Methods, devices, equipment and media for identifying welding condition categories

Abstract

This invention discloses a method, apparatus, device, and medium for identifying welding state categories. The method includes: acquiring candidate molten pool images during the target welding process and preprocessing the candidate molten pool images to obtain a target molten pool image; inputting the target molten pool image into a trained target welding state classification model to obtain the corresponding target welding state category; wherein the model hyperparameters in the target welding state classification model are determined based on a target hyperparameter scheme; the target hyperparameter scheme is determined by partitioning and searching within the hyperparameter solution space based on a sample dataset, a basic region database, and a region proxy model; the hyperparameter solution space is constructed by encoding the initial hyperparameters of the basic welding state classification model. This improves the accuracy of welding state category identification.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of artificial intelligence technology, and in particular to a method, apparatus, device and medium for identifying welding status categories. Background Technology

[0002] With the rapid development of intelligent manufacturing and industrial automation technologies, online monitoring and quality control of the welding process have become crucial for ensuring the safety and reliability of engineering structures. In fields such as oil and gas pipelines, pressure vessels, and large equipment manufacturing, welding quality directly affects the service performance and lifespan of structures. However, in actual welding processes, due to factors such as arc interference, metal spatter, material inhomogeneity, and fluctuations in process parameters, the welding process exhibits high dynamism and uncertainty. Traditional quality assessment methods relying on human experience are insufficient to achieve stable and accurate real-time judgment.

[0003] Existing technologies typically employ visual perception-based welding process monitoring methods. These methods, using molten pool images as the primary information carrier, analyze the molten pool's morphology, size changes, and dynamic behavior to identify welding conditions and determine quality trends, representing a crucial technological path for achieving intelligent welding. This type of problem can often be modeled as a multi-category classification task, such as categorizing welding conditions into stable states, fluctuating states, excessively large molten pools, excessively small molten pools, incomplete fusion tendencies, and burn-through risks, thus providing a basis for welding process control. Therefore, improving the accuracy of welding condition category identification is of paramount importance. Summary of the Invention

[0004] This invention provides a method, apparatus, device, and medium for identifying welding condition categories, in order to improve the accuracy of welding condition category identification.

[0005] According to one aspect of the present invention, a method for identifying welding condition categories is provided, comprising: Acquire candidate molten pool images during the target welding process, and preprocess the candidate molten pool images to obtain the target molten pool image; The target molten pool image is input into the trained target welding state classification model to obtain the corresponding target welding state category; The model hyperparameters in the target welding state classification model are determined based on the target hyperparameter scheme; the target hyperparameter scheme is determined by partitioning and searching within the hyperparameter solution space based on the sample dataset, the basic region database, and the region proxy model; the hyperparameter solution space is constructed by encoding the initial hyperparameters of the basic welding state classification model.

[0006] According to another aspect of the present invention, a welding condition category identification device is provided, comprising: The target molten pool image determination module is used to acquire candidate molten pool images during the target welding process and preprocess the candidate molten pool images to obtain the target molten pool image. The target welding state determination module is used to input the target molten pool image into the trained target welding state classification model to obtain the corresponding target welding state category; The model hyperparameters in the target welding state classification model are determined based on the target hyperparameter scheme; the target hyperparameter scheme is determined by partitioning and searching within the hyperparameter solution space based on the sample dataset, the basic region database, and the region proxy model; the hyperparameter solution space is constructed by encoding the initial hyperparameters of the basic welding state classification model.

[0007] According to another aspect of the present invention, an electronic device is provided, comprising: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by one or more processors, the one or more processors are able to execute any of the welding state category identification methods provided in the embodiments of the present invention.

[0008] According to another aspect of the present invention, a computer-readable storage medium is provided, which stores computer instructions for causing a processor to execute and implement any of the welding state category identification methods provided in the embodiments of the present invention.

[0009] This invention provides a welding state recognition scheme. It involves acquiring candidate molten pool images during the target welding process and preprocessing these images to obtain a target molten pool image. The target molten pool image is then input into a trained target welding state classification model to obtain the corresponding target welding state category. The model hyperparameters in the target welding state classification model are determined based on a target hyperparameter scheme. This target hyperparameter scheme is determined through a partitioned search within the hyperparameter solution space, using a sample dataset, a basic region database, and a region proxy model. The hyperparameter solution space is constructed by encoding the initial hyperparameters of the basic welding state classification model. The above scheme improves the accuracy of the determined target welding state category by inputting the preprocessed target molten pool image into the target welding state classification model after adjusting the model hyperparameters based on the target hyperparameter scheme. This improves the accuracy of welding state category identification, i.e., it enhances the accuracy of welding state category recognition for the target welding process. Simultaneously, by using a sample dataset, a basic region database, and a region proxy model, a partitioned search is performed in the hyperparameter solution space to determine the target hyperparameter scheme, further improving the accuracy of the determined scheme. Furthermore, by adjusting the model hyperparameters in the basic welding state classification model based on the target hyperparameter scheme, the accuracy and reliability of the obtained target welding state classification model are improved, thus enhancing the accuracy of the target welding state category determination based on the target welding state classification model and ultimately improving the accuracy of welding state category recognition.

[0010] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart of a welding state category identification method provided in Embodiment 1 of the present invention; Figure 2 This is a flowchart of a welding state category identification method provided in Embodiment 2 of the present invention; Figure 3 This is a schematic diagram of a welding status category identification device provided in Embodiment 4 of the present invention; Figure 4This is a schematic diagram of the structure of an electronic device that implements a welding state category identification method according to Embodiment 5 of the present invention. Detailed Implementation

[0013] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0014] Example 1 Figure 1 This is a flowchart of a welding state category identification method provided in Embodiment 1 of the present invention. This embodiment can be applied to the case of welding state category identification in the welding process. The method can be executed by a welding state category identification device, which can be implemented in software and / or hardware and can be configured in an electronic device that carries the welding state identification function.

[0015] See Figure 1 The welding condition category identification method shown includes: S110. Obtain candidate molten pool images during the target welding process, and preprocess the candidate molten pool images to obtain the target molten pool image.

[0016] The target welding process refers to the welding process that requires welding status identification. The candidate molten pool image refers to the molten pool image used for welding status category identification of the target welding process. The molten pool image is a two-dimensional grayscale image acquired during the welding process by capturing the molten pool area using a visual sensing device.

[0017] The target molten pool image refers to the image obtained after preprocessing the candidate molten pool image. For example, preprocessing may include denoising, enhancement, and normalization, etc., and this embodiment of the invention does not specifically limit these processes.

[0018] Specifically, candidate molten pool images corresponding to the target welding process are obtained, and the candidate molten pool images are preprocessed to obtain the target molten pool image.

[0019] S120. Input the target molten pool image into the trained target welding state classification model to obtain the corresponding target welding state category.

[0020] The target weld condition classification model can be used to determine the target weld condition category corresponding to the target molten pool image. The target weld condition category refers to the type of weld condition corresponding to the target molten pool image. For example, the weld condition category may include stable state, fluctuating state, excessively large molten pool, excessively small molten pool, lack of fusion tendency, and burn-through risk.

[0021] Among them, the model hyperparameters in the target welding state classification model are determined based on the target hyperparameter scheme; the target hyperparameter scheme is determined by partitioning and searching within the hyperparameter solution space based on the sample dataset, the basic region database, and the region proxy model; the hyperparameter solution space is constructed by encoding the initial hyperparameters of the basic welding state classification model.

[0022] The target hyperparameter scheme refers to the configuration scheme corresponding to the model hyperparameters in the target welding condition classification model. That is, by adjusting the model hyperparameters in the basic welding condition classification model through the target hyperparameter scheme, the target welding condition classification model can be obtained. Model hyperparameters refer to the hyperparameters in the welding condition classification model.

[0023] The sample dataset can be used to train and validate the welding condition classification model. The basic region database refers to the pre-set original database corresponding to the solution spaces of each region. The region surrogate model refers to the surrogate model corresponding to the region solution space. The hyperparameter solution space refers to the overall solution space corresponding to the initial hyperparameters of the basic welding condition classification model. The initial hyperparameters refer to the hyperparameter categories and corresponding hyperparameter values ​​corresponding to the model structure of the basic welding condition classification model. The hyperparameter values ​​can be the possible values ​​corresponding to the hyperparameter categories, and can be set by technicians according to their needs or experience.

[0024] For example, the sample dataset includes a sample training set and a sample validation set. The sample training set includes training molten pool images and training real weld state categories. The training molten pool images are input into the target weld state classification model to obtain the target predicted weld state category; based on the target predicted weld state category and the training real weld state category, the target model loss value of the target weld state classification model is determined; the target model loss value is then used to train the target weld state classification model.

[0025] Here, the target predicted welding state category refers to the output of the target welding state classification model during training. The target model loss value refers to the model loss value corresponding to the target welding state classification model.

[0026] For example, training is stopped when the number of training iterations of the target welding state classification model reaches a preset target training iteration threshold, resulting in a well-trained target welding state classification model. This embodiment of the invention does not impose any limitation on the size of the preset target training iteration threshold; it can be set by a technician based on experience or needs. For example, the preset target training iteration threshold can be 100.

[0027] This invention provides a welding state category recognition scheme. It involves acquiring candidate molten pool images during the target welding process and preprocessing these images to obtain the target molten pool image. The target molten pool image is then input into a trained target welding state classification model to obtain the corresponding target welding state category. The model hyperparameters in the target welding state classification model are determined based on a target hyperparameter scheme. This target hyperparameter scheme is determined through a partitioned search within the hyperparameter solution space, using a sample dataset, a basic region database, and a region proxy model. The hyperparameter solution space is constructed by encoding the initial hyperparameters of the basic welding state classification model. The above scheme improves the accuracy of the determined target welding state category by inputting the preprocessed target molten pool image into the target welding state classification model after adjusting the model hyperparameters based on the target hyperparameter scheme. This improves the accuracy of welding state category identification, i.e., it enhances the accuracy of welding state category recognition for the target welding process. Simultaneously, by using a sample dataset, a basic region database, and a region proxy model, a partitioned search is performed in the hyperparameter solution space to determine the target hyperparameter scheme, further improving the accuracy of the determined scheme. Furthermore, by adjusting the model hyperparameters in the basic welding state classification model based on the target hyperparameter scheme, the accuracy and reliability of the obtained target welding state classification model are improved, thus enhancing the accuracy of the target welding state category determination based on the target welding state classification model and ultimately improving the accuracy of welding state category recognition.

[0028] Example 2 Figure 2This is a flowchart of a welding state category recognition method provided in Embodiment 2 of the present invention. Based on the above embodiments, this embodiment further adds the following steps: "Obtaining a sample dataset and initial hyperparameters of a basic welding state classification model, and constructing a hyperparameter solution space based on the initial hyperparameters; wherein, the sample dataset includes a sample training set and a sample validation set; the sample training set includes training molten pool images and training real welding state categories; the sample validation set includes validation molten pool images and validation real welding state categories; sampling a preset number of initial hyperparameter schemes from the hyperparameter solution space, and constructing an initial welding state recognition model corresponding to the initial hyperparameter schemes; training the initial welding state recognition model based on the sample training set, and then..." The molten pool image is input into a trained initial welding state classification model to obtain the initial verification predicted welding state category. Based on the verification actual welding state category and the initial verification predicted welding state category, the initial classification accuracy corresponding to the corresponding initial hyperparameter scheme is determined. Based on a preset partitioning strategy, the initial hyperparameter scheme is classified according to its initial classification accuracy, and based on the classification results and the basic region database, the initial region database corresponding to the corresponding region solution space is determined. Each initial region database is iteratively updated to obtain the target region database corresponding to the corresponding region solution space, and the target hyperparameter scheme is determined based on the target region database to improve the determination mechanism of the target hyperparameter scheme. It should be noted that for parts not detailed in this embodiment, please refer to the descriptions in other embodiments.

[0029] See Figure 2 The welding condition category identification method shown includes: S210. Obtain the sample dataset and the initial hyperparameters of the basic welding state classification model, and construct the hyperparameter solution space based on the initial hyperparameters; wherein, the sample dataset includes a sample training set and a sample validation set; the sample training set includes training molten pool images and training real welding state categories; the sample validation set includes validation molten pool images and validation real welding state categories.

[0030] The sample dataset can be used to train and validate the welding condition classification model. For example, the sample dataset includes a training set and a validation set. The training set can be used to train the welding condition classification model. The validation set can be used to validate the performance of the trained welding condition classification model.

[0031] For example, the sample training set includes training molten pool images and training real weld state categories. Training molten pool images refer to the sample molten pool images input into the weld state classification model to be trained. Training real weld state categories refer to the actual weld state categories corresponding to the training molten pool images. Training real weld state categories can be understood as the labels corresponding to the training molten pool images.

[0032] For example, the sample validation set includes validation molten pool images and validation true weld state categories. Validation molten pool images refer to sample molten pool images input into the trained weld state classification model. Validation true weld state categories refer to the actual weld state categories corresponding to the validation molten pool images. Validation true weld state categories can be understood as the labels corresponding to the validation molten pool images.

[0033] For example, the training molten pool image and the verification molten pool image refer to the molten pool images corresponding to the sample welding process. The sample welding process refers to the welding process that has already undergone welding state category identification.

[0034] The basic welding condition classification model refers to the pre-set, original welding condition classification model that has not undergone hyperparameter adjustments. Initial hyperparameters refer to the hyperparameter categories and corresponding hyperparameter values ​​corresponding to the model structure of the basic welding condition classification model. The hyperparameter values ​​can be the possible values ​​corresponding to the hyperparameter categories and can be set by technicians based on their needs or experience.

[0035] The hyperparameter solution space refers to the global solution space corresponding to the initial hyperparameters of the basic welding state classification model.

[0036] For example, the encoding method and value range of each architecture hyperparameter (i.e. hyperparameter category) of the basic welding state classification model are determined. Different hyperparameter categories correspond to different encoding methods and value ranges. Based on the encoding method and value range corresponding to the hyperparameter category, the hyperparameter values ​​under that hyperparameter category are encoded, and the hyperparameter solution space is constructed based on the encoded initial hyperparameters.

[0037] For example, if the hyperparameters of the welding state classification model include the number of convolutional layer channels, the kernel size, and the activation function type, the initial hyperparameters can include 16 convolutional layer channels, 32 convolutional layer channels, 64 convolutional layer channels, 1×1 convolutional kernel size, 3×3 convolutional kernel size, 5×5 convolutional kernel size, 7×7 convolutional kernel size, and activation functions such as Sigmoid (S-shaped growth curve), ReLU (linear rectified function), and Tanh (hyperbolic tangent function).

[0038] Continuing the previous example, for the number of channels in a convolutional layer, direct encoding can be used. For example, when the number of channels in a convolutional layer is set to 16, 32, and 64, the encoded values ​​are 16, 32, and 64, respectively. For the kernel size, index encoding can be performed based on a preset size set. For example, kernel sizes such as 1×1, 3×3, 5×5, and 7×7 can be encoded as 0, 1, 2, and 3, respectively. For activation functions, categorical encoding can be performed according to a preset function list. For example, Sigmoid, ReLU, and Tanh can be mapped to 0, 1, and 2, respectively. Here, the preset size set refers to a pre-set list used to record the encoded values ​​corresponding to different kernel sizes. The preset function list refers to a pre-set list used to record the encoded values ​​corresponding to different activation functions.

[0039] For example, for any hyperparameter category, based on the encoded values ​​corresponding to each hyperparameter value under that category, determine the upper bound and lower bound encoded values ​​for that hyperparameter category; based on the upper bound encoded values ​​for each hyperparameter category, determine the upper bound vector of the hyperparameter solution space; based on the lower bound encoded values ​​for each hyperparameter category, determine the lower bound vector of the hyperparameter solution space; and based on the upper and lower bound vectors, determine the hyperparameter solution space. Here, the upper bound encoded value refers to the maximum encoded value under any hyperparameter category; correspondingly, the lower bound encoded value refers to the minimum encoded value under any hyperparameter category. The upper bound vector can be understood as the set of upper bound encoded values ​​corresponding to different hyperparameter categories. The lower bound vector can be understood as the set of lower bound encoded values ​​corresponding to different hyperparameter categories.

[0040] S220. Sample a preset number of initial hyperparameter schemes from the hyperparameter solution space, and construct an initial welding state recognition model corresponding to the initial hyperparameter schemes.

[0041] In this embodiment of the invention, the size of the preset quantity threshold is not limited and can be set by a technician based on experience or needs. An initial hyperparameter scheme refers to a hyperparameter configuration scheme sampled from the hyperparameter solution space, which can be used to partition the hyperparameter solution space. Any initial hyperparameter scheme includes various hyperparameter categories and their corresponding encoded hyperparameter values, and the number of encoded hyperparameter values ​​under the same hyperparameter category in any initial hyperparameter scheme is one.

[0042] The initial welding state identification model refers to the model obtained by adjusting the hyperparameters of the basic welding state classification model based on the initial hyperparameter scheme. Each initial hyperparameter scheme corresponds to an initial welding state classification model.

[0043] For example, based on a preset sampling method, a preset number of initial hyperparameter schemes are sampled from the hyperparameter solution space; for any initial hyperparameter scheme, the model hyperparameters in the basic welding state classification model are adjusted according to the initial hyperparameter scheme to obtain the initial welding state classification model corresponding to the initial hyperparameter scheme.

[0044] The embodiments of the present invention do not impose any limitations on the preset sampling method, which can be set by technicians based on experience or needs. For example, the preset sampling method can be random sampling or Latin hypercube sampling.

[0045] S230. Train the initial welding state recognition model based on the sample training set, and input the verification molten pool image into the trained initial welding state classification model to obtain the initial verification predicted welding state category.

[0046] The initial verification prediction of the welding state category refers to the prediction result output after the verification molten pool image is input into the trained initial welding state classification model.

[0047] For example, the training molten pool image is input into the initial welding state classification model to obtain the initial training predicted welding state category; based on the initial training predicted welding state category and the training real welding state category, the initial model loss value is determined; and the initial welding state classification model is trained based on the initial model loss value. Here, the initial training predicted welding state category refers to the prediction result output after inputting the training molten pool image into the initial welding state classification model to be trained. The initial model loss value refers to the loss value corresponding to the initial welding state classification model.

[0048] For example, training is stopped when the initial welding state classification model reaches a preset training count threshold, resulting in a well-trained initial welding state classification model. This embodiment of the invention does not limit the size of the preset training count threshold; it can be set by a technician based on experience or needs. For example, the preset training count threshold can be 5.

[0049] For example, for any initial welding state classification model, the training molten pool image is input into the initial welding state classification model to obtain the initial training predicted welding state category; based on the initial training predicted welding state category and the training real welding state category, the initial model loss value of the initial welding state classification model is determined; the initial welding state classification model is trained based on the initial model loss value; when the number of training times reaches a preset training time threshold, training is stopped, and the trained initial welding state classification model is obtained.

[0050] For example, for any initial welding state classification model, the verification molten pool image is input into the trained initial welding state classification model to obtain the initial verification predicted welding state category corresponding to the initial welding state classification model.

[0051] S240. Based on the verified actual welding state category and the initial verified predicted welding state category, determine the initial classification accuracy corresponding to the corresponding initial hyperparameter scheme.

[0052] The initial classification accuracy refers to the classification accuracy corresponding to the initial welding state classification model.

[0053] For example, for any initial hyperparameter scheme, the initial classification accuracy corresponding to the initial hyperparameter scheme is determined based on the verified actual welding state category and the initial verified predicted welding state category output by the initial welding state classification model corresponding to the initial hyperparameter scheme.

[0054] S250. Based on the preset partitioning strategy, classify the initial hyperparameter scheme according to the initial classification accuracy corresponding to the initial hyperparameter scheme, and determine the initial region database corresponding to the corresponding region solution space based on the classification results and the basic region database.

[0055] The preset partitioning strategy refers to a pre-set strategy used to classify the initial hyperparameter scheme. For example, the preset partitioning strategy includes a threshold for the number of regions in the regional solution space and a method for classifying the initial hyperparameter scheme. This embodiment of the invention does not specifically limit the size of the threshold for the number of regions; it can be set by a technician based on experience or needs. Alternatively, the threshold can be adaptively determined based on the changing trends of the cluster compactness and inter-class discriminancy of the initial hyperparameter scheme, using a combination of squared error and elbow method.

[0056] The basic region database refers to the pre-set original database corresponding to the solution space of each region. The region solution space is a subspace of the hyperparameter solution space. The classification result refers to the result obtained by classifying the initial hyperparameter schemes. The initial region database is the database obtained by storing the initial hyperparameter schemes into the basic region database of the corresponding region solution space based on the classification result.

[0057] For example, if the number of initial hyperparameter schemes is N and the threshold for the number of region spaces is M; the initial hyperparameter schemes are sorted from largest to smallest according to their initial classification accuracy; based on the sorting result, taking the initial hyperparameter scheme k with the highest initial classification accuracy as the center, the initial hyperparameter scheme k and the closest one to it are selected. (Rounded down) initial hyperparameter schemes are grouped into one class; among the remaining initial hyperparameter schemes, the initial hyperparameter scheme q with the highest initial classification accuracy is re-determined, and the initial hyperparameter scheme q and the closest one to it are grouped together. The initial hyperparameter schemes are classified into one class; the above process is repeated until M-1 classes are formed; the remaining initial hyperparameter schemes are classified into the same class; in summary, the classification result is obtained.

[0058] For example, based on the initial hyperparameter scheme within any class in the classification results, the corresponding region solution space is determined from the hyperparameter solution space; the initial hyperparameter scheme and the corresponding initial classification accuracy within each class in the classification results are stored in the corresponding basic region database to obtain the initial region database corresponding to the corresponding region solution space.

[0059] For example, if the classification results include classes A, B, and M, the intra-class upper and lower bounds of class A are determined based on the initial hyperparameter scheme within class A. The subspace defined by the intra-class upper and lower bounds of class A is then determined from the hyperparameter solution space; this subspace is the region solution space AA corresponding to class A. Similarly, the region solution space BB corresponding to class B and the region solution space MM corresponding to class M can be obtained. The initial hyperparameter scheme and corresponding initial classification accuracy within class A are stored in the basic region database corresponding to class A, resulting in the initial region database AAA corresponding to region solution space AA. Similarly, the initial region database BBB corresponding to region solution space BB and the initial region database MMM corresponding to region solution space MM can be obtained. The intra-class upper bound can include the maximum encoded hyperparameter value corresponding to each hyperparameter category in any intra-class initial hyperparameter scheme. The intra-class lower bound can include the minimum encoded hyperparameter value corresponding to each hyperparameter category in any intra-class initial hyperparameter scheme.

[0060] It should be noted that, in this embodiment of the invention, the initial hyperparameter scheme is classified by adopting a preset partitioning strategy. Essentially, this partitions the hyperparameter solution space to obtain regional solution spaces, so that subsequent searches can be performed in each regional solution space, i.e., partitioned search.

[0061] S260. Iteratively update the database of each initial region to obtain the target region database corresponding to the solution space of the corresponding region, and determine the target hyperparameter scheme based on the target region database.

[0062] The target region database refers to the database obtained after iteratively updating the initial region database.

[0063] For example, the candidate hyperparameter scheme with the highest candidate classification accuracy is selected from the target region database corresponding to all region solution spaces, and this is taken as the target hyperparameter scheme. That is, from all candidate hyperparameter schemes stored in the target region database, the candidate hyperparameter scheme with the highest candidate classification accuracy is selected as the target hyperparameter scheme. Here, candidate classification accuracy refers to the classification accuracy stored in the target region database. Candidate hyperparameter scheme refers to the hyperparameter configuration scheme stored in the target region database.

[0064] In one optional embodiment, each initial region database is iteratively updated to obtain a target region database corresponding to the corresponding region solution space. This includes: for any database iterative update process corresponding to any region solution space, determining the current region proxy model for the current iteration based on the current region hyperparameter scheme and corresponding current classification accuracy stored in the region database to be updated in the current iteration, and a preset initial region proxy model; sampling a preset population size threshold of reference region hyperparameter schemes from the region solution space, and determining the corresponding reference classification prediction results based on the reference region hyperparameter schemes and the current region proxy model; generating an initial region population including the reference region hyperparameter schemes and reference classification prediction results, and determining the reference region database obtained by the region solution space in the current iteration based on the initial region population; and using the reference region database obtained by the region solution space in the last iteration as the target region database corresponding to the region solution space.

[0065] Here, "current iteration" refers to the current database iteration update process. The database to be updated in the region's solution space refers to the initial database in the current iteration. It should be noted that if the current iteration is the first database iteration update process corresponding to the region's solution space, then the database to be updated in the region is the initial region database corresponding to the region's solution space; if the current iteration is not the first database iteration update process corresponding to the region's solution space, then the database to be updated in the region is the reference region database obtained from the previous iteration corresponding to the region's solution space. The previous iteration refers to the database iteration update process adjacent to the current iteration.

[0066] Here, the current region hyperparameter scheme refers to the hyperparameter configuration scheme already stored in the database of the region to be updated. The current classification accuracy refers to the classification accuracy corresponding to the hyperparameter configuration scheme already stored in the database of the region to be updated.

[0067] The initial region proxy model refers to a pre-set basic proxy model corresponding to the region solution space. This embodiment of the invention does not impose any limitations on the initial region proxy model; it can be set by those skilled in the art based on experience or needs. For example, the initial region proxy model can be a radial basis function model.

[0068] The current regional proxy model refers to the regional proxy model obtained by training the initial regional proxy model based on the current regional hyperparameter scheme and the current classification accuracy in the solution space of the region.

[0069] For example, the current hyperparameter scheme of the solution space of the region is input into the initial region surrogate model to obtain the current classification prediction result; based on the current classification prediction result and the current classification accuracy, the current model loss value corresponding to the solution space of the region is determined; the initial region surrogate model is trained based on the current model loss value to obtain the current region surrogate model of the solution space of the region in the current iteration. Here, the current classification prediction result refers to the classification accuracy corresponding to the current hyperparameter scheme predicted by the initial region surrogate model. The current model loss value refers to the model loss value corresponding to the initial region surrogate model of the solution space of the region.

[0070] For example, training is stopped when the current model loss value is less than a preset agent loss threshold, resulting in a trained current region agent model; or, training is stopped when the number of training iterations of the initial region agent model reaches a preset agent training iterations threshold, resulting in a trained current region agent model. This embodiment of the invention does not impose any limitations on the size of the preset agent loss threshold and the preset agent training iterations threshold; these can be set by those skilled in the art based on experience or needs.

[0071] This invention does not impose any limitation on the size of the preset population size threshold; it can be set by a technician based on experience or needs. The preset population size threshold and the preset number threshold can be different, and this invention does not specifically limit this.

[0072] The reference region hyperparameter scheme refers to a hyperparameter configuration scheme that is not a current region hyperparameter scheme and is sampled from the region solution space. The reference region hyperparameter scheme should not be the same as the hyperparameter configuration scheme already stored in the region database corresponding to each region solution space.

[0073] For example, based on a preset sampling method, a preset population size threshold of reference region hyperparameter schemes are sampled from the solution space of the region.

[0074] The reference classification prediction result refers to the classification accuracy corresponding to the reference region hyperparameter scheme predicted based on the current region surrogate model. For example, for any reference region hyperparameter scheme in the solution space of the region, the reference region hyperparameter scheme is input into the current region surrogate model of the solution space of the region to obtain the reference classification prediction result corresponding to the reference region hyperparameter scheme.

[0075] The initial region population refers to the population that includes the hyperparameter scheme of the reference region in the current iteration of the solution space of that region, as well as the corresponding reference classification prediction results. The reference region database refers to the region database obtained after the current iteration of the solution space of that region. In other words, the reference region database is the region database obtained after updating the region database to be updated in the current database iteration of the solution space of that region.

[0076] For example, for any region solution space, the reference region database obtained by updating the region solution space in the last database iteration is used as the target region database of the region solution space.

[0077] For example, when the number of database iterations for updating the initial region database reaches a preset database iteration threshold, the database iteration update is stopped, and the target region database corresponding to the solution space of the corresponding region is obtained. This embodiment of the invention does not impose any limitation on the size of the preset database iteration threshold; it can be set by those skilled in the art based on experience or needs.

[0078] Understandably, for any database iterative update process corresponding to any region solution space, based on the current region hyperparameter scheme and current classification accuracy in the current iteration's database of regions to be updated, and the initial region surrogate model, the current region surrogate model for the current iteration of the region solution space is determined. A reference region hyperparameter scheme is sampled from the region solution space. Based on the reference region hyperparameter scheme and the current region surrogate model, a reference classification prediction result is determined. An initial region population including the reference region hyperparameter scheme and the reference classification prediction result is generated. Based on the initial region population, a reference region database obtained in the current iteration of the region solution space is determined. The reference region database obtained in the last iteration of the region solution space is used as the target region database corresponding to the region solution space. This realizes the iterative update of the initial region database corresponding to any region solution space, improves the accuracy of the target region database obtained from the iterative update, improves the performance of the target welding state classification model determined based on the target hyperparameter scheme determined by each target region database, and thus improves the accuracy of determining the welding state category based on the target welding state classification model.

[0079] In one optional embodiment, determining the reference region database of the solution space of the region obtained in the current iteration based on the initial region population includes: performing population iteration update on the initial region population to obtain a reference region population; selecting a reference population hyperparameter scheme from the initial population hyperparameter schemes in the reference region population according to a preset selection ratio, and determining the target region hyperparameter scheme of the solution space of the region obtained in the current iteration based on the reference population hyperparameter scheme; updating the region database to be updated according to the target region hyperparameter scheme to obtain the reference region database of the solution space of the region obtained in the current iteration.

[0080] The reference region population refers to the region population obtained after iteratively updating the initial region population corresponding to the solution space of that region. The initial population hyperparameter scheme refers to the hyperparameter configuration scheme included in the reference region population.

[0081] In this embodiment of the invention, the selection ratio of the preset scheme is not limited in any way and can be set by a technician based on experience or needs. For example, a reference number is determined based on the initial number of initial population hyperparameter schemes and the preset scheme selection ratio; a reference number of reference population hyperparameter schemes with larger classification prediction results are selected from the initial population hyperparameter schemes. Here, the initial number refers to the number of initial population hyperparameter schemes. The reference number refers to the number of reference population hyperparameter schemes.

[0082] The reference population hyperparameter scheme refers to the initial population hyperparameter scheme with the largest classification prediction result selected from the population in the reference region. For example, based on the classification prediction results of each initial population hyperparameter scheme in the reference region population, the initial population hyperparameter schemes are sorted from largest to smallest; a reference number is determined based on the initial number and the preset scheme selection ratio; and the first reference number of initial population hyperparameter schemes are used as the reference population hyperparameter schemes.

[0083] Among them, the target region hyperparameter scheme refers to the hyperparameter configuration scheme determined based on the reference population hyperparameter scheme.

[0084] It is understandable that by iteratively updating the initial population corresponding to the solution space of the region, the accuracy of the determined reference population is improved, the accuracy of the target region hyperparameter scheme determined based on the reference population hyperparameter scheme in the reference population is improved, and the accuracy of updating the database of the region to be updated according to the target region hyperparameter scheme and obtaining the reference region database of the solution space of the region in the current iteration is improved.

[0085] In an optional embodiment, determining the target region hyperparameter scheme in the solution space of the region obtained in the current iteration, based on the reference population hyperparameter scheme, includes: determining the parameter value frequency corresponding to each reference hyperparameter value of the same reference hyperparameter category in the reference population hyperparameter scheme; determining the target hyperparameter value from the reference hyperparameter values ​​of the corresponding reference hyperparameter category based on the parameter value frequency; and determining the target region hyperparameter scheme based on the reference hyperparameter category and the corresponding target hyperparameter value.

[0086] Here, the reference hyperparameter category refers to the hyperparameter category in the reference population hyperparameter scheme. The reference hyperparameter value refers to the hyperparameter value corresponding to the reference hyperparameter category. The parameter value frequency refers to the number of times each reference hyperparameter value under any reference hyperparameter category appears in the reference population hyperparameter scheme. The target hyperparameter value refers to the reference hyperparameter value corresponding to the maximum parameter value frequency under any reference hyperparameter category.

[0087] For example, for any reference hyperparameter category, determine the frequency of parameter values ​​corresponding to each reference hyperparameter value under that reference hyperparameter category; and take the reference hyperparameter value corresponding to the highest parameter value frequency as the target hyperparameter value corresponding to that reference hyperparameter category.

[0088] For example, consider the following hyperparameter schemes: reference population hyperparameter scheme u1 includes 256 channels, kernel size 1, and activation function type 1; reference population hyperparameter scheme u2 includes 128 channels, kernel size 1, and activation function type 2; reference population hyperparameter scheme u3 includes 128 channels, kernel size 2, and activation function type 1; and reference population hyperparameter scheme u4 includes 128 channels, kernel size 0, and activation function type 0. Here, the number of channels, kernel size, and activation function type are all reference hyperparameter categories. The hyperparameter values ​​under each reference hyperparameter category, such as 256, 128, 1, 2, and 0, are the reference hyperparameter values ​​corresponding to the respective reference hyperparameter category.

[0089] Continuing the previous example, for the number of channels, the reference hyperparameter value 256 corresponds to a parameter value frequency of 1, and the reference hyperparameter value 128 corresponds to a parameter value frequency of 3. Therefore, the reference hyperparameter value 128 is taken as the target hyperparameter value corresponding to the number of channels. For the kernel size, the reference hyperparameter value 0 corresponds to a parameter value frequency of 1, the reference hyperparameter value 1 corresponds to a parameter value frequency of 2, and the reference hyperparameter value 2 corresponds to a parameter value frequency of 1. Therefore, the reference hyperparameter value 1 is taken as the target hyperparameter value corresponding to the kernel size. For the activation function type, the reference hyperparameter value 0 corresponds to a parameter value frequency of 1, the reference hyperparameter value 1 corresponds to a parameter value frequency of 2, and the reference hyperparameter value 2 corresponds to a parameter value frequency of 1. Therefore, the reference hyperparameter value 1 is taken as the target hyperparameter value corresponding to the activation function type. This generates a target region hyperparameter scheme that includes 128 channels, 1 kernel size, and 1 activation function type.

[0090] It is understandable that by determining the target hyperparameter value for the corresponding reference hyperparameter category based on the frequency of parameter values ​​corresponding to each reference hyperparameter value under the same reference hyperparameter category in the reference population hyperparameter scheme, and generating a target region hyperparameter scheme based on the reference hyperparameter category and the corresponding target hyperparameter value, the accuracy of the generated target region hyperparameter scheme is improved, which in turn improves the accuracy of the final target hyperparameter scheme. This ensures the accuracy of the target welding state classification model obtained by subsequent model hyperparameter tuning based on the target hyperparameter scheme when identifying welding state categories.

[0091] In one optional embodiment, the initial regional population is iteratively updated to obtain a reference regional population, including: for any population iterative update process, determining a corresponding current population experiment scheme based on the current population hyperparameter scheme in the population to be updated in the current population iteration; determining the corresponding current experiment classification prediction result based on the current population experiment scheme and the current regional proxy model; determining the current population update scheme obtained by the current population iteration based on the current scheme classification prediction result corresponding to the current population hyperparameter scheme and the current experiment classification prediction result corresponding to the current population experiment scheme; updating the population to be updated in the region based on the current population update scheme and the corresponding current update classification prediction result to obtain the current regional population obtained by the current population iteration; and using the current regional population obtained by the last population iteration as the reference regional population.

[0092] Here, the population to be updated in the region refers to the initial population of the solution space in the current population iteration. It should be noted that if the current population iteration is the first population iteration update process corresponding to the solution space of this region, then the population to be updated in the region is the initial population; if the current population iteration is a non-first population iteration update process corresponding to the solution space of this region, then the population to be updated in the region is the current population obtained from the previous population iteration. The previous population iteration refers to the population iteration update process adjacent to the current population iteration.

[0093] Here, the current population hyperparameter scheme refers to the hyperparameter configuration scheme in the population of the region to be updated. The current population experimental scheme refers to the experimental hyperparameter scheme obtained after mutation and crossover processing of the current population hyperparameter scheme.

[0094] For example, based on the mutation operator in the differential evolution algorithm, the current mutation vector, i.e., the current population mutation scheme, is determined according to the current population hyperparameter scheme; and based on the crossover operator in the differential evolution algorithm, the current experimental vector, i.e., the current population experimental scheme, is determined according to the current population hyperparameter scheme and the current population mutation scheme. Here, the current population mutation scheme refers to the mutated hyperparameter scheme obtained after mutating the current population hyperparameter scheme.

[0095] The embodiments of this invention do not impose any limitations on the mutation operator used, which can be set by those skilled in the art based on experience or needs. For example, the mutation operator can be a difference mutation based on the current individual and the best individual (DE / current-to-best / 1).

[0096] Here, the current experimental classification prediction result refers to the classification accuracy corresponding to the current population experimental scheme predicted based on the current regional proxy model. The current population update scheme refers to the hyperparameter configuration scheme for updating the population in the region to be updated, obtained from the current population iteration. The current scheme classification prediction result refers to the classification prediction result corresponding to the current population hyperparameter scheme in the population in the region to be updated. The current update classification prediction result refers to the classification prediction result corresponding to the current population update scheme.

[0097] For example, for any current population hyperparameter scheme, the current scheme classification prediction result of the current population hyperparameter scheme is compared with the current experimental classification prediction result of the current population experimental scheme corresponding to the current population hyperparameter scheme. If the current scheme classification prediction result is greater than the current experimental classification prediction result, then the current population hyperparameter scheme is taken as the corresponding current population update scheme, and the current scheme classification prediction result of the current population hyperparameter scheme is taken as the current update classification prediction result. If the current scheme classification prediction result is less than or equal to the current experimental classification prediction result, then the current population experimental scheme corresponding to the current population hyperparameter scheme is taken as the corresponding current population update scheme, and the current experimental classification prediction result of the current population experimental scheme is taken as the current update classification prediction result.

[0098] It should be noted that during any population iteration update process, each current population hyperparameter scheme in the population to be updated region has a corresponding current population update scheme obtained from the current population iteration and the corresponding current update classification prediction result.

[0099] The current regional population refers to the regional population obtained after updating the population in the region to be updated based on the current population update scheme and the corresponding current update classification prediction results.

[0100] For example, for any region solution space, the current region population obtained in the last population iteration of the region solution space is used as the reference region population of the region solution space.

[0101] For example, when the number of population iterations for updating the initial region population reaches a preset population iteration threshold, the population iteration is stopped, and the reference region population corresponding to the solution space of the corresponding region is obtained. This embodiment of the invention does not impose any limitation on the size of the preset population iteration threshold; it can be set by those skilled in the art based on experience or needs.

[0102] Understandably, for any population iteration update process, based on the current population hyperparameter scheme in the population to be updated in the current population iteration, the corresponding current population test scheme is determined, and based on the current region proxy model, the current test classification prediction result is determined. Based on the current scheme classification prediction result and the current test classification prediction result, the current population update scheme is determined. Based on the current population update scheme and the corresponding current update classification prediction result, the population in the region to be updated is updated to obtain the reference region population. The above process is repeated until the current region population obtained in the last population iteration is the reference region population. This improves the accuracy of the determined reference region population, improves the accuracy of the subsequent determination of the target region hyperparameter scheme based on the reference region population, and thus improves the accuracy of the final determined target hyperparameter scheme. This ensures the accuracy of the target welding state classification model obtained by subsequent model hyperparameter tuning based on the target hyperparameter scheme when identifying welding state categories.

[0103] In one optional embodiment, updating the database of regions to be updated according to the target region hyperparameter scheme to obtain a reference region database of the region solution space in the current iteration includes: constructing a region welding state recognition model corresponding to the target region hyperparameter scheme, and training the region welding state recognition model according to the sample training set; inputting the verification molten pool image into the trained region welding state classification model to obtain the region verification predicted welding state category; determining the region classification accuracy corresponding to the target region hyperparameter scheme according to the verification real welding state category and the region verification predicted welding state category; and storing the target region hyperparameter scheme and the corresponding region classification accuracy in the database of regions to be updated to obtain a reference region database of the region solution space in the current iteration.

[0104] The regional welding condition classification model refers to the model obtained by adjusting the hyperparameters of the basic welding condition classification model based on the target region hyperparameter scheme. Each target region hyperparameter scheme corresponds to a regional welding condition classification model.

[0105] Among them, the region verification prediction of the welding state category refers to the prediction result output after inputting the verification molten pool image into the trained region welding state classification model. The region classification accuracy refers to the classification accuracy corresponding to the region welding state classification model.

[0106] For example, the training molten pool image is input into the regional welding state classification model to obtain the regional training predicted welding state category; based on the regional training predicted welding state category and the training true welding state category, the regional model loss value is determined; and the regional welding state classification model is trained based on the regional model loss value. Here, the regional training predicted welding state category refers to the prediction result output after inputting the training molten pool image into the regional welding state classification model to be trained. The regional model loss value refers to the loss value corresponding to the regional welding state classification model.

[0107] For example, when the number of training iterations of the regional welding state classification model reaches a preset training iteration threshold, training is stopped, and a well-trained regional welding state classification model is obtained.

[0108] For example, the verification molten pool image is input into the trained regional welding state classification model to obtain the corresponding regional verification predicted welding state category; based on the verification real welding state category and the regional verification predicted welding state category output by the regional welding state classification model corresponding to the target region hyperparameter scheme, the regional classification accuracy corresponding to the target region hyperparameter scheme is determined.

[0109] Understandably, by training the regional welding state classification model based on the target region hyperparameter scheme using the sample training set, and then verifying the performance of the trained regional welding state classification model using the sample validation set, the regional classification accuracy is obtained. Based on the target region hyperparameter scheme and the corresponding regional classification accuracy, the database of regions to be updated is updated, and the reference regional database obtained in the current iteration of the corresponding regional solution space is obtained, thereby improving the accuracy of the determined reference regional database.

[0110] S270. Obtain candidate molten pool images during the target welding process, and preprocess the candidate molten pool images to obtain the target molten pool image.

[0111] S280. Input the target molten pool image into the trained target welding state classification model to obtain the corresponding target welding state category.

[0112] Among them, the model hyperparameters in the target welding state classification model are determined based on the target hyperparameter scheme; the target hyperparameter scheme is determined by partitioning and searching within the hyperparameter solution space based on the sample dataset, the basic region database, and the region proxy model; the hyperparameter solution space is constructed by encoding the initial hyperparameters of the basic welding state classification model.

[0113] This invention provides a welding state category recognition scheme. It involves adding initial hyperparameters to a sample dataset and a basic welding state classification model, and constructing a hyperparameter solution space based on these initial hyperparameters. The sample dataset includes a training set and a validation set. The training set includes training molten pool images and training real welding state categories. The validation set includes validation molten pool images and validation real welding state categories. A preset threshold number of initial hyperparameter schemes are sampled from the hyperparameter solution space, and an initial welding state recognition model corresponding to each initial hyperparameter scheme is constructed. The initial welding state recognition model is trained using the training set, and the validation molten pool images are input into the trained model. In the initial welding state classification model, the initial verification and predicted welding state categories are obtained. Based on the verification actual welding state categories and the initial verification and predicted welding state categories, the initial classification accuracy corresponding to the corresponding initial hyperparameter scheme is determined. Based on the preset partitioning strategy, the initial hyperparameter schemes are classified according to their initial classification accuracy. Based on the classification results and the basic region database, the initial region database corresponding to the corresponding region solution space is determined. Each initial region database is iteratively updated to obtain the target region database corresponding to the corresponding region solution space. Based on the target region database, the target hyperparameter scheme is determined, thus improving the determination mechanism of the target hyperparameter scheme. The above scheme classifies the initial hyperparameter scheme according to a preset partitioning strategy and the initial classification accuracy corresponding to the initial hyperparameter scheme. Based on the classification results and the basic region database, it determines the initial region database corresponding to the corresponding region solution space. Iterative updates are performed on each initial region database to obtain the target region database corresponding to the corresponding region solution space. This achieves partitioned search of the hyperparameter solution space, improves the accuracy of the target hyperparameter scheme determined based on the target region database corresponding to each region solution space, and thus improves the performance of the target welding state classification model determined based on the target hyperparameter scheme, and improves the accuracy of welding state category identification based on the target welding state classification model.

[0114] Example 3 This invention provides an optional example based on the above embodiments. It should be noted that for parts not described in detail in this invention's embodiments, please refer to the descriptions in other embodiments.

[0115] In existing technologies, convolutional neural networks (CNNs) have achieved significant results in image classification tasks due to their excellent feature extraction capabilities, and are gradually being extended to the field of welding condition recognition. However, model performance is largely constrained by the configuration of network hyperparameters. When general-purpose CNNs are directly applied to multi-classification tasks of weld pools, they often fail to achieve ideal performance due to factors such as scene complexity, blurred class boundaries, and uneven data distribution. Therefore, it is necessary to conduct targeted hyperparameter optimization of deep classification models for actual welding conditions to obtain a more suitable model configuration with potentially superior performance.

[0116] For example, in order to construct a deep convolutional neural network model for multiple classification of welding states (i.e., welding state classification model) and its hyperparameter adaptive optimization system, the following key issues need to be addressed: (1) How to construct a welding state classification system for the welding process, and realize the complete process integration from molten pool image acquisition, data preprocessing, and state classification recognition, so as to meet the requirements of system stability, real-time performance and scalability in practical engineering applications; (2) In the welding state classification task, how to realize the adaptive configuration of hyperparameters of the deep classification network to improve the model's adaptability to complex working conditions, thereby achieving high stability and real-time performance of welding state recognition; (3) In the case of a wide variety of hyperparameters in the deep convolutional neural network architecture, a large number of them, and mutual coupling, how to coordinately optimize the main architecture hyperparameters under the constraint of limited computing resources, avoid the inefficiency and insufficient global optimality caused by manual parameter tuning, and achieve efficient overall optimization of network structure configuration.

[0117] This invention addresses the challenges in welding process monitoring, such as difficulties in identifying welding state categories, insufficient adaptability of classification models, and low efficiency in hyperparameter optimization. It proposes a system for multi-class classification of welding states, with its core being a hyperparameter optimization method for a deep classification network based on partitioned proxy-assisted evolutionary optimization. The system uses molten pool images acquired during the welding process as primary input. By constructing a multi-class classification model for welding states (i.e., a welding state classification model) and adaptively optimizing the hyperparameters of a deep convolutional neural network, it achieves high-precision identification of welding state categories (including stable states, fluctuating states, energy anomalies, and defect precursors).

[0118] This invention proposes a novel welding state multi-classification system, the core of which is a deep classification network hyperparameter optimization method based on partitioned surrogate-assisted differential evolution (PSADE) to achieve high-precision discrimination of welding state categories.

[0119] For example, in the offline phase, the welding condition multi-classification system first acquires a sufficient and balanced number of molten pool condition images through the molten pool image acquisition module, and then performs denoising, enhancement, and normalization processing through the data preprocessing module. Based on the preprocessed data, a multi-attribute labeled molten pool condition dataset is constructed, and the model hyperparameters are adaptively searched through the PSADE-driven deep neural network optimization module to obtain the optimal classification model, which is then stored in the model management module. In the online phase, the system calls the optimal model to classify the input molten pool images, realizing real-time identification of welding stability, energy input status, and defect precursors; simultaneously, the classification results can guide the welding parameter adjustment module to dynamically adjust the process parameters.

[0120] For example, the molten pool image acquisition module is connected to the data preprocessing module, the data preprocessing module is connected to the dataset construction module and the state classification module, the dataset construction module is connected to the deep neural network optimization module, the deep neural network optimization module is connected to the optimal network model training and storage module, the optimal network model training and storage module is connected to the state classification module, and the state classification module is connected to the welding parameter adjustment module.

[0121] The system includes several modules: a melt pool image acquisition module for acquiring candidate melt pool images, as well as training and validation melt pool images; a data preprocessing module for preprocessing candidate melt pool images and sending them to the state classification module; and a dataset construction module for preprocessing training and validation melt pool images and sending the preprocessed images to the dataset construction module. The dataset construction module then uses the preprocessed training and validation melt pool images to construct training and validation sets.

[0122] The deep neural network optimization module adaptively adjusts the hyperparameters of the basic welding state classification model based on the sample training set, sample validation set, and a deep classification network hyperparameter optimization method using partitioned proxy-assisted evolutionary optimization, to obtain the target welding state classification model. The optimal network model training and storage module trains the target welding state classification model based on the sample training set and stores the trained target welding state classification model.

[0123] The state classification module identifies multiple welding states from the input target molten pool image based on the target welding state classification model to obtain the corresponding target welding state category. The welding parameter adjustment module adjusts the welding process parameters according to the target welding state category.

[0124] For example, a hyperparameter optimization method for deep classification networks based on partitioned proxy-assisted evolutionary optimization can be implemented through a hyperparameter encoding unit, an initial sampling unit, a hyperparameter solution space partitioning unit, a partitioned proxy model construction unit, a proxy-assisted evolutionary optimization unit, an information fusion unit, and a result output unit.

[0125] Step 1: Determine the hyperparameter solution space through the hyperparameter encoding unit.

[0126] For example, the encoding method and value range of each architecture hyperparameter of the basic welding state classification model are determined. For the number of channels in the convolutional layer, the actual number of channels is used for direct encoding; for example, 16, 32, and 64 channels are encoded as 16, 32, and 64 respectively. For the kernel size, the optional size is encoded incrementally from 0, such as 1×1, 3×3, 5×5, 7×7… corresponding to 0, 1, 2, 3…. For the activation function, the candidate function list is encoded sequentially from 0, such as Sigmoid, ReLU, Tanh… corresponding to 0, 1, 2… respectively. After completing the above encoding, the overall solution space S composed of n hyperparameters is obtained: S=[UB,LB]; ; ; Where S represents the hyperparameter solution space; UB represents the upper bound vector of the hyperparameter solution space; and LB represents the lower bound vector of the hyperparameter solution space. This represents the upper bound encoding value corresponding to the nth hyperparameter category; This represents the lower bound encoding value corresponding to the nth hyperparameter category.

[0127] Step 2: Determine the initial hyperparameter scheme and the corresponding initial classification accuracy through the initial sampling unit.

[0128] For example, N initial hyperparameter schemes are collected and evaluated using Latin hypercube sampling (LHS) across the entire hyperparameter solution space S, ensuring that the initial hyperparameter schemes fully cover the solution space. The evaluation process involves configuring the corresponding hyperparameters for the selected model architecture, i.e., determining the initial welding state classification model corresponding to each initial hyperparameter scheme, and training the model using the sample training set. The training process is based on stochastic gradient descent (SGD), with the learning rate and momentum set to 0.01 and 0.9, respectively. Data augmentation strategies such as random pruning and horizontal flipping are used to improve the model's generalization ability. To reduce computational overhead, the number of training epochs is set to 5. After training, a classification test is performed on the sample validation set, and the resulting initial classification accuracy is used as the performance evaluation result of the corresponding initial hyperparameter scheme.

[0129] Step 3: Determine the initial region database corresponding to the region solution space by dividing the hyperparameter solution space into units.

[0130] For example, to maintain search diversity and reduce the risk of the search getting trapped in local optima, the hyperparameter solution space is divided into M sub-regions (i.e., regional solution spaces), that is, the N initial hyperparameter schemes obtained in step two are clustered. Specifically, firstly, the initial hyperparameter schemes in step two are ranked according to their performance evaluation results (i.e., initial classification accuracy); then, taking the optimal initial hyperparameter scheme k as the center, the nearest... The first initial hyperparameter scheme is assigned to the first category, denoted as initial region database 1 (DB1). The same process is repeated among the remaining initial hyperparameter schemes: the current optimal initial hyperparameter scheme q is selected, and the scheme closest to it is... The initial hyperparameter schemes are assigned to the second class, denoted as initial region database 2 (DB2). This process is iterated until the first M-1 classes are formed, and finally the remaining initial hyperparameter schemes are assigned to the Mth class (DBM). During the clustering process, a combination of the sum of squared errors (SSE) and the elbow method is introduced to adaptively determine the value of M based on the changing trends of cluster compactness and inter-class discriminability, thereby completing the adaptive partitioning of the hyperparameter solution space.

[0131] Step 4: Construct the regional proxy model corresponding to the solution space of each region through the regional proxy model construction unit.

[0132] For example, in the hyperparameter optimization process, a large number of hyperparameter configuration schemes need to be repeatedly evaluated, while directly training and testing a real model would incur significant time overhead. To improve search efficiency, a surrogate model can be introduced to replace most of the expensive real evaluation processes. In this embodiment of the invention, a radial basis function (RBF) model is used to construct a regional surrogate model for each region's solution space in step three. The RBF model is widely used for approximate modeling of complex objective functions due to its simplicity and efficiency. This model utilizes radial basis functions (such as Gaussian kernels, cubic kernels, or polynomial kernels) built based on sample distances to achieve nonlinear mapping learning from the input space to performance indicators. Due to its good approximation properties, RBF can still achieve stable fitting results even with a limited number of samples, thus exhibiting high adaptability in high-dimensional nonlinear optimization scenarios.

[0133] For example, based on the initial hyperparameter scheme of the region solution space and the set of initial classification accuracies. ;in, This represents the set of initial hyperparameter schemes and initial classification accuracies for the solution space of the m-th region. This represents the y-th initial hyperparameter scheme for the solution space of the m-th region; Let Hm represent the initial classification accuracy corresponding to the y-th initial hyperparameter scheme in the solution space of the m-th region; Hm represents the number of initial hyperparameter schemes in the solution space of the m-th region. RBF surrogate model It can be represented as: In the formula, x represents the hyperparameter configuration scheme that needs to be used to predict classification accuracy in the regional agent model; This represents the classification prediction result output by the regional agent model; It is a distance function; Radial basis functions; vector Here are the weighting coefficients of the radial basis functions, and T denotes transpose; This is a correction term used to improve the approximation accuracy of the model.

[0134] Step 5: Through the agent-assisted evolutionary optimization unit, the initial regional population corresponding to each regional solution space is updated iteratively to obtain the reference regional population corresponding to the solution space of the corresponding region.

[0135] For example, in this embodiment of the invention, the Differential Evolutionary Algorithm (DE) is used as the core optimization algorithm to search the solution space of each region divided in step three. By combining the partition proxy model in step four, all real evaluations in the optimization process are replaced. The core idea of ​​DE is to perform mutation operations on the difference vectors between individuals in the population, and combine crossover and selection mechanisms to gradually guide the population to converge to the global optimum.

[0136] For example, taking any region solution space as an example, DE first constructs the initial region population of the region solution space in the initial stage. ,in, This represents the initial population g corresponding to the solution space of this region; This represents the i-th hyperparameter configuration scheme in the initial regional population g. D represents the number of hyperparameters in any hyperparameter configuration scheme, i.e., the dimension of the optimization problem; Rg represents the classification prediction result of the i-th hyperparameter configuration scheme in the initial regional population g; Rg represents the preset population size threshold of the initial regional population g corresponding to the solution space of this region. Each round of regional population iteration sequentially performs core operations such as mutation, crossover, and selection to continuously drive the solution towards a better region. This embodiment of the invention employs the DE / current-to-best / 1 mutation operator, which combines the current hyperparameter scheme of the current population in the region population to be updated during the current population iteration. The optimal hyperparameter scheme for the current population in the region to be updated. And two random hyperparameter schemes for the current population in the region to be updated. and the current population hyperparameter scheme Evolutionary information to generate the current population hyperparameter scheme The corresponding current population mutation scheme aims to balance global exploration with local development: ; in, Indicates the current population hyperparameter scheme The corresponding current mutation vector, i.e., the current population mutation scheme; F represents the scaling factor, which usually ranges from [0.4, 1]. This is a rounding function used to ensure the discreteness of hyperparameter configuration. Wherein, It can classify and predict the hyperparameter scheme of the current population corresponding to the largest current scheme in the population to be updated.

[0137] For example, the combination of cross operators and Evolutionary information generates the current test vector The components (i.e., the current population experimental scheme) are: ; in, Represents the current test vector The j-th component in; Represents the current mutation vector The j-th component in; Indicates the current population hyperparameter scheme The j-th component; CR∈[0,1] determines the crossover probability. from The probability of inheriting each component; A random integer between 1 and D, to ensure At least from We obtain one component; rand(0,1) represents a random decimal between 0 and 1. Based on the components in the determined current trial vector, we determine the current trial vector.

[0138] For example, the selection operator is determined by comparison. and Based on the classification prediction results, those with better performance are retained for the next generation, thereby driving the population to converge towards better regions. A regional surrogate model is used in the selection process to replace all true evaluations; its mathematical expression is: ; in, Indicates the current population hyperparameter scheme The corresponding current population update scheme; Indicates the current population trial protocol The corresponding current test classification prediction results; Indicates the current population hyperparameter scheme The corresponding classification prediction results for the current scheme.

[0139] For example, the population iteration update process in step five may include mutation operation, crossover operation, determination of classification prediction results based on the regional proxy model of the regional solution space, and selection operation. The above process is executed iteratively until the number of population iteration updates reaches the preset population iteration number threshold, at which point the population iteration update stops.

[0140] Step 6: Determine the hyperparameter scheme of the target region through the information fusion unit, and update the database of the region to be updated corresponding to the solution space of the region based on the hyperparameter scheme of the target region and the corresponding region classification accuracy to obtain the reference region database.

[0141] For example, after completing the search process for each partition, for any region's solution space, a high-performing initial population hyperparameter scheme is selected from the reference region population in that region's solution space according to a proportion ρ (i.e., a preset scheme selection proportion) as the reference population hyperparameter scheme. The reference population hyperparameter scheme reflects the characteristics of potential optimal solutions within the region's solution space. To fully explore the effective information contained therein, PSADE introduces a statistically driven ensemble mechanism to perform distribution analysis on the value characteristics of the reference population hyperparameter scheme across various decision variable dimensions, and uses a voting-style decision rule to determine the combination of values ​​with the highest dominance across different variables, i.e., the target region hyperparameter scheme corresponding to that region's solution space. The target region hyperparameter scheme constructed in this way is evaluated through the real model training and validation process in step two, i.e., based on the region welding state classification model, the region classification accuracy corresponding to the target region hyperparameter scheme is determined. This ensemble process can achieve the synergistic integration of beneficial features while maintaining the stability of the search direction, which not only helps improve the algorithm's convergence efficiency but also alleviates the interference of surrogate model prediction bias on the search process, thereby enhancing the robustness of the overall optimization process.

[0142] For example, the process of steps four to six is ​​repeated to achieve iterative updates of the initial region database corresponding to each region's solution space. When the number of database iterations reaches a preset database iteration threshold, the database iteration updates are stopped, and the target region database corresponding to each region's solution space is obtained.

[0143] Step 7: Determine the target hyperparameter scheme through the result output unit.

[0144] For example, the algorithm determines the stopping condition by setting a preset threshold for the number of database iterations. When the preset threshold is reached, the algorithm stops and outputs multiple hyperparameter configuration schemes from the target database. The optimal hyperparameter configuration, i.e., the target hyperparameter scheme, can be selected based on the number of parameters or the performance of the corresponding model on the test set.

[0145] The scope of application of this invention includes: it is applicable to multi-classification tasks of welding states based on various deep classification models, and is particularly suitable for welding process monitoring scenarios where visual information of the molten pool is the main input. It can be applied to typical industrial applications such as welding stability identification, energy input state discrimination, and defect precursor warning, and also has the ability to expand to welding defect classification, weld formation quality assessment, and industrial visual inspection under complex working conditions.

[0146] The application prospects of this invention include: in intelligent welding systems, by performing multi-category real-time identification of welding states, adaptive control and quality early warning of the welding process can be achieved. For example, when the system identifies an excessively large molten pool or a risk of burn-through, it can automatically reduce the heat input; when it identifies a tendency for incomplete fusion, it can dynamically adjust the welding speed or current parameters. Combined with the hyperparameter optimization method proposed in this invention, the stability of the classification model under complex working conditions can be significantly improved, thereby promoting the transformation of the welding process from the traditional post-detection mode to a real-time process control mode, which has significant engineering application value and promising prospects for promotion.

[0147] The solution provided in this invention addresses the needs of welding process state recognition and intelligent control. It proposes a welding state multi-classification system integrating deep learning and hyperparameter adaptive optimization. Its technical features include: using the molten pool image as the core information carrier, modeling the welding state as a multi-class classification problem, and introducing a deep neural network hyperparameter optimization method based on Partition Proxy-Assisted Evolutionary Optimization (PSADE) to adaptively search and collaboratively optimize the model structure parameters; simultaneously, by constructing an integrated process of "data acquisition—preprocessing—model optimization—state classification—parameter adjustment," it achieves effective connection between offline optimization and online application, forming a closed-loop control mechanism. Based on the above technical solution, this invention can significantly improve the accuracy and robustness of welding state recognition, reduce the computational cost of model training and optimization, and improve the real-time performance and stability of the system under complex working conditions, thereby realizing the transformation of the welding process from experience-driven to intelligent and adaptive control.

[0148] Example 4 Figure 3This is a schematic diagram of a welding condition category identification device provided in Embodiment 4 of the present invention. This embodiment is applicable to situations where welding condition category identification is performed during the welding process. This method can be executed by a welding condition category identification device, which can be implemented in software and / or hardware and can be configured in an electronic device that carries the welding condition identification function.

[0149] like Figure 3 As shown, the device includes: a target molten pool image determination module 310 and a target welding state determination module 320. Among them, The target molten pool image determination module 310 is used to acquire candidate molten pool images during the target welding process and preprocess the candidate molten pool images to obtain the target molten pool image; The target welding state determination module 320 is used to input the target molten pool image into the trained target welding state classification model to obtain the corresponding target welding state category; The model hyperparameters in the target welding state classification model are determined based on the target hyperparameter scheme; the target hyperparameter scheme is determined by partitioning and searching within the hyperparameter solution space based on the sample dataset, the basic region database, and the region proxy model; the hyperparameter solution space is constructed by encoding the initial hyperparameters of the basic welding state classification model.

[0150] This invention provides a welding state category recognition scheme. It involves acquiring candidate molten pool images during the target welding process and preprocessing these images to obtain the target molten pool image. The target molten pool image is then input into a trained target welding state classification model to obtain the corresponding target welding state category. The model hyperparameters in the target welding state classification model are determined based on a target hyperparameter scheme. This target hyperparameter scheme is determined through a partitioned search within the hyperparameter solution space, using a sample dataset, a basic region database, and a region proxy model. The hyperparameter solution space is constructed by encoding the initial hyperparameters of the basic welding state classification model. The above scheme improves the accuracy of the determined target welding state category by inputting the preprocessed target molten pool image into the target welding state classification model after adjusting the model hyperparameters based on the target hyperparameter scheme. This improves the accuracy of welding state category identification, i.e., it enhances the accuracy of welding state category recognition for the target welding process. Simultaneously, by using a sample dataset, a basic region database, and a region proxy model, a partitioned search is performed in the hyperparameter solution space to determine the target hyperparameter scheme, further improving the accuracy of the determined scheme. Furthermore, by adjusting the model hyperparameters in the basic welding state classification model based on the target hyperparameter scheme, the accuracy and reliability of the obtained target welding state classification model are improved, thus enhancing the accuracy of the target welding state category determination based on the target welding state classification model and ultimately improving the accuracy of welding state category recognition.

[0151] Optionally, the target hyperparameter scheme is determined based on the following means: The hyperparameter solution space determination module is used to obtain the initial hyperparameters of the sample dataset and the basic welding state classification model, and construct the hyperparameter solution space based on the initial hyperparameters; wherein, the sample dataset includes a sample training set and a sample validation set; the sample training set includes training molten pool images and training real welding state categories; the sample validation set includes validation molten pool images and validation real welding state categories; The initial model construction module is used to sample a preset number of initial hyperparameter schemes from the hyperparameter solution space and construct an initial welding state recognition model corresponding to the initial hyperparameter scheme; The initial verification and prediction welding state category determination module is used to train the initial welding state recognition model based on the sample training set, and input the verification molten pool image into the trained initial welding state classification model to obtain the initial verification and prediction welding state category. The initial classification accuracy determination module is used to determine the initial classification accuracy corresponding to the corresponding initial hyperparameter scheme based on the verified actual welding state category and the initial verified predicted welding state category. The initial region database determination module is used to classify the initial hyperparameter scheme based on a preset partitioning strategy and the initial classification accuracy corresponding to the initial hyperparameter scheme, and determine the initial region database corresponding to the corresponding region solution space based on the classification results and the basic region database. The target hyperparameter scheme determination module is used to perform database iterative updates on each of the initial region databases to obtain the target region database corresponding to the solution space of the corresponding region, and to determine the target hyperparameter scheme based on the target region database.

[0152] Optionally, the target hyperparameter scheme determination module includes: The current region proxy model determination unit is used to determine the current region proxy model for any database iteration update process corresponding to any region solution space, based on the current region hyperparameter scheme and corresponding current classification accuracy stored in the database of the region to be updated in the current iteration, as well as the preset initial region proxy model. The reference classification prediction result determination unit is used to sample a preset population size threshold of reference region hyperparameter schemes from the solution space of the region, and determine the corresponding reference classification prediction result based on the reference region hyperparameter schemes and the current region proxy model; The reference region database determination unit is used to generate an initial region population including the reference region hyperparameter scheme and the reference classification prediction results, and to determine the reference region database obtained in the current iteration based on the initial region population. The target region database determination unit is used to take the reference region database obtained in the last iteration of the solution space of the region as the target region database corresponding to the solution space of the region.

[0153] Optionally, the reference area database determining unit includes: A reference region population determination subunit is used to perform iterative population updates on the initial region population to obtain the reference region population. The target region hyperparameter scheme determination subunit is used to select a reference population hyperparameter scheme from the initial population hyperparameter schemes in the reference region population according to the preset scheme selection ratio, and determine the target region hyperparameter scheme obtained in the current iteration based on the reference population hyperparameter scheme; A reference region database determination sub-unit is used to update the region database to be updated according to the target region hyperparameter scheme, so as to obtain the reference region database obtained in the current iteration of the solution space of the region.

[0154] Optionally, the target region hyperparameter scheme determination sub-unit is specifically used for: Determine the frequency of parameter values ​​corresponding to each reference hyperparameter value of the same reference hyperparameter category in the reference population hyperparameter scheme; Based on the frequency of the parameter values, the target hyperparameter value is determined from the reference hyperparameter values ​​of the corresponding reference hyperparameter category; The target region hyperparameter scheme is determined based on the reference hyperparameter category and the corresponding target hyperparameter value.

[0155] Optionally, the reference region population determination subunit is specifically used for: For any population iterative update process, the corresponding current population experimental scheme is determined based on the current population hyperparameter scheme in the population of the region to be updated in the current population iteration. Based on the current population experiment plan and the current regional proxy model, determine the corresponding current experiment classification prediction results; Based on the classification prediction results of the current scheme corresponding to the current population hyperparameter scheme and the classification prediction results of the current experiment scheme corresponding to the current population, the current population update scheme obtained by the current population iteration is determined. Based on the current population update scheme and the corresponding current update classification prediction results, the population in the area to be updated is updated to obtain the current population in the current area obtained by iterative current population update. The current region population obtained from the last population iteration is used as the reference region population.

[0156] Optionally, the reference region database determines the sub-units, specifically for: Construct a regional welding state recognition model corresponding to the hyperparameter scheme of the target region, and train the regional welding state recognition model based on the sample training set; The verified molten pool image is input into the trained regional welding state classification model to obtain the regional verification predicted welding state category; Based on the verified actual welding state category and the regional verified predicted welding state category, determine the regional classification accuracy corresponding to the target region hyperparameter scheme; The hyperparameter scheme of the target region and the corresponding region classification accuracy are stored in the database of the region to be updated, thus obtaining the reference region database of the solution space of the region in the current iteration.

[0157] The welding state category identification device provided in the embodiments of the present invention can execute the welding state category identification method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects for executing each welding state category identification method.

[0158] In the technical solution of the present invention, the collection, storage, use, processing, transmission, provision and disclosure of candidate melt pool images and the like all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0159] Example 5 Figure 4 This is a schematic diagram of an electronic device for implementing a welding state category identification method according to Embodiment 5 of the present invention. The electronic device 410 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workbenches, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (such as helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0160] like Figure 4 As shown, the electronic device 410 includes at least one processor 411 and a memory, such as a read-only memory (ROM) 412 or a random access memory (RAM) 413, communicatively connected to the at least one processor 411. The memory stores computer programs executable by the at least one processor. The processor 411 can perform various appropriate actions and processes based on the computer program stored in the ROM 412 or loaded from storage unit 418 into the RAM 413. The RAM 413 may also store various programs and data required for the operation of the electronic device 410. The processor 411, ROM 412, and RAM 413 are interconnected via a bus 414. An input / output (I / O) interface 415 is also connected to the bus 414.

[0161] Multiple components in electronic device 410 are connected to I / O interface 415, including: input unit 416, such as keyboard, mouse, etc.; output unit 417, such as various types of displays, speakers, etc.; storage unit 418, such as disk, optical disk, etc.; and communication unit 419, such as network card, modem, wireless transceiver, etc. Communication unit 419 allows electronic device 410 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0162] Processor 411 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 411 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 411 performs the various methods and processes described above, such as the welding condition category identification method.

[0163] In some embodiments, the welding condition category identification method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 418. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 410 via ROM 412 and / or communication unit 419. When the computer program is loaded into RAM 413 and executed by processor 411, one or more steps of the welding condition category identification method described above may be performed. Alternatively, in other embodiments, processor 411 may be configured to perform the welding condition category identification method by any other suitable means (e.g., by means of firmware).

[0164] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0165] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0166] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0167] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0168] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0169] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0170] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0171] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for identifying welding condition categories, characterized in that, include: Acquire candidate molten pool images during the target welding process, and preprocess the candidate molten pool images to obtain the target molten pool image; The target molten pool image is input into the trained target welding state classification model to obtain the corresponding target welding state category; The model hyperparameters in the target welding state classification model are determined based on the target hyperparameter scheme; the target hyperparameter scheme is determined by partitioning and searching within the hyperparameter solution space based on the sample dataset, the basic region database, and the region proxy model; the hyperparameter solution space is constructed by encoding the initial hyperparameters of the basic welding state classification model.

2. The method according to claim 1, characterized in that, The target hyperparameter scheme is determined based on the following method: Obtain the initial hyperparameters of the sample dataset and the basic welding state classification model, and construct the hyperparameter solution space based on the initial hyperparameters; wherein, the sample dataset includes a sample training set and a sample validation set; the sample training set includes training molten pool images and training real welding state categories; the sample validation set includes validation molten pool images and validation real welding state categories; A preset number of initial hyperparameter schemes are sampled from the hyperparameter solution space, and an initial welding state recognition model corresponding to the initial hyperparameter scheme is constructed. The initial welding state recognition model is trained based on the sample training set, and the verification molten pool image is input into the trained initial welding state classification model to obtain the initial verification predicted welding state category. Based on the verified actual welding state category and the initial verified predicted welding state category, determine the initial classification accuracy corresponding to the corresponding initial hyperparameter scheme; Based on the preset partitioning strategy, the initial hyperparameter scheme is classified according to the initial classification accuracy corresponding to the initial hyperparameter scheme, and the initial region database corresponding to the corresponding region solution space is determined according to the classification results and the basic region database. Each initial region database is iteratively updated to obtain the target region database corresponding to the solution space of the corresponding region, and the target hyperparameter scheme is determined based on the target region database.

3. The method according to claim 2, characterized in that, The step of iteratively updating each of the initial region databases to obtain the target region database corresponding to the solution space of the corresponding region includes: For any database iteration update process corresponding to any region solution space, the current region proxy model in the current iteration is determined based on the current region hyperparameter scheme and corresponding current classification accuracy stored in the database of the region to be updated in the current iteration, as well as the preset initial region proxy model. Sample a preset population size threshold of reference region hyperparameter schemes from the solution space of the region, and determine the corresponding reference classification prediction results based on the reference region hyperparameter schemes and the current region proxy model; Generate an initial region population including the reference region hyperparameter scheme and the reference classification prediction results, and determine the reference region database obtained in the current iteration based on the initial region population; The reference region database obtained in the last iteration of the solution space of this region is used as the target region database corresponding to the solution space of this region.

4. The method according to claim 3, characterized in that, The step of determining the reference region database for the solution space of the region obtained in the current iteration based on the initial region population includes: The initial region population is iteratively updated to obtain the reference region population; According to the preset scheme, a reference population hyperparameter scheme is selected from the initial population hyperparameter schemes in the reference region population, and the target region hyperparameter scheme obtained in the current iteration is determined based on the reference population hyperparameter scheme. The database of the region to be updated is updated according to the hyperparameter scheme of the target region to obtain the reference region database of the solution space of the region in the current iteration.

5. The method according to claim 4, characterized in that, The step of determining the hyperparameter scheme of the target region in the current iteration based on the reference population hyperparameter scheme includes: Determine the frequency of parameter values ​​corresponding to each reference hyperparameter value of the same reference hyperparameter category in the reference population hyperparameter scheme; Based on the frequency of the parameter values, the target hyperparameter value is determined from the reference hyperparameter values ​​of the corresponding reference hyperparameter category; The target region hyperparameter scheme is determined based on the reference hyperparameter category and the corresponding target hyperparameter value.

6. The method according to claim 4, characterized in that, The step of iteratively updating the initial region population to obtain the reference region population includes: For any population iterative update process, the corresponding current population experimental scheme is determined based on the current population hyperparameter scheme in the population of the region to be updated in the current population iteration. Based on the current population experiment plan and the current regional proxy model, determine the corresponding current experiment classification prediction results; Based on the classification prediction results of the current scheme corresponding to the current population hyperparameter scheme and the classification prediction results of the current experiment scheme corresponding to the current population, the current population update scheme obtained by the current population iteration is determined. Based on the current population update scheme and the corresponding current update classification prediction results, the population in the area to be updated is updated to obtain the current population in the current area obtained by iterative current population update. The current region population obtained from the last population iteration is used as the reference region population.

7. The method according to claim 4, characterized in that, The step of updating the database of the region to be updated according to the hyperparameter scheme of the target region to obtain the reference region database of the solution space of the region in the current iteration includes: Construct a regional welding state recognition model corresponding to the hyperparameter scheme of the target region, and train the regional welding state recognition model based on the sample training set; The verified molten pool image is input into the trained regional welding state classification model to obtain the regional verification predicted welding state category; Based on the verified actual welding state category and the regional verified predicted welding state category, determine the regional classification accuracy corresponding to the target region hyperparameter scheme; The hyperparameter scheme of the target region and the corresponding region classification accuracy are stored in the database of the region to be updated, thus obtaining the reference region database of the solution space of the region in the current iteration.

8. A welding condition category identification device, characterized in that, include: The target molten pool image determination module is used to acquire candidate molten pool images during the target welding process and preprocess the candidate molten pool images to obtain the target molten pool image. The target welding state determination module is used to input the target molten pool image into the trained target welding state classification model to obtain the corresponding target welding state category; The model hyperparameters in the target welding state classification model are determined based on the target hyperparameter scheme; the target hyperparameter scheme is determined by partitioning and searching within the hyperparameter solution space based on the sample dataset, the basic region database, and the region proxy model; the hyperparameter solution space is constructed by encoding the initial hyperparameters of the basic welding state classification model.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement a welding state category identification method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements a welding state category identification method as described in any one of claims 1-7.

Images