Optimization control method for processing parameters of selective laser melting forming system based on deep learning

Abstract

The invention provides a deep learning-based optimal control method for processing parameters of a selective laser melting forming system, which comprises the following steps of: 1, processing parts by using the selective laser melting forming system, acquiring data, constructing a data set, and dividing the data set into a training set and a test set; step 2, building a deep convolutional neural network model for predicting defects generated when the system is used for processing parts; step 3, using the training set to train the deep convolutional neural network model; step 4, using the trained deep convolutional neural network model to predict defects in the processing process according to the processing image obtained in real time; and 5, according to the predicted defects, the machining parameters of the selective laser melting forming system are optimized. According to the method, the time cost and the material cost caused by a trial-and-error method can be remarkably reduced, the machining quality of parts is remarkably improved, defects are reduced, and rapid development of laser additive manufacturing is promoted.

Description

Technical Field

[0001] The present invention relates to a method for optimizing and controlling system processing parameters, and in particular to a method for optimizing and controlling processing parameters of a laser selective melting forming system based on deep learning. Background Art

[0002] This section merely provides background information related to the present disclosure and is not necessarily prior art.

[0003] In recent years, laser additive manufacturing technology has developed rapidly and has been proven to have advantages in providing design freedom for lightweight components with complex geometries. Although as a disruptive technology, laser additive manufacturing technology has demonstrated strong manufacturing capabilities, the repeatability, durability and reliability of the printing process still face major challenges. In the preparation of metal materials, laser additive manufacturing technology mainly uses the heat energy generated by the laser beam to sinter, melt and solidify the added materials, or to induce chemical curing reactions by emitting light quanta of specific wavelengths in tank photopolymerization. In the field of laser additive manufacturing, laser selective melting technology can produce very complex and fine metal parts compared to other laser additive manufacturing technologies, which is suitable for industrial fields with high precision requirements.

[0004] Selective laser melting is dominated by complex physical processes, including the absorption and transfer of laser energy, material evaporation, remelting and solidification, molten pool fluid dynamics, and microstructural evolution of epitaxial growth and nucleation. During the printing process, defects such as cracks, pores, deformation and insufficient melting often occur, which will affect the mechanical and functional properties of the manufactured parts. The size and number of these defects are often related to the laser energy density and its related parameters (such as scanning speed, scanning spacing, spot size, powder layer thickness, etc.).

[0005] At present, the optimization of laser selective melting forming process parameters in engineering is focused on the traditional trial and error method, which seeks the best process parameters through a large number of processing tests. This method is time-consuming, labor-intensive and costly, greatly increasing production costs. In addition, the trial and error method is usually difficult to design and optimize the additive manufacturing process parameters over a wide range, and it is almost impossible to optimize multiple parameters in parallel. At present, although some scholars have proposed some methods to optimize parameters using machine learning, they only focus on the physical properties of the final parts rather than the defects that occur during the manufacturing process of the parts, and the latter is the determining factor of the former.

[0006] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore may include information that does not constitute the prior art known to ordinary technicians in the field. Summary of the invention

[0007] Purpose of the invention: The technical problem to be solved by the present invention is to provide an optimization control method for processing parameters of a laser selective melting forming system based on deep learning in response to the shortcomings of the prior art.

[0008] In order to solve the above technical problems, the present invention discloses a method for optimizing control of processing parameters of a laser selective melting forming system based on deep learning, comprising the following steps:

[0009] Step 1, using a laser selective melting forming system to process parts, collect data, construct a data set, and divide the data set into a training set and a test set;

[0010] Step 2, building a deep convolutional neural network model to predict defects generated when the system processes parts;

[0011] Step 3, using the training set to train the deep convolutional neural network model;

[0012] Step 4: Use the trained deep convolutional neural network model to predict defects based on the processing images acquired in real time during the processing;

[0013] Step 5: Optimize the processing parameters of the laser selective melting forming system according to the predicted defects.

[0014] Furthermore, the step of constructing a data set in step 1 includes the following steps:

[0015] Step 1-1, using a laser selective melting forming system to process a part, during the processing, numbering each layer structure of the part, photographing and recording relevant data of each processing image during the processing, the relevant data including: layer structure number, spatial position coordinates of the current laser head and processing parameters;

[0016] Step 1-2, after processing is completed, detecting defects and recording the spatial locations of the defects in the parts, the defects including: cracks, pores, deformation and insufficient melting;

[0017] Step 1-3, through the spatial position of the defect in the part, the detected defect is matched with the processing image taken during the processing and its related data to form a defect data, which is added to the data set; the corresponding relationship between the defect and the processing image is: one defect corresponds to multiple processing images.

[0018] Furthermore, the defects described in step 1 include: cracks, pores, deformation and insufficient melting.

[0019] Furthermore, the deep convolutional neural network model described in step 2 has the following specific structure:

[0020] The input is a normal convolution layer with a convolution kernel size of 3x3 and a stride of 2;

[0021] After the ordinary convolution layer, batch normalization BN and activation function Swish are sequentially connected;

[0022] Then connect the mobile flip bottleneck convolution MBConv group;

[0023] After the mobile flip bottleneck convolution MBConv group, a 1×1 convolution layer is connected;

[0024] Then connect batch normalization BN and activation function Swish in sequence;

[0025] Then, an average pooling layer and a fully connected layer are connected in sequence as the output end.

[0026] Furthermore, the activation function Swish described in step 2 is as follows:

[0027] Swish(x)=x·σ(x)

[0028] Where x represents the input and σ(x) is the Sigmoid function, as follows:

[0029]

[0030] Furthermore, the mobile flip bottleneck convolution MBConv group described in step 2 includes:

[0031] Seven groups of moving flip bottleneck convolution MBConv structures are connected in sequence, and a random inactivation layer is connected after the last moving flip bottleneck convolution MBConv structure; wherein the moving flip bottleneck convolution MBConv structure includes:

[0032] A 1×1 ordinary convolution layer for dimensionality increase, followed by batch normalization (BN) and activation function Swish;

[0033] A k×k depthwise separable convolution, where k represents the size of the convolution kernel, followed by batch normalization (BN) and activation function Swish;

[0034] An SE module;

[0035] A 1×1 ordinary convolution layer for dimensionality reduction, followed by batch normalization (BN) and a linear activation function;

[0036] Furthermore, the mobile flip bottleneck convolution MBConv group described in step 2 further includes:

[0037] The feature pyramid network FPN is introduced after the first, third and fifth mobile flip bottleneck convolution MBConv structures to perform multi-scale feature fusion;

[0038] The linear activation function adopts:

[0039] y=x

[0040] Among them, x represents input and y represents output.

[0041] Furthermore, the training of the deep convolutional neural network model described in step 3 includes:

[0042] Step 3-1, using the training set obtained in step 1, the deep convolutional neural network model is trained using the K-fold crossover method;

[0043] Step 3-2, using the test set obtained in step 1, verify the trained deep convolutional neural network model;

[0044] Step 3-3, according to the test results of step 3-2, adjust the input resolution coefficient of the deep convolutional neural network model, the depth and width of the network model, and repeat steps 3-1 and 3-2 again until the prediction accuracy of the deep convolutional neural network model is higher than the threshold, and the training is completed.

[0045] Furthermore, the processing parameters include: scanning speed, scanning interval, spot size and powder layer thickness.

[0046] Further, the threshold value in step 3-3 is 90%.

[0047] Beneficial effects:

[0048] The present invention uses a combination of optical detection and ultrasonic detection to comprehensively detect defects in parts manufactured using the laser selective melting forming technology and locate their positions, thereby corresponding to the processed images; the data set obtained in this way has extremely high accuracy, and the model can obtain higher accuracy after being trained with high-quality data.

[0049] The present invention uses a deep convolutional neural network as a training model. In laser additive manufacturing defect detection, the size and shape of defects vary greatly. By introducing the feature pyramid network FPN, the model can learn more information from multi-scale feature maps: low-resolution high-level features can help the network learn more global contextual information (such as the global shape of the crack); high-resolution low-level features can help the network focus on details (such as tiny pores or local details). It further improves the accuracy of defect detection, effectively reduces the time and economic costs of manual trial and error, helps companies accelerate the research and development of new products, and also contributes to the rapid development of laser selective melting forming technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more clear.

[0051] Figure 1 It is a flow chart of the present invention for real-time control of process parameters of the laser selective melting forming system.

[0052] Figure 2 It is a structural diagram of the deep convolutional neural network model in the present invention.

[0053] Figure 3 This is a diagram of the moving flip bottleneck convolution structure in the present invention.

[0054] Figure 4 It is a processing image taken by the high-resolution microscopic imaging system of the present invention.

[0055] Figure 5 It is a schematic diagram of the process of model training convergence in the present invention. DETAILED DESCRIPTION

[0056] Aiming at the shortcomings of the laser selective melting forming process parameter optimization method, the present invention proposes a method of applying deep learning to obtain the optimal process parameters. The present invention predicts the defects in the laser selective melting forming process in real time during the processing and adjusts the process parameters to reduce the forming defects of the product.

[0057] The technical solution of the present invention is as follows:

[0058] S1, build laser selective melting forming system;

[0059] S2, constructing a data set; wherein the data set is a marked printing process image with detailed defects; the correspondence between defects and images in the data set is one-to-many, that is, one defect corresponds to multiple groups of images during laser selective melting; the data set is divided into a training set and a test set; the defects include cracks, pores, deformation and insufficient melting;

[0060] S3, build a deep convolutional neural network model; the model adopts the EfficientNet network and uses NAS (Neural Architecture Search) search technology to consider the input resolution coefficient, the depth and width of the network at the same time to build a more efficient deep learning network structure.

[0061] S4, training EfficientNet model;

[0062] S41, using the training set obtained in S2, adopts K-fold cross training EfficientNet model;

[0063] S42, using the test set obtained in S2 to verify the trained model;

[0064] S43, according to the test results in S42, adjust the input resolution coefficient of the model, the depth and width of the network, and repeat steps S41 and S42 again;

[0065] S5, importing the final deep learning model into the computing control system, inputting the processing image acquired in real time during the processing into the deep learning model, and then outputting the optimized processing parameters;

[0066] The laser selective melting forming system in step S1 includes not only core components such as an optical system, a powder forming system, a gas circulation purification system, and a computer control system, but is also equipped with a high-resolution microscopic imaging system for capturing microscopic images during the processing.

[0067] In step S2, the proportions of the training set and the test set in all data sets are: 80% for the training set and 20% for the test set;

[0068] The specific process of constructing the data set in step S2 includes the following steps:

[0069] S21, using the laser selective melting forming system to process parts, marking each layer of the structure with a serial number and corresponding to the processing image captured by the microscopic imaging system; the layer number here is not the number of the picture, but only one of the data corresponding to the picture. In addition to the layer number, each picture also has other corresponding data, including the xyz coordinates of the current laser head, the current processing parameters, whether the picture shows defects and the defect category. A picture and its corresponding data (layer number, xyz coordinates, processing parameters, whether defects are displayed and defect category) together constitute a piece of data.

[0070] S22, uses high-resolution cameras to detect surface defects of parts; uses ultrasonic waves to detect internal defects of parts;

[0071] S23, matching the defects with the images captured during the processing to form a set of defect data; the remaining data is normal data.

[0072] like Figure 2 As shown, the specific structure of the model in step S3 is as follows:

[0073] The model first consists of a normal convolutional layer with a kernel size of 3×3 and a stride of 2, followed by BN (Batch Normalization, Ioffe S. Batch normalization: Accelerating deep network training by reducing internal covariate shift [J]. arXiv preprint arXiv:1502.03167, 2015.) and an activation function Swish, where the Swish activation function formula is:

[0074] Swish(x)=x·σ(x)

[0075] Where σ(x) is the Sigmoid function, defined as:

[0076]

[0077] Then there are 7 groups of MBConv (Mobile Inverted Residual BottleneckBlock, Sandler M, Howard A, Zhu M, et al. Mobilenetv2: Inverted residuals and linear bottlenecks [C] / / Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 4510-4520.) structures. The MBConv structure is as follows Figure 3 As shown, it mainly includes:

[0078] A 1×1 ordinary convolution for dimensionality increase, followed by BN and Swish activation functions;

[0079] A k×k Depthwise Conv (Chollet F.Xception: Deeplearning with depthwise separable convolutions [C] / / Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 1251-1258.), k = 3 or 5, followed by BN and Swish activation function.

[0080] A SE module (Hu J, Shen L, Sun G. Squeeze-and-excitation networks [C] / / Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 7132-7141.); The SE module is an attention mechanism that can effectively improve the accuracy of the model. It includes an average pooling layer (Avg pooling), a fully connected layer (Full Connect, FC), a linear rectifier function (Relu), a second fully connected layer, and a hard restricted linear unit (Hard Sigmod, H-Sig)

[0081] A 1×1 ordinary convolution is used for dimensionality reduction, followed by BN and linear activation function. The role of linear activation y=x is to: ensure that the dimension after the bottleneck layer is consistent with the residual connection. Maintain the integrity and effectiveness of information during the compression process. Ensure that the network structure maintains stable and efficient gradient flow when compressing dimensions. Do not introduce unnecessary complexity, but ensure information flow and computational efficiency through linear mapping. Therefore, although this 1x1 convolution layer is mathematically linear, it is still an indispensable part of the MBConv structure. Removing it will cause dimension mismatch or affect residual connection, thereby affecting the training and performance of the model.

[0082] The last layer of the MBConv structure is a Dropout (random inactivation) layer (Srivastava N, Hinton G, Krizhevsky A, et al. Dropout: a simple way to prevent neural networks from overfitting [J]. The journal of machine learning research, 2014, 15 (1): 1929-1958.);

[0083] Feature Pyramid Networks (FPN) are introduced after the first, third, and fifth MBConv structures to perform multi-scale feature fusion, so that the model can learn more defect information from multi-scale feature maps, especially defects with large differences in size and shape.

[0084] After the 7 groups of MBConv structures, there is a 1×1 convolution layer, followed by BN and activation function Swish, followed by an average pooling layer and a fully connected layer. Finally, the model outputs whether the current image indicates a defect and the type of defect through the last fully connected layer.

[0085] In step S5, the model is placed in a computing system and runs. The high-resolution microscopic imaging system transmits the captured images to the computer as model input. After calculation, the model determines whether the current processed image indicates a part defect. If so, the processing parameters are adjusted according to the model output to avoid the defect.

[0086] In step S5, the processing parameters include scanning speed, scanning interval, spot size, and powder layer thickness.

[0087] Example:

[0088] Figure 1 The present invention discloses a method for optimizing and controlling the technical parameters of laser selective melting based on deep learning, which comprises the following steps:

[0089] S1. Build a laser selective melting forming system. First, an additional high-resolution microscopic imaging system is installed on the existing laser selective melting forming equipment to capture the processing image, and the imaging system transmits the captured image to the computer control system of the laser selective melting forming equipment.

[0090] S2. Construct a data set. First, use the laser selective melting forming system built in S1 to process parts. During the processing, the high-resolution microscopic imaging system will capture images of the processing process. The processed images will be transmitted to the computing device, and the corresponding layer number, XYZ coordinates of the laser head, scanning speed, scanning spacing, spot size, powder layer thickness and other processing parameters will be recorded in the computing device at the same time. Table 1 shows 10 sets of sample data.

[0091] Table 1.10 groups of sample data

[0092]

[0093]

[0094] like Figure 4 The processing image of the layer numbered 87 and the sequence numbered 14035, i.e., the image numbered 87-14035 in Table 1, shows a pore defect.

[0095] After the parts are processed, a high-resolution camera is used to capture the surface of the parts to detect surface defects and their types. In addition, ultrasonic waves are used to detect defects and their types inside the parts. Note that the spatial location of the defects will be marked during the inspection, which will correspond to the images of the parts during the processing.

[0096] After repeatedly processing a variety of different parts many times, a large number of related part defects and processing process images can be obtained. After cleaning, these data constitute the data set required for model training.

[0097] S3. Build a deep convolutional neural network model. Use EfficientNet as the basic model architecture. EfficientNet is a network structure optimized by NAS. It can effectively improve the performance of the model when considering the balance between input resolution coefficient, network depth and width. Its specific network structure is as follows:

[0098] First, it passes through a normal convolution layer with a 3x 3 convolution kernel and a stride of 2, using the Swish activation function. The definition of this activation function is as follows:

[0099]

[0100] Subsequently, the model contains 7 groups of MBConv structures, which mainly include:

[0101] A 1x 1 convolution is used for dimensionality increase, followed by batch normalization (BN) and Swish activation function.

[0102] A kxk depthwise convolution (Depthwise Conv), where k can be 3 or 5, followed by batch normalization (BN) and Swish activation function.

[0103] A SE module to implement the channel attention mechanism.

[0104] A 1x1 convolution is used for dimensionality reduction, followed by batch normalization and linear activation (activation function y=x).

[0105] In addition, due to the large differences in the size and shape of defects in laser additive manufacturing defect detection, Feature Pyramid Networks (FPN) is introduced after the first, third, and fifth MBConv structures to perform multi-scale feature fusion, so that the model can learn more information from multi-scale feature maps.

[0106] The last layer of the MBConv structure is a Dropout layer to prevent the model from overfitting.

[0107] The 7 groups of MBConv structures are followed by a 1×1 convolution layer, followed by BN and activation function Swish. The convolution layer is followed by an average pooling layer and a fully connected layer.

[0108] Finally, the model passes through the final fully connected layer and outputs whether the current image indicates a defect and the type of defect.

[0109] S4. Train the EfficientNet model. During the training process, the K-fold cross-validation method is used to optimize the model. The steps are as follows:

[0110] S41. Use the constructed data set for training to ensure that the model can learn various defect characteristics that may occur during the processing process from multiple angles.

[0111] S42. Use the test set to verify the model performance and evaluate its recognition accuracy and recall rate for various defects (cracks, pores, deformation and insufficient melting) in actual processing images.

[0112] S43. According to the verification results, the input resolution coefficient (r) of the model is adjusted, the depth (d) and width (w) of the network structure are changed, and repeated training and verification are performed to continuously optimize the parameters, thereby obtaining the best model parameters. The convergence process of the final model is as follows: Figure 5 As shown in Figure 2, the model converged after the 35th round of training. The best r used by the model is 1.6, d = 5.1, w = 3.3,

[0113] S5. Apply the model in the computer control system. The trained model is finally deployed in the computer control module of the laser selective melting forming system. The real-time processing image is captured by a high-resolution microscopic imaging system and input into the model. The model determines whether it indicates defects through calculation and adjusts the processing parameters in real time according to the detection results, including scanning speed, scanning spacing, spot size, and powder layer thickness.

[0114] In a specific implementation, the present application provides a computer storage medium and a corresponding data processing unit, wherein the computer storage medium can store a computer program, and when the computer program is executed by the data processing unit, the invention content of the method for optimizing the control of processing parameters of a laser selective melting forming system based on deep learning provided by the present invention and some or all of the steps in each embodiment can be a magnetic disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0115] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present invention can be implemented by means of computer programs and their corresponding general hardware platforms. Based on such an understanding, the technical solutions in the embodiments of the present invention can be essentially or partly contributed to the prior art in the form of computer programs, i.e., software products, which can be stored in a storage medium and include several instructions for enabling a device including a data processing unit (which can be a personal computer, a server, a single-chip microcomputer, an MCU or a network device, etc.) to execute the methods described in various embodiments of the present invention or certain parts of the embodiments.

[0116] The present invention provides an idea and method for optimizing the control method of processing parameters of a laser selective melting forming system based on deep learning. There are many methods and ways to implement the technical solution. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the scope of protection of the present invention. All components not specified in this embodiment can be implemented by existing technologies.

Claims

1. A method for optimizing control of processing parameters of a laser selective melting forming system based on deep learning, characterized in that: The following steps are involved: Step 1, using a laser selective melting forming system to process parts, collect data, construct a data set, and divide the data set into a training set and a test set; Step 2, building a deep convolutional neural network model to predict defects generated when the system processes parts; Step 3, using the training set to train the deep convolutional neural network model; Step 4: Use the trained deep convolutional neural network model to predict defects based on the processing images acquired in real time during the processing; Step 5: Optimize the processing parameters of the laser selective melting forming system according to the predicted defects.

2. According to the method for optimizing the processing parameters of the laser selective melting forming system based on deep learning in claim 1, it is characterized in that: The construction of the dataset described in step 1 includes the following steps: Step 1-1, using a laser selective melting forming system to process a part, during the processing, numbering each layer structure of the part, photographing and recording relevant data of each processing image during the processing, the relevant data including: layer structure number, spatial position coordinates of the current laser head and processing parameters; Step 1-2, after processing is completed, detecting defects and recording the spatial locations of the defects in the parts, the defects including: cracks, pores, deformation and insufficient melting; Step 1-3, through the spatial position of the defect in the part, the detected defect is matched with the processing image taken during the processing and its related data to form a defect data, which is added to the data set; the corresponding relationship between the defect and the processing image is: one defect corresponds to multiple processing images.

3. According to the method for optimizing the processing parameters of the laser selective melting forming system based on deep learning in claim 2, it is characterized in that: Defects described in step 1, including: cracks, porosity, deformation and lack of melting.

4. According to claim 3, a method for optimizing and controlling processing parameters of a laser selective melting forming system based on deep learning is characterized in that: The deep convolutional neural network model described in step 2 has the following specific structure: The input is a normal convolution layer with a convolution kernel size of 3x3 and a stride of 2; After the ordinary convolution layer, batch normalization BN and activation function Swish are sequentially connected; Then connect the mobile flip bottleneck convolution MBConv group; After the mobile flip bottleneck convolution MBConv group, a 1×1 convolution layer is connected; Then connect batch normalization BN and activation function Swish in sequence; Then, an average pooling layer and a fully connected layer are connected in sequence as the output end.

5. According to claim 4, a method for optimizing and controlling processing parameters of a laser selective melting forming system based on deep learning is characterized in that: The activation function Swish described in step 2 is as follows: Swish(x)=x·σ(x) Where x represents the input and σ(x) is the Sigmoid function, as follows:

6. The method for optimizing and controlling processing parameters of a laser selective melting forming system based on deep learning according to claim 5 is characterized in that: The mobile flip bottleneck convolution MBConv group described in step 2 includes: Seven groups of moving flip bottleneck convolution MBConv structures are connected in sequence, and a random inactivation layer is connected after the last moving flip bottleneck convolution MBConv structure; wherein the moving flip bottleneck convolution MBConv structure includes: A 1×1 ordinary convolution layer for dimensionality increase, followed by batch normalization (BN) and activation function Swish; A k×k depthwise separable convolution, where k represents the size of the convolution kernel, followed by batch normalization (BN) and activation function Swish; An SE module; A 1×1 ordinary convolution layer for dimensionality reduction, followed by batch normalization (BN) and a linear activation function.

7. The method for optimizing and controlling processing parameters of a laser selective melting forming system based on deep learning according to claim 6 is characterized in that: The mobile flip bottleneck convolution MBConv group described in step 2 also includes: The feature pyramid network FPN is introduced after the first, third and fifth mobile flip bottleneck convolution MBConv structures to perform multi-scale feature fusion; The linear activation function adopts: y=x Among them, x represents input and y represents output.

8. The method for optimizing and controlling processing parameters of a laser selective melting forming system based on deep learning according to claim 7 is characterized in that: The training of the deep convolutional neural network model described in step 3 includes: Step 3-1, using the training set obtained in step 1, the deep convolutional neural network model is trained using the K-fold crossover method; Step 3-2, using the test set obtained in step 1, verify the trained deep convolutional neural network model; Step 3-3, according to the test results of step 3-2, adjust the input resolution coefficient of the deep convolutional neural network model, the depth and width of the network model, and repeat steps 3-1 and 3-2 again until the prediction accuracy of the deep convolutional neural network model is higher than the threshold, and the training is completed.

9. The method for optimizing and controlling processing parameters of a laser selective melting forming system based on deep learning according to claim 8, characterized in that: The processing parameters include: scanning speed, scanning spacing, spot size and powder layer thickness.

10. The method for optimizing and controlling processing parameters of a laser selective melting forming system based on deep learning according to claim 9, characterized in that: The threshold value described in step 3-3 is 90%.

Images