D i )=exp(- (D i ) 2 / 2α 2 ), f2(P i )=exp(- (1- (P i ) 2 ) / 2β 2 ), D i is the distance between the i-th point cloud data in the neighborhood set and the target point cloud data, P i is the dot product between the normal vector corresponding to the i-th point cloud data in the neighborhood set and the normal vector corresponding to the target point cloud data, α is the distance weight, β is the normal consistency weight, Z i is the noise probability corresponding to the target point cloud data, and γ is the noise weight, and α, β and γ constitute the filtering kernel function weights in the filtering parameter vector; Performing weighted summation on all the point cloud data in the neighborhood set and the corresponding filtering weight, dividing the result by the sum of all the filtering weights, obtaining a noise reduction point cloud data, and replacing the target point cloud data with the noise reduction point cloud data to complete the noise reduction operation.

3. The laser processing adaptive path planning method based on probe data analysis according to claim 2, characterized in that, Determining a high residual aggregation region and a low error edge region based on the point cloud residual error corresponding to the point cloud data in the point cloud data set comprises the following steps: Constructing a high residual set by selecting all the point cloud data with a residual error higher than a residual threshold value, which reflects the degree of insufficient noise reduction, performing connected component extraction on the high residual set, constructing a plurality of connected sets, each connected set comprising a plurality of point cloud data, and selecting a region corresponding to a connected set with a number of point cloud data higher than a clustering threshold value as the high residual aggregation region. All point cloud data selected by the region label as the edge region constitutes an edge point set, and the root mean square error of all point cloud residuals in the edge point set is calculated as an edge fluctuation amplitude value. If the edge fluctuation amplitude value is lower than the fluctuation amplitude value at the current iteration, the region corresponding to the edge point set is recorded as a low-error edge region.

4. The laser processing adaptive path planning method based on probe data analysis according to claim 3, characterized in that, The filtering parameter vector corresponding to the point cloud data in the high residual aggregation region and the low-error edge region is adjusted, specifically including the following steps: For the point cloud data in the high residual aggregation region, the filtering parameter vector used by the point cloud data in the last iteration to perform the denoising operation is subjected to Hadamard product operation with the high residual aggregation adjustment vector to complete the adjustment of the filtering parameter vector corresponding to the point cloud data in the high residual aggregation region. The adjusted filtering parameter vector is used for the denoising operation in the current iteration; For the point cloud data in the low-error edge region, the filtering parameter vector used by the point cloud data in the last iteration to perform the denoising operation is subjected to Hadamard product operation with the low-error edge region adjustment vector to complete the adjustment of the filtering parameter vector corresponding to the point cloud data in the low-error edge region. The adjusted filtering parameter vector is used for the denoising operation in the current iteration; The high residual aggregation adjustment vector and the low-error edge region adjustment vector are both simulated by the sparrow search algorithm.

5. The laser processing adaptive path planning method based on probe data analysis according to claim 4, characterized in that, Based on all multi-dimensional feature vectors, the point cloud data set is labeled, and the corresponding region label and noise probability are added to the point cloud data, specifically including the following contents: the multi-dimensional feature vector corresponding to the point cloud data is sent into the region labeling network and the noise analysis network for processing to obtain the corresponding region label and noise probability. The region label includes flat region, curvature region and edge region. The region labeling network and the noise analysis network are both based on BP neural network and include input layer, hidden layer and output layer.

6. The laser processing adaptive path planning method based on probe data analysis according to claim 5, wherein, The filtering parameter vector is determined according to the region label corresponding to the point cloud data, specifically including the following contents: the point cloud data and the corresponding region label are spliced at the beginning and the end to obtain a filtering parameter analysis vector, and then the filtering parameter analysis vector is sent into the filtering parameter analysis network for processing to obtain the filtering parameter vector. The filtering parameter analysis network is also based on BP neural network and includes input layer, hidden layer and output layer.

7. A laser processing adaptive path planning system based on probe data analysis, characterized by, The system applies the laser processing adaptive path planning method based on probe data analysis according to any one of claims 1-6, including: A point cloud data acquisition module for acquiring a point cloud data set of a workpiece to be processed by a probe; A point cloud data labeling module for determining a multi-dimensional feature vector corresponding to each point cloud data based on the point cloud data set, and labeling the point cloud data set based on all multi-dimensional feature vectors to add a corresponding region label and noise probability to the point cloud data; A point cloud data denoising module for performing denoising operation on all point cloud data in the point cloud data set based on the corresponding filtering parameter vector to obtain a reconstructed point cloud data set; A filtering parameter vector adjustment module for determining a high residual aggregation region and a low-error edge region based on the point cloud residual corresponding to the point cloud data in the point cloud data set, and adjusting the filtering parameter vector corresponding to the point cloud data in the high residual aggregation region and the low-error edge region. A laser machining path determination module for determining a laser machining path based on the optimized point cloud dataset. A laser machining path determination module for determining a laser machining path based on the optimized point cloud dataset.

Images