A dynamic molten pool size detection and forming direction determination method

Abstract

The present invention discloses a method for dynamic melt pool size detection and forming direction discrimination, comprising: collecting dynamic melt pool images, performing image segmentation, and generating a dynamic melt pool image source data set; generating a dynamic melt pool image data set; constructing a melt pool size detection convolutional neural network with a direction discrimination function, improving the loss function, adding a melt pool forming direction loss, and performing sufficient training to obtain a melt pool size detection convolutional neural network with a direction discrimination function; collecting real-time dynamic melt pool images, and using the trained melt pool size detection convolutional neural network to perform size detection and forming direction discrimination on the real-time dynamic melt pool images, thereby obtaining the length, width, and forming direction information of the melt pool morphology. The present invention designs a melt pool size detection convolutional neural network with direction discrimination, which improves the detection method of traditional target recognition neural networks and has the characteristics of fast processing speed, high recognition accuracy, and simple and fast detection method.

Description

Technical Field

[0001] The present invention relates to the technical fields of machine learning, machine vision and metal additive manufacturing, and in particular to a method for dynamic molten pool size detection and forming direction discrimination. Background Art

[0002] In the past three decades, metal additive manufacturing technology has made great progress and has been widely used in various fields, such as aerospace, medicine, and shipbuilding. It is an important part of intelligent manufacturing and has provided strong impetus for the development of modern manufacturing and the transformation and upgrading of traditional manufacturing.

[0003] However, due to the "additive" nature of metal additive manufacturing, the parts produced by it are different from those produced by traditional methods and are prone to quality problems and defects. Therefore, how to improve the quality and level of production is currently a major research direction in metal additive manufacturing. The molten pool in the metal additive manufacturing process can directly reflect the forming state and quality of the produced parts. Therefore, monitoring, information collection, and analysis and feedback of the molten pool morphology play an important role in controlling the production process and improving production quality. Currently, the detection of molten pool morphology is mainly divided into detection based on traditional algorithms and detection based on instance segmentation neural networks. However, the detection accuracy of traditional algorithms is low and has large limitations. The detection accuracy of instance segmentation neural networks is high but the speed is relatively slow. Summary of the Invention

[0004] Purpose of the invention: In order to overcome the deficiencies in the prior art, a method for dynamic molten pool size detection and forming direction discrimination is provided, which has the characteristics of fast processing speed, high recognition accuracy, and simple and quick detection method.

[0005] Technical solution: To achieve the above-mentioned purpose, the present invention provides a method for dynamic molten pool size detection and forming direction determination, comprising the following steps:

[0006] S1: Collect dynamic melt pool images of metal additive manufacturing and perform image segmentation on them to generate a dynamic melt pool image source dataset;

[0007] S2: Use the double labeling method to mark the size of the dynamic melt pool image and the forming direction information of the dynamic melt pool image to create a dynamic melt pool image dataset;

[0008] S3: Build a convolutional neural network for melt pool size detection with direction discrimination function, improve the loss function, add melt pool forming direction loss, and fully train it using a dynamic melt pool image dataset to obtain a trained convolutional neural network for melt pool size detection with direction discrimination function;

[0009] S4: Collect real-time dynamic melt pool images, use the trained melt pool size detection convolutional neural network to perform size detection and forming direction discrimination on the real-time dynamic melt pool images, and obtain the length, width and forming direction information of the melt pool morphology.

[0010] Furthermore, the step S1 is specifically as follows:

[0011] A1: Using a CCD camera to capture dynamic melt pool images during metal additive manufacturing;

[0012] A2: With the center of the dynamic melt pool image as the center point, a square area with a side length of half the dynamic melt pool image is cut out to extract the center image of the melt pool, and image segmentation is performed. All dynamic melt pool images are segmented using this image segmentation method to generate a dynamic melt pool image source dataset.

[0013] Furthermore, the step S2 is specifically as follows:

[0014] B1: Treat the edge of the molten pool as an approximate ellipse. Use two rectangular boxes to mark the major and minor axes of the ellipse. The diagonal positions of the rectangular boxes are the positions of the major and minor axes to be marked. The two rectangular boxes together mark the size of the edge of the molten pool.

[0015] B2: Divide a 360° circle into 36 equal parts, with F1 as the 0° direction and clockwise as F i (i=1,2,…,36), according to the tail direction of the ellipse of the molten pool edge, the forming direction information of the dynamic molten pool image is annotated with the closest direction as D, and one-hot processing is performed on it to obtain the processed label direction f i ;

[0016] B3: According to the annotation method of steps B1 and B2, annotate all dynamic melt pool images in the dynamic melt pool image source dataset to generate a dynamic melt pool image dataset.

[0017] Furthermore, the method for constructing the convolutional neural network for detecting the melt pool size with the direction discrimination function in step S3 is as follows:

[0018] Based on the YoloX target recognition neural network, its Decoupled Head is improved. In the Decoupled Head, an additional direction prediction branch is added to predict the forming direction of the molten pool. The starting point of the direction prediction branch is located after the 1×1 convolution integration channel of the Decoupled Head, that is, it is located at the same starting point as the other classification task branch and regression task branch. The direction prediction branch is similar to the classification task branch and contains 3 convolution channels of different sizes corresponding to 3 different scales of input, namely, 80×80×256 input convolution channel, 40×40×512 input convolution channel, and 20×20×1024 input convolution channel. Finally, the three convolution channels jointly output the prediction probability of the molten pool forming direction in all i directions, which is recorded as The output of the classification task is the prediction of the major axis and minor axis, and the output of the regression task is the pixel coordinates of the two endpoints of a diagonal line of the major axis prediction box and the minor axis prediction box.

[0019] Furthermore, in step S3, the original loss function L of YoloX is yolox On the basis of , the molten pool forming direction loss is added, including the forming direction prediction loss L f and forming direction limitation loss L fl ;

[0020] Prediction loss in forming direction L f The specific calculation method is:

[0021] First use the Sigmoid function to calculate Get p i , the formula is as follows:

[0022]

[0023] Calculate x again i With label direction f i BCELoss is used as the prediction loss in the forming direction L f , the formula is as follows:

[0024]

[0025] Forming direction limitation loss L fl The specific calculation method is:

[0026]

[0027] Among them, x1, y1 and x2, y2 are the pixel coordinates of the two endpoints of a diagonal line of the long axis prediction box;

[0028] Finally, the loss function of the convolutional neural network for melt pool size detection with direction discrimination is obtained:

[0029] L=L yolox +L f +L fl .

[0030] Furthermore, in step S3, the dynamic melt pool image dataset is used to fully train the constructed melt pool size detection convolutional neural network with direction discrimination function. The training configuration is to use an Nvidia3090 graphics card, use the Adam optimizer, the initial learning rate is 0.001, the number of training rounds is 200 rounds, and the learning rate is reduced to 0.99 of the previous round in each round, and finally a trained melt pool size detection convolutional neural network with direction discrimination function is obtained.

[0031] Furthermore, the specific structure of the convolutional channels of the direction prediction branch of the melt pool size detection convolutional neural network built in step S3 with three different scale inputs is as follows:

[0032] The convolution channel with 80×80×256 input is composed of two identical convolution modules and a direction prediction convolution layer in sequence. The convolution module contains a standard convolution layer with an input channel of 256, an output channel of 256, a convolution kernel of 3, and a stride of 1, a batch normalization layer with a channel of 256, and a silu activation layer. The direction prediction convolution layer is a standard convolution layer with an input channel of 256, an output channel of 36, a convolution kernel of 1, and a stride of 1.

[0033] The convolution channel of 40×40×512 input is composed of two identical convolution modules and a direction prediction convolution layer in sequence. The convolution module contains a standard convolution layer with an input channel of 512, an output channel of 512, a convolution kernel of 3, and a stride of 1, a batch normalization layer with a channel of 512, and a silu activation layer. The direction prediction convolution layer is a standard convolution layer with an input channel of 512, an output channel of 36, a convolution kernel of 1, and a stride of 1.

[0034] The convolution channel with a 20×20×1024 input is composed of two identical convolution modules and a direction prediction convolution layer in sequence. The convolution module contains a standard convolution layer with an input channel of 1024, an output channel of 1024, a convolution kernel of 3, and a stride of 1, a batch normalization layer with a channel of 1024, and a silu activation layer. The direction prediction convolution layer is a standard convolution layer with an input channel of 1024, an output channel of 36, a convolution kernel of 1, and a stride of 1.

[0035] Furthermore, the step S4 is specifically as follows:

[0036] D1: Use a CCD camera to capture real-time dynamic melt pool images during the metal additive manufacturing process. The captured real-time dynamic melt pool images are input into a trained melt pool size detection convolutional neural network with direction discrimination function. The real-time dynamic melt pool images are subjected to size detection and forming direction discrimination. The pixel coordinates of the two end points of a diagonal line of the long and short axis prediction boxes of the melt pool are output, as well as the predicted melt pool forming direction.

[0037] D2: Let x1, y1 and x2, y2 be the pixel coordinates of the two end points of a diagonal line of the long axis prediction box, and the pixel distance of the long axis is Combined with the CCD camera calibration, the actual length of the long axis, that is, the length of the molten pool morphology, is obtained. Similarly, the actual length of the short axis, that is, the width of the molten pool morphology, is obtained.

[0038] Beneficial effects: Compared with the prior art, in order to improve the detection accuracy of the molten pool morphology while maintaining a high speed, the present invention designs a molten pool size detection convolutional neural network with direction discrimination, improves the detection method of the traditional target recognition neural network, and detects the same target twice, so that it not only has the ability of the traditional target recognition neural network to detect the molten pool position, but also can adapt to the characteristics of the constantly changing molten pool morphology, and distinguish the forming direction of the dynamic molten pool morphology, thereby making the recognition accuracy more accurate; and because the present invention is based on the target recognition neural network for improvement, the detection speed is faster than that of the instance segmentation neural network; and at the same time, it combines the molten pool forming direction discrimination, so that the two-step algorithm of dynamic molten pool size detection and forming direction judgment is integrated into one step, which is simpler and faster. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic flow diagram of the method of the present invention;

[0040] Figure 2 This is a partial dynamic molten pool image display diagram provided by an embodiment of the present invention;

[0041] Figure 3 Schematic diagram of a dynamic melt pool image segmentation method provided by an embodiment of the present invention;

[0042] Figure 4 Schematic diagram of the double labeling method provided by an embodiment of the present invention;

[0043] Figure 5 A schematic diagram of an improved Decoupled Head provided in an embodiment of the present invention;

[0044] Figure 6 Schematic diagram of the output results of dynamic molten pool size detection and forming direction discrimination provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The present invention is further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, modifications of various equivalent forms of the present invention made by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0046] The present invention provides a method for dynamic molten pool size detection and forming direction determination, such as Figure 1 As shown, the following steps are included:

[0047] Step 1: Use a CCD camera to capture the dynamic melt pool image of metal additive manufacturing, and perform image segmentation on the dynamic melt pool image to generate a dynamic melt pool image source dataset;

[0048] Step 2: Use the double labeling method to mark the size of the dynamic melt pool image and the forming direction information of the dynamic melt pool image to create a dynamic melt pool image dataset;

[0049] Step 3: Build a convolutional neural network for melt pool size detection with direction discrimination function, improve the loss function, add melt pool forming direction loss, and use dynamic melt pool image dataset for full training to obtain a trained convolutional neural network for melt pool size detection with direction discrimination function;

[0050] Step 4: Collect real-time dynamic melt pool images, use the trained melt pool size detection convolutional neural network to perform size detection and forming direction discrimination on the real-time dynamic melt pool images, and obtain the length, width and forming direction information of the melt pool morphology.

[0051] In this embodiment, the above method is applied as follows:

[0052] Step 1 includes:

[0053] Step 1.1: Use a CCD camera to capture dynamic melt pool images during metal additive manufacturing, refer to Figure 2 , Figure 2 This is a partial dynamic molten pool image display diagram provided by an embodiment of the present invention;

[0054] Step 1.2: With the center of the dynamic melt pool image as the center point, a square area with a side length of half the dynamic melt pool image is cut out to extract the center image of the melt pool and perform image segmentation. Figure 3 , Figure 3 This is a schematic diagram of the dynamic melt pool image segmentation method provided by an embodiment of the present invention. This image segmentation method is used to segment all dynamic melt pool images to produce a dynamic melt pool image source data set.

[0055] Step 2 includes:

[0056] Step 2.1: The double labeling method is as follows: consider the edge of the molten pool as an approximate ellipse, use two rectangular boxes to mark the major axis and minor axis of the ellipse of the molten pool edge respectively, and the diagonal position of the rectangular box is the position of the major axis and minor axis to be marked. The two rectangular boxes jointly mark the size of the edge of the molten pool, refer to Figure 4 , Figure 4 is a schematic diagram of the double labeling method provided by an embodiment of the present invention, wherein the red solid line frame marks the long axis of the molten pool, the red dotted line represents the long axis of the molten pool, the blue solid line frame marks the short axis of the molten pool, and the blue dotted line represents the short axis of the molten pool;

[0057] Step 2.2: Divide a 360° circle into 36 equal parts, with F1 as the 0° direction and clockwise as F i (i=1,2,…,36), according to the tail end direction of the ellipse of the molten pool edge, the forming direction information of the dynamic molten pool image is annotated as D with the closest direction, and one-hot processing is performed on it [see Fu Yixian, Lu Tianliang, Ma Zeliang. CNN malicious code detection technology based on one-hot [J]. Computer Applications and Software, 2020, 37(01): 304-308+333.], and the processed label direction f is obtained. i ;

[0058] Step 2.3: According to the annotation methods of steps 2.1 and 2.2, annotate all dynamic melt pool images in the dynamic melt pool image source dataset to create a dynamic melt pool image dataset.

[0059] Step 3 includes:

[0060] Step 3.1: The construction method of the convolutional neural network for melt pool size detection with direction discrimination function is as follows: Based on the YoloX target recognition neural network [see Zhao Hui, Zhao Yao, Jin Linlin, Dong Lanfang, Xiao Xiao. Research and implementation of small target fireworks detection technology based on YOLOX [J / OL]. Journal of Graphics: 1-8 [2022-06-07].], improve its Decoupled Head, refer to Figure 5 , Figure 5Schematic diagram of the improved Decoupled Head provided for an embodiment of the present invention. In the Decoupled Head, an additional direction prediction branch is added to predict the forming direction of the molten pool. The starting point of the direction prediction branch is located after the 1×1 convolution integration channel of the Decoupled Head, that is, it is located at the same starting point as the other classification task branches and regression task branches. The direction prediction branch is similar to the classification task branch, and contains 3 convolution channels of different sizes corresponding to 3 different scales of input, namely, a convolution channel with 80×80×256 input, a convolution channel with 40×40×512 input, and a convolution channel with 20×20×1024 input. Finally, the three convolution channels jointly output the predicted probability of the molten pool forming direction in all i directions, which is recorded as The classification task outputs the prediction of the major axis and minor axis, and the regression task outputs the pixel coordinates of the two endpoints of a diagonal of the major axis prediction box and the minor axis prediction box. Regardless of which diagonal endpoint pixel coordinates of the output prediction box are used, it will not affect the final detection results of the major axis and minor axis lengths. Therefore, there is no need to consider whether the diagonal of the output prediction box is on the major axis or minor axis of the melt pool.

[0061] The specific structure of the convolution channel of the direction prediction branch with three different scale inputs is:

[0062] The convolutional pipeline for an 80×80×256 input is composed of two identical convolutional modules followed by a direction prediction convolutional layer. The convolutional module includes a standard convolutional layer with 256 input channels, 256 output channels, a convolution kernel of 3, and a stride of 1; a batch normalization layer with 256 channels; and a silu activation layer. The direction prediction convolutional layer is a standard convolutional layer with 256 input channels, 36 output channels, a convolution kernel of 1, and a stride of 1.

[0063] The convolutional pipeline for a 40×40×512 input is composed of two identical convolutional modules followed by a direction prediction convolutional layer. The convolutional module includes a standard convolutional layer with 512 input channels, 512 output channels, a convolution kernel of 3, and a stride of 1; a batch normalization layer with 512 channels; and a silu activation layer. The direction prediction convolutional layer is a standard convolutional layer with 512 input channels, 36 output channels, a convolution kernel of 1, and a stride of 1.

[0064] The convolutional pipeline for a 20×20×1024 input consists of two identical convolutional modules followed by a direction prediction convolutional layer. The convolutional module includes a standard convolutional layer with 1024 input channels, 1024 output channels, a convolution kernel of 3, and a stride of 1; a batch normalization layer with 1024 channels; and a silu activation layer. The direction prediction convolutional layer is a standard convolutional layer with 1024 input channels, 36 output channels, a convolution kernel of 1, and a stride of 1.

[0065] Step 3.2: In YoloX original loss function L yolox On the basis of , the molten pool forming direction loss is added, including the forming direction prediction loss L f and forming direction limitation loss L fl . Forming direction prediction loss L f The specific calculation method is:

[0066] First use the Sigmoid function to calculate Get p i , the formula is as follows:

[0067]

[0068] Calculate x again i With label direction f i BCELoss is used as the prediction loss in the forming direction L f , the formula is as follows:

[0069]

[0070] Forming direction limitation loss L fl The specific calculation method is:

[0071]

[0072] Among them, x1, y1 and x2, y2 are the pixel coordinates of the two endpoints of a diagonal line of the long axis prediction box.

[0073] Finally, the loss function of the convolutional neural network for melt pool size detection with direction discrimination is obtained:

[0074] L=L yolox +L f +L fl

[0075] Step 3.3: Use the dynamic melt pool image dataset to fully train the constructed melt pool size detection convolutional neural network with direction discrimination. The training configuration uses an Nvidia3090 graphics card, uses the Adam optimizer, has an initial learning rate of 0.001, and trains 200 rounds. The learning rate is reduced to 0.99 of the previous round in each round, and finally a trained melt pool size detection convolutional neural network with direction discrimination is obtained.

[0076] Step 4 includes:

[0077] Step 4.1: Use a CCD camera to capture real-time dynamic melt pool images during the metal additive manufacturing process. Input the captured real-time dynamic melt pool images into the trained melt pool size detection convolutional neural network with direction discrimination. Perform size detection and forming direction discrimination on the real-time dynamic melt pool images, and output the pixel coordinates of the two end points of a diagonal line of the long axis and short axis prediction boxes of the melt pool, as well as the predicted melt pool forming direction. Figure 6 , Figure 6 This is a schematic diagram of the output results of dynamic molten pool size detection and forming direction discrimination provided by an embodiment of the present invention. The red box is the molten pool major axis prediction box, the "D=70" above indicates that the forming direction of the molten pool is 70°, and the green box is the molten pool minor axis prediction box.

[0078] Step 4.2: Let x1, y1 and x2, y2 be the pixel coordinates of the two end points of a diagonal line of the long axis prediction box, and the pixel distance of the long axis is refer to Figure 6 , Figure 6 This is a schematic diagram of the output results of dynamic melt pool size detection and forming direction discrimination provided by an embodiment of the present invention. The red dashed line represents the detected major axis of the melt pool, while the green dashed line represents the detected minor axis. Combined with CCD camera calibration, the actual length of the major axis, i.e., the length of the melt pool topography, is obtained. Similarly, the actual length of the minor axis, i.e., the width of the melt pool topography, is obtained.

[0079] Based on the above scheme, this embodiment compares and analyzes the technical solutions of the present invention and existing methods. The present invention specifically targets melt pool images of changing directions collected on-site, adding an additional direction determination branch for melt pool detection. Existing methods primarily focus on melt pools with fixed directions and do not consider scenarios with changing directions. The key to the present invention lies in the successful introduction of end-to-end determination of melt pool direction within the target recognition network, completing melt pool direction and size detection in a single step. This is something that existing target recognition networks and existing melt pool detection methods lack. Therefore, the present invention not only possesses the ability of traditional target recognition neural networks to detect melt pool position, but also adapts to the ever-changing nature of melt pool morphology and discerns the forming direction of dynamic melt pool morphology, thereby achieving more accurate recognition. Furthermore, because the present invention is based on an improved target recognition neural network, it is faster than instance segmentation neural networks in detection speed. Furthermore, by incorporating melt pool forming direction discrimination, the two-step algorithm for dynamic melt pool size detection and forming direction determination is integrated into a single step, making it simpler and faster.

Claims

1. A dynamic molten pool size detection and forming direction determination method, characterized in that: The steps include: S1: Collect dynamic melt pool images of metal additive manufacturing and perform image segmentation on them to generate a dynamic melt pool image source dataset; S2: Use the double labeling method to mark the size of the dynamic melt pool image and the forming direction information of the dynamic melt pool image to create a dynamic melt pool image dataset; S3: Build a convolutional neural network for melt pool size detection with direction discrimination function, improve the loss function, add melt pool forming direction loss, and fully train it using a dynamic melt pool image dataset to obtain a trained convolutional neural network for melt pool size detection with direction discrimination function; S4: Collect real-time dynamic melt pool images, use the trained melt pool size detection convolutional neural network to perform size detection and forming direction discrimination on the real-time dynamic melt pool images, and obtain the length, width and forming direction information of the melt pool morphology.

2. A method for dynamic molten pool size detection and forming direction determination according to claim 1, characterized in that: The step S1 is specifically as follows: A1: Using a CCD camera to capture dynamic melt pool images during metal additive manufacturing; A2: With the center of the dynamic melt pool image as the center point, a square area with a side length of half the dynamic melt pool image is cut out to extract the center image of the melt pool, and image segmentation is performed. All dynamic melt pool images are segmented using this image segmentation method to generate a dynamic melt pool image source dataset.

3. A dynamic molten pool size detection and forming direction determination method according to claim 1, characterized in that: The step S2 is specifically as follows: B1: Treat the edge of the molten pool as an approximate ellipse. Use two rectangular boxes to mark the major and minor axes of the ellipse. The diagonal positions of the rectangular boxes are the positions of the major and minor axes to be marked. The two rectangular boxes together mark the size of the edge of the molten pool. B2: Divide a 360° circle into 36 equal parts, with F1 as the 0° direction and clockwise as F i (i=1,2,…,36), according to the tail direction of the ellipse of the molten pool edge, the forming direction information of the dynamic molten pool image is annotated with the closest direction as D, and one-hot processing is performed on it to obtain the processed label direction f i ; B3: According to the annotation method of steps B1 and B2, annotate all dynamic melt pool images in the dynamic melt pool image source dataset to generate a dynamic melt pool image dataset.

4. A method for dynamic molten pool size detection and forming direction determination according to claim 1, characterized in that: The method for constructing the convolutional neural network for detecting the molten pool size with the direction discrimination function in step S3 is as follows: Based on the YoloX target recognition neural network, its Decoupled Head is improved. In the Decoupled Head, an additional direction prediction branch is added to predict the forming direction of the molten pool. The starting point of the direction prediction branch is located after the 1×1 convolution integration channel of the Decoupled Head, and is located at the same starting point as the other classification task branches and regression task branches. The direction prediction branch is similar to the classification task branch and contains 3 convolution channels of different sizes corresponding to 3 different scales of input, namely, 80×80×256 input convolution channel, 40×40×512 input convolution channel, and 20×20×1024 input convolution channel. Finally, the three convolution channels jointly output the prediction probability of the molten pool forming direction in all i directions, which is recorded as The output of the classification task is the prediction of the major axis and minor axis, and the output of the regression task is the pixel coordinates of the two endpoints of a diagonal line of the major axis prediction box and the minor axis prediction box.

5. A method for dynamic molten pool size detection and forming direction determination according to claim 4, characterized in that: In step S3, the original loss function L of YoloX is yolox On the basis of , the molten pool forming direction loss is added, including the forming direction prediction loss L f and forming direction limitation loss L fl ; Prediction loss in forming direction L f The specific calculation method is: First use the Sigmoid function to calculate Get p i , the formula is as follows: Calculate x again i With label direction f i BCELoss is used as the prediction loss in the forming direction L f , the formula is as follows: Forming direction limitation loss L fl The specific calculation method is: Among them, x1, y1 and x2, y2 are the pixel coordinates of the two endpoints of a diagonal line of the long axis prediction box; D is the forming direction information of the dynamic melt pool image; Finally, the loss function of the convolutional neural network for melt pool size detection with direction discrimination is obtained: L=L yolox +L f +L fl 。 6. A method for dynamic molten pool size detection and forming direction determination according to claim 1, characterized in that: In step S3, the dynamic melt pool image dataset is used to fully train the constructed melt pool size detection convolutional neural network with direction discrimination function. The training configuration is to use an Nvidia3090 graphics card, use an Adam optimizer, have an initial learning rate of 0.001, and train 200 rounds. The learning rate is reduced to 0.99 of the previous round in each round, and finally a trained melt pool size detection convolutional neural network with direction discrimination function is obtained.

7. A method for dynamic molten pool size detection and forming direction determination according to claim 4, characterized in that: The specific structure of the convolutional channels of the direction prediction branch of the melt pool size detection convolutional neural network built in step S3 with three different scale inputs is as follows: The convolution channel with 80×80×256 input is composed of two identical convolution modules and a direction prediction convolution layer in sequence. The convolution module contains a standard convolution layer with an input channel of 256, an output channel of 256, a convolution kernel of 3, and a stride of 1, a batch normalization layer with a channel of 256, and a silu activation layer. The direction prediction convolution layer is a standard convolution layer with an input channel of 256, an output channel of 36, a convolution kernel of 1, and a stride of 1. The convolution channel of 40×40×512 input is composed of two identical convolution modules and a direction prediction convolution layer in sequence. The convolution module contains a standard convolution layer with an input channel of 512, an output channel of 512, a convolution kernel of 3, and a stride of 1, a batch normalization layer with a channel of 512, and a silu activation layer. The direction prediction convolution layer is a standard convolution layer with an input channel of 512, an output channel of 36, a convolution kernel of 1, and a stride of 1. The convolution channel with a 20×20×1024 input is composed of two identical convolution modules and a direction prediction convolution layer in sequence. The convolution module contains a standard convolution layer with an input channel of 1024, an output channel of 1024, a convolution kernel of 3, and a stride of 1, a batch normalization layer with a channel of 1024, and a silu activation layer. The direction prediction convolution layer is a standard convolution layer with an input channel of 1024, an output channel of 36, a convolution kernel of 1, and a stride of 1.

8. A method for dynamic molten pool size detection and forming direction determination according to claim 1, characterized in that: The step S4 is specifically as follows: D1: Use a CCD camera to capture real-time dynamic melt pool images during the metal additive manufacturing process. The captured real-time dynamic melt pool images are input into a trained melt pool size detection convolutional neural network with direction discrimination function. The real-time dynamic melt pool images are subjected to size detection and forming direction discrimination. The pixel coordinates of the two end points of a diagonal line of the long and short axis prediction boxes of the melt pool are output, as well as the predicted melt pool forming direction. D2: Let x1, y1 and x2, y2 be the pixel coordinates of the two end points of a diagonal line of the long axis prediction box, and the pixel distance of the long axis is Combined with the CCD camera calibration, the actual length of the long axis is equal to the length of the molten pool morphology. Similarly, the actual length of the short axis is equal to the width of the molten pool morphology.

Images