Real-time neutron tube target film defect detection system based on AI vision

Abstract

The invention relates to the technical field of target film defect detection, and discloses a neutron tube target film defect real-time detection system based on AI vision, comprising: a data acquisition module for acquiring a real image and manufacturing process parameters of a target film surface; the theoretical texture generation module is used for generating a theoretical texture image by using the trained generative adversarial network model; the space-time comparative analysis module is used for matching the theoretical texture image with a real image, extracting a structure consistency feature, a texture direction consistency feature and an optical flow vector change feature, and positioning a mismatched region; the defect identification module is used for constructing a texture evolution trend curve, comparing the texture evolution trend curve with a preset multi-class defect evolution template curve, and identifying a defect type corresponding to the mismatched area; and the process defect backstepping module is used for predicting a process parameter deviation type which causes the mismatch area defect and generating a corresponding process adjustment strategy. According to the invention, real-time detection of neutron tube target film defects is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of target film defect detection, and particularly relates to a real-time detection system for neutron tube target film defects based on AI vision. Background Art

[0002] The neutron tube is a key device in the radiation detection system, and the quality of its internal target film directly affects the neutron response sensitivity and device stability. The quality detection of existing neutron tube target films mostly relies on manual visual inspection or traditional image processing methods for defect identification.

[0003] The manual visual inspection method has problems such as low efficiency and strong subjectivity, and it is difficult to adapt to mass production; traditional image processing methods are mainly based on static features such as edge detection and gray distribution analysis, and it is difficult to accurately identify recessive defects with complex textures, weak defects or caused by dynamic process disturbances. In addition, most existing detection methods only focus on the image itself, lack the association with process parameters in the manufacturing process, and it is difficult to locate the cause of defects and intelligently optimize manufacturing parameters, and an effective feedback control mechanism cannot be formed. Summary of the Invention

[0004] The present invention provides a real-time detection system for neutron tube target film defects based on AI vision, which solves the technical problems of low accuracy, inability to identify defects caused by process disturbances, and lack of intelligent feedback ability linked with process parameters in related technologies.

[0005] The present invention provides a real-time detection system for neutron tube target film defects based on AI vision, including: A data acquisition module for real-time acquiring the real image of the target film surface and manufacturing process parameters during the manufacturing process of the neutron tube target film; A theoretical texture generation module for generating a theoretical texture image corresponding to the current manufacturing process parameters by using a trained generative adversarial network model based on the manufacturing process parameters; A spatio-temporal contrast analysis module for frame-level matching of the theoretical texture image and the real image in time according to the process time axis recorded during the manufacturing process, establishing a spatial mapping relationship in combination with the target film manufacturing process trajectory model, and performing multi-scale difference calculation on the same region in the theoretical texture image and the real image, and extracting structural consistency features, texture direction consistency features and optical flow vector change features for locating the mismatch region; A defect identification module for receiving the mismatch region, identifying the defect type corresponding to the mismatch region by constructing a texture evolution trend curve and comparing the curve with a preset multi-class defect evolution template curve, and outputting the defect position, defect type and confidence score; A process defect reverse inference module, which is used to reverse predict the type of process parameter deviation that causes defects in the mismatch area based on the texture residual map of the mismatch area, and combine with the process-texture mapping model to generate corresponding process adjustment strategies.

[0006] Further, the generator of the generative adversarial network model takes manufacturing process parameters as input. The manufacturing process parameters include: spraying speed, spin coating speed, nozzle position, and temperature. The generator performs weighted fusion on the spatial distributions of the nozzle position and temperature through a spatio-temporal attention mechanism to generate a theoretical texture image that matches the spatial resolution of the target film manufacturing equipment. The training data set of the generative adversarial network model includes simulated texture images generated by a physical simulation platform and real images collected during the actual manufacturing process.

[0007] Further, the discriminator of the generative adversarial network model is a multi-scale discriminator, and its input includes: Multi-scale pyramids of the generated theoretical texture images and corresponding real images; The difference vector between the current manufacturing process parameters and the preset process standard parameters, and the difference vector is converted into an attention mask through a fully connected layer.

[0008] Further, the spatio-temporal contrast analysis module includes: A time axis synchronization unit, which is used to divide the theoretical texture image sequence and the real image sequence into multiple time periods according to the spraying trajectory based on the process time axis and the process time stamps recorded by the manufacturing equipment. Each time period corresponds to one reciprocating motion cycle of the nozzle, and the frame indexes of the theoretical texture image and the real image are aligned through time stamp interpolation. A spatial mapping unit, which establishes an affine transformation matrix between the theoretical texture image coordinate system and the real image coordinate system based on the parameterized equation of the spraying path in the target film manufacturing process trajectory model, and performs alignment through bilinear interpolation. A dynamic compensation unit adjusts the texture density distribution of the theoretical texture image according to the real-time change value of the spraying speed, so that the local texture density deviation between the spatially mapped theoretical texture image and the real image is less than the first preset threshold.

[0009] Further, the structural consistency feature is extracted according to the first registration region of the theoretical texture image and the real image in the spatial position. Among them, by calculating the ratio of the product of the brightness means of the theoretical texture image and the real image in the registration region to the sum of the squares of their brightness means, the first intermediate feature is obtained; by calculating the ratio of the product of the brightness standard deviations of the theoretical texture image and the real image in the registration region to the sum of the squares of their brightness standard deviations, the second intermediate feature is obtained; by calculating the ratio of the joint covariance of the theoretical texture image and the real image in the registration region to the product of their standard deviations, the third intermediate feature is obtained. Multiply the first intermediate feature, the second intermediate feature and the third intermediate feature, and take the negative logarithm to obtain the structural consistency feature; The texture direction consistency feature is extracted according to the local texture direction information of the theoretical texture image and the real image in the second registration region. Among them, the main texture directions of the theoretical texture image and the real image in the second registration region are respectively extracted by the structure tensor method, and the absolute value of the angle difference between the two is calculated to obtain the fourth intermediate feature. Map the fourth intermediate feature through the cosine function to obtain the texture direction consistency feature; The optical flow vector change feature is extracted according to the local motion trend between consecutive frames of the theoretical texture image and the real image. Among them, the inter-frame optical flow vectors are respectively calculated for the theoretical texture image and the real image to obtain the theoretical texture optical flow vector and the real optical flow vector at the first target position. The ratio of the difference between the theoretical texture optical flow vector and the real optical flow vector to the modulus of the theoretical texture optical flow vector is used to obtain the optical flow vector change feature.

[0010] Further, the positioning of the mismatch region is based on the structural consistency feature, the texture direction consistency feature and the optical flow vector change feature. The positioning process includes: For each image block of the theoretical texture image and the real image, calculate three feature values respectively, and construct a three-channel texture difference feature map; According to the preset feature deviation weight coefficient, perform weighted fusion on the three features to obtain the texture deviation score; Mark the image blocks with the texture deviation score greater than the second preset threshold as mismatch regions.

[0011] Further, the defect recognition module constructs a texture evolution trend curve based on the image sequence extracted from the mismatch region. The texture evolution trend curve is based on the temporal features of the texture direction entropy and the contrast of the gray-level co-occurrence matrix, and uses the spraying speed as the weight coefficient for weighted modeling. The calculation formula of the texture evolution trend curve is: , where S represents the function value of the texture evolution trend curve, represents the spraying speed, represents the preset standard speed, represents the texture direction entropy of the current frame, t represents the index of the frame, Represents the contrast of the gray-level co-occurrence matrix of the current frame, and respectively represent the maximum values of the texture direction entropy and the contrast of the gray-level co-occurrence matrix in the historical data.

[0012] Furthermore, the texture evolution trend curve of the mismatch region and the multi-class defect evolution template curves are respectively calculated for the dynamic time warping distance to obtain the matching distance. If the matching distance is less than the third preset threshold and the corresponding confidence score is greater than the fourth preset threshold, it is determined that the mismatch region is the corresponding defect type, where the confidence score is constructed based on the matching distance, and the defect types include: uneven sputtering, texture fracture, poor adhesion, and particle impurity deposition.

[0013] Furthermore, the process defect backtracking module includes: a residual map segmentation unit, a parameter deviation calculation unit, and a deviation type judgment unit; The residual map segmentation unit is used to divide the texture residual map of the mismatch region into multiple sub-regions according to the manufacturing process trajectory, and each sub-region corresponds to a control stage of the spraying path, where the texture residual map is obtained by performing pixel-by-pixel difference calculation on the mismatch region; The parameter deviation calculation unit is used to reverse-derive the process parameter deviation amount that causes the abnormality based on the texture residual map of the mismatch region, including: extracting the texture residual values for each sub-region and calculating the deviation values of each manufacturing process parameter in combination with the sensitivity coefficients in the process-texture mapping model; The deviation type determination unit is used to classify and judge the deviation amount of the manufacturing process parameters. When any deviation amount exceeds the fifth preset range, it is determined that the manufacturing process parameter is abnormal.

[0014] The beneficial effects of the present invention are as follows: By constructing a spatio-temporal registration model of the theoretical texture image and the real image and integrating multiple features such as structural consistency, direction consistency, and optical flow change, the present invention realizes high-precision identification of target film defects. At the same time, by introducing a process-texture mapping model, it realizes reverse reasoning of process parameter deviation based on the texture residual map, and then generates a process optimization strategy, with the advantages of accurate defect identification, clear abnormal cause tracing, and intelligent manufacturing control, significantly improving the manufacturing quality and stability of the target film. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a schematic diagram of the modules of a real-time detection system for target film defects of a neutron tube based on AI vision according to the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0016] Reference will now be made to exemplary embodiments to discuss the subject matter described herein. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein. Without departing from the scope of protection of the content of this specification, changes can be made to the functions and arrangements of the elements discussed. Each example may omit, substitute, or add various processes or components as needed. Additionally, the features described relative to some examples can also be combined in other examples.

[0017] It should be noted that unless otherwise defined, the technical terms or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by those of ordinary skill in the art to which the present invention pertains. The "first", "second", and similar terms used in one or more embodiments of the present invention do not denote any order, quantity, or importance, but are only used to distinguish different components. Words such as "including" or "comprising" mean that the elements or objects appearing before this word cover the elements or objects listed after this word and their equivalents, without excluding other elements or objects. Words such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Upper", "lower", "left", "right", etc. are only used to represent relative position relationships. When the absolute position of the object being described changes, the relative position relationship may also change accordingly.

[0018] As Figure 1 shown, a real-time detection system for defects in the neutron tube target film based on AI vision includes: A data acquisition module 101 for acquiring real-time images of the target film surface and manufacturing process parameters during the manufacturing process of the neutron tube target film; A theoretical texture generation module 102 for generating a theoretical texture image corresponding to the current manufacturing process parameters using a trained generative adversarial network model based on the manufacturing process parameters; A spatio-temporal contrast analysis module 103 for frame-level matching of the theoretical texture image and the real image in terms of time according to the process timeline recorded during the manufacturing process, establishing a spatial mapping relationship in combination with the target film manufacturing process trajectory model, and calculating multi-scale differences for the same region in the theoretical texture image and the real image, extracting structural consistency features, texture direction consistency features, and optical flow vector change features for locating the mismatched region; A defect recognition module 104 for receiving the mismatched region, identifying the defect type corresponding to the mismatched region by constructing a texture evolution trend curve and comparing this curve with a preset multi-class defect evolution template curve, and outputting the defect position, defect type, and confidence score; The process defect reverse inference module 105 is used to reverse predict the type of process parameter deviation that causes defects in the mismatch area based on the texture residual map of the mismatch area, and generate a corresponding process adjustment strategy.

[0019] In an embodiment of the present invention, the generative adversarial network model includes: a generator and a discriminator; The generator of the generative adversarial network model takes manufacturing process parameters as input. The manufacturing process parameters include: spraying speed, spin coating speed, nozzle position, and temperature. The generator performs weighted fusion on the spatial distribution of the nozzle position and temperature through a spatio-temporal attention mechanism to generate a theoretical texture image that matches the spatial resolution of the target film manufacturing equipment. Among them, the spraying speed is calculated by combining the feedback signal of the click controller with a linear encoder installed on the spraying slide rail, and the unit of the spraying speed is mm / s; the spin coating speed is collected by the drive controller of the spin coating platform, and the unit of the spin coating speed is rps; the nozzle position is obtained by acquiring the coordinates of the nozzle in three-dimensional space; the temperature is collected by arranging a thermocouple array in the target film spraying area to obtain the temperature in the manufacturing area. The training data set of the generative adversarial network model includes simulated texture images generated by a physical simulation platform and real images collected during the actual manufacturing process.

[0020] In an embodiment of the present invention, during the model training stage, simulation images of the target film spraying process generated by numerical simulation platforms such as COMSOL are introduced to construct a texture evolution simulation diagram containing various typical process parameter combinations; it is combined with the images of the actual production process to form a mixed data set for training a conditional generative adversarial network, so that the generated theoretical texture image has higher reliability in terms of physical consistency and structural fidelity.

[0021] In an embodiment of the present invention, the discriminator of the generative adversarial network model is a multi-scale discriminator, and its input includes: Multi-scale pyramids of the generated theoretical texture image and the corresponding real image. Specifically, the generated theoretical texture image and the real image are respectively downsampled by a Gaussian pyramid to generate image pyramids with different resolutions, which are respectively input into the sub-discriminators of the corresponding scales to achieve multi-level feature extraction from the global structure to local details; The difference vector between the current manufacturing process parameters and the preset process standard parameters. The difference vector is converted into an attention mask through a fully connected layer to enhance the sensitivity of the discriminator to the process parameter deviation area. Specifically, calculate the difference vector between the current process parameters and the preset process standard parameters, input the difference vector into the fully connected layer, and then convert it into a spatial attention map through a Reshape operation to generate an attention mask.

[0022] In one embodiment of the present invention, the spatio-temporal contrast analysis module includes: A time axis synchronization unit, which divides the theoretical texture image sequence and the real image sequence into multiple time periods according to the process timestamps of the manufacturing equipment along the spraying trajectory. Each time period corresponds to one reciprocating motion cycle of the nozzle, and aligns the frame indices of the theoretical texture image and the real image through timestamp interpolation; A spatial mapping unit, which precisely registers the theoretical texture image and the real image spatially, constructs a parametric equation of the spraying path based on the target film manufacturing process trajectory, and establishes an affine transformation matrix from the theoretical texture image coordinate system to the real image coordinate system. To avoid interpolation errors, bilinear interpolation is used to reconstruct the gray values of non-integer pixel points during the coordinate mapping process to obtain the mapped theoretical texture image; Among them, the parametric equation of the spraying path is expressed as: , represents the position of the nozzle, t is the process timestamp, and respectively represent the coordinate positions of the nozzle in the x and y axis directions at time t; By extracting the nozzle position, spraying direction, and scanning length recorded by the manufacturing equipment, an affine transformation matrix from the theoretical image coordinate system to the real image coordinate system is constructed and expressed in the following form: , where M is a 3×3 affine transformation matrix, a and b are the coordinate positions of the nozzle position in the x and y axis directions of the real image coordinate system, and u and v are the coordinate positions of the nozzle position in the x and y axis directions of the theoretical image coordinate system; A dynamic compensation unit adjusts the texture density distribution of the theoretical texture image according to the real-time change value of the spraying speed, so that the local texture density deviation between the spatially mapped theoretical texture image and the real image is less than the first preset threshold, reduces the misregistration rate of the image texture caused by process fluctuations, and improves the contrast accuracy.

[0023] In one embodiment of the present invention, the structural consistency feature is extracted according to the first registration region of the theoretical texture image and the real image in the spatial position. Among them, by calculating the ratio of the product of the brightness means of the theoretical texture image and the real image in the registration region to the sum of the squares of their brightness means, the first intermediate feature is obtained; by calculating the ratio of the product of the brightness standard deviations of the theoretical texture image and the real image in the registration region to the sum of the squares of their brightness standard deviations, the second intermediate feature is obtained, and by calculating the ratio of the joint covariance of the theoretical texture image and the real image in the registration region to the product of their standard deviations, the third intermediate feature is obtained. Multiply the first intermediate feature, the second intermediate feature, and the third intermediate feature, and take the negative logarithm to obtain the structural consistency feature; The calculation formula of the structural consistency feature is: , where, Indicates the structural consistency feature, which is used to measure the overall consistency of the theoretical texture image and the real image in the local structural form, and assist in identifying the macroscopic texture mismatch caused by the fluctuation of the spraying thickness or the deformation of the area. and respectively represent the average brightness of the theoretical texture image and the real image in the first registration area. and respectively represent the standard deviation of the brightness of the theoretical texture image and the real image in the first registration area. represents the covariance of the theoretical texture image and the real image in the first registration area. 、 and respectively represent the first constant, the second constant and the third constant, which are used to avoid the denominator being zero. The texture direction consistency feature is extracted according to the local texture direction information of the theoretical texture image and the real image in the second registration area. Among them, the main texture directions of the theoretical texture image and the real image in the second registration area are respectively extracted by the structure tensor method, and the absolute value of the included angle difference between the two is calculated to obtain the fourth intermediate feature, and the fourth intermediate feature is mapped through the cosine function to obtain the texture direction consistency feature. The calculation formula of the texture direction consistency feature is: , where represents the texture direction consistency feature, which is used to detect the deviation degree of the theoretical texture and the real texture in the main direction, and locate the directional defects caused by abnormal spin coating trajectories or nozzle yaw. and respectively represent the main texture directions of the real image and the theoretical texture image in the second registration area, which are calculated by the structure tensor method. represents and the absolute value of the difference. The optical flow vector change feature is extracted according to the local motion trend between consecutive frames of the theoretical texture image and the real image. Among them, the inter-frame optical flow vectors of the theoretical texture image and the real image are respectively calculated to obtain the theoretical texture optical flow vector and the real optical flow vector at the first target position, and the ratio of the difference between the theoretical texture optical flow vector and the real optical flow vector to the modulus length of the theoretical texture optical flow vector is used to obtain the optical flow vector change feature. The calculation formula of the optical flow vector change feature is: , where represents the optical flow vector change feature, which is used to compare the deviation degree of the dynamic behavior between consecutive frames of the theoretical texture image and the real image, and capture the motion inconsistency caused by process anomalies such as sudden deposition and adhesion failure. represents the optical flow vector between consecutive frames of the real image. Represents the optical flow vector between theoretical texture image frames. The optical flow vector refers to the motion vector of the same pixel in the image that changes with the time frame. It indicates the movement direction and speed of the pixel and is calculated by the traditional optical flow algorithm. represents a very small positive number, express The Euclidean norm of .

[0024] In one embodiment of the present invention, the positioning of the mismatching area is based on the structural consistency feature, the texture direction consistency feature and the optical flow vector change feature, and the positioning process includes: For each image block of the theoretical texture image and the real image, three eigenvalues ​​are calculated respectively, and a three-channel texture difference feature map is constructed. The three features are used as multi-dimensional descriptors of the image block. According to the preset feature deviation weight coefficient, the three features are weighted and fused to obtain the texture deviation score, which is used to indicate the comprehensive deviation degree of the image block in structure, direction and dynamic behavior; The image blocks whose texture deviation scores are greater than a second preset threshold are marked as mismatch areas.

[0025] Through this three-feature fusion judgment mechanism, the system can accurately identify the texture offset between the theoretical texture image and the real image in terms of structure, directionality and dynamic behavior, and improve the robustness and accuracy of mismatch area detection.

[0026] In one embodiment of the present invention, the defect recognition module constructs a texture evolution trend curve based on an image sequence extracted from the mismatch area. The texture evolution trend curve is based on the temporal characteristics of the texture direction entropy and the grayscale co-occurrence matrix contrast, and the spraying speed is used as a weight coefficient for weighted modeling. The calculation formula of the texture evolution trend curve is: , where S represents the function value of the texture evolution trend curve, Indicates the spraying speed, Indicates the preset standard speed in mm / s. represents the texture direction entropy of the current frame, t represents the index of the frame, represents the gray-level co-occurrence matrix contrast of the current frame, and They represent the maximum values ​​of texture direction entropy and gray-level co-occurrence matrix contrast in historical data respectively.

[0027] Different from traditional texture features which are only based on static calculation of images, the texture evolution trend curve constructed by the present invention combines the changes in spraying speed during the manufacturing process. By dynamically adjusting the relative weights of texture direction entropy and grayscale co-occurrence matrix contrast, the trend curve is more time-process consistent when capturing defect evolution patterns, thereby improving the accuracy and discrimination of defect type identification.

[0028] In one embodiment of the present invention, the texture evolution trend curve of the mismatch region and the evolution template curves of multiple types of defects are respectively calculated for the dynamic time warping distance to obtain the matching distance. If the matching distance is less than the third preset threshold and the corresponding confidence score is greater than the fourth preset threshold, it is determined that the mismatch region is the corresponding defect type. The confidence score is constructed based on the matching distance, and the defect types include: uneven sputtering, texture fracture, poor adhesion, and particle impurity deposition.

[0029] Specifically, the dynamic time warping method is used to calculate the temporal shape similarity between the texture evolution trend curve and each defect evolution template curve. The calculation formula for the matching distance is: , where DTW represents the matching distance, and min represents the operation of taking the minimum value. represents the i-th defect evolution template curve, and i represents the index of the defect evolution template curve; The calculation formula for the confidence score is: , represents the confidence score of the i-th defect evolution template. represents the mean value of the matching distances of all defect evolution templates.

[0030] In one embodiment of the present invention, the process defect backtracking module includes: a residual map segmentation unit, a parameter deviation calculation unit, and a deviation type determination unit; The residual map segmentation unit is used to divide the texture residual map of the mismatch region into multiple sub-regions according to the manufacturing process trajectory. Each sub-region corresponds to a control stage of the spraying path. The texture residual map is obtained by performing pixel-by-pixel difference calculation on the mismatch region. The parameter deviation calculation unit is used to reverse-derive the process parameter deviation amount that causes the abnormality based on the texture residual map of the mismatch region, including: extracting the texture residual values for each sub-region and calculating the deviation values of each manufacturing process parameter in combination with the sensitivity coefficients in the process-texture mapping model. Among them, the texture residual value is obtained through the pixel gray difference of the sub-region of the texture residual map. The process-texture mapping model is used to represent the relationship between manufacturing process parameters and texture, including the sensitivity coefficients of manufacturing process parameters in each sub-region. The sensitivity coefficient reflects the average response amplitude caused by the change of manufacturing process parameters to the texture direction consistency feature in the sub-region. The deviation value of the manufacturing process parameter is obtained by weighted accumulation of the texture residual values of multiple sub-regions and the corresponding sensitivity coefficients. The deviation type determination unit is used to classify and judge the deviation amount of the manufacturing process parameter. When any deviation amount exceeds the fifth preset range, it is determined that the manufacturing process parameter is abnormal. Among them, the fifth preset range is preferably set to the range of 3 times the standard deviation. The types of process parameter deviations include the deviated manufacturing process parameters and the deviation directions. For example, the types of process parameter deviations include, but are not limited to: "too high spraying speed", "too low temperature", etc. Corresponding process adjustment strategies are generated according to the types of process parameter deviations. The process adjustment strategies include, but are not limited to: reducing the spraying speed by 5% and extending the downward section time; raising the temperature of the spraying temperature zone to 580 °C and extending the constant temperature section; reducing the spraying speed and at the same time increasing the time of the uniform rotation coating section.

[0031] In an embodiment of the present invention, by dividing the mismatch region into sub-regions according to the manufacturing process trajectory and combining the sensitivity coefficients in the process-texture mapping model, the accurate backtracking of process parameter deviations is realized. This method can effectively identify the potential process anomaly types that cause image defects, improve the interpretability and control response ability of the manufacturing process, and has high intelligent diagnosis and closed-loop optimization value.

[0032] The embodiments of the present invention have been described above, but the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of this embodiment, those of ordinary skill in the art can also make many forms, all of which fall within the protection scope of this embodiment.

Claims

1. A real-time detection system for neutron tube target film defects based on AI vision, characterized in that: include: A data acquisition module is used to obtain the real image of the target film surface and manufacturing process parameters in real time during the manufacturing process of the neutron tube target film; A theoretical texture generation module, used to generate a theoretical texture image corresponding to the current manufacturing process parameters based on the manufacturing process parameters using a trained generative adversarial network model; The spatiotemporal comparison analysis module is used to match the theoretical texture image with the real image at the frame level in time according to the process timeline recorded in the manufacturing process, and to establish a spatial mapping relationship in combination with the target film manufacturing process trajectory model, and to perform multi-scale difference calculations on the same area in the theoretical texture image and the real image, extracting structural consistency features, texture direction consistency features, and optical flow vector change features for locating mismatch areas; A defect recognition module is used to receive the mismatch area, identify the defect type corresponding to the mismatch area by constructing a texture evolution trend curve, compare the curve with preset multiple types of defect evolution template curves, and output the defect position, defect type and confidence score; The process defect inference module is used to reversely predict the type of process parameter deviation that causes defects in the mismatch area based on the texture residual map of the mismatch area and the process-texture mapping model, and generate the corresponding process adjustment strategy.

2. According to the AI ​​vision-based real-time detection system for neutron tube target film defects in claim 1, it is characterized in that: The generator of the generative adversarial network model takes manufacturing process parameters as input, and the manufacturing process parameters include: spraying speed, spin coating speed, nozzle position and temperature. The generator performs weighted fusion on the spatial distribution of nozzle position and temperature through a spatiotemporal attention mechanism to generate a theoretical texture image that matches the spatial resolution of the target film manufacturing equipment; The training dataset for the generative adversarial network model includes simulated texture images generated by a physical simulation platform and real images collected during the actual manufacturing process.

3. According to the AI ​​vision-based real-time detection system for neutron tube target film defects in claim 1, it is characterized in that: The discriminator of the generative adversarial network model is a multi-scale discriminator, and its input includes: The generated theoretical texture image and the multi-scale pyramid of the corresponding real image; The difference vector between the current manufacturing process parameters and the preset process standard parameters is converted into an attention mask through a fully connected layer.

4. According to the AI ​​vision-based real-time detection system for neutron tube target film defects in claim 1, it is characterized in that: The spatiotemporal comparison analysis module includes: A time axis synchronization unit is used to divide the theoretical texture image sequence and the real image sequence into multiple time periods according to the spraying trajectory based on the process time axis and according to the process timestamp recorded by the manufacturing equipment. Each time period corresponds to a reciprocating motion cycle of the nozzle, and the frame index of the theoretical texture image and the real image are aligned by timestamp interpolation. The spatial mapping unit establishes the affine transformation matrix between the theoretical texture image coordinate system and the real image coordinate system based on the parameterized equation of the spraying path in the target film manufacturing process trajectory model, and aligns them through bilinear interpolation; The dynamic compensation unit adjusts the texture density distribution of the theoretical texture image according to the real-time change value of the spraying speed, so that the local texture density deviation between the theoretical texture image after spatial mapping and the real image is less than a first preset threshold.

5. The real-time detection system for neutron tube target film defects based on AI vision according to claim 1 is characterized in that: The structural consistency feature is extracted according to the first registration area of ​​the theoretical texture image and the real image in the spatial position, wherein the first intermediate feature is obtained by calculating the ratio of the product of the brightness mean of the theoretical texture image and the real image in the registration area to the sum of the squares of their brightness mean; the second intermediate feature is obtained by calculating the ratio of the product of the brightness standard deviation of the theoretical texture image and the real image in the registration area to the sum of the squares of their brightness standard deviations; the third intermediate feature is obtained by calculating the ratio of the joint covariance of the theoretical texture image and the real image in the registration area to the product of their standard deviations; the first intermediate feature, the second intermediate feature and the third intermediate feature are multiplied, and the negative logarithm is taken to obtain the structural consistency feature; The texture direction consistency feature is extracted according to the local texture direction information of the theoretical texture image and the real image in the second registration area, wherein the main texture directions of the theoretical texture image and the real image in the second registration area are respectively extracted by the structure tensor method, and the absolute value of the angle difference between the two is calculated to obtain a fourth intermediate feature, and the fourth intermediate feature is mapped by a cosine function to obtain a texture direction consistency feature; The optical flow vector change feature is extracted based on the local motion trend of the theoretical texture image and the real image between consecutive frames, wherein the inter-frame optical flow vector is calculated for the theoretical texture image and the real image respectively, and the theoretical texture optical flow vector and the real optical flow vector at the first target position are obtained, and the optical flow vector change feature is obtained by the ratio of the difference between the theoretical texture optical flow vector and the real optical flow vector to the modulus of the theoretical texture optical flow vector.

6. The AI ​​vision-based real-time neutron tube target film defect detection system according to claim 5, characterized in that: The positioning of the mismatched area is based on the structural consistency feature, texture direction consistency feature and optical flow vector change feature. The positioning process includes: For each image block of the theoretical texture image and the real image, three eigenvalues ​​are calculated respectively, and a three-channel texture difference feature map is constructed; According to the preset feature deviation weight coefficient, the three features are weighted and fused to obtain the texture deviation score; The image blocks whose texture deviation scores are greater than a second preset threshold are marked as mismatch areas.

7. The AI ​​vision-based real-time neutron tube target film defect detection system according to claim 2, characterized in that: The defect recognition module constructs a texture evolution trend curve based on the image sequence extracted from the mismatch area. The texture evolution trend curve is based on the temporal characteristics of the texture direction entropy and the grayscale co-occurrence matrix contrast, and the spraying speed is used as a weight coefficient for weighted modeling. The calculation formula of the texture evolution trend curve is: , where S represents the function value of the texture evolution trend curve, Indicates the spraying speed, Indicates the preset standard speed. represents the texture direction entropy of the current frame, t represents the index of the frame, represents the gray-level co-occurrence matrix contrast of the current frame, and They represent the maximum values ​​of texture direction entropy and gray-level co-occurrence matrix contrast in historical data respectively.

8. The AI ​​vision-based real-time neutron tube target film defect detection system according to claim 1, characterized in that: The texture evolution trend curve of the mismatch area and the multiple defect evolution template curves are respectively subjected to dynamic time warping distance calculation to obtain the matching distance. If the matching distance is less than the third preset threshold and the corresponding confidence score is greater than the fourth preset threshold, the mismatch area is determined to be the corresponding defect type, wherein the confidence score is constructed based on the matching distance, and the defect types include: uneven sputtering, texture fracture, poor adhesion and particle impurity deposition.

9. The AI ​​vision-based real-time neutron tube target film defect detection system according to claim 1, characterized in that: The process defect inference module includes: a residual graph segmentation unit, a parameter deviation calculation unit and a deviation type judgment unit; The residual map segmentation unit is used to divide the texture residual map of the mismatch area into a plurality of sub-areas according to the manufacturing process trajectory, each sub-area corresponds to a control stage of the spraying path, wherein the texture residual map is obtained by performing pixel-by-pixel difference calculation on the mismatch area; The parameter deviation calculation unit is used to reversely deduce the process parameter deviation causing the abnormality based on the texture residual map of the mismatch area, including: extracting the texture residual value for each sub-area, and calculating the deviation value of each manufacturing process parameter in combination with the sensitivity coefficient in the process-texture mapping model; The deviation type determination unit is used to classify and determine the deviation amount of the manufacturing process parameter, and when any deviation amount exceeds a fifth preset range, it is determined that the manufacturing process parameter is abnormal.

Images