Industrial defect detection automation system based on generative adversarial network

Abstract

The invention discloses an industrial defect detection automation system based on a generative adversarial network, and particularly relates to the technical field of industrial product welding defect detection. The system comprises a data acquisition and preprocessing module, a model setting and training module, a welding defect detection module, a welding defect response module and a database. According to the method, the defect image of the target industrial product is generated according to the welding parameters and defect types of the target industrial product, then the defect image is compared with the actual welding image of the target industrial product, and detection of the welding defects of the target industrial product under different welding conditions is achieved. According to the method, all welding parameters of the target industrial product are considered, the repairing defect image of the target industrial product is generated according to the generator, then the repairing defect image is compared with the actual welding image of the repaired target industrial product, and a reliable basis is provided for repairing effect evaluation of the target industrial product.

Description

Technical Field

[0001] The present invention relates to the technical field of welding defect detection for industrial products, and specifically relates to an automated system for industrial defect detection based on a generative adversarial network. Background Art

[0002] In industrial production, welding is a crucial process, widely used in fields such as manufacturing, construction, and energy. However, due to the complexity and variability of the welding process, various defects are likely to occur, such as pores, cracks, incomplete penetration, over-welding, etc. These defects are often difficult to detect by the naked eye. Traditional manual inspection methods are easily affected by subjective factors, and have low detection efficiency and high costs. With the development of computer vision and deep learning technologies, welding defect detection methods based on generative adversarial networks (GANs) have gradually received attention. GANs have achieved remarkable results in fields such as image generation, data augmentation, and defect detection. Therefore, it is necessary to study its application in the field of welding defect detection.

[0003] In the prior art, the welding defect detection of industrial products can meet certain requirements, but there are also certain potential risks. Specifically, on the one hand, in the prior art, ultrasonic signals are often emitted to the welding part and the reflected waves are received, and whether there are defects is judged according to the changes in the reflected signals. However, this ultrasonic detection method has a limited detection range, is applicable to certain types of defects, cannot be widely applied to all welding defects, has strict detection conditions, requires a flat welding surface without factors such as coatings that affect signal propagation, and it is difficult to meet the actual detection requirements of industrial products with different welding defects under different conditions in actual production, resulting in inaccuracy in the welding defects of industrial products and difficulty in meeting the actual industrial product defect requirements.

[0004] On the other hand, there are also some methods in the prior art that use computer vision to detect welding defects of industrial products. However, the anti-reconstruction of the images of industrial products with defects after repair under different welding parameters is ignored, resulting in difficulty in providing the repaired defect images of industrial products as a comparison reference for evaluating the repair effect of the repaired industrial products, causing non-uniformity in evaluating the defect repair effect of the target industrial products and difficulty in meeting the detection and evaluation of product defects in actual industrial production. Summary of the Invention

[0005] The purpose of the present invention is to provide an automated system for industrial defect detection based on a generative adversarial network, which solves the problems in the background art.

[0006] To solve the above technical problems, the present invention provides an industrial defect detection automation system based on a generative adversarial network. The system includes: a data acquisition and preprocessing module, which is used to extract the welding images and welding parameters of historical industrial products from a database, perform standardized processing on the welding images of historical industrial products, and label the welding images of historical industrial products with various defect types;

[0007] A model setting and training module, which uses a generative adversarial network model to set the generator and discriminator, and trains the generative adversarial network model according to the preprocessed welding images and welding parameters of historical industrial products to optimize the parameters of the generator and discriminator;

[0008] A welding defect detection module, which is based on the welding parameters and defect types of the target industrial product as the conditional vector of the generative adversarial network model to generate a defect image of the target industrial product, and compares it with the actual welding image of the target industrial product to obtain the defect detection result of the target industrial product;

[0009] A welding defect response module, which is based on the defect result of the target industrial product. If the target industrial product has a defect, it reconstructs the defect repair image of the target industrial product through the generative adversarial network model based on the welding parameters and defect types of the target industrial product, and uses it as a reference standard for evaluating the defect repair effect of the target industrial product to evaluate the defect repair effect of the target industrial product;

[0010] A database, which is used to store the welding images and welding parameters of historical industrial products, store the feature image slices of various defect types, store the initial parameter values and learning rates of the generator, and store the initial parameter values and learning rates of the discriminator.

[0011] Compared with the prior art, the advantages of the present invention are as follows: First, the present invention uses a generative adversarial network model, and the generator and discriminator play against each other through adversarial training. In this process, the generator and discriminator are continuously optimized. Eventually, the generator can generate more and more real welding defect images that conform to the conditional vector, while the discriminator can more accurately distinguish between the actual welding image and the generated defect image.

[0012] Second, the present invention generates a defect image of the target industrial product according to the welding parameters and various defect types of the target industrial product, and then compares it with the actual welding image of the target industrial product, which can realize the detection of welding defects of the target industrial product under different welding conditions, avoiding the risk that some defect types are difficult to detect in the prior art, ensuring both the high efficiency of the welding defect detection of the target industrial product and meeting the scientific nature of the welding defect detection of the target industrial product.

[0013] Thirdly, the present invention considers various parameters of the welding of the target industrial product, generates a repaired defect image of the target industrial product according to the generator, and then compares it with the actual welding image of the repaired target industrial product, providing a reliable basis for the evaluation of the repair effect of the target industrial product, thereby providing a perfect automated system for the welding defect detection of the target industrial product. Brief Description of the Drawings

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0015] Figure 1 It is the system structure connection diagram of the present invention. Detailed Embodiments

[0016] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0017] Refer to Figure 1 As shown, the present invention provides an automated system for industrial defect detection based on a generative adversarial network. The system includes: a data acquisition and preprocessing module, which is used to extract the welding images and various welding parameters of historical industrial products from the database, perform standardization processing on the welding images of historical industrial products, and label various defect types of the welding images of historical industrial products.

[0018] In a specific embodiment of the present invention, the various welding parameters of the industrial products are specifically: the temperature during the welding process collected by a temperature sensor, the voltage during the welding process collected by a voltage sensor, and the current during the welding process collected by a current sensor.

[0019] Based on the collected temperature, voltage, and current during the welding process, various welding parameters of industrial products are constituted.

[0020] It should be noted that in a specific embodiment, the influence of the current during the welding process on the welding defects of industrial products is as follows: too large or too small current will cause welding quality problems. Too large current will cause the weld seam to be too wide, the molten pool to be too large, and even cause burn-through. Too small current may result in incomplete welding and an insecure weld seam.

[0021] The influence of voltage during the welding process on welding defects of industrial products is as follows: The welding voltage determines the length of the arc, which in turn affects the shape of the molten pool. Too high or too low voltage will affect the stability of the welding process. High voltage may lead to excessive arc light and poor control of the molten pool, resulting in irregular welds. Too low voltage may lead to unstable arcs and small molten pools, causing poor welding.

[0022] The influence of temperature during the welding process on welding defects of industrial products is as follows: Excessive welding temperature poses a risk of overheating the heat-affected zone, resulting in adverse changes in the microstructure of the material and a decrease in toughness. When the material is overheated, grain coarsening may occur, affecting the mechanical properties of the material, especially reducing the tensile strength and fatigue strength of the welded joint. When the welding temperature is too low, the fusion between the welding metal and the base material is poor, resulting in a risk of incomplete fusion areas in the weld, and defects such as pores and inclusions appear. At low temperatures, the molten pool cannot flow effectively and expand sufficiently, resulting in a reduction in the strength and reliability of the welded joint.

[0023] In a specific embodiment of the present invention, for the standardization processing of welding images of historical industrial products, the specific analysis method is as follows: A1: Scale the welding images of historical industrial products to a fixed size.

[0024] A2: Use a Gaussian kernel to smooth the welding images of industrial products to remove high-frequency noise.

[0025] It should be noted that in a specific embodiment, through the built-in method of Opencv, the Gaussian kernel is convolved with the welding images of industrial products in two dimensions for smoothing processing to obtain the welding images of industrial products with high-frequency noise removed.

[0026] It should be noted that in a specific embodiment, the bicubic interpolation calculation method is adopted, and cubic polynomials are used for calculation in both the horizontal and vertical directions. For a certain pixel position in the target image, a set of intermediate values are first calculated through horizontal interpolation, and then interpolation is performed in the vertical direction to finally obtain the value of the target pixel. Scale the welding images of historical industrial products to a fixed size, and the specific method steps are as follows:

[0027] A1.1: Assume that we have a pixel in the scaled welding image of industrial products where , represents the abscissa of the pixel in the scaled welding image of industrial products, represents the ordinate of the pixel in the scaled welding image of industrial products, and the corresponding pixel in the unscaled welding image of industrial products , where , represents the abscissa of the pixel in the welding image of each industrial product before scaling, represents the ordinate of the pixel in the welding image of each industrial product before scaling, represents the maximum value of the pixel coordinates in the welding image of each industrial product before scaling, that is represents the size of the welding image of each industrial product before scaling.

[0028] A1.2: When calculating the pixel value of the welding image of each industrial product after scaling, first interpolate the surrounding pixels in the horizontal direction. Assume that we have 4 adjacent pixel points in the row: , using the formula of the cubic polynomial: , calculate the horizontal interpolation of the pixel point in the th column of the th row in the welding image of each industrial product before scaling, where represents the partial derivative of with respect to , represents the horizontal distance between the pixel point in the th row and

[0029] A1.3: Through the formula: , , , calculate the horizontal interpolation coefficient of each pixel point in the welding image of each industrial product before scaling.

[0030] A1.4: According to the analysis method of the horizontal interpolation of the pixel point in the th row and th column in the welding image of each industrial product before scaling, assume that there are 4 adjacent pixel points in the th column in the welding image of each industrial product before scaling, and perform analysis to obtain the vertical interpolation of the pixel point in the th row and th column in the welding image of each industrial product before scaling.

[0031] A1.5: Based on the horizontal interpolation of the pixel point in the th row and , the vertical interpolation of the pixel at the th row and the th column in the welding image of each industrial product before scaling , through the formula: , calculate the pixels in the welding image of each industrial product after scaling , according to the analysis method of obtaining the pixels in the welding image of each industrial product after scaling, analyze each pixel point in the welding image of each industrial product before scaling, and obtain the welding image of each industrial product scaled to a fixed size.

[0032] In a specific embodiment of the present invention, the annotation of each defect type for the welding images of historical industrial products is as follows: by extracting the characteristic image slices of each defect type from the database, and through image recognition technology, sequentially compare the standardized welding images of historical industrial products with the characteristic image slices of each defect type to obtain the defective areas in the standardized welding images of historical industrial products, and extract the defect types corresponding to the defective areas in the standardized welding images of historical industrial products, so as to annotate each defect type for the welding images of historical industrial products;

[0033] The said each defect type includes: porosity, crack, incomplete penetration, overwelding.

[0034] It should be noted that in a specific embodiment, the standardized welding images of historical industrial products are annotated with each defect type. For example, the porosity defect in the standardized welding images of historical industrial products is marked as JR1, the crack defect in the standardized welding images of historical industrial products is marked as JR2, the incomplete penetration defect in the standardized welding images of historical industrial products is marked as JR3, and the overwelding defect in the standardized welding images of historical industrial products is marked as JR4, so as to annotate each defect type for the standardized welding images of historical industrial products. The annotated standardized welding images of industrial products are recorded as the training data set.

[0035] After preprocessing the welding images of historical industrial products, the present invention obtains the annotated welding images of industrial products, providing an initial data set for the training of the subsequent generative adversarial network model. By using image recognition technology and based on the characteristic image slices of each defect type, the standardized welding images of historical industrial products are annotated with each defect type, improving the annotation efficiency of the welding images of historical industrial products, thus ensuring the stability and efficiency of the industrial defect automatic monitoring system based on the generative adversarial network.

[0036] The model setting and training module uses a generative adversarial network model to set up the generator and discriminator, and trains the generative adversarial network model according to the preprocessed welding images and welding parameters of historical industrial products, so as to optimize the parameters of the generator and discriminator.

[0037] In a specific embodiment of the present invention, the generator and the discriminator are specifically: the generator is used to generate a defect image similar to the true defect type of the industrial product and a defect repair image similar to the defect-free industrial product according to the welding parameters and the defect type as conditional vectors.

[0038] Set the initial parameters of the generator to and set the learning rate of the generator to where represents the value of the initial parameters of the generator extracted from the database, represents the value of the learning rate of the generator extracted from the database.

[0039] The discriminator is used to determine whether the defect images and defect repair images of industrial products generated by the generator are realistic, and to judge whether the industrial products have defects and extract the defect types of the defective industrial products.

[0040] Set the initial parameters of the discriminator to and set the learning rate of the discriminator to , where represents the value of the initial parameters of the discriminator extracted from the database, represents the value of the learning rate of the discriminator extracted from the database.

[0041] It should be noted that in a specific embodiment, the generative adversarial network model specifically refers to: the generative adversarial network (GAN) is a deep learning model composed of two parts: a generator (Generator) and a discriminator (Discriminator). These two parts are jointly trained in an adversarial manner. The generator attempts to generate realistic data, while the discriminator attempts to distinguish between the generated data and the real data. This adversarial training process enables the generator to generate samples that increasingly approximate the real data.

[0042] The core idea of the generative adversarial network is to train the model through a zero-sum game process: the generator and the discriminator compete against each other until the generator can generate samples that are realistic enough for the discriminator to accurately distinguish between true and false data. Specifically, the generator hopes to generate fake data similar to the real data by continuously adjusting its parameters, while the discriminator hopes to be able to distinguish between real data and generated data.

[0043] In a specific embodiment of the present invention, the training of the generative adversarial network model is performed to optimize the parameters of the generator and the discriminator. The specific analysis method is as follows: Each parameter type and the welding parameter combinations of each product in the historical industry are respectively combined to form a set of conditional vectors, which are input into the generator to obtain each set of generated images.

[0044] Each set of generated images and the labeled welding images of each product in the historical industry are respectively input into the discriminator to obtain the results of judging whether each set of generated images is a real image.

[0045] Based on the results of judging whether each set of generated images is a real image, optimization measures for the parameters of the generator and the discriminator are implemented.

[0046] The optimization steps of the generator include: B1: Denote a set of welding parameters and defect types as a conditional vector , and obtain a set of generated images through the generator, denoted as .

[0047] B2: Through the output result of the discriminator, obtain the probability that a set of generated images is predicted as real data .

[0048] B3: By minimizing the loss function of the generator, calculate the loss of the generator , and update the parameters of the generator through an optimization algorithm .

[0049] It should be noted that in a specific embodiment, the loss function of the generator is: , where represents the total number of conditional vectors composed of each welding parameter and defect type, represents the mathematical expectation of each group of input conditional vectors.

[0050] It should be noted that in a specific embodiment, the generator trains the discriminator by minimizing such that tends to 1, is the loss for the generator to make close to 1.

[0051] It should be noted that in a specific embodiment, the specific method for updating the parameters of the generator through the optimization algorithm is: B3.1: Through the backpropagation algorithm formula: , calculate the gradient of the loss of the generator with respect to the generator parameter , where represents the identifier of the partial derivative, represents the loss of the generator with respect to the generator parameter The identifier for partial derivative calculation.

[0052] B3.2: Through the stochastic gradient descent optimization algorithm: , obtain the updated parameters of the generator .

[0053] It should be noted that in a specific embodiment, the learning rate of the generator is artificially set by a technician according to the stochastic gradient descent optimization algorithm and the actual working effect of the generator, and is stored in the database.

[0054] The optimization steps of the discriminator are as follows: C1: Denote a set of welding images of historical industrial products as , input them into the discriminator, and obtain the probability that a set of welding images of each historical industrial product is predicted as real data .

[0055] C2: Based on the probability that a set of generated images is predicted as real data , calculate and The total cross-entropy loss , and update the parameters of the discriminator through the optimization algorithm .

[0056] It should be noted that in a specific embodiment, the calculation of and The total cross-entropy loss is specifically as follows: , where represents each group of welding images of historical industrial products input into the discriminator, represents the mathematical expectation of each group of welding images of historical industrial products input into the discriminator.

[0057] It should be noted that in a specific embodiment, represents the prediction of the discriminator for real data. For each real sample input, the goal of the discriminator is to make as close to 1 as possible, indicating that the probability of it judging this sample as a real sample is the largest, and the loss is . When is closer to 1, the loss is smaller.

[0058] represents the prediction of the discriminator for generated data. For each generated sample input, the goal of the discriminator is to make as close to 0 as possible, indicating that the probability of it judging this sample as a false sample is the largest, and the loss is . When is closer to 0, the loss is smaller.

[0059] It should be noted that in a specific embodiment, the parameters of the discriminator are updated by optimizing the algorithm. The specific method is as follows: C3.1: Through the backpropagation algorithm formula: , calculate the gradient of the loss of the discriminator with respect to the parameters of the discriminator where represents the loss of the discriminator and is the identifier for taking the partial derivative of the discriminator parameters

[0060] C3.2: Through the stochastic gradient descent optimization algorithm: , obtain the updated parameters of the discriminator .

[0061] It should be noted that in a specific embodiment, the learning rate of the discriminator is manually set by a technician according to the stochastic gradient descent optimization algorithm and the actual working effect of the discriminator, and is stored in the database.

[0062] The present invention adopts a generative adversarial network model, and the generator and the discriminator play against each other through adversarial training. In this process, the generator and the discriminator are continuously optimized. Finally, the generator can generate more and more real welding defect images that conform to the conditional vector, while the discriminator can more accurately distinguish the actual welding image from the generated defect image.

[0063] The welding defect detection module, based on the welding parameters and defect types of the target industrial product, serves as the conditional vector of the generative adversarial network model to generate the defect image of the target industrial product, and compares it with the actual welding image of the target industrial product to obtain the defect detection result of the target industrial product.

[0064] In a specific embodiment of the present invention, the specific analysis method for generating the defect image of the target industrial product is as follows: sequentially taking pores, cracks, incomplete penetration, and overwelding as defect types, and combining the welding parameters of the target industrial product to form four groups of conditional vectors , where , represents the number of each group of conditional vectors.

[0065] It should be noted that in a specific embodiment, the specific method for forming the four groups of conditional vectors is as follows: for example, taking pores, cracks, incomplete penetration, and overwelding as respectively, and taking the temperature during the welding process in the welding parameters of the target industrial product as , the voltage during the welding process as and the current during the welding process as .

[0066] The welding parameters and defect types of the target industrial product are formed into four groups of conditional vectors, namely: , , , .

[0067] The four groups of conditional vectors are sequentially input into the generator to obtain four groups of defect images of the target industrial product .

[0068] In a specific embodiment of the present invention, the method for specifically analyzing the obtained defect detection result of the target industrial product is as follows: F1: Obtain the actual welding image of the target industrial product through an industrial camera, denoted as .

[0069] F2: Sequentially input the four groups of defect images of the target industrial product and the actual welding image of the target industrial product into the discriminator.

[0070] F3: Determine whether the defect features in the four groups of defect images of the target industrial product exist in the actual welding image of the target industrial product . If they exist, it is determined that the target industrial product has defects, and at the same time, the defect type of the target industrial product is extracted. If they do not exist, it is determined that the target industrial product has no defects.

[0071] F4: Output the defect detection result of the target industrial product.

[0072] It should be noted that in a specific embodiment, the defect detection result of the target industrial product includes: whether the target industrial product has welding defects. If there are welding defects, the defect type of the target industrial product is extracted, that is, including one of porosity, crack, incomplete penetration, and overwelding, denoted as , where , represents the number of each defect type.

[0073] The present invention generates defect images of the target industrial product according to the welding parameters and various defect types of the target industrial product, and then compares them with the actual welding image of the target industrial product, which can realize the detection of welding defects of the target industrial product under different welding conditions, avoid the risk that it is difficult to detect some defect types in the prior art, ensure both the high efficiency of the welding defect detection of the target industrial product and meet the scientific nature of the welding defect detection of the target industrial product.

[0074] Welding defect response module. Based on the defect results of the target industrial product, if there are defects in the target industrial product, then based on the welding parameters and defect types of the target industrial product, a defect repair image of the target industrial product is reconstructed through a generative adversarial network model, which is used as a reference standard for evaluating the welding defect repair of the target industrial product to evaluate the defect repair effect of the target industrial product.

[0075] In a specific embodiment of the present invention, the specific analysis method for reconstructing the defect repair image of the target industrial product is as follows: Based on the welding parameters of the target industrial product, the defect type is set as a null value to form an inverted condition vector , which is input into the generator to obtain the defect repair image of the target industrial product .

[0076] In a specific embodiment of the present invention, the specific analysis method for evaluating the defect repair effect of the target industrial product is as follows: S1: Take an actual welding image of the repaired target industrial product through an industrial camera .

[0077] S2: Input the defect repair image of the target industrial product and the actual welding image of the repaired target industrial product into the discriminator.

[0078] S3: Obtain the output eigenvalue of the discriminator. If the output eigenvalue is 1, it is determined that the repair effect of the target industrial product is qualified; if the output eigenvalue is 0, it is determined that the repair effect of the target industrial product is unqualified.

[0079] The present invention considers the welding parameters of the target industrial product, generates a repaired defect image of the target industrial product according to the generator, and then compares it with the actual welding image of the repaired target industrial product, providing a reliable basis for evaluating the repair effect of the target industrial product, thereby providing a perfect automated system for the welding defect detection of the target industrial product.

[0080] Database, used to store the welding images and welding parameters of historical industrial products, store the characteristic image slices of each defect type, store the initial parameter values and learning rates of the generator, and store the initial parameter values and learning rates of the discriminator.

[0081] The above content is only an example and explanation of the concept of the present invention. Those skilled in the art of this technology make various modifications or supplements to the described specific embodiments or use similar methods to replace them. As long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, they should all fall within the protection scope of the present invention.

Claims

1. An automated industrial defect detection system based on generative adversarial networks, characterized in that: The system comprises: The data acquisition and preprocessing module is used to extract the welding images and welding parameters of various historical industrial products from the database, perform standardization processing on the welding images of various historical industrial products, and mark the welding images of various historical industrial products with various defect types; The model setting and training module uses the generative adversarial network model to set the generator and discriminator. According to the pre-processed welding images and welding parameters of various historical industrial products, the generative adversarial network model is trained to optimize the parameters of the generator and discriminator. The welding defect detection module uses the welding parameters and defect types of the target industrial product as the condition vector of the generative adversarial network model to generate a defect image of the target industrial product, compare it with the actual welding image of the target industrial product, and obtain the defect detection result of the target industrial product; The welding defect response module is based on the defect results of the target industrial product. If the target industrial product has defects, the defect repair image of the target industrial product is reconstructed through the generative adversarial network model based on the welding parameters and defect types of the target industrial product. This image is used as a reference standard for the evaluation of the welding defect repair of the target industrial product, and the defect repair effect of the target industrial product is evaluated.

2. The industrial defect detection automation system based on generative adversarial network according to claim 1 is characterized in that: The welding parameters of the industrial products are as follows: The temperature during welding collected by the temperature sensor, the voltage during welding collected by the voltage sensor, and the current during welding collected by the current sensor; Based on the collected temperature, voltage and current during the welding process, the welding parameters of various industrial products are constructed.

3. The industrial defect detection automation system based on generative adversarial network according to claim 1 is characterized in that: The specific analysis method of the welding image standardization processing of various historical industrial products is as follows: A1: Scale welding images of historical industrial products to Fixed size; A2: Use Gaussian kernel to smooth the welding images of various industrial products and remove high-frequency noise.

4. The industrial defect detection automation system based on generative adversarial network according to claim 3 is characterized in that: The specific analysis method for marking each defect type on the welding images of each product of the historical industry is as follows: By extracting characteristic image slices of each defect type from the database, and using image recognition technology, sequentially comparing the standardized welding images of each industrial product of the historical industry with the characteristic image slices of each defect type, the defective areas in the standardized welding images of each industrial product of the historical industry are obtained, and the defect types corresponding to the defective areas in the standardized welding images of each industrial product of the historical industry are extracted, and the welding images of each product of each industry in the history are marked with each defect type; The defect types include: pores, cracks, incomplete penetration, and excessive welding.

5. The industrial defect detection automation system based on generative adversarial network according to claim 4 is characterized in that: The generator and discriminator are specifically: The generator is used to generate a defect image similar to the real defect type of the industrial product and a defect repair image similar to the defect-free industrial product based on various welding parameters and defect types as condition vectors; Set the initial parameters of the generator to , the learning rate of the generator is set to ,in represents the value of the generator's initial parameters extracted from the database, represents the value of the learning rate of the generator extracted from the database; The discriminator is used to determine whether the defect image and defect repair image of the industrial product generated by the generator are realistic, and to determine whether the industrial product has defects and extract the defect type of the defective industrial product; Set the initial parameters of the discriminator to , the learning rate of the discriminator is set to ,in represents the value of the initial parameters of the discriminator extracted from the database, Represents the value of the discriminator learning rate extracted from the database.

6. The industrial defect detection automation system based on generative adversarial network according to claim 5 is characterized in that: The training of the generative adversarial network model is performed to optimize the parameters of the generator and the discriminator. The specific analysis method is as follows: Combine the welding parameters of each parameter type and each product of the historical industry to form a set of conditional vectors , input into the generator, and obtain each group of generated images; Input each group of generated images and the annotated welding images of historical industrial products into the discriminator to obtain the result of judging whether each group of generated images is a real image; Based on the results of judging whether each group of generated images is a real image, optimization measures for the parameters of the generator and the discriminator are implemented; The optimization steps of the generator include: B1: Recording a set of welding parameters and defect types as condition vectors , a set of generated images is obtained through the generator, denoted as ; B2: The output of the discriminator is used to obtain the probability that a set of generated images are predicted to be real data. ; B3: By minimizing the loss function of the generator, the loss of the generator is calculated , through the optimization algorithm, update the parameters of the generator ; The optimization steps of the discriminator are: C1: a set of historical industrial product welding images are recorded as , input the discriminator, and obtain the probability that a set of welding images of historical industrial products are predicted to be real data ; C2: The probability of predicting a set of generated images as real data ,calculate and The total cross entropy loss is , through the optimization algorithm, update the parameters of the discriminator .

7. The industrial defect detection automation system based on generative adversarial network according to claim 6 is characterized in that: The specific analysis method of generating the defect image of the target industrial product is as follows: Porosity, cracks, incomplete penetration, and excessive welding are taken as defect types in turn, and combined with the welding parameters of the target industrial products to form four sets of conditional vectors. ,in , Indicates the number of each group of conditional vectors; The four sets of conditional vectors Input into the generator in sequence to obtain four groups of defect images of target industrial products .

8. The industrial defect detection automation system based on generative adversarial network according to claim 7 is characterized in that: The defect detection result of the target industrial product is obtained, and the specific analysis method is as follows: F1: The actual welding image of the target industrial product is obtained through an industrial camera, denoted as ; F2: Four groups of defect images of target industrial products are sequentially and actual welding images of target industrial products Input to the discriminator; F3: Determine the actual welding image of the target industrial product Are there defective images of four groups of target industrial products in If the defect features exist, it is determined that the target industrial product has defects, and the defect type of the target industrial product is extracted. If not, it is determined that the target industrial product does not have defects; F4: Output the defect detection results of the target industrial product.

9. The industrial defect detection automation system based on generative adversarial network according to claim 8 is characterized in that: The defect repair image of the target industrial product is reconstructed, and the specific analysis method is as follows: Based on the welding parameters of the target industrial product, the defect type is set to a null value to form an inverse condition vector , input into the generator, and obtain the defect repair image of the target industrial product .

10. The industrial defect detection automation system based on generative adversarial network according to claim 9 is characterized in that: The specific analysis method for evaluating the defect repair effect of the target industrial product is as follows: S1: Use an industrial camera to capture the actual welding image of the target industrial product after repair ; S2: Repair the defect image of the target industrial product and actual welding images of target industrial products after repair Input to the discriminator; S3: Obtain the output feature value of the discriminator. If the output feature value is 1, it is determined that the repair effect of the target industrial product is qualified. If the output feature value is 0, it is determined that the repair effect of the target industrial product is unqualified.

Images