Selective laser melting process monitoring method combining plume dynamic characteristics and graph convolutional network
Through the combination of plume dynamic features and graph convolution network, graph structure data is constructed and graph convolution deep learning model is used, which solves the problem of inaccurate plume monitoring in the existing technology, and realizes efficient and accurate monitoring of the selected laser melting process and accurate identification of melting state.
Patent Information
- Application Number
- CN202410987198.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-23
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2044-07-23
AI Technical Summary
The existing plume-based selection laser melting process monitoring methods cannot carefully and accurately describe the plume morphology and instantaneous dynamics under processing conditions, resulting in inaccurate identification of melting states and high error detection rate.
Using a method of combining plume dynamic features with graph convolution networks, we can collect and process plume images, build graph structure data, and use graph convolution deep learning model to extract spatiotemporal features and melting state classification, including spatiotemporal gated convolution module and residual graph convolution module improve the model's ability to extract time-dimensional features.
It improves the accuracy of melting state recognition, realizes efficient and accurate monitoring of the selected laser melting process, can effectively describe the dynamic changes and intrinsic correlation of plume flow, and reduces the false detection rate.
Smart Images

Figure CN118982502A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of selective laser melting process monitoring, and specifically is a selective laser melting process monitoring method combining plume dynamic characteristics with a graph convolutional network. Background Art
[0002] The high degree of customization and excellent processing accuracy of selective laser melting show its advantages in different fields such as aerospace, automotive industry, and medical and health sectors. Since selective laser melting is a fast and non-stationary process, and it is difficult to directly measure the quality of the powder after melting and forming online, the use of process signals such as sound and optics for selective laser melting process monitoring has become one of the key issues in recent years. Among them, collecting optical signal characteristics during the processing process by high-speed cameras is an important method for monitoring the selective laser melting process. This method can capture the plume image composed of metal vapor and plasma during the processing process. Existing studies have shown that there is a close relationship between plumes and process stability. The feasibility and effectiveness of plume-based selective laser melting process monitoring have been proven in many cases, but they are only in the theoretical stage and have not been put into actual process monitoring applications.
[0003] At the same time, most of the current plume-based selective laser melting process monitoring methods only calculate simple statistical features of the plume, such as plume length, area, brightness, etc., and input these features into shallow machine learning models for melting state identification. These simple statistical features cannot describe the plume morphology under each processing condition in detail and accurately, and cannot reflect the instantaneous dynamics and dynamic behavior of the plume. At the same time, simple machine learning models generally only take the plume features at the current moment as input, and do not contain dynamic information about the changes in plume features at continuous moments. These models also usually do not have the ability to extract time series signal features, so there is a high false detection rate, and the identification of the melting state is not accurate. Summary of the invention
[0004] In order to solve the deficiencies of the above-mentioned prior art, the present invention proposes a method for monitoring the selective laser melting process by combining plume dynamic characteristics with a graph convolutional network, in order to improve the accuracy of melting state recognition and solve the problem of low classification accuracy in the melting state recognition of selective laser melting, thereby realizing efficient and accurate monitoring of the selective laser melting process.
[0005] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:
[0006] The characteristic of the selective laser melting process monitoring method combining plume dynamic characteristics with graph convolutional network of the present invention is that it is carried out according to the following steps:
[0007] Step 1: Collect near-infrared images of the selected area laser melting process and crop them to retain the plume part, and convert the cropped images into grayscale images to obtain the plume image dataset. in, represents the nth plume image, y n represents the nth plume image I n Corresponding to the category label of the molten state, C, H and W represent the number of channels, height and width of the plume image; n = 1, 2, …, N, N represents the total number of plume images;
[0008] Step 2, threshold segmentation is performed on the plume image and the plume dynamic feature V is established;
[0009] Each plume feature point corresponding to the plume feature point value set in V is used as each node of the graph, each plume feature value in V is used as the feature on the corresponding node, and A and D are used as the adjacency matrix and degree matrix of the graph respectively, so as to construct graph structure data;
[0010] Step 3: Establish a graph convolution network based on graph convolution deep learning, including: spatiotemporal feature extraction module and feature output module;
[0011] Step 3.1: The spatiotemporal feature extraction module is composed of L residual graph convolution modules;
[0012] When l = 1, the plume dynamic features V, adjacency matrix A and degree matrix D are input into the lth residual graph convolution module for processing to obtain the lth plume spatiotemporal feature map H l ;
[0013] When l = 2, 3, ..., L, the l-1th plume spatiotemporal characteristic map H l-1 , the adjacency matrix A and the degree matrix D are input into the lth residual graph convolution module for processing, and the lth plume spatiotemporal feature map H is obtained. l , so that the Lth residual graph convolution module outputs the Lth plume spatiotemporal graph H L , and serves as a coding feature of plume motion;
[0014] Step 3.2: The feature output module includes c linear layers and one Sigmoid layer;
[0015] For H L After the flattening operation, the one-dimensional plume motion encoding feature is obtained, and then it is processed by c linear layers and a Sigmoid layer in sequence to obtain the melting state probability vector Q with a dimension of 1×num, where num is the number of categories of the melting state category label;
[0016] Step 4. A cross entropy loss function is constructed based on Q and the true melting state label to update the parameters of the graph convolutional network, thereby obtaining a melting state classification model, which is used to identify the plume image in the selective laser melting process and obtain the melting state category of the current processing process, thereby realizing the monitoring of the selective laser melting process.
[0017] The method for monitoring a selective laser melting process by combining plume dynamic features with a graph convolutional network according to the present invention is also characterized in that the plume dynamic features V, the adjacency matrix A and the degree matrix D in step 2 are obtained according to the following steps:
[0018] Step 2.1: Use the OTSU algorithm to calculate the nth plume image I n Perform threshold segmentation to obtain the nth image segmentation contour of the plume and the molten pool, and define its boundary points as in, is the coordinates of the i-th boundary point of the n-th image segmentation contour on the x-axis and y-axis, i = 1, 2, ..., P n , P n Represents the total number of boundary points on the nth image segmentation contour, and the minimum y-axis coordinate in the nth image segmentation contour is recorded as
[0019] Set the y-axis coordinate to arrive The contour within the range is regarded as the nth plume image I n The molten pool contour is set to The contour of n The plume profile of ; where h is the pixel-level height of the melt pool;
[0020] The centroid of the melt pool contour in the nth image segmentation contour is recorded as
[0021] Step 2.2: Take the centroid of the melt pool contour As a starting point, we will follow the nth plume image I n The positive direction of the x-axis is defined as the starting vector;
[0022] After the starting vector is rotated counterclockwise with a step length of m°, I n M rotation vectors;
[0023] Combine the M rotation vectors with the nth plume image I n The intersection points on the plume contour are defined as plume feature points, thus obtaining the plume feature point set F n ={f n,u |u=1,...,M}, where f n,urepresents the u-th plume feature point, that is, the u-th rotation vector and I n The intersection of the plume contours;
[0024] Construct the adjacency matrix as A = {a uv |u=1,..,M,v=1,..,M}, where, a uv represents the degree of the angle between the u-th rotation vector and the v-th rotation vector among the M rotation vectors;
[0025] The degree matrix D = {d uu |u=1,..,M} is defined as a diagonal matrix, where d uu represents the element in the uth row and uth column of the degree matrix D, and
[0026] Calculate the plume feature point to the centroid of the melt pool contour The Euclidean distance of n The plume eigenvalue set S n ={s n,v |v=1,...,M}, where s n,v represents the vth plume characteristic value, that is, the vth plume characteristic point f u,v To the centroid of the melt pool contour The Euclidean distance of
[0027] Get the nth plume image I n To the n+tth plume image I n+t The plume characteristic values are stacked according to the time dimension to obtain the plume dynamic characteristics V = {S j |j=n,..,n+t}, where S j Table 1 The jth plume image I j The set of plume eigenvalues .
[0028] Each residual graph convolution module in step 3.1 includes a graph convolution layer, a spatiotemporal gated convolution module and a residual edge module;
[0029] Step 3.1.1, the graph convolution layer in the lth residual graph convolution module uses formula (1) to obtain the lth plume spatial domain feature gcn l :
[0030]
[0031] In formula (1), I e Representation and The same dimension unit matrix, when l = 1, let H l-1 =V,W l is the weight to be learned in the graph convolution layer of the lth residual graph convolution module;
[0032] Step 3.1.2: The spatiotemporal gated convolution module in the lth residual graph convolution module is used for gcn l Process and output the lth plume time domain feature tcn l ;
[0033] Step 3.1.3: The residual edge module in the lth residual graph convolution module uses the convolution layer, batch normalization layer and ReLU activation function to sequentially l-1 Process and output the lth plume residual feature res l ;
[0034] Step 3.1.4: TCN l and res l After adding, we get H l .
[0035] The step 3.1.2 includes:
[0036] Step a: The spatiotemporal gated convolution module uses a 1D convolutional layer to convolutionally map the gcn l Processing is performed to obtain the lth 1D convolution feature
[0037] Step b: The spatiotemporal gated convolution module uses the gated linear unit GLU to convolution gcn l and After processing, the lth plume time domain feature tcn is obtained l ;
[0038] The linear gating unit GLU converts gcn into l Average split into two dilated convolution input features and in, After a dilated convolution layer and a Sigmoid function, the lth gated probability feature is obtained. After processing by another dilated convolutional layer, we get the lth gated numerical feature
[0039] Will and After addition, we get the lth gated convolution feature
[0040] Will and After multiplication, we get the lth plume time domain feature tcn l ;
[0041] An electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the selective laser melting process monitoring method, and the processor is configured to execute the program stored in the memory.
[0042] The present invention provides a computer-readable storage medium, and the computer-readable storage medium stores a computer program, which is characterized in that when the computer program is run by a processor, the steps of the selective laser melting process monitoring method are executed.
[0043] Compared with the prior art, the present invention has the following beneficial effects:
[0044] 1. The present invention proposes a method for monitoring the selective laser melting process based on plume information. The method adopts an advanced graph convolution deep learning model, takes the plume dynamic map obtained in the real-time processing process as input, and takes the final melting state as a label, thereby solving the mapping problem between the plume signal and the final melting quality, and realizing the prediction of the melting state through an optical monitoring platform, thereby realizing online selective laser melting forming quality monitoring.
[0045] 2. The present invention proposes a method for establishing plume dynamic features with a graph structure. By combining the plume image features at multiple time steps, the two-dimensional image signal is converted into graph structure data that can effectively describe the dynamic changes of the plume, and the nodes, features, adjacency matrix and degree matrix of the graph structure are defined. The feature contains the plume change information under the time series, and improves the intrinsic correlation between the dynamic behavior of the plume and the melting state.
[0046] 3. The present invention designs a graph convolution network based on graph convolution to realize the classification of melting state. The model uses time-gated convolution to improve the model's ability to extract time dimension features, and realizes the feature extraction of plume through the residual graph convolution module. The model has a high accuracy rate for melting state classification, thereby effectively improving the accuracy of selective laser melting process monitoring. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] Figure 1 It is a flow chart of the method for monitoring the selective laser melting process based on plume dynamic characteristics and graph convolutional network of the present invention;
[0048] Figure 2 is a schematic diagram of the experimental platform and monitoring equipment of the present invention;
[0049] Figure 3 It is a schematic diagram of the plume image collected by the present invention and the single-track electron microscope image corresponding to the five types of melting state labels;
[0050] Figure 4 It is a structural diagram of the graph convolutional network of the present invention;
[0051] Figure 5 It is a structural diagram of the residual graph convolution module of the present invention;
[0052] Figure 6 It is a structural diagram of the time-gated convolution module of the present invention. DETAILED DESCRIPTION
[0053] In this embodiment, a method for monitoring a selective laser melting process is provided by combining plume dynamic characteristics with a graph convolutional network. The specific process is as follows: Figure 1 As shown, follow the steps below:
[0054] Step 1: Collect near-infrared images of the selected area laser melting process and crop them to retain the plume part. In this example, the experimental platform is as follows: Figure 2 As shown, the cropped image is converted into a grayscale image to obtain the plume image dataset in, represents the nth plume image, y n represents the nth plume image I n Corresponding to the category label of the molten state, C, H and W represent the number of channels, height and width of the plume image; n = 1, 2, …, N, N represents the total number of plume images.
[0055] In this embodiment, the model of the selective laser melting equipment used is the YLR-200-SM-AC laser with a laser power range of 0-200w. The high-speed camera model of the plume image acquisition equipment is FASTCAMMiniUX50 / 100 (near-infrared camera), and an external bandpass filter of 700-1000nm is installed. The angle between the camera and the platform is 30°, and the camera sampling frequency is 3000PFS. The metal powder material used in the embodiment is 304L stainless steel powder, and by setting different process parameters, a single-track melting experiment is carried out to simulate five different melting states. The collected plume images are 3 channels with an image size of 120×160. A total of 2,800 plume images are collected to form a data set. The five melting states are normal melting, slightly overmelting, overmelting, slightly undermelting, and undermelting. The plume image examples and the electron microscope images of the melt channels of the five melting states are shown in the figure. Figure 3 shown.
[0056] Step 2, threshold segmentation is performed on the plume image and the plume dynamic feature V is established;
[0057] Step 2.1: Use the OTSU algorithm to calculate the nth plume image I n Perform threshold segmentation to obtain the nth image segmentation contour of the plume and the molten pool, and define its boundary points as in, is the coordinates of the i-th boundary point of the n-th image segmentation contour on the x-axis and y-axis, i = 1, 2, ..., P n , P n Represents the total number of boundary points on the nth image segmentation contour, and the minimum y-axis coordinate in the nth image segmentation contour is recorded as
[0058] Set the y-axis coordinate to arrive The contour within the range is regarded as the nth plume image I n The molten pool contour is set to The contour of n The plume profile of the molten pool; where h is the pixel-level height of the molten pool. In this embodiment, h=20;
[0059] The centroid of the melt pool contour in the nth image segmentation contour is recorded as
[0060] Step 2.2: Take the centroid of the melt pool contour As a starting point, we will follow the nth plume image I n The positive direction of the x-axis is defined as the starting vector;
[0061] After the starting vector is rotated counterclockwise with a step size of m°, I is obtained. n M rotation vectors.
[0062] Combine the M rotation vectors with the nth plume image I n The intersection points on the plume contour are defined as plume feature points, thus obtaining the plume feature point set F n ={f n,u |u=1,...,M}, where f n,u represents the u-th plume feature point, that is, the u-th rotation vector and I n The intersection of the plume contours.
[0063] Construct the adjacency matrix as A = {a uv |u=1,..,M,v=1,..,M}, where, a uv represents the degree of the angle between the u-th rotation vector and the v-th rotation vector among the M rotation vectors;
[0064] The degree matrix D = {d uu |u=1,..,M} is defined as a diagonal matrix, where d uu represents the element in the uth row and uth column of the degree matrix D, and
[0065] Calculate the plume feature point to the centroid of the melt pool contour The Euclidean distance ofn The plume eigenvalue set S n ={s n,v |v=1,...,M}, where s n,v represents the vth plume characteristic value, that is, the vth plume characteristic point f u,v To the centroid of the melt pool contour The Euclidean distance of
[0066] In this example, the angular step of the vector rotation is m°=10°, and a total of 19 rotation vectors are obtained, that is, M=19.
[0067] Get the nth plume image I n To the n+tth plume image I n+t The plume characteristic values are stacked according to the time dimension to obtain the plume dynamic characteristics V = {S j |j=n,..,n+t}, where S j Table 1 The jth plume image I j In this example, t=5.
[0068] The plume feature points corresponding to the plume feature point value set in V are used as nodes of the graph, the plume feature values in V are used as features on the corresponding nodes, and A and D are used as the adjacency matrix and degree matrix of the graph respectively, so as to construct graph structure data.
[0069] Step 3: Establish a graph convolutional network based on graph convolutional deep learning. In this example, the graph convolutional network structure used is as follows: Figure 4 As shown, it includes: a spatiotemporal feature extraction module and a feature output module;
[0070] Step 3.1: The spatiotemporal feature extraction module consists of L residual graph convolution modules;
[0071] When l = 1, the plume dynamic features V, adjacency matrix A and degree matrix D are input into the lth residual graph convolution module for processing to obtain the lth plume spatiotemporal feature map H l ; In this embodiment, the dimensions of V are 1×5×19, and the dimensions of A and D are 19×19.
[0072] When l = 2, 3, ..., L, the l-1th plume spatiotemporal characteristic map H l-1 , the adjacency matrix A and the degree matrix D are input into the lth residual graph convolution module for processing, and the lth plume spatiotemporal feature map H is obtained. l , so that the Lth residual graph convolution module outputs the Lth plume spatiotemporal graph H L, and used as plume motion encoding features; in this example, the number of residual graph convolution modules is 2, that is, L = 2, where the dimension of the plume spatiotemporal feature map output by the first residual graph convolution module is 32×3×19, and the dimension of the plume spatiotemporal feature map output by the second residual graph convolution module is 64×1×19, that is, H L The dimensions are 64×1×19.
[0073] Each residual graph convolution module in step 3.1 contains a graph convolution layer, a spatiotemporal gated convolution module, and a residual edge module. In this example, the residual graph convolution module used is as follows: Figure 5 As shown;
[0074] Step 3.1.1, the graph convolution layer in the lth residual graph convolution module uses formula (1) to obtain the lth plume spatial domain feature gcn l :
[0075]
[0076] In formula (1), I e Representation and The same dimension unit matrix, when l = 1, let H l-1 =V,W l is the weight to be learned in the graph convolution layer of the lth residual graph convolution module.
[0077] Step 3.1.2: The spatiotemporal gated convolution module in the lth residual graph convolution module is used for gcn l Process and output the lth plume time domain feature tcn l In this embodiment, the spatiotemporal gated convolution module is as follows: Figure 6 As shown:
[0078] Step a: The spatiotemporal gated convolution module uses a 1D convolutional layer to convolutionally map the gcn l Processing is performed to obtain the lth 1D convolution feature In this embodiment, the convolution kernel size of the 1D convolution layer is 1×1 and the stride is 1.
[0079] Step b: The spatiotemporal gated convolution module uses the gated linear unit GLU to convolution gcn l and After processing, the lth plume time domain feature tcn is obtained l ;
[0080] The linear gating unit GLU converts gcn according to the number of channels l Average split into two dilated convolution input features and in, After a dilated convolution layer and a Sigmoid function, the lth gated probability feature is obtained. After processing by another dilated convolutional layer, we get the lth gated numerical feature
[0081] Will and After addition, we get the lth gated convolution feature
[0082] Will and After multiplication, we get the lth plume time domain feature tcn l ;
[0083] In this embodiment, the convolution kernel size of the dilated convolution is 2×1, and the step size is 1.
[0084] Step 3.1.3: The residual edge module in the lth residual graph convolution module uses the convolution layer, batch normalization layer and ReLU activation function to sequentially l-1 Process and output the lth plume residual feature res l ; In this embodiment, the convolution kernel size of the convolution layer is 2×1, and the step size is 1.
[0085] Step 3.1.4: TCN l and res l After adding, we get H l .
[0086] Step 3.2: The feature output module contains c linear layers and one Sigmoid layer;
[0087] For H L After the flattening operation is performed, a one-dimensional plume motion coding feature is obtained, and then it is processed by c linear layers and one Sigmoid layer in sequence to obtain a melting state probability vector Q with a dimension of 1×num, where num is the number of categories of the melting state category label; in this embodiment, the one-dimensional plume motion coding feature dimension obtained after the flattening operation is 1×1216, the number of linear layers is 2, the output dimension of the first linear layer is 1×512, and the output dimension of the second linear layer is 1×5, that is, the dimension of Q is 1×5, where the melting state categories include normal melting, slightly overmelting, overmelting, slightly undermelting, and undermelting, that is, num=5.
[0088] Step 4: A cross entropy loss function is constructed based on Q and the true melting state label, which is used to update the parameters of the graph convolutional network, thereby obtaining a melting state classification model, which is used to identify the plume image in the selective laser melting process and obtain the melting state category of the current process, thereby realizing the monitoring of the selective laser melting process. In this embodiment, 2800 plume images are collected to form a data set, and are divided into a training set, a test set, and a validation set according to a ratio of 6:2:2. The Adam optimizer is used, and the initial learning rate is 0.01, where the learning rate is reduced by 10% every 10 iterations, and the training is iterated 200 times in total.
[0089] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the tool wear state monitoring method, and the processor is configured to execute the program stored in the memory.
[0090] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and when the computer program is executed by a processor, the steps of the method for monitoring the melting state of selected area laser melting are executed.
Claims
1. A method for monitoring a selective laser melting process by combining plume dynamic characteristics with a graph convolutional network, characterized in that: Follow these steps: Step 1: Collect near-infrared images of the selected area laser melting process and crop them to retain the plume part, and convert the cropped images into grayscale images to obtain the plume image dataset. in, represents the nth plume image, y n represents the nth plume image I n Corresponding to the category label of the molten state, C, H and W represent the number of channels, height and width of the plume image; n = 1, 2, …, N, N represents the total number of plume images; Step 2, threshold segmentation is performed on the plume image and the plume dynamic feature V is established; Each plume feature point corresponding to the plume feature point value set in V is used as each node of the graph, each plume feature value in V is used as the feature on the corresponding node, and A and D are used as the adjacency matrix and degree matrix of the graph respectively, so as to construct graph structure data; Step 3: Establish a graph convolution network based on graph convolution deep learning, including: spatiotemporal feature extraction module and feature output module; Step 3.1: The spatiotemporal feature extraction module is composed of L residual graph convolution modules; When l = 1, the plume dynamic features V, adjacency matrix A and degree matrix D are input into the lth residual graph convolution module for processing to obtain the lth plume spatiotemporal feature map H l ; When l = 2, 3, ..., L, the l-1th plume spatiotemporal characteristic map H l-1 , the adjacency matrix A and the degree matrix D are input into the lth residual graph convolution module for processing, and the lth plume spatiotemporal feature map H is obtained. l , so that the Lth residual graph convolution module outputs the Lth plume spatiotemporal graph H L , and serves as a coding feature of plume motion; Step 3.2: The feature output module includes c linear layers and one Sigmoid layer; For H L After the flattening operation, the one-dimensional plume motion encoding feature is obtained, and then it is processed by c linear layers and a Sigmoid layer in sequence to obtain the melting state probability vector Q with a dimension of 1×num, where num is the number of categories of the melting state category label; Step 4. A cross entropy loss function is constructed based on Q and the true melting state label to update the parameters of the graph convolutional network, thereby obtaining a melting state classification model, which is used to identify the plume image in the selective laser melting process and obtain the melting state category of the current processing process, thereby realizing the monitoring of the selective laser melting process.
2. According to claim 1, a method for monitoring a selective laser melting process combining plume dynamic characteristics with a graph convolutional network is characterized in that: The plume dynamic characteristics V, adjacency matrix A and degree matrix D in step 2 are obtained according to the following steps: Step 2.1: Use the OTSU algorithm to calculate the nth plume image I n Perform threshold segmentation to obtain the nth image segmentation contour of the plume and the molten pool, and define its boundary points as in, is the coordinates of the i-th boundary point of the n-th image segmentation contour on the x-axis and y-axis, i = 1, 2, ..., P n , P n Represents the total number of boundary points on the nth image segmentation contour, and the minimum y-axis coordinate in the nth image segmentation contour is recorded as Set the y-axis coordinate to arrive The contour within the range is regarded as the nth plume image I n The molten pool contour is set to The contour of n The plume profile of ; where h is the pixel-level height of the melt pool; The centroid of the melt pool contour in the nth image segmentation contour is recorded as Step 2.2: Take the centroid of the melt pool contour As a starting point, we will follow the nth plume image I n The positive direction of the x-axis is defined as the starting vector; After the starting vector is rotated counterclockwise with a step length of m°, I n M rotation vectors; Combine the M rotation vectors with the nth plume image I n The intersection points on the plume contour are defined as plume feature points, thus obtaining the plume feature point set F n ={f n,u |u=1,...,M}, where f n,u represents the u-th plume feature point, that is, the u-th rotation vector and I n The intersection of the plume contours; Construct the adjacency matrix A = {a uv |u=1,..,M,v=1,..,M}, where, a uv represents the degree of the angle between the u-th rotation vector and the v-th rotation vector among the M rotation vectors; The degree matrix D = {d uu |u=1,..,M} is defined as a diagonal matrix, where d uu represents the element in the uth row and uth column of the degree matrix D, and Calculate the plume feature point to the centroid of the melt pool contour The Euclidean distance of n The plume eigenvalue set S n ={s n,v |v=1,...,M}, where s n,v represents the vth plume characteristic value, that is, the vth plume characteristic point f u,v To the centroid of the melt pool contour The Euclidean distance of Get the nth plume image I n To the n+tth plume image I n+t The plume characteristic values are stacked according to the time dimension to obtain the plume dynamic characteristics V = {S j |j=n,..,n+t}, where S j Table 1 The jth plume image I j The set of plume eigenvalues .
3. According to claim 1, the method for monitoring the selective laser melting process by combining plume dynamic characteristics with graph convolutional networks is characterized in that: Each residual graph convolution module in step 3.1 includes a graph convolution layer, a spatiotemporal gated convolution module and a residual edge module; Step 3.1.1, the graph convolution layer in the lth residual graph convolution module uses formula (1) to obtain the lth plume spatial domain feature gcn l : In formula (1), I e Representation and The same dimension unit matrix, when l = 1, let H l-1 =V,W l is the weight to be learned in the graph convolution layer of the lth residual graph convolution module; Step 3.1.2: The spatiotemporal gated convolution module in the lth residual graph convolution module is used for gcn l Process and output the lth plume time domain feature tcn l ; Step 3.1.3: The residual edge module in the lth residual graph convolution module uses the convolution layer, batch normalization layer and ReLU activation function to sequentially l-1 Process and output the lth plume residual feature res l ; Step 3.1.4: TCN l and res l After adding, we get H l .
4. According to claim 3, the method for monitoring the selective laser melting process by combining plume dynamic characteristics with graph convolutional networks is characterized in that: The step 3.1.2 includes: Step a: The spatiotemporal gated convolution module uses a 1D convolutional layer to convolutionally map the gcn l Processing is performed to obtain the lth 1D convolution feature Step b: The spatiotemporal gated convolution module uses the gated linear unit GLU to convolution gcn l and After processing, the lth plume time domain feature tcn is obtained l ; The linear gating unit GLU converts gcn into l Average split into two dilated convolution input features and in, After a dilated convolution layer and a Sigmoid function, the lth gated probability feature is obtained. After processing by another dilated convolutional layer, we get the lth gated numerical feature Will and After addition, we get the lth gated convolution feature Will and After multiplication, we get the lth plume time domain feature tcn l .
5. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the method for monitoring the selective laser melting process as described in any one of claims 1 to 4, and the processor is configured to execute the program stored in the memory.
6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for monitoring a selective laser melting process according to any one of claims 1 to 4 are executed.
Citation Information
Patent Citations
Molten pool and splashing monitoring method in selective laser melting process based on target detection and application
CN115526852A
Heterogeneous graph convolution-based multi-modal comment sentiment analysis method and system
CN116258147A
Cited By
Combustible gas cloud imaging method based on refrigeration type infrared camera off-site sample learning
CN120014328A
