A real-time prediction method for the molten pool size in the selective laser melting process

Abstract

The present invention discloses a method for real-time prediction of molten pool size during selective laser melting, and the steps thereof include: 1. constructing a molten pool image data set; 2. extracting the length, width and area of the molten pool by processing the molten pool image; 3. extracting process features related to laser power, scanning speed, power density and scanning strategy; 4. constructing a molten pool size prediction model based on CNN and informer; 5. training the model. The present invention can overcome the defect that the existing models can only be used for predicting the average molten pool size of single-pass or multi-pass, and can perform relatively accurate real-time prediction of the molten pool size at multiple future moments under complex scanning strategies with different thermal histories.

Description

Technical Field

[0001] The present invention relates to the technical field of laser additive manufacturing process monitoring, and particularly relates to a method for real-time prediction of molten pool size during selective laser melting process. Background Art

[0002] Due to its advantages in manufacturing high-performance complex structural parts, selective laser melting technology has become one of the fastest-developing metal additive manufacturing processes. However, due to the lack of stability and repeatability during the manufacturing process, it is difficult to ensure product quality, which limits its wide application, especially in industries such as aerospace and medical that have high requirements for forming accuracy and performance. The molten pool is formed during the most basic subprocess of additive manufacturing. Its dynamic changes not only reflect the stability of the manufacturing process but also are the underlying cause of defects such as porosity and balling. Therefore, if the size of the molten pool can be predicted accurately in real time, the process parameters can be adjusted before an abnormal molten pool is formed, thereby maintaining a relatively stable melting state. This can reduce defects during the forming process and significantly improve the final quality of the part.

[0003] Regarding the problem of predicting the molten pool size during the selective laser melting process, there are currently physics-based modeling methods such as finite element analysis and computational fluid dynamics. However, these physical models have inherent assumptions, simplifications, and approximations, which reduce their accuracy and require a large amount of computational time, making it difficult to apply them to real-time monitoring processes. In addition, machine learning models such as Gaussian processes and multi-layer perceptrons are also used to predict the molten pool size during the forming process, but mainly with process parameters as inputs to predict the average molten pool size of single or multiple tracks. Generally, the scanning strategy automatically generated according to the three-dimensional model may be very complex. Changing manufacturing conditions such as the scanning strategy, laser power, and scanning speed will result in different thermal histories, and the molten pool size during the selective laser melting process is directly affected by the thermal history. In addition, the selective laser melting process has inherent randomness. For example, the same set of inputs (process parameters) may lead to different results, which may be related to environmental factors, system drift caused by long-term manufacturing, spatter generation, etc. Therefore, it is necessary to study a method that can capture the change in the molten pool size during the selective laser melting process in real time based on historical scanning conditions. Summary of the Invention

[0004] The present invention is proposed to solve the above-mentioned deficiencies of the prior art, and provides a method for real-time prediction of molten pool size during selective laser melting process, aiming to achieve real-time and accurate prediction of the molten pool size throughout the selective laser melting process under complex scanning strategies, thereby providing a basis for controlling the molten pool size and maintaining the stability of the melting process, and further reducing or even eliminating defects during the forming process, improving the geometric accuracy, surface quality, and mechanical properties of the manufactured parts, and ensuring the quality of the final product.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A real-time prediction method for the molten pool size during selective laser melting according to the present invention is characterized by comprising the following steps:

[0007] Step 1: Use a near-infrared high-speed camera to collect molten pool images during the selective laser melting process at a fixed frequency, and construct a molten pool image dataset I = {I1, I2, …, I i , … I N0}; where I i ∈R L×H represents the molten pool image at the i-th moment, and N0 is the number of molten pool image samples; L and H are the length and height of the image respectively;

[0008] Step 2: Extract the length, width and area of the molten pool by processing the molten pool image.

[0009] Step 2.1: Set an initial threshold, and perform binarization processing on the molten pool image I i at the i-th moment to obtain the molten pool boundary in the image I i ;

[0010] Obtain the width of the molten pool in the molten pool image I i according to the molten pool boundary, and compare it with the width of the actual scanning track to determine the optimal threshold of the molten pool boundary;

[0011] Perform binarization processing on all molten pool images according to the optimal threshold to obtain the optimal molten pool boundary of each molten pool image;

[0012] Step 2.2: Use the least square method to fit all the optimal molten pool boundaries into an ellipse, and convert the major axis length, minor axis length and ellipse area of the ellipse according to the instantaneous field of view angle of the near-infrared high-speed camera to obtain the length, width and area of the molten pool respectively, and use them as the size of the molten pool;

[0013] Take the molten pool image I i at the i-th moment to the molten pool image I i+k-1 at the i + k - 1-th moment as a sample, and take the molten pool size sequence of the molten pool image I i at the i-th moment to the molten pool image I i+k-1 at the i + k - 1-th moment as the label Y i,i+k-1 of the sample;

[0014] Step 3: Extract process features related to laser power, scanning speed, power density and scanning strategy;

[0015] Step 3.1: Calculate the laser power p at the (i - 1)-th moment corresponding to the label Y i,i+k-1 respectively by using equations (1) to (3)i-1 、The scanning speed v at the (i - 1)-th moment i-1 、The power density E at the (i - 1)-th moment i-1 :

[0016]

[0017]

[0018]

[0019] In equations (1) - (3), n is the number of laser scanning points between the (i - 1)-th moment and the i-th moment, and are respectively the laser power and the laser incident angle of the j-th scanning point at the (i - 1)-th moment, and are respectively the distance and the scanning time between the (j - 1)-th to the (j + 1)-th scanning points at the (i - 1)-th moment;

[0020] Step 3.2: Calculate the label Y using equations (4) - (9) i,i+k-1 The laser moving direction at the (i - 1)-th moment corresponding to The X-direction distance at the (i - 1)-th moment The Y-direction distance at the (i - 1)-th moment The Euclidean distance at the (i - 1)-th moment The laser power feature with scanning distance weight added at the (i - 1)-th moment The laser power density feature with scanning distance weight added at the (i - 1)-th moment

[0021]

[0022]

[0023]

[0024]

[0025]

[0026]

[0027] In equations (4) - (9), is the laser moving unit direction vector of the j-th scanning point at the (i - 1)-th moment, e0 is the X-axis unit direction vector of the working coordinate system, and g is the function for calculating the angle, is the position of the j-th scanning point at the (i - 1)-th moment in the X-axis direction of the working coordinate system, is the position of the j-th scanning point at the (i - 1)-th moment in the Y-axis direction of the working coordinate system; is the position of the 1st scanning point at the i-th moment in the X-axis direction of the working coordinate system, is the position of the 1st scanning point at the i-th moment in the Y-axis direction of the working coordinate system;

[0028] Step 3.3: Let the label Y i,i+k-1 All the process features at the (i - 1)-th moment corresponding to it are Calculate the process features from the (i - N)-th moment to the (i - 2)-th moment according to the process of Step 3.1 to Step 3.2, so as to obtain the process feature sequence of the label Y i,i+k-1 from the (i - N)-th moment to the (i - 1)-th moment corresponding to it is

[0029] Step 4: Construct a molten pool size prediction model, including: an extraction module for the region of interest of the molten pool, a CNN module for extracting the features of the molten pool image, and an informer module for fusing the process features and the molten pool features;

[0030] Step 4.1: The extraction module calculates the centroid of the molten pool image sequence I i,i+k-1 from the (i - N)-th moment to the (i - 1)-th moment corresponding to the label Y i-N,i-1 ={I i-N , I i-(N-1) , …, I i-1}, and determines a rectangular region with side lengths of l×h centered on the centroid of the molten pool as the sequence of regions of interest of the molten pool i-N,i-1 in the molten pool image sequence I

[0031] Step 4.2: The CNN module sequentially includes: two convolutional pooling layers and one fully connected layer;

[0032] Input the sequence of regions of interest of the molten pool from the (i - N)-th moment to the (i - 1)-th moment into the CNN module for processing, and obtain the sequence of molten pool image features i,i+k-1 corresponding to the label Y wherein, represents the feature of the region of interest of the molten pool at the (i - 1)-th moment;

[0033] Step 4.3: The informer module is used to process and to obtain the decoded feature

[0034] Step 4.4: After passing through a fully connected layer, the decoded feature outputs the predicted sequence of molten pool sizes

[0035] Step 5: Construct the loss function loss using Equation (10):

[0036]

[0037] In Equation (10), is the predicted value of the molten pool length sequence in , is the predicted value of the molten pool width sequence in , is the predicted value of the molten pool area sequence in , and λ is a parameter determining the relative importance of this loss term;

[0038] Train the molten pool size prediction model using the gradient descent method, and calculate the total loss function loss to update the model parameters. When the total loss function loss converges, obtain the trained molten pool size prediction model for real-time prediction of the molten pool size at multiple future moments during the selective laser melting process.

[0039] Another feature of the real-time prediction method for the molten pool size during the selective laser melting process according to the present invention is that the informer module in Step 4.3 includes: an encoder module and a decoder module;

[0040] Step 4.3.1: The encoder module includes: a preprocessing module, a multi-head probabilistic sparse self-attention mechanism module, and a distillation module;

[0041] Step a. The preprocessing module normalizes the process feature sequence i,i+k-1 corresponding to the label Y to obtain the normalized process feature sequence where represents the normalized process feature at the (i - 1)-th moment;

[0042] After splicing with the molten pool image feature sequence to obtain the fused feature sequence and performing Embedding operation and position encoding in sequence, obtain the high-dimensional fused feature vector with embedded position information

[0043] Step b. The multi-head probabilistic sparse self-attention mechanism module performs multi-head probabilistic sparse self-attention mechanism operation on to obtain the fused feature integrated with the attention mechanism, and performs residual connection with the input , then performs layer normalization operation, and finally inputs it into the feed-forward neural network and outputs the attention module feature

[0044] Step c: The distillation module performs attention module features in the time dimension After the one-dimensional convolution operation, the ELU activation function is used for nonlinear transformation, and then the maximum pooling operation is performed to finally obtain the encoding features output by the encoder.

[0045] Step 4.3.2: Label Y i,i+k-1 The corresponding fusion feature sequence from the iMth moment to the i-1th moment and the fusion feature sequence from the i-th moment to the i+k-1th moment Features obtained after splicing And input into the decoder module, after the Embedding operation, position encoding, and mask multi-head probabilistic sparse self-attention mechanism are processed in sequence, and then combined with the encoded feature Perform multi-head attention mechanism operation together to finally obtain the decoding features of the decoder output

[0046] 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 real-time prediction method, and the processor is configured to execute the program stored in the memory.

[0047] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the real-time prediction method when executed by a processor.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1. The present invention extracts process features related to laser power, scanning speed, power density and scanning strategy that contain complex characteristics of the scanning process, and combines machine learning methods to predict the molten pool size, thereby overcoming the problem that only the average molten pool size of a single channel or multiple channels can be predicted in the prior art, thereby greatly improving the real-time performance of molten pool prediction.

[0050] 2. The present invention predicts the molten pool size by fusing the molten pool image features and process features containing the inherent randomness of the forming process, thereby overcoming the problem in the prior art that the inherent randomness of the selective laser melting process cannot be considered, thereby improving the prediction accuracy of the molten pool size.

[0051] 3. By leveraging advanced technologies in the field of machine learning, the present invention extracts the features of the molten pool image through the CNN module, fuses them with the process features, and inputs them into the Informer module to predict the molten pool size sequence, avoiding the prediction of the molten pool size through the finite element analysis method and greatly reducing the calculation time. At the same time, the unique advantages of the Informer network in long-term sequence dependence and long-term sequence prediction are utilized to further improve the accuracy of the molten pool size sequence prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 are schematic diagrams of different scanning strategies;

[0053] Figure 2 is a schematic diagram of molten pool image monitoring;

[0054] Figure 3 is a diagram of the molten pool size prediction model of the present invention;

[0055] Figure 4 is a structural diagram of the CNN module. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0056] In this embodiment, a method for real-time prediction of the molten pool size during the selective laser melting process includes the following steps:

[0057] Step 1: Use a near-infrared high-speed camera to collect the molten pool images during the selective laser melting process at a fixed frequency and construct a molten pool image dataset where I i ∈R L×H represents the molten pool image at the i-th moment, and N0 is the number of molten pool image samples; L and H are the length and height of the image respectively; in this embodiment, L and H are 120 and 128 respectively;

[0058] In this embodiment, the image dataset used is collected by a high-speed camera with a near-infrared band-pass filter through coaxial monitoring. The model of the high-speed camera is Mikrotron EoSens S 3CL, and the sampling frequency is 2 kHz.

[0059] In a specific example, the laser wavelength of the selective laser melting system is 1070 nm, and the laser spot diameter is 85 μm. The powder material in the experiment is forged nickel alloy 625, and the average powder size is 30.6 μm. In the experiment, 12 rectangular parts with dimensions of 10 mm × 10 mm × 5 mm were fabricated using different scanning strategies. The different scanning strategies are based on Figure 1The three basic scanning strategies are obtained by changing the scanning pitch, laser power, single-track scanning direction, and remelting strategy. Considering the gray body radiation of nickel-based alloys at the solidus temperature (1500 °C) and the laser wavelength to be avoided, an 850 nm (±20 nm) band-pass filter is selected to capture the molten pool area as accurately as possible. Figure 2 Schematic diagram of the high-speed camera process monitoring adopted for this example.

[0060] Step 2: By processing the molten pool image, extract the length, width, and area of the molten pool;

[0061] Step 2.1: Set an initial threshold and perform binary processing on the molten pool image I at the i-th moment i to obtain the molten pool boundary in the image I i ;

[0062] Obtain the width of the molten pool in the molten pool image I according to the molten pool boundary, and compare it with the width of the actual scanning track to determine the optimal threshold of the molten pool boundary. In this embodiment, the determined optimal threshold is 0.5; i

[0063] Perform binary processing on all molten pool images according to the optimal threshold to obtain the optimal molten pool boundary of each molten pool image;

[0064] Step 2.2: Use the least squares method to fit all the optimal molten pool boundaries into an ellipse, and according to the instantaneous field of view angle of the near-infrared high-speed camera, after converting the major axis length, minor axis length, and ellipse area of the ellipse, obtain the length, width, and area of the molten pool respectively, and use them as the dimensions of the molten pool;

[0065] Take the molten pool image I at the i-th moment i to the molten pool image I at the i + k - 1-th moment i+k-1 as a sample, and take the molten pool size sequence of the molten pool image I at the i-th moment i to the molten pool image I at the i + k - 1-th moment i+k-1 as the label Y of the sample i,i+k-1 , in this embodiment, the value of k is 10;

[0066] Step 3: Extract the process characteristics related to laser power, scanning speed, power density, and scanning strategy;

[0067] Step 3.1: Calculate the laser power p at the (i - 1)-th moment corresponding to the label Y i,i+k-1 , the scanning speed v at the (i - 1)-th moment i-1 , and the power density E at the (i - 1)-th moment i-1 respectively using equations (1) to (3): i-1

[0068]

[0069]

[0070]

[0071] In formulas (1)-(3), n is the number of laser scanning points between the (i - 1)-th moment and the i-th moment. In this example, n is 50. and are respectively the laser power and the laser incident angle of the j-th scanning point at the (i - 1)-th moment. and are respectively the distance and the scanning time between the (j - 1)-th to the (j + 1)-th scanning points at the (i - 1)-th moment.

[0072] Step 3.2: Calculate label Y using formulas (4)-(9). i,i+k-1 The laser moving direction at the (i - 1)-th moment corresponding to The X-direction distance at the (i - 1)-th moment The Y-direction distance at the (i - 1)-th moment The Euclidean distance at the (i - 1)-th moment The laser power feature with scanning distance weight added at the (i - 1)-th moment The laser power density feature with scanning distance weight added at the (i - 1)-th moment

[0073]

[0074]

[0075]

[0076]

[0077]

[0078]

[0079] In formulas (4)-(9), is the laser moving unit direction vector of the j-th scanning point at the (i - 1)-th moment. e0 is the unit direction vector of the X-axis in the working coordinate system. g is a function for calculating the angle. The working coordinate system is determined by the machine manufacturer. First, an origin is determined in the horizontal direction, then the vertical upward direction is the Z-axis direction, and the two mutually perpendicular axes in the horizontal plane are the X-axis and the Y-axis. is the position of the j-th scanning point in the X direction of the working coordinate system at the (i - 1)-th moment. is the position of the j-th scanning point in the Y direction of the working coordinate system at the (i - 1)-th moment. is the position of the first scanning point at the i-th moment in the X direction of the working coordinate system, is the position of the first scanning point at the i-th moment in the Y direction of the working coordinate system;

[0080] Step 3.3: Let the label Y i,i+k-1 corresponding to all process features at the (i - 1)-th moment be Calculate the process features from the (i - N)-th moment to the (i - 2)-th moment according to the process of Step 3.1 to Step 3.2, so as to obtain the process feature sequence of the label Y i,i+k-1 corresponding to the (i - N)-th moment to the (i - 1)-th moment as In this embodiment, N is 50;

[0081] Step 4: Construct a molten pool size prediction model, Figure 3 is the constructed prediction model, including: a molten pool region of interest extraction module, a CNN module for extracting molten pool image features, and an informer module for fusing process features and molten pool features;

[0082] Step 4.1: The extraction module calculates the centroid of the molten pool image sequence I i,i+k-1 corresponding to the (i - N)-th moment to the (i - 1)-th moment i-N,i-1 ={I i-N , I i-(N-1) , …, I i-1}, and determines a rectangular region with side lengths of l×h centered on the centroid of the molten pool as the molten pool region of interest sequence i-N,i-1 In this embodiment, both l and h are 60; In this embodiment, both l and h are 60;

[0083] Step 4.2: The CNN module, as shown in Figure 4 , sequentially includes: two convolutional pooling layers and one fully connected layer. In this embodiment, the convolutional kernel size is 5×5 and the number is 32; input the molten pool region of interest sequence from the (i - N)-th moment to the (i - 1)-th moment into the CNN module for processing, and obtain the molten pool image feature sequence i,i+k-1 corresponding to the label Y wherein, represents the molten pool region of interest feature at the (i - 1)-th moment, and the dimension of the region of interest feature of each image is d. In this embodiment, d is 64;

[0084] Step 4.3: The informer module includes: an encoder module and a decoder module;

[0085] Step 4.3.1: The encoder module includes: a preprocessing module, a multi-head probabilistic sparse self-attention mechanism module, and a distillation module;

[0086] Step a: After the preprocessing module normalizes the process feature sequence corresponding to label Y, the normalized process feature sequence is obtained, where represents the normalized process feature at the (i - 1)-th moment. i,i+k-1 The corresponding process feature sequence After normalization, the normalized process feature sequence is obtained where, represents the normalized process feature at the (i - 1)-th moment;

[0087] Then, is concatenated with the molten pool image feature sequence to obtain the fused feature sequence. After performing the Embedding operation and positional encoding in sequence, the high-dimensional fused feature vector with embedded position information is obtained. with the molten pool image feature sequence After concatenation, the fused feature sequence is obtained And after performing the Embedding operation and positional encoding in sequence, the high-dimensional fused feature vector with embedded position information is obtained

[0088] Step b: After the multi-head probabilistic sparse self-attention mechanism module performs the multi-head probabilistic sparse self-attention mechanism operation on, the fused feature incorporating the attention mechanism is obtained, which is then subjected to residual connection with the input and layer normalization operation, and finally input into the feed-forward neural network to output the attention module feature. After performing the multi-head probabilistic sparse self-attention mechanism operation, the fused feature incorporating the attention mechanism is obtained And it is subjected to residual connection with the input After that, layer normalization operation is performed, and finally it is input into the feed-forward neural network to output the attention module feature

[0089] Step c: After the distillation module performs one-dimensional convolution operation on the attention module feature in the time dimension, then performs non-linear transformation using the ELU activation function, and then performs max pooling operation, the encoded feature output by the encoder is finally obtained. After performing one-dimensional convolution operation on the attention module feature in the time dimension, then performing non-linear transformation using the ELU activation function, and then performing max pooling operation, the encoded feature output by the encoder is finally obtained

[0090] Step 4.3.2: The fused feature sequence from the (i - M)-th moment to the (i - 1)-th moment and the fused feature sequence from the i-th moment to the (i + k - 1)-th moment corresponding to label Y are concatenated to obtain, which is input into the decoder module. Here, 0 is used for padding. After performing the Embedding operation, positional encoding, and masked multi-head probabilistic sparse self-attention mechanism processing in sequence, it is then subjected to the multi-head attention mechanism operation together with the encoded feature to finally obtain the decoded feature output by the decoder. i,i+k-1 The fused feature sequence from the (i - M)-th moment to the (i - 1)-th moment and the fused feature sequence from the i-th moment to the (i + k - 1)-th moment are concatenated to obtain which is input into the decoder module, where 0 is used for padding. After performing the Embedding operation, positional encoding, and masked multi-head probabilistic sparse self-attention mechanism processing in sequence, it is then subjected to the multi-head attention mechanism operation together with the encoded feature to finally obtain the decoded feature output by the decoder Both of the two attention mechanism modules include residual connection, layer normalization operation, and feed-forward neural network;

[0091] Step 4.3.3: After passing through a fully connected layer, the decoded feature outputs the predicted molten pool size sequence. After passing through a fully connected layer, the predicted molten pool size sequence is output

[0092] Step 5: Construct the loss function loss using Equation (10):

[0093]

[0094] In Equation (10), the first term is the mean absolute percentage error of the molten pool length, width, and area, and the second term is the loss term that constrains the geometric relationship of the molten pool length, width, and area, which is used to improve the geometric relationship consistency of the prediction results. For the predicted value of the molten pool length sequence in For the predicted value of the molten pool width sequence in For the predicted value of the molten pool area sequence in, λ is a parameter that determines the relative importance of this loss term. In this example, λ is 0.4;

[0095] Use the gradient descent method to train the molten pool size prediction model, and calculate the total loss function loss to update the model parameters. When the total loss function loss converges, obtain the trained molten pool size prediction model, which is used to perform real-time prediction of the molten pool size at multiple future moments during the selective laser melting process.

[0096] In this embodiment, an electronic device includes a memory and a processor. The memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0097] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium. When the computer program is run by the processor, it executes the steps of the above method.

[0098] In summary, this application overcomes the deficiency that the existing model can only be used for single-pass or multi-pass average molten pool size prediction, and can perform real-time and accurate prediction of the molten pool size during the entire selective laser melting process under complex scanning strategies.

Claims

1. A real-time prediction method for the molten pool size during selective laser melting, characterized in that, Including the following steps: Step 1: Use a near-infrared high-speed camera to collect molten pool images during the selective laser melting process at a fixed frequency, and construct a molten pool image dataset ; where represents the molten pool image at the i-th moment, is the number of molten pool image samples; L and H are the length and height of the image respectively; Step 2: By processing the molten pool image, extract the length, width and area of the molten pool, and use them as the dimensions of the molten pool; The molten pool image at the $i$-th moment to the molten pool image at the moment is used as a sample. The molten pool image at the $i$-th moment to the molten pool image at the moment is used as the label of the sample for the molten pool size sequence ; Step 3: Extract the process features related to laser power, scanning speed, power density and scanning strategy; Step 3.1: Calculate the laser power, the scanning speed, and the power density at the corresponding th moment, th moment, and th moment respectively using Equations (1) to (3): corresponding laser power , scanning speed , power density : (1) (2) (3) In formulas (1)-(3), n is the number of laser scanning points between the moment and the moment, and are respectively the laser power and the laser incident angle of the j-th scanning point at the moment, and are respectively the distance and the scanning time between the (j - 1)-th to the (j + 1)-th scanning points at the moment; Step 3.2: Calculate the laser movement direction corresponding to the time, the X-direction distance corresponding to the time, the Y-direction distance corresponding to the time, the Euclidean distance corresponding to the time, the laser power feature with scanning distance weight : (4) (5) (6) (7) (8) (9) In Formula (4) - Formula (9), is the laser movement unit direction vector of the j-th scanning point at the moment, is the unit direction vector of the X-axis of the working coordinate system, is the function for calculating the included angle, is the position of the j-th scanning point in the X-axis direction of the working coordinate system at the moment, is the position of the j-th scanning point in the Y-axis direction of the working coordinate system at the moment; is the position of the 1st scanning point in the X-axis direction of the working coordinate system at the i-th moment, is the position of the 1st scanning point in the Y-axis direction of the working coordinate system at the i-th moment; Step 3.3: Let the corresponding all process characteristics at the th moment be . Calculate the process characteristics from the th moment to the th moment according to the process of Step 3.1 to Step 3.2, so as to obtain the process characteristic sequence from the corresponding th moment to the th moment as ; Step 4: Build a molten pool size prediction model, including: an extraction module for the region of interest of the molten pool, a CNN module for extracting the features of the molten pool image, and an informer module for fusing the process features and the molten pool features; Step 4.1: The extraction module calculates the centroid of the molten pool for the image sequence of the molten pool from the moment corresponding to to the moment corresponding to , and determines, with the centroid of the molten pool as the center, a rectangular region with a side length of in the image sequence of the molten pool as the sequence of regions of interest of the molten pool ; in the image sequence of the molten pool as the sequence of regions of interest of the molten pool ; ​ Step 4.2: The CNN module sequentially includes: two convolutional pooling layers and one fully connected layer; From the moment to the sequence of regions of interest of the molten pool at the moment is input into the CNN module for processing and the label is obtained, and the corresponding sequence of molten pool image features , where represents the feature of the region of interest of the molten pool at the moment; Step 4.3: The informer module is used to and After processing, the decoded feature ; Step 4.4: Decode features After passing through a fully connected layer, output the predicted molten pool size sequence ; Step 5: Construct the total loss function using Equation (10) :[[]]END]] (10) In formula (10), is the predicted value of the molten pool length sequence in is the predicted value of the molten pool width sequence in is the predicted value of the molten pool area sequence in is the parameter of importance; Use the gradient descent method to train the molten pool size prediction model, and calculate the total loss function loss to update the model parameters. When the total loss function loss converges, obtain the trained molten pool size prediction model, which is used to perform real-time prediction on the molten pool sizes at multiple future moments during the selective laser melting process.

2. The real-time prediction method for the molten pool size during selective laser melting according to claim 1, characterized in that, The said Step 2 includes: Step 2.1: Set an initial threshold and perform binarization on the molten pool image at the i-th moment to obtain the molten pool boundary in the image ; Obtain the molten pool image based on the molten pool boundary to obtain the width of the molten pool in it, and compare it with the width of the actual scanning track, so as to determine the optimal threshold of the molten pool boundary; Perform binary processing on all molten pool images according to the optimal threshold to obtain the optimal molten pool boundary of each molten pool image; Step 2.2: Use the least squares method to fit all the optimal molten pool boundaries into an ellipse, and according to the instantaneous field of view angle of the near-infrared high-speed camera, after converting the major axis length, minor axis length and ellipse area of the ellipse, respectively obtain the length, width and area of the molten pool, and use them as the dimensions of the molten pool.

3. The real-time prediction method for the molten pool size during selective laser melting according to claim 2, characterized in that,The informer module in the said Step 4.3 includes: an encoder module and a decoder module; Step 4.3.1: The encoder module includes: a preprocessing module, a multi-head probabilistic sparse self-attention mechanism module and a distillation module; Step a, the preprocessing module processes the label corresponding process feature sequence after normalization to obtain the normalized process feature sequence , where represents the normalized process feature at the th moment; Combine with the molten pool image feature sequence to obtain a fused feature sequence after splicing . After performing Embedding operation and positional encoding in sequence, a high-dimensional fused feature vector with embedded position information is obtained ; Step b. After the multi-head probabilistic sparse self-attention mechanism module performs the multi-head probabilistic sparse self-attention mechanism operation on it obtains the fused feature integrated with the attention mechanism , and after performing a residual connection with the input , it performs layer normalization operation, and finally inputs it into the feed-forward neural network and outputs the attention module feature ; Step c: After the distillation module performs a one-dimensional convolution operation on the attention module features in the time dimension, it then performs a non-linear transformation using the ELU activation function, and then performs a max pooling operation to finally obtain the encoded features output by the encoder ; ; Step 4.3.2: Label The corresponding Time to Fusion feature sequence at each moment and from the i-th moment to the Fusion feature sequence at each moment Features obtained after splicing And input into the decoder module, after the Embedding operation, position encoding, and mask multi-head probabilistic sparse self-attention mechanism are processed in sequence, and then combined with the encoded feature Perform multi-head attention mechanism operation together to finally obtain the decoding features of the decoder output .

4. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program for supporting the processor to execute the real-time prediction method described in Claim 1 or 2 or 3, and the processor is configured to execute the program stored in the memory.

5. A computer-readable storage medium, on which a computer program is stored, characterized in that, When the computer program is run by the processor, it executes the steps of the real-time prediction method described in Claim 1 or 2 or 3.

Images