Semiconductor coating path planning and pose estimation method for heat generating clothing

Abstract

This invention discloses a method for semiconductor coating path planning and pose estimation for heat-generating clothing, relating to the field of adaptive control technology. The method divides the coating area, confirmed by a target detection method, into multiple units and calculates the center coordinates, center normal vector, and curvature of each unit. Some spraying parameters are preset. A Gaussian decay model is used to construct the thickness distribution of each unit to describe the relationship between the spraying direction, spray volume, and thickness at any position of the spray gun in each unit. Through iterative simulation using the alternating direction multiplier method, the optimal spraying direction and optimal spray volume for each unit are obtained, ensuring the overall thickness distribution conforms to the standard thickness. A reward function is designed, considering the total movement of the spray gun, the total direction adjustment, and the reasonable spraying distance of each unit, to guide reinforcement learning simulation training of the spray gun. This generates the optimal spraying sequence and the optimal spraying coordinates for each unit, achieving adaptive path planning and spraying parameter estimation based on the geometric characteristics of the coating area.

Description

Technical Field

[0001] This invention relates to the field of adaptive control technology, specifically to a method for semiconductor coating path planning and pose estimation for heat-generating clothing. Background Technology

[0002] In the process of semiconductor coating on insulators, traditional coating methods often lack precise consideration of the shape and size of the coating area. For coating areas with complex shapes and varying sizes, it is difficult to ensure that the coating material achieves uniform coverage and meets the required thickness. Currently, the spray gun is mostly controlled by manual experience or simple preset programs, which cannot adaptively adjust the spray parameters according to the coating area, resulting in inconsistent coating quality and limiting the development of semiconductor coating technology. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes a semiconductor coating path planning and pose estimation method for heat-generating clothing, in order to achieve adaptive planning of optimal spraying parameters and optimal spraying path based on the geometric characteristics of the coating area.

[0004] The technical solution to achieve the purpose of this invention is as follows:

[0005] A method for semiconductor coating path planning and pose estimation for heat-generating clothing includes the following steps:

[0006] Point cloud data is generated based on the calibrated coordinate system of the device pose unified scanner and spectrometer data. and spectral data Among them, I m and Let r be the gray value and coordinates of the m-th point, respectively. n and These are the reflectance vector and coordinates of the nth light spot, respectively.

[0007] Point cloud data The input point is a convolutional network, which aggregates the coordinates and gray values ​​of points in the neighborhood using dynamic graph convolution and generates an enhanced feature map F based on attention. I,a , spectral data The input spectral encoder uses a sampling algorithm to combine the spectral information mined by different dilated convolutional kernels with the enhanced feature map F. I,a Perform spatial alignment to generate an alignment feature map F r,a Based on the Pearson correlation coefficient, the feature map F is aligned. r,a With enhanced feature map F I,a The features are then stitched together to form a fused feature map F.

[0008] Computing point cloud data The normal vectors of M points are concatenated to form a normal vector graph. After being concatenated with the fused feature map F, a segmentation probability map p is generated through multilayer perceptron mapping, and a segmentation label vector ξ is generated based on conditional random field optimization to determine the coating region U. co ;

[0009] Coating area U co Divide the area into Q elements with overlapping bands between adjacent elements, and calculate the q-th element. center coordinates center normal vector and curvature γ q It also presets some injection parameters and constructs a description of the q-th unit based on the Gaussian decay model. Spray direction Injection quantity λ q Thickness distribution related to the coordinates of any point P The overall thickness distribution is generated by superimposing the thickness distributions of Q units. The overall thickness distribution is calculated through iterative simulation using the alternating direction multiplier method. Meets standard thickness d std The optimal injection direction and optimal injection quantity for each unit;

[0010] A reward function is designed to guide reinforcement learning simulation training of the spray gun by considering the total movement and direction adjustment of the spray gun, along with the reasonable spray distance for Q units. This generates the optimal spray sequence and the optimal spray coordinates for each unit, along with the center coordinates, spray direction, and spray speed v of each unit. la They are integrated into planning instructions and delivered to the spray guns for execution.

[0011] Furthermore, the processing steps of point convolutional networks include:

[0012] Extracting point cloud data The gray value I of the m-th point m and point coordinates And concatenate them into the vector of the m-th point.

[0013] Aggregating the neighborhood of the m-th point using dynamic graph convolution Extract the feature vector f of the m-th point from the information of the point vectors within the matrix. m The point feature vectors of M points are spliced ​​together to generate a point feature vector map F;

[0014] The point feature vector F is pooled and linearly transformed along the channel dimension, and then converted into a spatial attention weight vector w using the Sigmoid function. s,a ;

[0015] The point feature vector F is pooled and linearly transformed in the spatial dimension, and then converted into a channel attention weight vector w using the Sigmoid function. c,a;

[0016] Using spatial attention weight vector w s,a With channel attention weight vector w c,a The matrix multiplication result is used to enhance the boundary of the point feature vector map F, generating an enhanced feature map F. I,a .

[0017] Furthermore, the processing steps of the spectral encoder include:

[0018] Extracting spectral data The reflectivity vector r of the nth light point n ;

[0019] The reflectivity vector r of the nth light point is obtained by using K dilatational convolution kernels with different dilation rates. n Dilated convolutions are performed separately to mine and generate spectral feature vectors of the nth light point in K different fields of view, and then concatenated to form the multi-field feature vector fnr of the nth light point;

[0020] Point cloud data is obtained by interpolating and predicting based on the multi-view feature vectors of N light points using a sampling algorithm. The predicted multi-view feature vectors at the coordinates of M points are integrated into an aligned feature map F. r,a The sampling algorithms include nearest neighbor interpolation, bilinear interpolation, inverse distance weighted interpolation, and Kriging interpolation.

[0021] Furthermore, based on the Pearson correlation coefficient, the feature maps F are aligned. r,a With enhanced feature map F I,a The process of concatenating and generating a fused feature map F includes the following steps:

[0022] Extracting Enhanced Feature Map F I,a and alignment feature map F r,a Enhanced feature vector of the m-th point and predicting multi-view feature vectors

[0023] Calculate the average augmentation feature f of the m-th point. m I，a Average prediction of multi-view features Correlation coefficient α with Pearson m ;

[0024] Pearson correlation coefficients from M points are concatenated and transformed into a correlation weight vector w using the Sigmoid function. α ;

[0025] The predicted multi-view feature vector of the m-th point With the relevant weight vector w α The product of the corresponding relevant weights and the enhanced point feature vector Concatenate the data to generate the fused feature vector f at the m-th point. m ;

[0026] The fused feature vectors of M points are concatenated to form a fused feature map F.

[0027] Specifically, normal vector map The normal vector of the m-th point equals the neighborhood of the m-th point The local covariance matrix C of the interior point coordinates m The eigenvector corresponding to the smallest eigenvalue extracted through principal component analysis, singular value decomposition, or matrix decomposition.

[0028] Furthermore, the segmentation label vector ξ is generated based on conditional random fields, including the following steps:

[0029] Obtain the probability vector p of the m-th point from the segmented probability map p. m ;

[0030] Based on the negative log-likelihood form and the segmentation probability graph p, the point potential function φ of the m-th point is defined. m (ξ m =β), point potential function φ m (ξ m =β) reflects the label ξ of the m-th point m The confidence level is either 1 or 2, and the label ξ of the m-th point is... m The value 1 or 2 represents whether the m-th point is located in the coated area or the non-coated area, respectively.

[0031] Define the neighborhood of the m-th point and the m-th point. Label edge potential function of the i-th point The label ξ indicating the m-th point m Neighborhood of the m-th point The label ξ of the i-th point i Are the values ​​the same?

[0032] Based on the probability vector p of the m-th point m Initialize the label ξ with a relatively high probability. m Find the values ​​of , sum the potential function value of the m-th point, and the neighborhood of the m-th point and the m-th point. The edge potential function values ​​of other points within the boundary are used to obtain the energy value ε of the m-th point. m ;

[0033] The iterative conditional pattern algorithm is used to iteratively update the label value of each point to obtain the label values ​​of the M points when the sum of the energy values ​​of the M points is minimized, and then concatenate them into a segmentation label vector ξ.

[0034] Furthermore, the iterative conditional pattern algorithm traverses M points. When it reaches the m-th point, it keeps the labels of the remaining points unchanged and changes the label ξ of the m-th point. m The value of is updated to minimize the sum of the energy values ​​of the M points. After the labels of the M points are updated in each round, it is determined whether the absolute value of the error between the sum of the energy values ​​in each round and the sum of the energy values ​​in the previous round is less than the minimum value or whether the number of rounds in the next round is equal to the maximum number of rounds. If the conditions are met, the iteration ends; otherwise, the next round of iteration begins.

[0035] Furthermore, preset some injection parameters and calculate the q-th unit. center coordinates center normal vector and curvature γ q This includes the following steps:

[0036] Determine the material temperature T ma and jet speed v la ;

[0037] Clustering all cells located in the q-th cell using the cluster center estimation method The points in the solution are used to generate the q-th element. center coordinates

[0038] Calculate the q-th unit The coordinates of each point and the center coordinates The inverse distance is used to generate the q-th cell by weighting the normal vector of each point with the inverse distance. The center normal vector

[0039] For the q-th unit The center normal vector The average of the angles between the q-th cell and the center normal vector of each adjacent cell is taken as the q-th cell. curvature γ q .

[0040] Furthermore, the q-th unit is solved by clustering based on the cluster center estimation method. center coordinates Includes the following steps:

[0041] In the qth unit Randomly select one point from all points within the cluster as the cluster point;

[0042] Calculate the minimum distance for each non-clustered point, where the minimum distance is equal to the distance between the non-clustered point and its nearest existing clustered point;

[0043] Divide the square of the minimum distance of each non-clustered point by the sum of the squares of the minimum distances of all non-clustered points to generate the selection value of each non-clustered point. Set the non-clustered point with the highest selection value as the next clustered point, and stop when the total number of clustered points reaches the preset value.

[0044] In each iteration, the q-th unit is calculated. The distance between each non-clustered point in this round and the existing clustered points is calculated, and the points are classified into the class corresponding to the clustered point with the smallest distance. The average coordinates of all points in each class are used as the clustered point of each class in this round.

[0045] Calculate the average distance between the cluster points of all classes in the current round and the cluster points of the previous round, and compare it with the distance threshold to decide whether to start the next round of iteration or obtain the cluster points and the number of points of each class in the current round;

[0046] Based on the number of points in each category occupying the qth unit The ratio of the total number of points in each cluster is used to perform a weighted sum of the coordinates of the points in each cluster, generating the q-th cell. center coordinates

[0047] Furthermore, the q-th unit is constructed based on the Gaussian decay model. Spray direction Injection quantity λ q With thickness distribution The specific connections are as follows:

[0048] At a given material temperature T ma and jet speed v la Below, the thickness distribution within the element conforms to a Gaussian decay model, and the q-th element... thickness distribution Specifically as follows:

[0049]

[0050] Among them, s q and The qth unit The cell area and point coordinates P of the q-th cell center coordinates The square of the distance between them, ρ q For the coating material in the qth unit The flow coefficient is related to the injection velocity v. la and the qth unit curvature γ q Showing a positive correlation, Curvature correction for injection direction at point coordinate P With the qth unit The center normal vector The included angle, the curvature correction of the injection direction at point coordinate P. For the q-th unit Spray direction Subjected to the q-th unit center coordinates The curvature effect γ of the vector direction to point P q Offset generation.

[0051] Furthermore, the thickness distributions of Q units are superimposed, and the optimal injection direction and optimal injection quantity for each unit are obtained using the alternating direction multiplier method, including the following steps:

[0052] The thickness distributions of Q units are summed to form the coating region U. co Overall thickness distribution

[0053] The objective function is constructed as the overall thickness distribution. With standard thickness d std The mean square error is used to construct an optimization problem by minimizing the objective function, and the constraint requires that the injection amount of Q units be non-negative;

[0054] By using the alternating direction multiplier method, the constraints are introduced into the objective function through Q Lagrange multiplier terms to construct a Lagrange augmented function. The injection direction, injection quantity, and Lagrange multipliers of the Q units are initialized and used as optimization variables. With the goal of minimizing the Lagrange augmented function, the optimal injection direction and optimal injection quantity are solved iteratively.

[0055] Furthermore, the simulation training steps for reinforcement learning include:

[0056] Define execution step t as incremented by 1 after executing the spray of one unit;

[0057] Define the action A(t) of execution step t as the cell number Nu(t) of the cell the spray gun is about to move to and the spray coordinate P. t ;

[0058] Define the state ψ(t) of execution step t. The state ψ(t) includes the unit number Nu(t-1) related to the spray gun before the execution of action A(t), and the spray coordinate P. t-1 Spray direction and center coordinates

[0059] Define the reward δ(t) for execution step t. The reward δ(t) is the immediate reward of the spray gun after performing action A(t) in state ψ(t). The specific formula for reward δ(t) is as follows:

[0060]

[0061] Where δ(0) is the fixed reward after the spray gun performs action A(t), ω1, ω2 and ω3 are the movement balance coefficient, angle balance coefficient and constraint balance coefficient, respectively, d(P(t),P(t-1)) represents the distance the spray gun moves from the spray coordinate P(t-1) to the spray coordinate P(t), and θ(t,t-1) is the distance the spray gun moves from the spray direction. Rotate to the direction of spray Adjust the direction and angle. Let the injection coordinates P(t) and the center coordinates be... The injection distance, τ() is the penalty function, applied only at the injection distance. When the distance exceeds the adjustable range, the maximum positive integer is taken.

[0062] Define the value G(t) of execution step t, which is equal to the value G(t-1) of execution step t-1, the reward δ(t) of execution step t, and the weighted sum of the maximum value maxG(t+1) that can be obtained by action A(t+1) based on state ψ(t+1) after executing action A(t);

[0063] By iteratively selecting different actions for each execution step to update the value G(Q), the spraying sequence of the spray gun and the spraying coordinates of Q units when the value G(Q) is maximized are selected as the optimal spraying sequence and the optimal spraying coordinates of Q units.

[0064] Compared with existing technologies, this invention analyzes the correlation between positional and grayscale information of point cloud data through point convolutional networks, mines spectral information of spectral data through spectral encoders, and achieves spatial alignment through sampling algorithms. Based on Pearson correlation coefficient, it generates a fused feature map by combining the outputs of point convolutional networks and spectral encoders. After concatenation with the normal vector map, it generates a segmentation probability map through multilayer perceptron mapping and optimizes the generation of segmentation label vectors based on conditional random fields to determine the coating area. The coating area is divided into multiple units, and the center coordinates, center normal vector, and curvature of each unit are calculated. Some spraying parameters are preset, and the thickness distribution of each unit is constructed based on a Gaussian decay model to describe the correlation between the spraying direction, spray volume, and thickness at any position of the spray gun in each unit. Through iterative simulation using the alternating direction multiplier method, the optimal spraying direction and optimal spray volume of each unit that makes the overall thickness distribution conform to the standard thickness are obtained. The reward function is designed by considering the total movement of the spray gun, the total direction adjustment, and the reasonable spraying distance of each unit to guide the reinforcement learning simulation training of the spray gun, generating the optimal spraying sequence and the optimal spraying coordinates of each unit, realizing adaptive path planning and spraying parameter estimation based on the geometric characteristics of the coating area. Attached Figure Description

[0065] Figure 1 A flowchart for a semiconductor coating path planning and pose estimation method for heat-generating clothing;

[0066] Figure 2 Diagram of a point convolutional network model;

[0067] Figure 3 Here is a flowchart of the Conditional Random Field (CRF) process.

[0068] Figure 4 This is a flowchart of the cluster center estimation method. Detailed Implementation

[0069] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0070] Example 1

[0071] like Figure 1 As shown, a specific embodiment of the present invention discloses a method for semiconductor coating path planning and pose estimation for heat-generating clothing, including the following steps:

[0072] Point cloud data was generated by acquiring insulator data using a scanner and a spectrometer, and then performing coordinate system 1 based on the device pose. and spectral data The device pose includes scanning pose and illumination pose, which are obtained using the Zhang Zhengyou calibration method. m and Let r be the gray value and coordinates of the m-th point, respectively. n and Let M and N be the reflectance vector and coordinates of the nth light point, respectively, and the total number of points M is much greater than the total number of light points N.

[0073] Point cloud data The input point convolutional network aggregates the coordinates and gray values ​​of points in the neighborhood using dynamic graph convolution and enhances the boundary gradient changes based on attention to generate an enhanced feature map F. I,a , spectral data The input spectral encoder mines spectral information from different fields of view using dilated convolutional kernels with varying dilation rates, and then applies this information based on a sampling algorithm and enhanced feature map F. I,a Achieve spatial alignment and generate alignment feature map F r,a Based on the Pearson correlation coefficient, the feature map F is aligned. r,a With enhanced feature map F I,a The feature map F is generated by splicing the features together.

[0074] Calculate point cloud data separately The normal vectors of M points are concatenated row by row to form a normal vector graph. After being concatenated with the fused feature map F, a segmentation probability map p is generated through multilayer perceptron mapping. The segmentation probability map p is then adjusted and optimized based on a conditional random field. The segmentation label vector ξ is generated by minimizing the energy function, and the coating region U is determined. co ;

[0075] Using regular shapes to form the coating area U co The system is divided into Q units with a fixed-width overlap between adjacent units. Preset injection parameters and calculate the values ​​for the q-th unit. center coordinates center normal vector and curvature γ q The q-th unit is constructed based on the Gaussian decay model. Spray direction Injection quantity λ q With thickness distribution The correlation between the thickness distribution of Q units is superimposed, and the optimal spraying direction and optimal spraying amount of each unit are obtained by using the alternating direction multiplier method, so that the overall thickness distribution of the coating conforms to the standard thickness d. std Where P represents the coating area U co The coordinates of any point in the array;

[0076] A reward function is designed to guide reinforcement learning training of the spray gun by considering the total movement and direction adjustment of the spray gun, along with the reasonable spray distance for Q units. This generates the optimal spray sequence and the optimal spray coordinates for each unit. The optimal spray coordinates, center coordinates, spray direction, and spray speed v for each unit are then considered. la The commands are integrated into unit commands and arranged into planning commands according to the optimal spray sequence, and then delivered to the spray gun for execution.

[0077] like Figure 2 As shown, the further processing steps of the point convolutional network include:

[0078] Extracting point cloud data The gray value I of the m-th point m and point coordinates And concatenate them column by column to form the vector of the m-th point. The dimension is 1×4;

[0079] Aggregating the neighborhood of the m-th point using dynamic graph convolution Extract the point feature vector of the m-th point, which is sensitive to geometric structure, from the information of the point vectors within the matrix. M1 represents the number of point feature channels. The specific formula for dynamic graph convolution is as follows:

[0080]

[0081] in, The kernel is a dynamic graph convolution kernel, and convolution is performed with a stride of 1. * indicates convolution. The point vector of the m-th point With neighboring regions The point vector of the i-th point inside The learning weight matrix between them;

[0082] Concatenate the feature vectors of M points row by row to obtain a point feature vector map.

[0083] The point feature vector F is input into the spatial attention branch to focus on the importance differences of M point coordinates. In the channel dimension, it is transformed into a spatial attention weight vector through pooling and linear transformation, and then using the sigmoid function.

[0084] The point feature vector F is input into the channel attention branch to focus on the importance of the M1 first channels. In the spatial dimension, it is transformed into a channel attention weight vector through pooling and linear transformation and then using the sigmoid function.

[0085] The spatial attention weight vector w s,a With channel attention weight vector w c,a Perform matrix multiplication to obtain the attention weight graph Based on attention weight map W att Boundary enhancement is performed on the point feature vector map F to highlight the boundary gradient change features, generating an enhanced feature map F. I,a =W att ·F, Enhanced Feature Map in, Let be the enhanced feature vector of the m-th point, and · denotes the dot product.

[0086] Furthermore, the processing steps of the spectral encoder include:

[0087] Extracting spectral data The reflectivity vector of the nth light point L represents the total number of wavelengths;

[0088] The reflectivity vector r of the nth light point is obtained by using K dilatational convolution kernels with different dilation rates. n Dilated convolutions are performed separately to mine and generate the spectral feature vectors of the nth light spot in K different fields of view, and the spectral feature vector of the nth light spot in the kth field of view. The calculation formula is as follows:

[0089]

[0090] in, For the k-th dilated convolution kernel The convolution weights in the j-th dimension, h k For the k-th dilated convolution kernel The expansion rate, r n [(j×h k[)mod(L)] represents the reflectivity vector r of the nth light point. n (j×h) k The reflectance of L dimensions mod(L), where mod represents the modulo operation, k = 1, ..., K;

[0091] The spectral feature vectors of the nth light spot in K fields of view are concatenated column by column to form the multi-field feature vector of the nth light spot. M2 = K × L is the number of multi-view feature channels;

[0092] Point cloud data is obtained by interpolating and predicting based on the multi-view feature vectors of N light points using a sampling algorithm. Predicted multi-view feature vectors at the coordinates of M points, generating and enhancing feature map F I,a Alignment feature map that maintains spatial alignment in, The coordinates of the m-th point The predicted multi-view feature vectors at the location are sampled using algorithms including nearest neighbor interpolation, bilinear interpolation, inverse distance weighted interpolation, and kriging interpolation.

[0093] Furthermore, based on the Pearson correlation coefficient, the feature maps F are aligned. r,a With enhanced feature map F I,a The process of concatenating and generating a fused feature map F includes the following steps:

[0094] Extract the enhanced feature map F respectively I,a and alignment feature map F r,a Enhanced feature vector of the m-th point and predicting multi-view feature vectors

[0095] Calculate the average augmentation feature f of the m-th point respectively. m I,a and average prediction of multi-view features Specifically as follows:

[0096]

[0097] in, and The enhanced feature vectors of the m-th point are respectively and predicting multi-view feature vectors The enhanced point feature of the a-th point feature channel and the predicted multi-view feature of the b-th multi-view feature channel;

[0098] Calculate the Pearson correlation coefficient α at the m-th point. m The details are as follows:

[0099]

[0100] Obtain the Pearson correlation coefficients of M points and concatenate them into a correlation coefficient vector α = [α1, ..., α2]. m ,…,α M The ] is transformed into a relevant weight vector through the Sigmoid function. in, The relevant weight of the m-th point;

[0101] The predicted multi-view feature vector of the m-th point With relevant weights The product result and the enhanced point feature vector Concatenate columns to generate the fused feature vector of the m-th point. The fused feature vectors of M points are concatenated row by row to form a fused feature map. Where M3 = M1 + M2 is the number of fused feature channels.

[0102] Specifically, normal vector map in, The normal vector of the m-th point is equal to the neighborhood of the m-th point. The local covariance matrix C of the interior point coordinates m The eigenvector corresponding to the smallest eigenvalue, and the local covariance matrix C. m The eigenvalues ​​and eigenvectors in the matrix can be obtained through principal component analysis, singular value decomposition, or matrix decomposition. This is because the eigenvalues ​​and eigenvectors of the m-th point are derived from the neighborhood of the m-th point. The local surface formed by all points can be approximated as a plane. Points within the plane exhibit the least variation in the direction of the normal vector, which is reflected in the fact that points within the tangent plane have the smallest variance in the direction of the normal vector. The local covariance matrix C m The eigenvector corresponding to the smallest eigenvalue reflects the direction of the minimum variance, which is the normal vector of the m-th point.

[0103] like Figure 3 As shown, further, the segmentation probability map p is adjusted and optimized based on the conditional random field, and the segmentation label vector ξ is generated by minimizing the energy function, including the following steps:

[0104] From the segmentation probability map Obtain the probability vector p of the m-th point m =[p m,1 ,p m,2 ], where p m,1 and p m,2 The labels ξ of the m-th point are respectively m =1 and ξ m The probability of ξ = 2, label ξ m =1 and ξ m=2 indicates that the m-th point is located in the coated area and the non-coated area, respectively;

[0105] Based on the negative log-likelihood form combined with the segmentation probability map p output by the multilayer perceptron, the point potential function φ of the m-th point is defined. m (ξ m =β) = -log(p m,β ), where β∈{1,2}, and the point potential function φ m (ξ m =β) reflects the label ξ of the m-th point m Confidence level of the value;

[0106] Define the neighborhood of the m-th point and the m-th point. The edge potential function of the i-th point inside When the label ξ of the m-th point m Neighborhood of the m-th point The label ξ of the i-th point i When the values ​​are the same, the boundary potential function When the label ξ of the m-th point m Neighborhood of the m-th point The label ξ of the i-th point i The boundary potential function with different values

[0107] Based on the probability vector p of the m-th point m Initialize the label ξ with a relatively high probability. m The value of is used to determine the potential function value of the m-th point and the neighborhood of the m-th point. Sum the edge potential function values ​​of other points within the boundary to obtain the energy value ε of the m-th point. m ;

[0108] With the objective of minimizing the sum of energy values ​​of M points, an iterative conditional mode algorithm is used to iteratively update the label value of each point and recalculate the sum of energy values ​​of the M points until the convergence condition is met. Then, the label values ​​of the M points are determined and concatenated column-wise to form a segmented label vector.

[0109] Furthermore, the iterative conditional pattern algorithm traverses M points in a certain order. When traversing to the m-th point, it keeps the labels of all points except the m-th point unchanged and calculates the label ξ of the m-th point. m The sum of the energy values ​​of M points when the value is 1 or 2; select the value with the smaller sum of energy values ​​for the label ξ. mThe update process is as follows: After the labels of all M points are updated in each round, it is determined whether the absolute value of the error between the sum of energy values ​​in each round and the sum of energy values ​​in the previous round is less than the minimum value or whether the number of rounds in the next round is equal to the maximum number of rounds. If the absolute value of the error between the sum of energy values ​​in each round and the sum of energy values ​​in the previous round is less than the minimum value or the number of rounds in the next round is equal to the maximum number of rounds, the iteration ends. Otherwise, the next round of iteration begins.

[0110] Furthermore, preset some injection parameters and calculate the q-th unit. center coordinates center normal vector and curvature γ q This includes the following steps:

[0111] Determine the material temperature T based on the coating material. ma The spray speed v is determined based on the spray gun model. la ;

[0112] Get all elements located in the q-th cell The points in the middle construct the q-th unit Given a set of points, the q-th unit is determined by clustering using the cluster center estimation method. center coordinates

[0113] Calculate the q-th unit The coordinates of each point and the center coordinates The q-th cell is obtained by multiplying the product of the normal vector of each point and the inverse distance, and then dividing the sum of the products of the normal vector and the inverse distance by the sum of the inverse distances. The center normal vector

[0114] For the q-th unit The center normal vector The average of the angles between the q-th cell and the center normal vector of each adjacent cell is taken as the q-th cell. curvature γ q curvature γ q Reflecting the q-th unit The degree of tilt.

[0115] like Figure 4 As shown, further, the q-th unit is solved by clustering based on the cluster center estimation method. center coordinates Includes the following steps:

[0116] In the qth unit Randomly select a point from the set of points as the first cluster point;

[0117] Calculate the q-th unit The minimum distance between each non-clustered point in the point set, where the minimum distance between each non-clustered point is equal to the distance between each non-clustered point and its nearest existing cluster point;

[0118] The selection value of each non-clustered point is generated by dividing the square of the minimum distance of each non-clustered point by the sum of the squares of the minimum distances of all non-clustered points. The non-clustered point with the highest selection value is set as the next clustered point. This process continues until the total number of clustered points reaches a preset value. Each clustered point represents a class.

[0119] In each iteration, the q-th unit is calculated. The distance between each non-clustered point in the current round and the existing clustered points in the point set is calculated, and the points are classified into the class corresponding to the clustered point with the smallest distance. The average point coordinates of all points in each class are used as the clustered point of each class in the current round.

[0120] Calculate the average distance between the cluster points of all classes in the current round and the cluster points of the previous round, and determine whether it is less than or equal to the distance threshold. If it is greater than the distance threshold, start a new round of iteration. If it is less than or equal to the distance threshold, obtain the point coordinates of the cluster points of each class in the current round and the number of points of each class.

[0121] The number of points in each category occupies the q-th unit. The ratio of the total number of points in the point set is used as the weight to perform a weighted sum of the point coordinates of the cluster points in each class, generating the q-th unit. center coordinates The cluster point is essentially the q-th unit. The surface is distributed with multiple centroids, and the number of points in each category reflects the size of the local surface controlled by the centroid. The q-th element can be effectively determined by weighted summation. The center point.

[0122] Furthermore, the q-th unit is constructed based on the Gaussian decay model. Spray direction Injection quantity λ q With thickness distribution The specific connections are as follows:

[0123] At the preset material temperature T ma Under these conditions, the coating material exhibits fluid properties, and at a spray velocity v la Under the applied initial kinetic energy, the coating material is sprayed into the coating area, resulting in spray deposition that affects the coating thickness. Specifically, the coating thickness tends to be thicker at the center and thinner at the edges, and the thickness distribution within the unit cell conforms to a Gaussian decay model. The q-th unit... thickness distribution Specifically as follows:

[0124]

[0125] Among them, s q and The qth unit The cell area and point coordinates P of the q-th cell center coordinates The square of the distance between them, ρ q For the coating material in the qth unit The flow coefficient is related to the injection velocity v. la and the qth unit curvature γ q Directly related, as follows:

[0126] ρ q =ρ0(T ma )·(1+η v v la +η γ ·|γ q |),

[0127] Wherein, ρ0(T ma ) is at material temperature T ma Initial value of the flow coefficient of the coating material, η v and η γ These are the velocity correction coefficients and curvature correction coefficients, which can be obtained through experimental calibration and fitting. (The q-th unit...) Flow coefficient ρ q This reflects that, under the same molten state, the flowability of the coating material increases with increasing initial kinetic energy and surface curvature, θ q o,la (P) represents the q-th unit. Spray direction Curvature correction injection direction at point coordinate P With the qth unit The center normal vector The angle between the spray gun and the spray direction, where the angle is due to the spray direction of the spray gun. For the q-th unit center coordinates The actual direction acting at point coordinate P will be affected by the q-th unit. curvature γ q This causes a deviation, and the curvature corrects the jet direction. It can be approximated as the q-th unit. center coordinates The tangential vector from point P to point P.

[0128] Furthermore, the thickness distributions of Q units are superimposed, and the optimal injection direction and optimal injection quantity for each unit are obtained using the alternating direction multiplier method, including the following steps:

[0129] The thickness distributions of Q units are summed to form the coating region U. co Overall thickness distribution Where, P∈U co Because there are overlapping bands between the elements and the edge thickness of the elements is relatively thin, the standard thickness d can be achieved by stacking the thicknesses between the overlapping bands. std Requirements;

[0130] The objective function is constructed as the overall thickness distribution. With standard thickness d std The mean square error is used to construct an optimization problem by minimizing the objective function, and the constraint requires that the injection amount of Q units be non-negative;

[0131] The Lagrange augmented function is constructed by introducing the constraints into the objective function through Q Lagrange multiplier terms using the alternating direction multiplier method. The injection direction, injection amount, and Lagrange multipliers of the Q units are initialized and used as optimization variables. The goal is to minimize the Lagrange augmented function. In each iteration, the Lagrange multipliers and injection amount from the previous iteration are fixed, and the injection direction for the current iteration is updated using gradient descent. Then, the Lagrange multipliers and injection direction from the previous iteration are fixed, and the injection amount for the current iteration is updated using gradient descent. Finally, the injection direction and injection amount for the current iteration are fixed, and the Lagrange multipliers for the current iteration are updated using gradient descent. The error between the current and previous Lagrange augmented function values ​​is calculated, and it is determined whether it is less than the error convergence value. If it is greater than or equal to the error convergence value, the next iteration continues; otherwise, if it is less than the error convergence value, the injection direction and injection amount of the Q units in the current iteration are taken as the optimal injection direction and optimal injection amount, respectively.

[0132] Furthermore, the simulation training steps for reinforcement learning include:

[0133] Define execution step t. After each unit spray is executed, execution step t is incremented by 1, where t∈{1,…,q,…,Q};

[0134] Define the action A(t) for execution step t, where action A(t) is the cell number Nu(t) and the spray coordinate P of the cell to which the spray gun will move in execution step t to perform spraying. t Nu(t)∈{1,…,q,…,Q};

[0135] Define the state ψ(t) for execution step t. The state ψ(t) includes the cell number Nu(t-1) where the spray gun is located before the execution of action A(t), and the spray coordinate P. t-1 The injection direction of the unit with unit number Nu(t-1) and center coordinates The specific value is determined by the action A(t-1) in step t-1, that is, the state ψ(t) is obtained by performing action A(t-1) in state ψ(t-1);

[0136] Define the reward δ(t) for execution step t. The reward δ(t) is the immediate reward obtained by the spray gun after performing action A(t) in state ψ(t). The specific formula for reward δ(t) is as follows:

[0137]

[0138] Where δ(0) is the fixed reward after the spray gun performs action A(t), ω1, ω2 and ω3 are the movement balance coefficient, angle balance coefficient and constraint balance coefficient, respectively, d(P(t),P(t-1)) represents the distance the spray gun moves from the spray coordinate P(t-1) to the spray coordinate P(t) after performing action A(t), and θ(t,t-1) represents the distance the spray gun moves from the spray direction after performing action A(t). Rotate to the direction of spray Adjust the direction and angle. Let the injection coordinates P(t) and the center coordinates of the element with element number Nu(t) be... The spray distance between them, τ() is the penalty function, when the spray distance When within the adjustable distance range, the penalty function value is 0; when the spray distance... When the distance exceeds the adjustable range, the penalty function value takes the largest positive integer value. According to the reward δ(t), after the spray gun performs action A(t), the larger the spray gun's movement distance d(P(t),P(t-1)), the larger the direction adjustment angle θ(t,t-1), or the larger the spray distance... If the value is outside the range, the reward δ(t) decreases.

[0139] Define the value G(t) of execution step t. The value G(t) is the comprehensive value generated by considering past gains, the immediate gains from action A(t), and future gains. The specific formula is as follows:

[0140] G(t)=(1-ω4)×G(t-1)+ω4×(δ(t)+ω5·maxG(t+1));

[0141] Where ω4 and ω5 are the learning rate and discount factor, respectively, reflecting the importance of past and future returns. maxG(t+1) is the maximum value obtained by action A(t+1) based on state ψ(t+1) after executing action A(t). This is because the sprayed units do not need to be sprayed again. Each time an action is executed, the number of selectable units for the next execution step decreases. The value G(t) considers the comprehensive returns and guides the spray gun not to focus solely on the reward δ(t).

[0142] By iteratively selecting different actions for each execution step to update the value G(Q), the spraying sequence of the spray gun and the spraying coordinates of Q units when the value G(Q) is maximized are selected as the optimal spraying sequence and the optimal spraying coordinates of Q units.

[0143] This invention discloses a method for semiconductor coating path planning and pose estimation for heat-generating clothing. It analyzes the correlation between positional and grayscale information in point cloud data using a point convolutional network, mines spectral information from spectral data using a spectral encoder, and achieves spatial alignment through a sampling algorithm. Based on the Pearson correlation coefficient, a fused feature map is generated by combining the outputs of the point convolutional network and the spectral encoder. This map is then concatenated with a normal vector map and mapped using a multilayer perceptron to generate a segmentation probability map. A segmentation label vector is then generated based on conditional random fields to determine the coating region. The coating region is divided into multiple units, and the center coordinates, center normal vector, and curvature of each unit are calculated. Pre-set some spraying parameters, construct the thickness distribution of each unit based on the Gaussian decay model to describe the relationship between the spraying direction, spray volume and thickness at any position of the spray gun in each unit. Through simulation iteration using the alternating direction multiplier method, the optimal spraying direction and optimal spray volume of each unit that makes the overall thickness distribution conform to the standard thickness are obtained. The reward function is designed by taking into account the total movement of the spray gun, the total direction adjustment, and the reasonable spraying distance of each unit to guide the reinforcement learning simulation training of the spray gun, generate the optimal spraying sequence and the optimal spraying coordinates of each unit, and realize adaptive path planning and spraying parameter estimation for the geometric characteristics of the coating area.

[0144] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for semiconductor film path planning and pose estimation for heat-generating clothing, characterized by Includes the following steps: Confirming a coating area based on a target detection method and divided into units and has an overlapping zone between adjacent units, the center coordinates, central normal vector and curvature of each unit are calculated, and the material temperature and spraying speed are preset, and the center coordinates are the target positions of the spray gun spraying. Considering the effect of cell curvature on the flowability of the coating material, a thickness distribution for each cell is constructed based on a Gaussian decay model. This thickness distribution describes the spray direction, spray volume, and point coordinates of the spray gun in each cell. The relationship between coating thickness and superposition The thickness distribution of each unit generates the overall thickness distribution. Through simulation and iteration using the alternating direction multiplier method, the optimal injection direction and optimal injection amount of each unit that makes the overall thickness distribution conform to the standard thickness are calculated. Taking into account the total movement and directional adjustment of the spray gun The design of a reward function for the reasonable spray distance of each unit guides the reinforcement learning simulation training of the spray gun, generating the optimal spray sequence and the optimal spray coordinates for each unit. The spray coordinates are the positions of the spray gun when it sprays. Among them, the target detection method is used to confirm the coating area. This includes the following steps: Based on the calibrated coordinate system of the unified device pose scanner and spectrometer, point cloud data and spectral data are generated. Point cloud data is input into a point convolutional network. Dynamic graph convolution is used to aggregate the coordinates and gray values ​​of the points in the neighborhood of each point in the point cloud data and generate an enhanced feature map based on attention. Spectral data is input into a spectral encoder. The spectral information mined by different dilated convolutional kernels is spatially aligned with the enhanced feature map through a sampling algorithm to generate an aligned feature map. The aligned feature map and the enhanced feature map are concatenated into a fused feature map based on the Pearson correlation coefficient. The normal vectors of all points in the point cloud data are calculated and concatenated into a normal vector map. This map is then concatenated with the fused feature map and mapped through a multilayer perceptron to generate a segmentation probability map. Finally, a segmentation label vector is generated based on a conditional random field to determine the coating area. .

2. The semiconductor coating path planning and pose estimation method for heat-generating clothing as described in claim 1, characterized in that, The thickness distribution of each unit, constructed based on the Gaussian attenuation model, includes: Under a given material temperature and injection velocity, the first Unit thickness distribution It conforms to the Gaussian decay model, as follows: ， in, , and The first Unit The amount of spray, unit area and point coordinates With the Unit center coordinates The square of the distance, For the coating material in the first Unit The flow coefficient, and the injection velocity and the Unit curvature Showing a positive correlation, Point coordinates Curvature correction injection direction With the Unit The center normal vector The included angle, coordinates of the point Curvature correction injection direction By the Unit Spray direction Curvature Affects offset generation .

3. The semiconductor coating path planning and pose estimation method for heat-generating clothing as described in claim 1, characterized in that, The processing steps of the point convolutional network include: Extracting the first point cloud data The grayscale value and coordinates of the nth point are concatenated to form the nth point. A point vector; Aggregating the first by using dynamic graph convolution The point vectors in the neighborhood of the i-th point are used to generate the i-th point. The point feature vectors of each point are concatenated to generate a point feature vector map. The point feature vector map is pooled and linearly transformed in the channel dimension and spatial dimension respectively, and spatial attention weight vector and channel attention weight vector are generated based on the Sigmoid function mapping. The boundary enhancement of the point feature vector map is performed using the matrix multiplication result of the spatial attention weight vector and the channel attention weight vector to generate an enhanced feature map.

4. The semiconductor coating path planning and pose estimation method for heat-generating clothing as described in claim 1, characterized in that, The processing steps of the spectral encoder include: Extracting the first spectral data The reflectance vector of a single light spot; use The number of hollow convolution kernels with different expansion rates is the first. The reflectance vectors of the n light points are each subjected to dilated convolution to mine and generate the nth light point. A point of light The spectral feature vectors of different views are concatenated to form the first... Multi-view feature vector of a single light point; Based on sampling algorithm The multi-view feature vectors of each light point are interpolated and predicted. The predicted multi-view feature vectors at the coordinates of all points in the point cloud data are obtained and integrated into an aligned feature map. The sampling algorithms include nearest neighbor interpolation, bilinear interpolation, inverse distance weighted interpolation, and kriging interpolation.

5. The semiconductor coating path planning and pose estimation method for heat-generating clothing as described in claim 1, characterized in that, The simulation training steps of the reinforcement learning include: Define execution step Rewards ,award For the spray gun to be in condition Execute action Immediate rewards afterwards The formula is as follows: ， in, Perform actions for the spray gun The fixed reward afterwards , and These are the moving balance coefficient, the angular balance coefficient, and the constraint balance coefficient, respectively. Indicates the spray gun's spray coordinates To the jet coordinates The distance traveled For the spray gun from the spray direction Rotate to the direction of spray Adjust the direction and angle. For the spray coordinates With center coordinates The spray distance, As a penalty function, it applies only at the spray distance. When the distance exceeds the adjustable range, the value is taken as the largest positive integer, where, For action Unit number of the unit you are about to visit ; Define execution step value ,value Equal to execution step value Execution Steps Rewards With the execution of actions Later in state Based on the action Maximum value available The weighted sum.

6. The semiconductor coating path planning and pose estimation method for heat-generating clothing as described in claim 1, characterized in that, The process of generating segmentation label vectors based on conditional random fields includes the following steps: Obtain the first from the segmentation probability map A probability vector of points; Based on the negative log-likelihood form and the segmentation probability graph, the first... The point potential function at the nth point reflects the point potential function of the nth point. The confidence level of the label for each point is either 1 or 2. The label of the nth point can be either 1 or 2, representing the nth point, respectively. The points are located in the coating area. Or non-coated areas; Definition of the first The point and the first Within the neighborhood of point i, the first The label edge potential function of each point is used to indicate whether the label values ​​of two points are the same; According to the The larger probability among the probability vectors of the points is used to initialize the first... The values ​​of the labels at the nth point are summed. The point potential function value of the nth point and the nth point The point and the first The boundary potential function values ​​of other points in the neighborhood of the nth point are obtained. Energy value at each point; The iterative conditional pattern algorithm is used to iteratively update the label value of each point to find the label value of each point that minimizes the sum of the energy values ​​of all points, and then concatenates them into a segmented label vector.

7. The semiconductor coating path planning and pose estimation method for heat-generating clothing as described in claim 1, characterized in that, The calculation of the center coordinates, center normal vector, and curvature of each element includes the following steps: Clustering all clusters located at the first cluster center using the cluster center estimation method Unit The points in the middle are used to solve for the generation of the first... Unit center coordinates ; Calculate the first Unit The coordinates of each point and the center coordinates The inverse distance is used to generate the first point by weighting the normal vector of each point with the inverse distance. Unit The center normal vector ; For the first Unit The center normal vector The average of the angles between the vector and the center normal vector of each adjacent cell is taken as the first value. Unit curvature .

8. The semiconductor coating path planning and pose estimation method for heat-generating clothing as described in claim 7, characterized in that, Solving the first cluster using the cluster center estimation method Unit center coordinates This includes the following steps: In the Unit Randomly select one point from all points within the cluster as the cluster point; Calculate the minimum distance for each non-clustered point, where the minimum distance is equal to the distance between the non-clustered point and its nearest existing clustered point; Divide the square of the minimum distance of each non-clustered point by the sum of the squares of the minimum distances of all non-clustered points to generate the selection value of each non-clustered point. Set the non-clustered point with the highest selection value as the next clustered point, and stop when the total number of clustered points reaches the preset value. In each iteration, calculate the... Unit The distance between each non-clustered point in this round and the existing clustered points is calculated, and the points are classified into the class corresponding to the clustered point with the smallest distance. The average coordinates of all points in each class are used as the clustered point of each class in this round. Calculate the average distance between the cluster points of all classes in the current round and the cluster points of the previous round, and compare it with the distance threshold to decide whether to start the next round of iteration or obtain the cluster points and the number of points of each class in the current round; Based on the number of points in each category as a percentage of the total number of points... Unit The ratio of the total number of points within each cluster is used to perform a weighted sum of the coordinates of the points in each cluster, generating the first cluster. Unit center coordinates .

9. The semiconductor coating path planning and pose estimation method for heat-generating clothing as described in claim 1, characterized in that, The process of simulating and iterating using the alternating direction multiplier method to calculate the optimal injection direction and optimal injection amount for each unit that makes the overall thickness distribution conform to the standard thickness includes the following steps: The objective function is constructed as the mean square error between the overall thickness distribution and the standard thickness. The optimization problem is then established by minimizing this objective function, and constraints are set. The injection volume of each unit is non-negative; The constraints are applied using the alternating direction multiplier method. Introducing Lagrange multipliers into the objective function to construct the Lagrange augmented function, initializing... The injection direction, injection quantity, and Lagrange multipliers of each unit are used as optimization variables. The goal is to minimize the Lagrange augmented function, and the optimal injection direction and optimal injection quantity are solved iteratively.

Images