SLM process parameter analysis method based on machine learning

Abstract

The invention relates to the field of parameter optimization, in particular to an SLM process parameter analysis method based on machine learning. The invention discloses an SLM process parameter analysis method based on machine learning. The method comprises the following steps: constructing a workpiece image feature vector; constructing a workpiece prediction image; and outputting an SLM process optimization parameter set and the like. According to the method, the SLM process control parameter change and the corresponding workpiece image change are subjected to sequential analysis, the influence of the SLM process control parameter change on the workpiece is learned, meanwhile, the workpiece change on the premise that the SLM process control parameter is not changed is predicted, then the defect of the workpiece is analyzed based on the predicted workpiece change, and the defect of the workpiece is analyzed. And the subsequent SLM process control parameters are optimized based on the analyzed defects, so that the defects are avoided, and the quality of the workpiece is improved.

Description

Technical Field

[0001] The present invention relates to the field of parameter optimization, and in particular to a SLM process parameter analysis method based on machine learning. Background Art

[0002] Selective laser melting (SLM) is an advanced additive manufacturing technology that is widely used in aerospace, medical, and automotive fields. However, the selection and control of process parameters in the SLM process are crucial and directly affect the quality and performance of the final product. Although the existing technology has been studied in terms of static parameter setting, in the dynamic process, the SLM system usually relies on fixed parameters set by humans and lacks the ability to adaptively adjust. This fixed parameter setting method often cannot cope with the uncertainties caused by factors such as material properties, environmental changes, and equipment status, resulting in frequent defects in the production process, such as pores, unfused, and warping. This not only increases the subsequent detection and repair costs, but also limits the application potential of SLM technology in the manufacture of complex workpieces. Summary of the invention

[0003] The present invention performs a temporal analysis on the changes in SLM process control parameters and the corresponding workpiece image changes. While studying the impact of the changes in SLM process control parameters on the workpiece, the present invention predicts the changes in the workpiece under the premise that the SLM process control parameters remain unchanged. The defects of the workpiece are then analyzed based on the predicted changes in the workpiece, and the subsequent SLM process control parameters are optimized based on the analyzed defects to avoid the occurrence of defects, thereby improving the quality of the workpiece.

[0004] A method for analyzing SLM process parameters based on machine learning, comprising:

[0005] Obtain the SLM process parameter set and workpiece monitoring image at the current monitoring time point, send the workpiece monitoring image to the image feature extraction model for processing, and output the workpiece image feature vector;

[0006] The SLM process parameter set at the current monitoring time point and the workpiece image feature vector are spliced ​​to form the workpiece timing prediction data, and the workpiece timing prediction data at the current monitoring time point and the previous N-1 workpiece timing prediction data are combined to form a workpiece timing prediction data set; the workpiece timing prediction data set is sent to the workpiece image prediction model for processing, and the workpiece prediction image is output;

[0007] The workpiece prediction image and the SLM process parameter set at the current monitoring time point are sent to the parameter adjustment model for processing, and the SLM process optimization parameter set is output;

[0008] The parameter adjustment model includes a standard image construction layer, a difference image construction layer, a difference image enhancement layer, a difference feature extraction layer, a first feature stitching layer and an SLM process optimization parameter set output layer, wherein the standard image construction layer is used to construct a corresponding standard image based on the workpiece prediction image; the difference image construction layer is used to construct a difference image based on the standard image and the standard image; the difference image enhancement layer is used to perform feature enhancement on the workpiece prediction image through the difference image to construct an enhanced defective workpiece image; the difference feature extraction layer is used to perform feature extraction on the enhanced defective workpiece image to construct a difference feature vector; the first feature stitching layer is used to stitch the SLM process parameter set at the current monitoring time point with the difference feature vector to construct a process parameter optimization analysis data set; the SLM process optimization parameter set output layer is used to process the process parameter optimization analysis data set and output M process optimization parameter set.

[0009] As a preferred aspect of the present invention, the workpiece prediction image and the SLM process parameter set at the current monitoring time point are sent to the parameter adjustment model for processing, and the SLM process optimization parameter set is output, specifically, including the following steps:

[0010] Send the workpiece prediction image to the standard image construction layer for processing to construct a standard image;

[0011] In the difference image construction layer, a subtraction operation is performed between the standard image and the artifact prediction image, and then a normalization operation is performed to construct a difference image;

[0012] In the difference image enhancement layer, the artifact prediction image is multiplied with the key weight matrix and the value weight matrix respectively to construct the artifact key feature map K and the artifact value feature map V. Then, the difference image is multiplied with the query weight matrix to construct the difference query feature map Q. The attention weight matrix ATT = (QK T / (D) 0.5 ), and finally, the attention weight matrix ATT is multiplied with the workpiece value feature map V to construct an enhanced defect workpiece image;

[0013] Send the enhanced defect workpiece image to the difference feature extraction layer for processing to construct a difference feature vector;

[0014] In the first feature concatenation layer, the SLM process parameter set at the current monitoring time point is concatenated with the difference feature vector to construct a process parameter optimization analysis data set;

[0015] The process parameter optimization analysis data set is sent to the SLM process optimization parameter set output layer for processing, which needs to pass through several hidden layers to output the SLM process optimization parameter set.

[0016] As a preferred aspect of the present invention, the workpiece image prediction model includes a timing analysis layer, a second feature stitching layer and a workpiece prediction image construction layer, wherein the timing analysis layer is used to perform timing analysis on the workpiece timing prediction data set to construct an SLM process timing analysis vector; the second feature stitching layer is used to stitch the SLM process parameter set at the current monitoring time point with the SLM process timing analysis vector to construct a workpiece prediction analysis vector; the workpiece prediction image construction layer is used to process the workpiece prediction analysis vector to construct a workpiece prediction image.

[0017] As a preferred aspect of the present invention, training the workpiece image prediction model specifically includes the following steps:

[0018] Obtain a workpiece image prediction training sample, the workpiece image prediction training sample includes N+1 workpiece time series prediction data, annotate the workpiece image prediction training sample with the workpiece monitoring image corresponding to the workpiece image feature vector of the last workpiece time series prediction data in the workpiece image prediction training sample, form a workpiece image prediction training set with all the annotated workpiece image prediction training samples, and then send the workpiece image prediction training set to the workpiece image prediction model with parameter initialization for training, use the first N workpiece time series prediction data in the workpiece image prediction training sample as the input of the time series analysis layer, use the SLM process parameter set of the last workpiece time series prediction data in the workpiece image prediction training sample as the input of the second feature splicing layer, calculate the workpiece image prediction loss value, and determine whether the workpiece image prediction loss value is within a first preset range. If the workpiece image prediction loss value is within the first preset range, output the trained workpiece image prediction model; otherwise, continue to train the workpiece image prediction model through the workpiece image prediction training set.

[0019] As a preferred aspect of the present invention, training is performed on the parameter adjustment model, specifically comprising the following steps:

[0020] Obtain a number of parameter adjustment training samples, the parameter adjustment training samples including N workpiece timing prediction data, annotate the parameter adjustment training samples through the process parameter optimization analysis data set; and form all the annotated parameter adjustment training samples into a parameter adjustment training set;

[0021] Acquire a number of image construction training samples, the image construction training samples include defective workpiece monitoring images and corresponding standard images, form all image construction training samples into an image construction training set, and then send the image construction training set to the parameter-initialized standard image construction layer for training, use the workpiece monitoring images in the image construction training samples as input, use the standard images in the image construction training samples as target output, calculate the image construction loss value, and determine whether the image construction loss value is within a second preset range. If the image construction loss value is within the second preset range, output the trained standard image construction layer; otherwise, continue to train the workpiece standard image construction layer through the image construction training set;

[0022] The parameters of the fixed standard image construction layer remain unchanged, the trained workpiece image prediction model and the parameter adjustment model are connected to form a fusion model, the parameter adjustment training set is sent to the fusion model for training, the fusion loss value is calculated, and it is determined whether the fusion loss value is within a third preset range. If the fusion loss value is within the third preset range, the trained parameter adjustment model is output; otherwise, the fusion model is continued to be trained through the parameter adjustment training set.

[0023] As a preferred aspect of the present invention, it also includes simulation optimization of parameters in the output layer of the SLM process optimization parameter set, which specifically includes the following steps:

[0024] Construct several network parameter simulation individuals, and combine all network parameter simulation individuals into a population set, and set the maximum number of iterations;

[0025] Calculate the fitness of the network parameters and simulate the corresponding individuals;

[0026] Based on the fitness of the network parameters simulating the corresponding individuals, the population set is iteratively updated through the group optimization algorithm;

[0027] Until the number of iterations reaches the maximum number of iterations, the network parameter simulation individual with the largest fitness is output as the parameter in the output layer of the SLM process optimization parameter set.

[0028] As a preferred aspect of the present invention, the fitness corresponding to the network parameter simulation individual is calculated. The specific calculation method is as follows: the output layer of the SLM process optimization parameter set in the parameter adjustment model is set through the network parameter simulation individual, and then the set parameter adjustment model is tested by constructing training samples through several images, and the accuracy rate obtained is used as the fitness corresponding to the network parameter simulation individual.

[0029] As a preferred aspect of the present invention, the standard image construction layer is established based on the U-net model, and the artifact prediction image construction layer is set based on the decoder in the U-net model.

[0030] The present invention has the following advantages:

[0031] The present invention performs a temporal analysis on the changes in SLM process control parameters and the corresponding workpiece image changes. While studying the impact of the changes in SLM process control parameters on the workpiece, the present invention predicts the changes in the workpiece under the premise that the SLM process control parameters remain unchanged. The defects of the workpiece are then analyzed based on the predicted changes in the workpiece, and the subsequent SLM process control parameters are optimized based on the analyzed defects to avoid the occurrence of defects, thereby improving the quality of the workpiece. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A schematic diagram of the process flow of the SLM process parameter analysis method based on machine learning adopted in an embodiment of the present invention.

[0033] Figure 2 It is a schematic diagram of the structure of the workpiece image prediction model adopted in the embodiment of the present invention.

[0034] Figure 3 It is a structural schematic diagram of the parameter adjustment model adopted in the embodiment of the present invention. DETAILED DESCRIPTION

[0035] In order to make the technical solutions and advantages in the embodiments of the present application more clearly understood, the exemplary embodiments of the present application are further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than an exhaustive list of all the embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflicting changes.

[0036] Embodiment 1, a SLM process parameter analysis method based on machine learning, comprising:

[0037] Obtain the SLM (Selective laser melting) process parameter set and workpiece monitoring image at the current monitoring time point, wherein the SLM process parameter set includes SLM process parameters such as laser power and scanning speed, which can be obtained through a series of sensors such as laser power meters, and these SLM process parameters can affect the quality of the formed workpiece; the workpiece monitoring image is a photographed image of the workpiece, which can be obtained through an industrial camera, and the quality of the workpiece can be reflected from the workpiece monitoring image; the workpiece monitoring image is sent to the image feature extraction model for processing, and the workpiece image feature vector is output, and the current state of the workpiece can be reflected through the workpiece image feature vector;

[0038] The SLM process parameter set at the current monitoring time point and the workpiece image feature vector are spliced ​​to form the workpiece timing prediction data, and the workpiece timing prediction data at the current monitoring time point and the previous N-1 workpiece timing prediction data are combined to form a workpiece timing prediction data set; the workpiece timing prediction data set is sent to the workpiece image prediction model for processing, and the workpiece prediction image is output; the workpiece image prediction model performs a temporal analysis on the workpiece timing prediction data set, and can learn the changes in the workpiece parameters, and then learn the impact of the changes in the workpiece parameters on the workpiece, thereby realizing the prediction of the workpiece image. The workpiece prediction image can reflect the subsequent workpiece quality, and then provide a reference for the adjustment of the SLM process parameter set;

[0039] The workpiece prediction image and the SLM process parameter set at the current monitoring time point are sent to the parameter adjustment model for processing, and the SLM process optimization parameter set is output. The SLM process optimization parameter set includes various adjusted SLM process parameters. The equipment that performs the SLM process can be controlled through the SLM process optimization parameter set, so as to achieve the effect of improving the workpiece quality;

[0040] For example, when the laser power is 75W and the scanning speed is 415mm / s, the grain structure of the bottom and side areas of the workpiece is relatively orderly and dense, showing a higher yield strength (645MPa) and greater ductility (53%). However, if the power and speed do not match (such as the power is too low or the speed is too fast), the molten pool will be too shallow, resulting in unfused areas in the bottom and middle layers of the workpiece, which will eventually affect the overall quality. When the input energy density (the combination of laser power and scanning speed) is too low (such as less than 70J / mm 3 ), the upper surface of the workpiece is prone to spheroidization, affecting the surface quality. When the energy density is too high, the depth of the molten pool increases, which may cause keyhole phenomenon in the deep area and produce porosity defects. Therefore, by predicting the laser power and scanning speed in advance, the fusion of the bottom and middle layers of the workpiece can be predicted, and the LM process parameters can be adjusted in advance.

[0041] See also Figure 3The parameter adjustment model includes a standard image construction layer, a difference image construction layer, a difference image enhancement layer, a difference feature extraction layer, a first feature stitching layer and an SLM process optimization parameter set output layer. The standard image construction layer is established based on the U-net model and is used to construct a corresponding standard image based on the workpiece prediction image. The standard image is a predicted workpiece image without defects, which can be used to determine the location and degree of defects. The defects here include pore defects, crack defects and deformation defects. In the SLM process, the laser melts the powder material layer by layer. As the number of layers increases, the heat inside the workpiece continues to accumulate, especially in workpieces with large sizes or complex geometric structures. This heat accumulation will cause the molten pool temperature to increase, which in turn affects the cooling rate and grain growth of the material, thereby causing these defects. The difference image construction layer is used to construct the corresponding standard image based on the standard image and the standard image. Construct a difference image, which can reflect the location and degree of defects; the difference image enhancement layer is used to enhance the features of the workpiece prediction image through the difference image to construct an enhanced defective workpiece image, which can highlight the defect features in the workpiece prediction image; the difference feature extraction layer is established based on a pre-trained convolutional neural network, and the VGG16 model is used in this embodiment to extract features from the enhanced defective workpiece image to construct a difference feature vector, which can characterize the location and degree of defects; the first feature splicing layer is used to splice the SLM process parameter set at the current monitoring time point with the difference feature vector to construct a process parameter optimization analysis data set; the SLM process optimization parameter set output layer is established based on a BP neural network, which is used to process the process parameter optimization analysis data set and output M process optimization parameter sets;

[0042] VGG16 is a deep convolutional neural network model proposed by the Computer Vision Group of Oxford University and is widely used in image classification tasks. The model consists of 16 layers, including 13 convolutional layers and 3 fully connected layers. In this embodiment, the output of the second-to-last fully connected layer is selected as the difference feature vector. The design concept is to improve the feature extraction capability by increasing the depth of the network. Each convolutional layer uses a 3x3 convolution kernel and introduces nonlinearity in combination with the ReLU activation function. The size of the feature map is gradually reduced through the maximum pooling operation to reduce the computational complexity. VGG16 has a simple structure and efficient parameter sharing, so that it uses the same convolution kernel size in multiple layers. Especially after pre-training on the ImageNet dataset, VGG16 shows good transfer learning capabilities, so it is often used as the basis for feature extraction and model fine-tuning.

[0043] The present application performs a time-series analysis on the changes in SLM process control parameters and the corresponding changes in the workpiece image. While learning the impact of the changes in the SLM process control parameters on the workpiece, it predicts the changes in the workpiece under the premise that the SLM process control parameters remain unchanged. The defects of the workpiece are then analyzed based on the predicted workpiece changes, and the subsequent SLM process control parameters are optimized based on the analyzed defects to avoid the occurrence of defects and thereby improve the quality of the workpiece.

[0044] The workpiece prediction image and the SLM process parameter set at the current monitoring time point are sent to the parameter adjustment model for processing, and the SLM process optimization parameter set is output. Specifically, the following steps are included:

[0045] The workpiece prediction image is sent to the standard image construction layer for processing to construct a standard image. It should be noted that the standard image construction layer is established with reference to the U-net model, including an encoder and a decoder, which can achieve the function of image reconstruction;

[0046] In the difference image construction layer, the standard image and the workpiece prediction image are subtracted and then normalized to construct a difference image. After the normalization operation, the non-zero position of the difference image indicates that the subsequent workpiece will have defects.

[0047] In the difference image enhancement layer, the artifact prediction image is multiplied with the key weight matrix and the value weight matrix respectively to construct the artifact key feature map K and the artifact value feature map V. Then, the difference image is multiplied with the query weight matrix to construct the difference query feature map Q. The attention weight matrix ATT = (QK T / (D) 0.5 ), and finally, the attention weight matrix ATT is multiplied with the workpiece value feature map V to construct an enhanced defect workpiece image; the query weight matrix, key weight matrix, and value weight matrix are all set based on the self-attention mechanism in the Transformer model;

[0048] The enhanced defect workpiece image is sent to the difference feature extraction layer for processing to construct a difference feature vector. It should be noted that the difference feature extraction layer performs operations such as convolution, pooling and full connection to perform image feature extraction operations;

[0049] In the first feature concatenation layer, the SLM process parameter set at the current monitoring time point is concatenated with the difference feature vector to construct a process parameter optimization analysis data set;

[0050] The process parameter optimization analysis data set is sent to the SLM process optimization parameter set output layer for processing, which needs to pass through several hidden layers to output the SLM process optimization parameter set.

[0051] See also Figure 2 The workpiece image prediction model includes a timing analysis layer, a second feature splicing layer and a workpiece prediction image construction layer, wherein the timing analysis layer is established based on the Transformer model, and is used to perform timing analysis on the workpiece timing prediction data set to construct an SLM process timing analysis vector. It should be noted that the SLM process timing analysis vector reflects the impact of workpiece parameter changes on the workpiece; the second feature splicing layer is used to splice the SLM process parameter set at the current monitoring time point with the SLM process timing analysis vector to construct a workpiece prediction analysis vector. Splicing the SLM process parameter set at the current monitoring time point with the SLM process timing analysis vector represents that the subsequent SLM control parameters remain unchanged; the workpiece prediction image construction layer is used to process the workpiece prediction analysis vector to construct a workpiece prediction image. The workpiece prediction image construction layer is set based on the decoder in the U-net model for image reconstruction.

[0052] Training the workpiece image prediction model includes the following steps:

[0053] Obtain a workpiece image prediction training sample, which includes N+1 workpiece timing prediction data. These workpiece timing prediction data are all obtained when the SLM process is actually performed. The workpiece image prediction training sample is annotated with the workpiece monitoring image corresponding to the workpiece image feature vector of the last workpiece timing prediction data in the workpiece image prediction training sample. All annotated workpiece image prediction training samples are composed of a workpiece image prediction training set. The workpiece image prediction training set is then sent to a parameter-initialized workpiece image prediction model for training. The first N workpiece timing prediction data in the workpiece image prediction training sample are used as inputs of the timing analysis layer. The SLM process parameter set of the last workpiece timing prediction data in the workpiece image prediction training sample is used as input of the second feature splicing layer. The workpiece image prediction loss value is calculated to determine whether the workpiece image prediction loss value is within a first preset range. The first preset range is set by the developer. If the workpiece image prediction loss value is within the first preset range, the trained workpiece image prediction model is output; otherwise, the workpiece image prediction model is continuously trained through the workpiece image prediction training set.

[0054] Training the parameter adjustment model includes the following steps:

[0055] Obtain several parameter adjustment training samples, which include N workpiece timing prediction data. Similarly, these workpiece timing prediction data are all obtained during the actual SLM process. The parameter adjustment training samples are annotated through the process parameter optimization analysis data set. It should be noted that the process parameter optimization analysis data set is the optimal control parameter set of the parameter adjustment training samples corresponding to the subsequent SLM process, which is obtained by expert experience, that is, the optimal control parameters adjusted by the actual operator; all the annotated parameter adjustment training samples form a parameter adjustment training set;

[0056] Obtain several sets of image construction training samples, the image construction training samples include defective workpiece monitoring images and corresponding standard images, the standard images are obtained by developers based on manual repair or software repair, and the standard images are workpiece images without defects, all image construction training samples are combined into an image construction training set, and then the image construction training set is sent to the parameter-initialized standard image construction layer for training, the workpiece monitoring images in the image construction training samples are used as input, the standard images in the image construction training samples are used as target output, the image construction loss value is calculated, and it is determined whether the image construction loss value is within a second preset range, which is also set by the developer. If the image construction loss value is within the second preset range, the trained standard image construction layer is output; otherwise, the workpiece standard image construction layer is continuously trained through the image construction training set;

[0057] The parameters of the fixed standard image construction layer remain unchanged, the trained workpiece image prediction model and the parameter adjustment model are connected to form a fusion model, the parameter adjustment training set is sent to the fusion model for training, the fusion loss value is calculated, and it is determined whether the fusion loss value is within a third preset range. The third preset range is also set by the developer. If the fusion loss value is within the third preset range, the trained parameter adjustment model is output; otherwise, the fusion model is continued to be trained through the parameter adjustment training set.

[0058] In order to further improve the performance of the parameter adjustment model, the simulation optimization of the parameters in the output layer of the SLM process optimization parameter set is also included, which specifically includes the following steps:

[0059] Construct several network parameter simulation individuals, and form a population set of all network parameter simulation individuals, and set the maximum number of iterations. It should be noted that the network parameter simulation individuals include the parameters in the output layer of the SLM process optimization parameter set, such as the number of hidden layers, the number of neurons in the hidden layer, and the learning rate;

[0060] The fitness corresponding to the network parameter simulation individual is calculated. The specific calculation method is as follows: the output layer of the SLM process optimization parameter set in the parameter adjustment model is set through the network parameter simulation individual, and then the set parameter adjustment model is tested by constructing training samples through several images. The obtained accuracy is used as the fitness corresponding to the network parameter simulation individual;

[0061] On the basis of simulating the fitness of individuals corresponding to network parameters, the population set is iteratively updated through the population optimization algorithm. The population optimization algorithm here can be a particle swarm optimization algorithm.

[0062] Until the number of iterations reaches the maximum number of iterations, the network parameter simulation individual with the largest fitness is output as the parameter in the output layer of the SLM process optimization parameter set.

[0063] Although the present invention has been described according to a limited number of embodiments, it will be apparent to those skilled in the art, with the benefit of the above description, that other embodiments may be envisioned within the scope of the invention thus described. In addition, it should be noted that the language used in this specification is selected primarily for readability and teaching purposes, rather than for explaining or defining the subject matter of the present invention. Therefore, many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the appended claims. The disclosure of the present invention is illustrative, not restrictive, with respect to the scope of the present invention, which is defined by the appended claims.

Claims

1. A SLM process parameter analysis method based on machine learning, characterized in that: include: Obtain the SLM process parameter set and workpiece monitoring image at the current monitoring time point, send the workpiece monitoring image to the image feature extraction model for processing, and output the workpiece image feature vector; The SLM process parameter set at the current monitoring time point and the workpiece image feature vector are spliced ​​to form the workpiece timing prediction data, and the workpiece timing prediction data at the current monitoring time point and the previous N-1 workpiece timing prediction data form a workpiece timing prediction data set; the workpiece timing prediction data set is sent to the workpiece image prediction model for processing, and the workpiece prediction image is output; the workpiece prediction image and the SLM process parameter set at the current monitoring time point are sent to the parameter adjustment model for processing, and the SLM process optimization parameter set is output; The parameter adjustment model includes a standard image construction layer, a difference image construction layer, a difference image enhancement layer, a difference feature extraction layer, a first feature splicing layer and an SLM process optimization parameter set output layer, wherein the standard image construction layer is used to construct a corresponding standard image based on the workpiece prediction image; the difference image construction layer is used to construct a difference image based on the standard image and the standard image; the difference image enhancement layer is used to enhance the features of the workpiece prediction image through the difference image to construct an enhanced defect workpiece image; The difference feature extraction layer is used to extract features from the enhanced defect workpiece image to construct a difference feature vector; The first feature splicing layer is used to splice the SLM process parameter set at the current monitoring time point with the difference feature vector to construct a process parameter optimization analysis data set; the SLM process optimization parameter set output layer is used to process the process parameter optimization analysis data set and output M process optimization parameter set.

2. The SLM process parameter analysis method based on machine learning according to claim 1, characterized in that: The workpiece prediction image and the SLM process parameter set at the current monitoring time point are sent to the parameter adjustment model for processing, and the SLM process optimization parameter set is output. Specifically, the following steps are included: Send the workpiece prediction image to the standard image construction layer for processing to construct a standard image; In the difference image construction layer, a subtraction operation is performed between the standard image and the artifact prediction image, and then a normalization operation is performed to construct a difference image; In the difference image enhancement layer, the artifact prediction image is multiplied with the key weight matrix and the value weight matrix respectively to construct the artifact key feature map K and the artifact value feature map V. Then, the difference image is multiplied with the query weight matrix to construct the difference query feature map Q. The attention weight matrix ATT = (QK T / (D) 0.5 ), and finally, the attention weight matrix ATT is multiplied with the workpiece value feature map V to construct an enhanced defect workpiece image; Send the enhanced defect workpiece image to the difference feature extraction layer for processing to construct a difference feature vector; In the first feature concatenation layer, the SLM process parameter set at the current monitoring time point is concatenated with the difference feature vector to construct a process parameter optimization analysis data set; The process parameter optimization analysis data set is sent to the SLM process optimization parameter set output layer for processing, which needs to pass through several hidden layers to output the SLM process optimization parameter set.

3. The SLM process parameter analysis method based on machine learning according to claim 2, characterized in that: The workpiece image prediction model includes a timing analysis layer, a second feature stitching layer and a workpiece prediction image construction layer, wherein the timing analysis layer is used to perform timing analysis on the workpiece timing prediction data set to construct an SLM process timing analysis vector; The second feature stitching layer is used to stitch the SLM process parameter set at the current monitoring time point with the SLM process timing analysis vector to construct a workpiece prediction analysis vector; the workpiece prediction image construction layer is used to process the workpiece prediction analysis vector to construct a workpiece prediction image.

4. The SLM process parameter analysis method based on machine learning according to claim 3, characterized in that: Training the workpiece image prediction model includes the following steps: Obtain a workpiece image prediction training sample, the workpiece image prediction training sample includes N+1 workpiece time series prediction data, annotate the workpiece image prediction training sample with the workpiece monitoring image corresponding to the workpiece image feature vector of the last workpiece time series prediction data in the workpiece image prediction training sample, form a workpiece image prediction training set with all the annotated workpiece image prediction training samples, and then send the workpiece image prediction training set to the workpiece image prediction model with parameter initialization for training, use the first N workpiece time series prediction data in the workpiece image prediction training sample as the input of the time series analysis layer, use the SLM process parameter set of the last workpiece time series prediction data in the workpiece image prediction training sample as the input of the second feature splicing layer, calculate the workpiece image prediction loss value, and determine whether the workpiece image prediction loss value is within a first preset range. If the workpiece image prediction loss value is within the first preset range, output the trained workpiece image prediction model; otherwise, continue to train the workpiece image prediction model through the workpiece image prediction training set.

5. The SLM process parameter analysis method based on machine learning according to claim 4, characterized in that: Training the parameter adjustment model includes the following steps: Obtain a number of parameter adjustment training samples, the parameter adjustment training samples including N workpiece timing prediction data, annotate the parameter adjustment training samples through the process parameter optimization analysis data set; and form all the annotated parameter adjustment training samples into a parameter adjustment training set; Acquire a number of image construction training samples, the image construction training samples include defective workpiece monitoring images and corresponding standard images, form all image construction training samples into an image construction training set, and then send the image construction training set to the parameter-initialized standard image construction layer for training, use the workpiece monitoring images in the image construction training samples as input, use the standard images in the image construction training samples as target output, calculate the image construction loss value, and determine whether the image construction loss value is within a second preset range. If the image construction loss value is within the second preset range, output the trained standard image construction layer; otherwise, continue to train the workpiece standard image construction layer through the image construction training set; The parameters of the fixed standard image construction layer remain unchanged, the trained workpiece image prediction model and the parameter adjustment model are connected to form a fusion model, the parameter adjustment training set is sent to the fusion model for training, the fusion loss value is calculated, and it is determined whether the fusion loss value is within a third preset range. If the fusion loss value is within the third preset range, the trained parameter adjustment model is output; otherwise, the fusion model is continued to be trained through the parameter adjustment training set.

6. The SLM process parameter analysis method based on machine learning according to claim 5, characterized in that: It also includes simulation optimization of parameters in the output layer of the SLM process optimization parameter set, which specifically includes the following steps: Construct several network parameter simulation individuals, and combine all network parameter simulation individuals into a population set, and set the maximum number of iterations; Calculate the fitness of the network parameters and simulate the corresponding individuals; Based on the fitness of the network parameters simulating the corresponding individuals, the population set is iteratively updated through the group optimization algorithm; Until the number of iterations reaches the maximum number of iterations, the network parameter simulation individual with the largest fitness is output as the parameter in the output layer of the SLM process optimization parameter set.

7. The SLM process parameter analysis method based on machine learning according to claim 6, characterized in that: The fitness corresponding to the network parameter simulation individual is calculated. The specific calculation method is as follows: the output layer of the SLM process optimization parameter set in the parameter adjustment model is set through the network parameter simulation individual, and then the set parameter adjustment model is tested by constructing training samples through several images. The obtained accuracy is used as the fitness corresponding to the network parameter simulation individual.

8. The SLM process parameter analysis method based on machine learning according to claim 7, characterized in that: The standard image construction layer is built based on the U-net model, and the artifact prediction image construction layer is set based on the decoder in the U-net model.

Images