Scanning galvanometer laser welding keyhole molten pool automatic identification method based on deep learning

Abstract

The invention discloses a scanning galvanometer laser welding keyhole molten pool automatic identification method based on deep learning, which comprises the following steps: acquiring original visual image information in a scanning galvanometer laser welding process, manufacturing a small amount of segmentation mark samples, and forming a semantic segmentation data set together with the original visual image; and then through training and verification, obtaining a semi-supervised generative adversarial semantic segmentation model capable of accurately segmenting an optimal attention mechanism of the keyhole molten pool. A segmentation model is used for segmenting an original visual image, then segmentation images containing a keyhole and a molten pool are obtained through traditional image processing algorithms including image channel segmentation based on pixel values, image morphological processing denoising, contour searching and the like, and then the segmentation images are multiplied by an original visual signal for mask operation, so that in the welding process, the key hole and the molten pool are obtained. And accurately identifying a keyhole and a molten pool. According to the method, the recognition accuracy of dynamic changes of the keyhole and the molten pool in the scanning galvanometer laser welding process can be improved, and it is guaranteed that the quality of machined products is controllable.

Description

Technical Field

[0001] The present invention relates to the technical field of welding keyhole and molten pool image recognition, and in particular to a method for automatic recognition of a keyhole molten pool in a scanning galvanometer laser welding process based on deep learning. Background Art

[0002] The development of intelligent technology has brought huge productivity to the industrial field, provided more solutions for the production of higher quality products, and promoted the rapid iteration of industrial manufacturing technology. Scanning galvanometer laser welding technology has been widely used in key fields such as aerospace, rail transportation, and new energy batteries due to its advantages in regulating micro-composition, controlling welding formation, and suppressing pore defects. Its automation and intelligence will bring huge economic benefits to these key industrial fields. As a key link in welding automation, welding online monitoring technology can realize the recording and analysis of physical signals such as sound, light, electricity, and heat that are closely related to welding quality, and provide the most direct welding status information for welding automation. Among them, the vision-based online monitoring technology not only has simple signal acquisition equipment, but also can directly obtain the monitored information. It has become the most widely used technology in the field of automated and intelligent welding technology. However, compared with arc and traditional non-scanning galvanometer laser welding, in the scanning galvanometer laser welding process, the laser beam stirs the molten pool at a frequency of hundreds of Hz, resulting in more dramatic changes in the shape and surface state of the keyhole molten pool in the visual signal at different times, and the visual image signal is more complex and irregular, with local highlights, steam plumes and spatters, etc., which brings great limitations to traditional image processing.

[0003] The dynamic behavior of the keyhole and the molten pool directly reflects the quality of scanning galvanometer laser welding. Especially in the scanning galvanometer laser welding process, the complex and highly dynamically changing keyhole is closely related to key parameters such as welding penetration, molten width and porosity defects. Therefore, the development of a high-precision and high-generalization online recognition method for keyholes and molten pools is of great significance for improving the automation level of scanning galvanometer laser welding technology and achieving high-quality scanning galvanometer laser welding. Existing laser welding visual monitoring methods mostly use traditional algorithms based on threshold segmentation, image morphological processing, edge detection, etc., which are difficult to achieve accurate segmentation of targets with complex surface states and geometric shapes and significant changes. In recent years, most image recognition algorithms using deep learning are also mainly aimed at non-scanning galvanometer laser welding with relatively stable welding vision. The molten pool and keyhole at different times of the swing laser are complex and have significant and drastic changes. It is difficult for traditional visual monitoring methods for scanning lasers to construct pixel-level precision target recognition. In addition, a large amount of labeled sample data is required for learning, which also increases the marking cost. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the present invention proposes a method for automatic recognition of the keyhole molten pool of scanning galvanometer laser welding based on deep learning. The method uses a semantic segmentation dataset consisting of a small number of labeled data and a large amount of unlabeled data to train a semi-supervised generative adversarial semantic segmentation model based on the attention mechanism, successfully segmenting the complex and violent keyhole molten pool area of ​​scanning galvanometer laser welding, and realizing the online recognition of the keyhole and molten pool targets that change highly dynamically during the welding process. The present invention can not only improve the recognition accuracy of the scanning galvanometer laser welding molten pool and keyhole, and achieve pixel-level recognition accuracy, but also use less labeled data to achieve higher recognition accuracy, and also reduce the heavy manual marking work.

[0005] In order to achieve the above object, the present invention adopts the following technical scheme:

[0006] A method for automatic identification of a keyhole molten pool in a scanning galvanometer laser welding process based on deep learning, comprising the following steps:

[0007] S1: Before welding, by changing the process parameters, the molten pool and keyhole visual signals with different surface states and geometric dimensions are generated; when collecting the welding visual signals, an auxiliary light source is used to illuminate the rear of the molten pool area to prevent the rear end of the molten pool from being too dark, and then the scanning galvanometer laser welding visual image signal is collected by the camera; the process parameters include welding speed, welding power, scanning amplitude, and scanning frequency;

[0008] S2: The acquired original welding images are divided into training set, validation set and test set. The image annotation software is used to annotate only the dataset of no more than 10% of the original welding visual images in the training set, and the molten pool and keyhole areas are annotated to produce real segmentation labels. Together with the original welding visual images, a semantic segmentation dataset consisting of labeled and unlabeled data is constructed. At the same time, a semi-supervised generative adversarial segmentation network based on the attention mechanism is used to construct a semantic segmentation model of the molten pool and keyhole of the scanning galvanometer laser welding.

[0009] S3: Use the training set and validation set to train and validate the semantic segmentation model of the scanning galvanometer laser welding molten pool and keyhole in S2, and finally save the optimal model; segment the original visual signal in the acquired test set to achieve accurate segmentation of the molten pool and keyhole; then use the image processing algorithm to obtain segmented images containing only the keyhole and molten pool areas from the segmented images, and then use the mask algorithm to multiply the acquired keyhole and molten pool segmentation images with the original welding image, and finally realize the online recognition of the keyhole and molten pool in the welding process.

[0010] Furthermore, S2 includes: constructing a scanning galvanometer laser welding semantic segmentation dataset ,in and They represent unlabeled welding images and labeled welding images, respectively. , and Represents the original welding images without labels at different times; welding images with labels , and Represents the original welding images with labels at different times, and Represents the true labels at different moments.

[0011] Furthermore, in S2, the constructed scanning galvanometer laser welding molten pool and keyhole semantic segmentation model is composed of an Attention U-Net network embedded with an attention module and a residual module as a generator, and a fully connected network as a discriminator; the generator is used to segment the image, and the discriminator is used to determine whether the input is a real label or a segmented image generated by the generator.

[0012] Furthermore, in S3, the semantic segmentation dataset is used to train and verify the semantic segmentation model of the scanning galvanometer laser welding molten pool and keyhole image. First, the dimension is The original RGB image X is input into the generator of the semi-supervised adversarial semantic segmentation network, that is, the Attention U-Net network S embedded with the attention module and the residual module, and then a dimension of Probability distribution image of , where H and W represent the height and width of the image respectively, and C represents the type of semantic segmentation, representing keyhole, melt pool, and background; the probability distribution image generated later and the true label are input to the discriminator, which recognizes that the input is a probability distribution image Or the real label image, the final dimension is Confidence map of .

[0013] Furthermore, in S3, the supervised training process uses images with real labels to alternately train the generator and discriminator networks, and the semi-supervised training process uses images with real labels and unlabeled original welding images at the same time; before semi-supervised learning, multiple supervised trainings are performed using labeled data to constrain the model from being affected by initial noise; then, in the semi-supervised training process, the generator network is trained using images with real labels and unlabeled original welding images.

[0014] Furthermore, the semi-supervised semantic segmentation model training of the scanning galvanometer laser welding keyhole molten pool is carried out. In the supervised training process, the process of generating network training is: by minimizing multiple loss functions Generate training, where , and Represent multi-class cross entropy loss and adversarial loss, respectively. Representation weight; adversarial loss It is used to promote the segmentation prediction probability map generated by the generation network to be closer to the actual label obtained by manual annotation, and is expressed by the following formula: ; In order to improve the performance of the generator, the multi-classification cross entropy loss Minimum;

[0015] The original welding image with true label Segmentation network via Attention U-Net Output a probability map ) and compare it with The true label is one-hot encoded. For comparison, the following formula is used: The discriminator training process is used to distinguish whether the input image is derived from a real labeled image or a probability distribution image of the generated network prediction output , using binary cross entropy loss Optimize the discriminator, ,in, is 0 or 1, corresponding to the input of the discriminator being the probability distribution image generated by the generator after segmentation Or the real label; and Represent the true label and the probability distribution image generated by segmentation respectively; and They are respectively the confidence of the probability distribution image and the real labeled image generated by the discriminator at the position (h, w) in the confidence map.

[0016] Furthermore, the semi-supervised semantic segmentation model of scanning galvanometer laser welding keyhole molten pool is trained. In the semi-supervised training process, the training of the discriminator is the same as its supervised training process. The other segmentation training is carried out by minimizing multiple loss functions Optimize, where , , and Represent multi-class cross entropy loss, adversarial loss and semi-supervised loss respectively, and Indicates two for reducing When the segmentation network is trained using the original welding images with true labels, optimization, , where I represents the indicator function of different confidence values, Express confidence Figure 2The threshold of the meta-operation, ; After continuous iterative training, the model with the highest accuracy will be saved.

[0017] Furthermore, S3 includes: using the optimal segmentation model obtained through training and verification to perform the segmentation task of the images in the original scanning galvanometer laser welding visual test set to obtain the corresponding segmented image; then using the image processing algorithm to obtain the keyhole and molten pool areas in the segmented image, including using image channel separation based on pixel values ​​to obtain keyhole and molten pool images, using closing operations to remove noise in the image, contour search to obtain the contours of the keyhole and molten pool, and contour filling to prevent the existence of holes inside the recognition area, and finally obtaining the keyhole and molten pool segmented images respectively.

[0018] Furthermore, in S3, the original welding image scanning galvanometer laser welding visual image is multiplied with the obtained keyhole and molten pool segmentation images respectively to perform mask operations, and finally the keyhole and molten pool images in the actual scanning galvanometer laser welding process are obtained, thereby realizing online automatic recognition of the keyhole and molten pool.

[0019] The present invention can realize online automatic recognition of the scanning galvanometer laser welding keyhole molten pool. Compared with the traditional image processing algorithm and CNN neural network method for identifying the scanning galvanometer laser welding keyhole molten pool, the present invention has the following beneficial effects:

[0020] 1. The semi-supervised approach adopted by the present invention, under the premise of using a small amount of labeled and a large amount of unlabeled scanning galvanometer laser welding visual image data, can accurately identify complex and significantly different keyholes and molten pools during the scanning galvanometer laser welding process.

[0021] 2. Since the semi-supervised generative adversarial segmentation network of the attention mechanism is embedded with the attention mechanism and the residual module, the generator can calculate the weights between different pixels in the image and find the relationship between different features in the image, especially in dealing with the complex and drastically changing welding visual signals of the scanning galvanometer laser, and can extract more detailed information from the image information. Therefore, the present invention can not only accurately identify the molten pool with complex surface state, but also accurately identify the small target and high-speed moving keyhole. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a flow chart of a method for automatic identification of a keyhole molten pool in scanning galvanometer laser welding based on deep learning according to the present invention.

[0023] Figure 2 This is a schematic diagram of the network structure of the semi-supervised adversarial scanning galvanometer laser welding image segmentation based on the attention mechanism proposed in the present invention.

[0024] Figure 3The original scanning galvanometer laser welding visual images under different welding process parameters of the test set collected by the present invention; a-a3 are the original welding images at different times of test 4#, b-b3 are the original welding images at different times of test 7#, and c-c3 are the original welding images at different times of test 12#.

[0025] Figure 4 The network accuracy and loss curve graph is generated during the optimization process of the semantic segmentation model of the present invention.

[0026] Figure 5 It is the segmentation results and confidence map output in real time by the generator and discriminator during the training process of the semantic segmentation model of the present invention.

[0027] Figure 6 The test set for this invention Figure 3 The segmentation results of the generator in the optimal semi-supervised generative adversarial segmentation model obtained; a-a3 are the segmentation result images of experiment 4# at different times, b-b3 are the segmentation result images of experiment 7# at different times, and c-c3 are the segmentation result images of experiment 12# at different times.

[0028] Figure 7 This is a flow chart of the method for automatically identifying the keyhole molten pool in the original scanning galvanometer laser welding image based on the segmentation results of the semi-supervised generative adversarial network of the present invention. DETAILED DESCRIPTION

[0029] In order to make the features, technical solutions and advantages of the present invention more understandable, the present invention is further described in detail in conjunction with the accompanying drawings and specific implementation schemes. In addition, the specific implementation schemes described herein are only used to explain the main points of the present invention and are not used to limit the present invention. The present invention can be appropriately improved according to specific actual use conditions.

[0030] like Figure 1 As shown, a method for automatic identification of a keyhole molten pool of a scanning galvanometer laser welding keyhole based on deep learning of the present invention comprises the following steps:

[0031] S1: Before welding, the process parameters including welding speed, welding power, scanning amplitude and scanning frequency are changed to produce welding states with significant differences, thereby obtaining rich welding visual image signals; during the welding visual signal acquisition process, an auxiliary light source is used to illuminate the rear of the molten pool area to prevent the rear end of the molten pool from being too dark, and then the scanning galvanometer laser welding visual image signal (i.e. Figure 1 (The acquisition of visual signals during the scanning galvanometer laser welding process).

[0032] S2: The original welding images are divided into training sets, validation sets, and test sets. The image annotation software is used to annotate the molten pool and keyhole areas of a small part of the original welding visual images in the training set, and the real segmentation labels are produced. Together with the original welding visual images, a semantic segmentation dataset consisting of labeled and unlabeled data is constructed. At the same time, a semi-supervised generative adversarial segmentation network based on the attention mechanism is used to construct a scanning galvanometer laser welding molten pool and keyhole semantic segmentation model, which consists of a generator S and a discriminator D.

[0033] S3: Use the training set and validation set to train and validate the semantic segmentation model of the scanning galvanometer laser welding molten pool and keyhole in S2, and finally save the optimal model. Segment the original visual signal in the test set to achieve accurate segmentation of the molten pool and keyhole; then use the image processing algorithm to obtain segmented images containing only the keyhole and molten pool area from the segmented image, and then use the mask algorithm to multiply the acquired keyhole and molten pool segmented images with the original welding image, and finally achieve online recognition of the keyhole and molten pool during the welding process.

[0034] Example:

[0035] This embodiment constructs a method for automatically identifying a keyhole molten pool in a scanning galvanometer laser welding process based on deep learning, including:

[0036] First, the scanning laser welding experimental parameters with different laser powers, welding speeds, swing amplitudes, and swing frequencies were designed. The specific parameter matching is shown in Table 1 (welding process parameter table of scanning laser welding experiment), and then a rich welding visual image signal is obtained. The visual image signals of some welding process parameters at different times are shown in Table 1. Figure 3 . At the same time, an aluminum alloy with a thickness of 8mm was selected.

[0037] surface

[0038]

[0039] Secondly, use image segmentation software to annotate the keyhole and melt pool in the obtained melt pool image to obtain the real label. The annotation scheme is shown in Table 1. Randomly select some experimental groups and annotate 100-200 images. Experiments 4#, 7#, and 12# in Table 1 are used as test sets, and then the annotated ( ) and unmarked ( ) visual images to build a scanning galvanometer laser welding molten pool and keyhole semantic segmentation model dataset, and divide the dataset into training set, validation set, and test set. and represent unlabeled welding images and labeled welding images respectively, and Represents the original welding images with labels at different times, and represents the true labels at different times, and Represents raw welding images without labels at different times.

[0040] Then, a semi-supervised adversarial semantic segmentation model of the molten pool and keyhole of the scanning galvanometer laser welding based on the attention mechanism is constructed. The Attention U-Net network embedded with the attention module and the residual module is used as the generator, and the fully connected network is used as the discriminator. Its structure and related parameters are as follows: Figure 2 As shown in the figure. The Attention U-Net network consists of two parts: 4 layers of downsampling and 4 layers of upsampling. The upsampling operation is as follows: First, the residual connection F fuses the image features extracted by the 3X3 convolution operation (Conv) and the normalization layer (BN) with the image features extracted by the 1X1 convolution operation (Conv), and then the fused image is average pooled to complete a downsampling operation; the upsampling operation is to expand the dimension of the final image obtained by downsampling through the upsampling operation, and splice it with the image obtained by the previous downsampling layer after the gating unit, and the spliced ​​image is connected with the image after the 1x1 convolution residual to complete an upsampling, and the same 4 upsamplings are performed to finally complete the segmentation task of the original image and obtain the segmented image. The discriminator consists of a 4-layer structure, the first layer includes padding (zero padding), a 3x3 convolution layer (Conv) with 128 channels and an activation layer (LeakRelu), the second layer includes padding (zero padding), a 3x3 convolution layer (Conv) with 256 channels and an activation layer (LeakRelu), and the third layer includes a @3x3 convolution layer (Conv) with 1 channel, followed by upsampling (Upsamplelayer) to output a confidence map with 1 channel. The generator is used to segment the image, and the discriminator is used to determine whether the input is a real label or a segmented image generated by the generator.

[0041] Then, the training set and validation set are used to optimize the model. The number of segmented images sent to the semi-supervised adversarial semantic network each time, i.e., the batch size, is set to 32, and the model iteration cycle epoch is set to 100. First, the model is supervised and trained using annotated welding visual images. Specifically, a welding process image with a dimension of (H: indicates the height of the image, W: indicates the width of the image) Original welding visual image X n Input into the generator model, and then generate a dimension of Probability distribution image of , where C represents the type of semantic segmentation, which in this invention represents keyhole, melt pool, and background. And calculate the segmentation result and the true label Multi-class cross entropy loss between and combat loss , and through the loss function Optimize the segmentation capability of the generator: ,in, and Represent multi-class cross entropy loss and adversarial loss respectively. , where h and w represent the coordinate values ​​in the height and width directions of the image respectively. represents the weight, which can be set to 0.02. Then the discriminator calculates the binary cross entropy adversarial loss is optimized, where ; and Represent the true label and the probability distribution image generated by segmentation, respectively. and The confidence of the probability distribution image and the real labeled image generated by the discriminator at the position (h, w) in the confidence map, respectively, enables it to more accurately judge whether the input is the real label or the image obtained by the generator segmentation. Therefore, after the initial training of the model with labeled data, the generator has a certain segmentation ability and the discriminator has a certain recognition ability, so that the entire generative adversarial network model has a certain degree of immunity to noise interference.

[0042] Then, semi-supervised training is performed using welding images with real labels and unlabeled images. The training parameter batchsize is 32 and the training is performed for 200 epochs. The training process of the discriminator is the same as the supervised training process. The generator is trained by calculating the minimum multiple loss function , improve the segmentation ability of the generator, where , , and They represent multi-class cross entropy loss, adversarial loss, and semi-supervised loss respectively. and Indicates two for reducing The weights can be set to 0.02 and 0.1 respectively. When the original welding images with real labels are used for segmentation network training, optimization, , where I represents the indicator function of different confidence values, c represents the channel of the image, The true label value of channel c at position (h, w) in the image, Represents the segmentation result. Express confidence Figure 2 The threshold of the meta-operation, ,in In the model training process, the segmentation accuracy of the generator and its loss are calculated as follows: Figure 4 As shown; during the training process, the segmentation effect of the model is recorded every 20 epochs from 0 to 120 epochs. The results are as follows Figure 5 As shown. It can be seen that as the training progresses, the accuracy of the model becomes higher and the loss becomes smaller, indicating that the original welding image is getting closer and closer to the real label after segmentation by the generator, and the noise in the segmentation result is getting less and less. Finally, the model with the highest accuracy and the smallest error is saved as the optimal model for scanning galvanometer laser welding keyhole molten pool image segmentation.

[0043] Then, the optimal model was tested using test sets 4#, 7#, and 12#. The final segmented image is as follows: Figure 6 shown.

[0044] Then the scanning galvanometer laser welding keyhole molten pool is extracted, and the process is as follows: Figure 7 Then, the image processing algorithm is used to obtain the keyhole and molten pool areas in the segmented image: the keyhole and molten pool images are obtained by image channel separation based on pixel values, the noise in the image is removed by closing operation, the contour search is used to obtain the contours of the keyhole and molten pool, and the contour is filled to prevent the existence of holes in the recognition area, and finally the keyhole and molten pool images are obtained respectively.

[0045] Furthermore, the original welding image scanning galvanometer laser welding visual image is multiplied with the keyhole and molten pool images obtained from the segmented image to perform mask operations, and finally the actual scanning galvanometer laser welding keyhole and molten pool images are obtained, such as Figure 7 As shown, online automatic identification of keyhole and molten pool is realized.

[0046] The present invention does not elaborate on some of the common technologies of those skilled in the art. The above-described embodiments are only descriptions of preferred implementations of the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementations described. Without departing from the design spirit of the present invention, various modifications and improvements made by ordinary technicians in this field to the technical solution of the present invention should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A method for automatic identification of keyhole molten pool in scanning galvanometer laser welding based on deep learning, characterized in that: The following steps are involved: S1: Before welding, by changing the process parameters, the molten pool and keyhole visual signals with different surface states and geometric dimensions are generated; when collecting the welding visual signals, an auxiliary light source is used to illuminate the rear of the molten pool area to prevent the rear end of the molten pool from being too dark, and then the scanning galvanometer laser welding visual image signal is collected by the camera; the process parameters include welding speed, welding power, scanning amplitude, and scanning frequency; S2: The original welding images are divided into training set, validation set and test set. The image annotation software is used to annotate the molten pool and keyhole areas of only 10% of the original welding visual images in the training set, and the real segmentation labels are produced. Together with the original welding visual images, a semantic segmentation dataset consisting of labeled and unlabeled images is constructed. At the same time, a semi-supervised generative adversarial segmentation network based on the attention mechanism is used to construct a semantic segmentation model for the molten pool and keyhole of scanning galvanometer laser welding. S3: Use the training set and validation set to train and validate the semantic segmentation model of the scanning galvanometer laser welding molten pool and keyhole in S2, and finally save the optimal model; segment the original visual signal in the acquired test set to achieve accurate segmentation of the molten pool and keyhole; then use the image processing algorithm to obtain segmented images containing only the keyhole and molten pool areas from the segmented images, and then use the mask algorithm to multiply the acquired keyhole and molten pool segmentation images with the original welding image, and finally realize the online recognition of the keyhole and molten pool in the welding process.

2. The method for automatic identification of keyhole molten pool in scanning galvanometer laser welding based on deep learning according to claim 1 is characterized in that: S2 includes: constructing a scanning galvanometer laser welding semantic segmentation dataset ,in and They represent unlabeled welding images and labeled welding images, respectively. , and Represents the original welding images without labels at different times; welding images with labels , and Represents the original welding images with labels at different times, and Represents the true labels at different moments.

3. The method for automatic identification of keyhole molten pool in scanning galvanometer laser welding based on deep learning according to claim 1 is characterized in that: In S2, the constructed scanning galvanometer laser welding molten pool and keyhole semantic segmentation model is composed of an Attention U-Net network embedded with an attention module and a residual module as a generator, and a fully connected network as a discriminator; the generator is used to segment the image, and the discriminator is used to determine whether the input is a real label or a segmented image generated by the generator.

4. The method for automatic identification of keyhole molten pool in scanning galvanometer laser welding based on deep learning according to claim 3 is characterized in that: In S3, the semantic segmentation dataset is used to train and verify the semantic segmentation model of the scanning galvanometer laser welding molten pool and keyhole image. First, the dimension is The original RGB image X is input into the generator of the semi-supervised adversarial semantic segmentation network, that is, the Attention U-Net network S embedded with the attention module and the residual module, and then a dimension of Probability distribution image of , where H and W represent the height and width of the image respectively, and C represents the type of semantic segmentation, representing keyhole, melt pool, and background; the probability distribution image generated later and the true label are input to the discriminator, which recognizes that the input is a probability distribution image Or the real label image, the final dimension is Confidence map of .

5. The method for automatic identification of keyhole molten pool in scanning galvanometer laser welding based on deep learning according to claim 3 is characterized in that: In S3, the supervised training process uses images with real labels to alternately train the generator and discriminator networks, and the semi-supervised training process uses images with real labels and unlabeled original welding images at the same time; before the semi-supervised learning, multiple supervised trainings are performed using labeled data to constrain the model from being affected by initial noise; The generative network is then trained using both true labeled images and unlabeled raw welding images in a semi-supervised training process.

6. The method for automatic identification of keyhole molten pool in scanning galvanometer laser welding based on deep learning according to claim 5 is characterized in that: The semi-supervised semantic segmentation model training of the scanning galvanometer laser welding keyhole molten pool is carried out. During the supervised training process, the process of generating network training is: by minimizing multiple loss functions Generate training, where , and Represent multi-class cross entropy loss and adversarial loss, respectively. Representation weight; adversarial loss It is used to promote the segmentation prediction probability map generated by the generation network to be closer to the actual label obtained by manual annotation, and is expressed by the following formula: ; In order to improve the performance of the generator, the multi-classification cross entropy loss Minimum; The original welding image with true label Segmentation network via Attention U-Net Output a probability map ) and compare it with The true label is one-hot encoded. For comparison, the following formula is used: The discriminator training process is used to distinguish whether the input image is derived from a real labeled image or a probability distribution image of the generated network prediction output , using binary cross entropy loss Optimize the discriminator, ,in, is 0 or 1, corresponding to the input of the discriminator being the probability distribution image generated by the generator after segmentation Or real label ; and They are respectively the confidence of the probability distribution image and the real labeled image generated by the discriminator at the position (h, w) in the confidence map.

7. The method for automatic identification of keyhole molten pool in scanning galvanometer laser welding based on deep learning according to claim 5 is characterized in that: The semi-supervised semantic segmentation model of scanning galvanometer laser welding keyhole molten pool is trained. In the semi-supervised training process, the training of the discriminator is the same as that in the supervised training process. The additional segmentation training is performed by minimizing multiple loss functions. Optimize, where , , and Represent multi-class cross entropy loss, adversarial loss and semi-supervised loss respectively, and Indicates two for reducing When the segmentation network is trained using the original welding images with true labels, optimization, , where I represents the indicator function of different confidence values, represents the threshold of the confidence map binary operation, ; After continuous iterative training, the model with the highest accuracy will be saved.

8. The method for automatic identification of keyhole molten pool in scanning galvanometer laser welding based on deep learning according to claim 3 is characterized in that: The S3 includes: using the optimal segmentation model obtained through training and verification to perform the segmentation task of the original scanning galvanometer laser welding test set visual image to obtain the corresponding segmented image; then using the image processing algorithm to obtain the keyhole and molten pool areas in the segmented image, including using the image channel separation algorithm based on pixel values ​​to obtain the keyhole and molten pool images; then using the closing operation to remove the noise in the image, the contour search to obtain the contour of the keyhole and the molten pool, and the contour filling to prevent the existence of holes inside the recognition area, and finally obtaining the keyhole and molten pool segmented images respectively.

9. The method for automatic identification of keyhole molten pool in scanning galvanometer laser welding based on deep learning according to claim 3, characterized in that: In S3, the original welding image scanning galvanometer laser welding visual image is multiplied with the obtained keyhole and molten pool segmentation images respectively to perform mask operations, and finally the keyhole and molten pool images in the actual scanning galvanometer laser welding process are obtained, thereby realizing online automatic recognition of the keyhole and molten pool.

Images