PVC adhesive defect detection method based on AI visual recognition and intelligent sorting system

Abstract

The embodiment of the application discloses a PVC sizing material defect detection method and an intelligent sorting system based on AI visual recognition. The method comprises the following steps: collecting continuous multi-frame visual imaging data of the PVC sizing material in the production conveying process; combining production conveying state monitoring information, analyzing the position offset trajectory and the shape contour change of the PVC sizing material, and extracting inter-frame feature correlation information; generating a cross-frame defect tracking link based on the inter-frame feature correlation information, tracking the shape abnormal area, and determining a suspected surface defect area; mining surface defect features through edge contour recognition; inputting the surface defect features into a layered reasoning mining model, reversely deducing internal implicit defect types, and mining distribution trends to generate a defect detection result. The embodiment of the application realizes full-dimensional accurate identification of PVC sizing material defects from the surface to the interior.

Description

TECHNICAL FIELD

[0001] The embodiment of the present application relates to the technical field of computer vision, and particularly relates to a PVC compound defect detection method based on AI visual recognition and an intelligent sorting system. BACKGROUND

[0002] PVC compound is the core raw material for producing pipe, profile, film and other products, and defects in the production process will directly affect the mechanical properties, weather resistance and service life of the final product. Therefore, defect detection is a core quality control link in the production process of PVC compound. At present, the mainstream defect detection technology for PVC compound in the industry mainly includes manual detection and machine vision detection. Manual detection relies on quality inspectors to judge defects through visual observation and tactile feeling, and is mainly used in small-batch and low-rate production scenarios. Machine vision detection collects surface images of the compound through an industrial camera and identifies surface defects by combining image processing algorithms, and has gradually become the mainstream detection method in high-speed and large-scale production lines.

[0003] In the field of machine vision detection, the existing technology mainly adopts a single-frame image independent analysis mode, extracts the shape and gray scale of the compound in the single-frame image, and compares and identifies the surface defects with a standard template. Some technologies combine continuous frame images for defect tracking, but are mainly based on pixel matching or simple position association, and do not deeply combine the conveying state information in the production process. At the same time, the defect recognition of the existing technology is mainly based on the surface shape, and can only identify visible surface defects, and cannot inversely deduce the type and distribution trend of internal implicit defects through surface characteristics, so as to provide effective basis for deep optimization of the production process.

[0004] Therefore, the existing single-frame independent analysis is easy to be disturbed by production conveying fluctuations, and the cross-frame defect tracking has the problem of insufficient continuity. Moreover, the existing technology cannot realize the associated recognition of surface defects and internal implicit defects, and cannot meet the requirements of high-end PVC compound production for defect detection accuracy and foresight. SUMMARY

[0005] The embodiment of the present application provides a PVC compound defect detection method based on AI visual recognition and an intelligent sorting system.

[0006] The embodiment of the present application provides a PVC compound defect detection method based on AI visual recognition and an intelligent sorting system.

[0007] Collecting continuous multi-frame visual imaging data of the PVC compound in the production conveying process;

[0008] The frame-to-frame feature association information is determined based on a position offset trajectory and a shape profile change of the PVC compound in the continuous multiple frames of visual imaging data, a previous frame of compound motion feature is taken as an input constraint of a defect tracking task of a next frame, and a cross-frame defect tracking link is generated based on the time sequence association logic information.

[0009] The frame-to-frame feature association information is determined based on a position offset trajectory and a shape profile change of the PVC compound in the continuous multiple frames of visual imaging data, a previous frame of compound motion feature is taken as an input constraint of a defect tracking task of a next frame, and a cross-frame defect tracking link is generated based on the time sequence association logic information.

[0010] The frame-to-frame feature association information is determined based on a position offset trajectory and a shape profile change of the PVC compound in the continuous multiple frames of visual imaging data, a previous frame of compound motion feature is taken as an input constraint of a defect tracking task of a next frame, and a cross-frame defect tracking link is generated based on the time sequence association logic information.

[0011] The frame-to-frame feature association information is determined based on a position offset trajectory and a shape profile change of the PVC compound in the continuous multiple frames of visual imaging data, a previous frame of compound motion feature is taken as an input constraint of a defect tracking task of a next frame, and a cross-frame defect tracking link is generated based on the time sequence association logic information.

[0012] The frame-to-frame feature association information is determined based on a position offset trajectory and a shape profile change of the PVC compound in the continuous multiple frames of visual imaging data, a previous frame of compound motion feature is taken as an input constraint of a defect tracking task of a next frame, and a cross-frame defect tracking link is generated based on the time sequence association logic information.

[0013] The frame-to-frame feature association information is determined based on a position offset trajectory and a shape profile change of the PVC compound in the continuous multiple frames of visual imaging data, a previous frame of compound motion feature is taken as an input constraint of a defect tracking task of a next frame, and a cross-frame defect tracking link is generated based on the time sequence association logic information.

[0014] The frame-to-frame feature association information is determined based on a position offset trajectory and a shape profile change of the PVC compound in the continuous multiple frames of visual imaging data, a previous frame of compound motion feature is taken as an input constraint of a defect tracking task of a next frame, and a cross-frame defect tracking link is generated based on the time sequence association logic information.

[0015] The frame-to-frame feature association information is determined based on a position offset trajectory and a shape profile change of the PVC compound in the continuous multiple frames of visual imaging data, a previous frame of compound motion feature is taken as an input constraint of a defect tracking task of a next frame, and a cross-frame defect tracking link is generated based on the time sequence association logic information.

[0016] The frame-to-frame feature association information is determined based on a position offset trajectory and a shape profile change of the PVC compound in the continuous multiple frames of visual imaging data, a previous frame of compound motion feature is taken as an input constraint of a defect tracking task of a next frame, and a cross-frame defect tracking link is generated based on the time sequence association logic information.

[0017] The embodiment of the application realizes the full-dimensional accurate identification of PVC compound defects from the surface to the inside. In detail, the cooperative analysis of production state monitoring information and visual imaging data effectively eliminates the interference of production conveying rate fluctuation on feature extraction, ensures the time sequence consistency and accuracy of inter-frame feature correlation information; the constraint mechanism of the motion feature of the previous frame compound to the tracking of the next frame defect greatly improves the space-time coherence of the cross-frame defect tracking link, and avoids the defect missed detection and misjudgment caused by independent frame analysis; the hierarchical analysis and semantic mapping of surface defect features provide multi-dimensional feature support for the reverse deduction of internal implicit defects, breaking through the limitation of traditional defect judgment relying only on surface morphology; combined with the internal implicit defect distribution trend mining of inter-frame time sequence change rule, the upgrade from "static recognition" to "dynamic prediction" of defects is realized, which can perceive the diffusion range and evolution direction of defects in advance. The embodiment of the application deeply integrates visual perception, time sequence correlation and hierarchical reasoning as a whole, and improves the accuracy and foresight of PVC compound defect detection. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the following will briefly introduce the drawings needed to be used in the embodiment or related art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor based on the above drawings.

[0019] Figure 1 The flow chart of a PVC compound defect detection method based on AI visual recognition provided by the embodiment of the present application.

[0020] Figure 2 The schematic diagram of the basic structure of a PVC compound defect detection system provided by the embodiment of the present application.

[0021] Figure 3 The functional module block diagram of a PVC compound defect detection device provided by the embodiment of the present application.

[0022] Figure 4 The schematic diagram of an intelligent sorting interactive scene provided by the embodiment of the present application. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0024] Please refer toFigure 1 , Figure 1 is a flowchart of a PVC compound defect detection method based on AI visual recognition provided by an embodiment of the present application. The method can be executed by a PVC compound defect detection system, or can be executed by the PVC compound defect detection system and a server. The method can include steps 110-160.

[0025] It should be understood that when implementing the embodiments of the present application, there are various physical meanings and parameters with different dimensions. Direct operation can easily lead to no clear physical meaning or calculation deviation due to dimension mismatch. However, this is not a fundamental obstacle. Those skilled in the art can use known signal processing and data analysis techniques to perform adaptive normalization or standardization preprocessing on multi-source heterogeneous parameters, convert all parameters to a unified scale or dimensionless range, and ensure the effectiveness of subsequent feature fusion, matrix operation and model reasoning. Normalization is a routine operation in the field of machine vision and industrial detection, which can eliminate the influence of dimension difference on algorithm performance and ensure the mathematical consistency and engineering realizability of the overall process.

[0026] In specific implementation, normalization technology is required to solve the dimension problem at different links. When extracting inter-frame feature correlation information, methods such as min-max normalization are used to convert parameters such as position offset trajectory and motion rate into relative values or standardized scores, avoiding weight imbalance and calculation errors, while enhancing the robustness of the algorithm; when constructing a hierarchical reasoning model or performing defect morphology vector set matching, the surface defect features are normalized and mapped to a standard vector space according to the feature engineering criteria, avoiding model convergence problems or mis-matching, and this technology is already implicit in the design of the deep learning framework; when mining the time series defect distribution trend, the related parameters are converted through time series standardization technology to ensure the accuracy of collaborative analysis with other variables. The above all rely on conventional signal normalization methods.

[0027] Therefore, it should be noted that the potential dimension problem in the embodiments of the present application can be effectively solved by adaptive normalization. As a core technology of data preprocessing and feature engineering, normalization has been mature and popular in the field of machine vision, industrial detection, etc., covering various methods such as scale transformation and unit unification, which can realize mathematical compatibility of different parameters and support subsequent correlation analysis and decision generation. Although the normalization details are not explicitly specified in the scheme, those skilled in the art can automatically implement and adjust based on existing technology to overcome dimension errors and ensure reliable detection results and the overall feasibility of the scheme. The above potential problems do not affect the essence of the scheme, but highlight the supporting role of conventional technical means.

[0028] Step 110: Collecting continuous multiple frames of visual imaging data of the PVC compound in the production and conveying process.

[0029] In this step, the industrial line array camera deployed above the production conveying pipeline is linked with the photoelectric sensor at the starting end of the pipeline. When the PVC compound triggers the photoelectric sensor, the camera starts continuous acquisition mode, with the sampling frame rate being linked and adapted to the production conveying rate of the PVC compound, to ensure that each frame of image can cover a preset length of the PVC compound area. Each frame of image collected is stored in uncompressed BMP format, with each single frame of image being a two-dimensional pixel matrix of fixed width and height, each element in the matrix corresponding to the RGB three-channel grayscale value of a pixel point, and each frame of image being additionally attached with metadata information such as frame serial number, acquisition timestamp, camera intrinsic and extrinsic parameters, and the like. All frames of image and corresponding metadata are integrated into a continuous image data stream according to the acquisition time sequence.

[0030] Step 120: In combination with the production conveying state monitoring information of the PVC compound, frame-to-frame feature correlation information is extracted by analyzing the position offset trajectory and morphological contour change of the PVC compound in the continuous multiple frames of visual imaging data.

[0031] Exemplarily, first, the PVC compound area is segmented frame by frame from the continuous image data stream, the background difference method is used to remove the background pixels such as the pipeline conveying belt, and only the pixel area corresponding to the PVC compound is retained, and then the top vertex coordinates of the minimum circumscribed rectangle of each PVC compound area are extracted as the morphological contour features; the displacement vector of the feature points in the PVC compound area in adjacent frames is calculated by using the optical flow method, and the position offset trajectory of the PVC compound between frames is fitted; the production conveying state monitoring information is synchronously accessed, which includes the conveying roller speed, the conveying belt tension, the PVC compound thickness detection value and other real-time collected parameters, the time scale deviation of the position offset trajectory is corrected in combination with the conveying roller speed, to avoid the trajectory distortion caused by the conveying rate fluctuation; finally, the overlapping area ratio of the PVC compound morphological contour in adjacent frames, the contour vertex coordinate deviation value are calculated, the position offset trajectory, the morphological contour change parameters and the corrected conveying state information are integrated into the frame-to-frame feature correlation information, which includes the frame pair serial number of each set of adjacent frames, the position offset vector set, the morphological contour overlapping ratio, the contour deviation feature vector, the conveying state correction coefficient and the like.

[0032] Step 130: Based on the frame-to-frame feature correlation information, the time sequence correlation logic information of the continuous multiple frames of visual imaging data is determined, the motion features of the previous frame of compound are taken as the input constraints of the defect tracking task of the next frame, and the cross-frame defect tracking link is generated in combination with the time sequence correlation logic information.

[0033] In the embodiment of the present application, firstly, the inter-frame feature association information is subjected to feature recognition in the time sequence dimension, and the position offset, the shape contour similarity and the motion rate change parameter of the PVC compound between different frames are extracted; a time sequence association feature matrix is generated based on the above parameters, the correlation strength between adjacent frames in the continuous multi-frame visual imaging data is quantified by the weight value of each element in the matrix, and then the time sequence association logic information is determined; subsequently, the motion feature of the compound corresponding to the previous frame is separated from the inter-frame feature association information, which includes the motion direction vector, the motion acceleration and the motion trajectory curvature, and is subjected to standardization processing to obtain the standardized motion feature parameter, which is used as the input constraint of the next frame defect tracking task, for limiting the area range and the motion trend prediction direction of the next frame defect tracking; finally, a frame jump relationship model is created based on the time sequence association logic information, the standardized motion feature parameter is introduced into the model, the corresponding relationship of the defect tracking target in adjacent frames is determined through the model, the defect tracking nodes in each frame are connected according to the corresponding relationship, and a cross-frame defect tracking link covering the continuous multi-frame visual imaging data is generated.

[0034] Step 131: The inter-frame feature association information is subjected to feature recognition in the time sequence dimension, and the position offset, the shape contour similarity and the motion rate change parameter of the PVC compound between different frames are extracted.

[0035] A sliding window feature traversal rule is adopted, a preset number of frame pairs are taken as the window size, the inter-frame feature association information is traversed window by window, the mean value and the variance of the position offset, the time sequence change curve of the shape contour similarity and the first-order difference parameter of the motion rate are extracted in each window, and the above parameters are integrated into a time sequence feature set, wherein the position offset is stored in the form of a two-dimensional vector group, the shape contour similarity is stored in the form of a continuous value sequence between 0 and 1, and the motion rate change parameter is stored in the form of a sequence of rate difference values.

[0036] Step 132: A time sequence association feature matrix is generated based on the position offset, the shape contour similarity and the motion rate change parameter, the correlation strength between adjacent frames in the continuous multi-frame visual imaging data is quantified by the time sequence association feature matrix, and the time sequence association logic information of the continuous multi-frame visual imaging data is determined.

[0037] The position offset, the shape contour similarity and the motion rate change parameter are respectively mapped to the same dimensional range, and then the three types of parameters are weighted and spliced according to a preset weight ratio to generate a time sequence correlation feature matrix, the rows of the matrix correspond to adjacent frame pairs, and the columns correspond to each type of feature parameter; the matrix is processed by using a matrix element normalization rule, each element is converted into a value between 0 and 1, and the value size represents the correlation strength between the corresponding frame pairs; based on the distribution characteristics of the correlation strength, an correlation strength threshold is set, the frame pairs with a correlation strength higher than the threshold are marked as strong correlation frame pairs, and the frame pairs with a correlation strength lower than the threshold are marked as weak correlation frame pairs, so as to determine the time sequence correlation logic information of the continuous multi-frame visual imaging data, and the strong correlation frame pairs represent that the PVC material motion and the shape characteristics between the corresponding frames have continuous inheritance, and the weak correlation frame pairs represent that there is possible interference or abnormality.

[0038] Step 133: The material motion characteristics corresponding to the previous frame are separated from the inter-frame feature correlation information, and the material motion characteristics include a motion direction vector, a motion acceleration and a motion trajectory curvature.

[0039] It can be understood that the fitting direction of the position offset vector set in the feature entry corresponding to the previous frame is located by the frame pair serial number in the inter-frame feature correlation information, the first derivative of the motion rate change parameter is extracted as the motion acceleration, and the curve curvature calculation value of the position offset trajectory is extracted as the motion trajectory curvature, and the three types of characteristics are integrated into the material motion characteristic set of the previous frame.

[0040] Step 134: The material motion characteristics are standardized to obtain standardized motion characteristic parameters, and the standardized motion characteristic parameters are used as input constraint conditions of the next frame defect tracking task, and the input constraint conditions are used to limit the region range and the motion trend prediction direction of the next frame defect tracking.

[0041] The motion direction vector, the motion acceleration and the motion trajectory curvature are respectively processed by using a Z-Score standardization rule, each characteristic parameter is converted into a standardized value with a mean value of 0 and a variance of 1, the standardized motion direction vector is mapped to a potential direction range of the next frame defect tracking, the motion acceleration is mapped to a rate change range of the defect tracking target, and the motion trajectory curvature is mapped to a boundary curvature range of the defect tracking region, and the three together constitute the input constraint conditions of the next frame defect tracking task.

[0042] Step 135: An inter-frame jump relationship model is created based on the time sequence correlation logic information, the input constraint conditions are introduced into the inter-frame jump relationship model, the corresponding relationship of the defect tracking target in adjacent frames is determined through the inter-frame jump relationship model, the defect tracking nodes in each frame are connected based on the corresponding relationship, and a cross-frame defect tracking link covering the continuous multi-frame visual imaging data is generated.

[0043] The embodiment of the application firstly constructs an initial model network of the inter-frame jump relationship model based on the inter-frame association strength and the time sequence arrangement rule in the time sequence association logic information, embeds an inter-frame feature matching module and a tracking target screening module in the initial model network, the inter-frame feature matching module is used to realize matching of PVC adhesive related features in adjacent frames, and the tracking target screening module is used to screen a target area meeting the defect tracking condition from the PVC adhesive area; subsequently, the input constraint condition corresponding to the standardized motion feature parameter is disassembled into a region range constraint parameter and a motion trend constraint parameter, the region range constraint parameter is used to define the effective area boundary of defect tracking in the next frame, and the motion trend constraint parameter is used to limit the potential motion direction and the motion rate range of the defect tracking target, and the disassembled constraint parameters are respectively introduced into the corresponding modules of the inter-frame jump relationship model; the first motion feature corresponding to the defect tracking node of the previous frame and the second motion feature of the PVC adhesive in the effective area of the next frame are extracted through the inter-frame feature matching module, the similarity of the first motion feature and the second motion feature is calculated and a feature similarity matrix is generated, the feature similarity matrix is corrected in combination with the shape contour similarity in the time sequence association logic information, and a corrected feature matching result is obtained; based on the corrected feature matching result, the region in the next frame which meets the preset requirement with the defect tracking node feature matching degree of the previous frame is screened as the defect tracking node of the next frame through the tracking target screening module, and the position mapping relationship and the feature inheritance relationship between adjacent frame defect tracking nodes are recorded synchronously; finally, taking the time sequence order of the continuous multiple frame visual imaging data as the axis, according to the position mapping relationship and the feature inheritance relationship, the defect tracking nodes in each frame of visual imaging data are connected in turn in a chain connection mode, and the connection path is smoothed and corrected in the connection process in combination with the position offset trajectory in the inter-frame feature association information to eliminate the connection deviation caused by inter-frame offset, and a cross-frame defect tracking link covering the continuous multiple frame visual imaging data and having time sequence continuity is generated.

[0044] Step 1351: constructing an initial model network of the inter-frame jump relationship model based on the inter-frame association strength and the time sequence arrangement rule in the time sequence association logic information, embedding an inter-frame feature matching module and a tracking target screening module in the initial model network, the inter-frame feature matching module is used to realize matching of PVC adhesive related features in adjacent frames, and the tracking target screening module is used to screen a target area meeting the defect tracking condition from the PVC adhesive area.

[0045] The inter-frame correlation strength in the time sequence correlation logic information is taken as an edge weight value, and a frame serial number is taken as a node to construct a time sequence directed graph as a basic structure of the initial model network. An inter-frame feature matching module is embedded in the initial model network, which is implemented based on a Siamese network architecture and includes two feature extraction sub-networks with the same structure, which are respectively used to extract feature vectors of PVC adhesive regions in adjacent frames, and feature matching is realized by calculating the cosine similarity of the feature vectors. A tracking target screening module is also embedded, which screens out regions deviating from the standard template in the PVC adhesive region as potential defect tracking target regions based on preset defect tracking condition rules.

[0046] Step 1352: The input constraint condition corresponding to the standardized motion feature parameter is disassembled into a region range constraint parameter and a motion trend constraint parameter; wherein the region range constraint parameter is used to define the effective region boundary of defect tracking in the next frame, and the motion trend constraint parameter is used to limit the potential motion direction and motion rate range of the defect tracking target. The disassembled constraint parameters are respectively introduced into the corresponding modules of the inter-frame jump relationship model.

[0047] Exemplarily, the motion direction vector and the motion trajectory curvature in the standardized motion feature parameter are integrated into the region range constraint parameter, which is converted into a polygon region boundary in the next frame with a preset point as the center through geometric transformation, and the boundary is the effective region of defect tracking. The motion acceleration and the motion rate change parameter in the standardized motion feature parameter are integrated into the motion trend constraint parameter, which is converted into the potential motion direction interval and the rate change interval of the defect tracking target. The region range constraint parameter is introduced into the tracking target screening module to limit the region range of screening potential tracking targets, and the motion trend constraint parameter is introduced into the inter-frame feature matching module for trend consistency verification in the feature matching process.

[0048] Step 1353: The first motion feature corresponding to the defect tracking node of the previous frame and the second motion feature of the PVC adhesive in the effective region of the next frame are extracted through the inter-frame feature matching module, the similarity of the first motion feature and the second motion feature is calculated to generate a feature similarity matrix, and the feature similarity matrix is corrected in combination with the shape contour similarity in the time sequence correlation logic information to obtain a corrected feature matching result.

[0049] It can be understood that the feature extraction sub-network of the inter-frame jump relationship model extracts the motion direction, acceleration, trajectory curvature and other features of the previous frame defect tracking node respectively, integrates them into a first motion feature vector, and simultaneously extracts the corresponding features of each potential tracking target region in the effective region of the next frame, and integrates them into a second motion feature vector set; the cosine similarity of the first motion feature vector and each second motion feature vector is calculated, all similarity values are arranged in the order of potential tracking targets to generate a feature similarity matrix; the shape contour similarity of the corresponding frame pair is extracted from the time sequence association logic information, which is used as a weight coefficient to weight and correct each element in the feature similarity matrix, to obtain a corrected feature matching result. The higher the shape contour similarity, the greater the weight coefficient, and the closer the corrected similarity value to the original calculated value.

[0050] Step 1354: Based on the corrected feature matching result, the tracking target screening module screens out the region in the next frame which has a feature matching degree with the defect tracking node of the previous frame satisfying a preset requirement as the defect tracking node of the next frame, and synchronously records the position mapping relationship and the feature inheritance relationship between adjacent frame defect tracking nodes.

[0051] In this step, a feature matching degree threshold is set, and the potential tracking target region with a value higher than the threshold in the corrected feature matching result is marked as the defect tracking node of the next frame; the coordinates of the defect tracking node of the previous frame and the coordinates of the defect tracking node of the next frame are recorded to generate a position mapping relationship; and the feature parameter difference value between the two nodes is recorded to generate a feature inheritance relationship, which is used to represent the feature change of the defect tracking target between frames.

[0052] Step 1355: Taking the time sequence order of the continuous multiple frame visual imaging data as the axis, according to the position mapping relationship and the feature inheritance relationship, the defect tracking nodes in each frame of visual imaging data are connected in turn in a chain connection mode, and the connection path is smoothed and corrected in the connection process in combination with the position offset trajectory in the inter-frame feature association information to eliminate the connection deviation caused by inter-frame offset, to generate a cross-frame defect tracking link covering the continuous multiple frame visual imaging data and having time sequence continuity.

[0053] In detail, according to the order of frame number from small to large, the defect tracking nodes in adjacent frames are connected by line segments in turn to form an initial cross-frame defect tracking link; the position offset trajectory of the corresponding frame pair is extracted from the inter-frame feature association information, and the connection path of the initial link is smoothed and corrected by using the Bezier curve fitting rule, so that the bending trend of the link is consistent with the actual motion trajectory of the PVC compound; after the correction is completed, a cross-frame defect tracking link containing all defect tracking node coordinates, node connection relationship and feature inheritance information is generated, each node in the link corresponds to a defect tracking region in a frame of image, and the connection relationship between the nodes represents the time sequence continuity of the tracking target.

[0054] Step 140: Track the morphological abnormal area of the PVC compound in the continuous multi-frame visual imaging data by using the cross-frame defect tracking link, and determine the suspected surface defect area from the morphological abnormal area by time sequence feature correlation analysis.

[0055] In this step, first, the PVC compound area corresponding to each frame in the continuous multi-frame visual imaging data is located in sequence according to the inter-frame jump order of the cross-frame defect tracking link, and the PVC compound area in each frame is morphologically compared based on a preset compound standard morphological template to mark the area with morphological deviation from the compound standard morphological template as an initial morphological abnormal area. Then, the morphological feature parameters of each initial morphological abnormal area are extracted, which include area, edge roughness, shape irregularity, and gray scale distribution uniformity. The time sequence correlation relationship of the initial morphological abnormal areas between frames is established based on the cross-frame defect tracking link, and the corresponding nodes connected in the link are marked as time sequence correlated initial morphological abnormal areas. The morphological feature parameter similarity of the initial morphological abnormal areas in adjacent frames is determined through the time sequence correlation relationship, and the initial morphological abnormal areas that are continuous in time sequence and have morphological feature parameter similarity satisfying a preset threshold are extracted. The extracted initial morphological abnormal areas are fitted with a time sequence trajectory to obtain the morphological evolution trend of the area in the continuous multi-frames. The initial morphological abnormal area whose morphological evolution trend matches a preset defect development rule is determined as a suspected surface defect area, and the position coordinates and morphological evolution parameters of the suspected surface defect area in each frame of visual imaging data are recorded synchronously.

[0056] Step 141: Locate the PVC compound area corresponding to each frame in the continuous multi-frame visual imaging data in sequence according to the inter-frame jump order of the cross-frame defect tracking link, and perform morphological comparison on the PVC compound area in each frame based on a preset compound standard morphological template to mark the area with morphological deviation from the compound standard morphological template as an initial morphological abnormal area.

[0057] Among them, the PVC compound area is located from the image of the corresponding frame in sequence according to the node order of the cross-frame defect tracking link; the compound standard morphological template is a set of morphological profile parameters of a defect-free PVC compound collected in advance, including standard area, standard edge roughness, standard shape regularity, etc.; the actual morphological parameters of the PVC compound area in each frame are compared one by one with the parameters in the standard morphological template, and when the deviation of any parameter of a certain area from the standard parameter exceeds the preset deviation range, the area is marked as an initial morphological abnormal area.

[0058] Step 142: Extract the morphological feature parameters of each initial morphological abnormal area, which include area, edge roughness, shape irregularity, and gray scale distribution uniformity.

[0059] Exemplarily, the pixel number of the initial morphological abnormal region is calculated by using a regional pixel counting rule, and is mapped as an actual area parameter of the region; the gradient value variance of the edge of the initial morphological abnormal region is calculated by using an edge pixel gradient calculation rule, and is taken as an edge roughness parameter; the shape irregularity is calculated by using a ratio of the area of the minimum circumscribed rectangle of the region contour to the actual area of the region, and the closer the ratio is to 1, the more regular the shape is, and the smaller the ratio is, the more irregular the shape is; the gray scale distribution uniformity is calculated by using a variance of the gray scale values of the pixels in the region, and the greater the variance is, the more uneven the gray scale distribution is.

[0060] Step 143: Establishing a time sequence association relationship of the initial morphological abnormal region between frames based on the defect tracking link across frames.

[0061] It can be understood that the initial morphological abnormal region node of the previous frame is associated with the defect tracking node corresponding to the next frame through the node connection relationship in the defect tracking link across frames, and is marked as a time sequence associated initial morphological abnormal region. The association relationship includes the position mapping information and the feature inheritance information between the nodes.

[0062] Step 144: Determining the morphological feature parameter similarity of the initial morphological abnormal region in adjacent frames through the time sequence association relationship, and extracting the initial morphological abnormal region which is continuous in time sequence and has a morphological feature parameter similarity satisfying a preset threshold.

[0063] Among them, the similarity of the area, the edge roughness, the shape irregularity and the gray scale distribution uniformity of the initial morphological abnormal region in adjacent frames is calculated by using a cosine similarity calculation rule, and the four similarity values are weighted and averaged according to a preset weight to obtain a comprehensive similarity; a comprehensive similarity threshold is set, and the initial morphological abnormal region of the adjacent frame whose comprehensive similarity is higher than the threshold is marked as a continuous association region; and a set of initial morphological abnormal regions which are continuously associated and have a continuous frame number satisfying a preset number requirement is extracted.

[0064] Step 145: Fitting a time sequence trajectory for the extracted initial morphological abnormal region, obtaining a morphological evolution trend of the extracted initial morphological abnormal region in continuous multiple frames, determining the initial morphological abnormal region whose morphological evolution trend matches a preset defect development rule as a suspected surface defect region, and synchronously recording the position coordinates and the morphological evolution parameters of the suspected surface defect region in each frame of visual imaging data.

[0065] The center coordinates, area, edge roughness and other parameters of the extracted initial morphological abnormal region are fitted by using a polynomial fitting rule, to obtain a fitting curve of each parameter with respect to the frame number, and the trend of the curve is the morphological evolution trend; the preset defect development rule is the morphological change rule of the defect-free PVC compound in the normal production process, including the rate range and trend direction of the parameter change; the morphological evolution trend of the extracted initial morphological abnormal region is compared with the preset defect development rule, and when the trend deviates from the preset rule and the deviation degree exceeds the preset range, the region is determined as a suspected surface defect region; the center coordinates, area, edge roughness and other parameters of the suspected surface defect region in each frame of image, and the time series change value of each parameter are recorded synchronously to form a morphological evolution parameter set.

[0066] Step 150: Based on the time series distribution information of the suspected surface defect region, feature mining processing is performed on the suspected surface defect region by edge contour recognition to obtain surface defect features containing feature level correlation and semantic mapping labels.

[0067] In the embodiments of the present application, firstly, the occurrence time, duration frame number and inter-frame position offset rule of the suspected surface defect region in the continuous multiple frames of visual imaging data are determined based on the time series distribution information of the suspected surface defect region, and the time window for feature mining is divided according to the time series distribution information, and the multiple frames of suspected surface defect regions in the same time window are taken as a feature mining unit; then a multi-scale edge detection algorithm is used to perform edge contour recognition on the suspected surface defect region in each feature mining unit, to extract the edge contour point set of each frame of suspected surface defect region, and to obtain the fusion edge contour corresponding to the feature mining unit through matching and fusion of the inter-frame edge contour point sets; feature mining processing is performed based on the fusion edge contour to extract the contour level feature, texture feature and gray gradient feature; wherein the contour level feature contains the nesting relationship of the main contour and the secondary contour, the texture feature contains the texture density and texture direction of the defect region, and the gray gradient feature contains the gray change rate of the defect region inside and the surrounding normal region; the extracted contour level feature, texture feature and gray gradient feature are subjected to level correlation analysis and a feature level correlation graph is generated, and each feature is given a corresponding semantic mapping label based on a preset defect semantic label library, and the feature level correlation graph and the semantic mapping label are fused to obtain surface defect features containing feature level correlation and semantic mapping labels.

[0068] Step 151: Based on the time series distribution information of the suspected surface defect region, the occurrence time, duration frame number and inter-frame position offset rule of the suspected surface defect region in the continuous multiple frames of visual imaging data are determined, and the time window for feature mining is divided according to the time series distribution information, and the multiple frames of suspected surface defect regions in the same time window are taken as a feature mining unit.

[0069] wherein the frame number corresponding to the occurrence time, the frame number corresponding to the duration frame number, and the coordinate change value corresponding to the inter-frame position offset are extracted from the recording information of the suspected surface defect region to determine the distribution characteristics thereof in time sequence; a sliding window division rule is used to divide the time window for feature mining according to the time sequence order with a preset frame number as the window size, so as to ensure that each suspected surface defect region is contained in at least one window; all suspected surface defect regions in the same time window are integrated into a feature mining unit, and the unit contains the position coordinates, shape parameters and inter-frame correlation information of each frame region.

[0070] Step 152: an edge contour recognition is performed on the suspected surface defect region in each feature mining unit by using a multi-scale edge detection algorithm, and an edge contour point set of each suspected surface defect region is extracted, and a fused edge contour corresponding to the feature mining unit is obtained through matching and fusion of the inter-frame edge contour point sets.

[0071] For example, a multi-scale improved version of the Canny edge detection algorithm is used to perform edge detection on each suspected surface defect region in the feature mining unit, a plurality of different Gaussian filter scales are set, and edge contour point sets under corresponding scales are extracted respectively; a feature point matching rule is used to match the edge contour point sets in adjacent frames to find the corresponding contour point pairs between frames; based on the matched contour point pairs, a weighted fusion rule is used to fuse the edge contour point sets of multiple frames, the coordinates of the corresponding contour points are weighted and averaged according to the weight of the frame number, and a fused edge contour point set is obtained, which is the fused edge contour corresponding to the feature mining unit.

[0072] Step 153: a feature mining processing is performed based on the fused edge contour to extract contour level features, texture features and gray gradient features; wherein the contour level features include the nesting relationship of the main contour and the secondary contour, the texture features include the texture density and the texture direction of the defect region, and the gray gradient features include the gray change rate of the defect region and the surrounding normal region.

[0073] For example, a contour nesting analysis rule is used to analyze the fused edge contour to identify the outermost main contour and the secondary contour inside the main contour, record the position nesting relationship of the main contour and the secondary contour, and construct the contour level features; a gray level co-occurrence matrix analysis rule is used to analyze the texture of the suspected surface defect region to calculate the energy, entropy and other parameters of the texture matrix, map them to the texture density and texture direction parameters, and construct the texture features; and a gradient calculation rule is used to calculate the gray values of the suspected surface defect region and the surrounding normal region to obtain the gray gradient mean value of the region and the gray gradient difference value between the region and the surrounding region, and construct the gray gradient features.

[0074] Step 154: Perform hierarchical correlation analysis on the extracted contour level features, texture features, and gray gradient features, and generate a feature hierarchical correlation graph, and assign corresponding semantic mapping labels to each feature based on a pre-set defect semantic label library, fuse the feature hierarchical correlation graph and the semantic mapping labels, and obtain surface defect features containing feature hierarchical correlation and semantic mapping labels.

[0075] For example, a graph structure is constructed according to a rule, contour level features, texture features, and gray gradient features are taken as nodes, and edges between the nodes represent the correlation between the features, such as the correlation between the secondary contour in the contour level features and the texture density in the texture features, and the weight of the edge represents the correlation strength, to generate a feature hierarchical correlation graph; the pre-set defect semantic label library is a pre-constructed corresponding relationship set of defect features and semantic labels, containing semantic labels such as “irregular main contour”, “high texture density”, “large gray gradient difference”, etc.; each feature extracted is matched with the labels in the defect semantic label library, and a corresponding semantic mapping label is assigned to each feature; each node in the feature hierarchical correlation graph is bound with its corresponding semantic mapping label, and surface defect features containing feature hierarchical correlation and semantic mapping labels are generated, which are stored in the form of graph structure data, containing node features, correlation edge information, and semantic label information.

[0076] Step 160: Input the surface defect features into a pre-set hierarchical reasoning mining model for hierarchical feature analysis, extract a defect morphology vector set corresponding to the surface defect features, inversely deduce the internal implicit defect type corresponding to the suspected surface defect area according to the defect morphology vector set and the internal implicit defect correlation logic, mine the internal implicit defect distribution trend combining the time sequence change rule in the inter-frame feature correlation information, and generate a defect detection result of the PVC compound based on the internal implicit defect type and the internal implicit defect distribution trend.

[0077] It can be understood that first, the surface defect features are input into the hierarchical reasoning mining model, and the defect morphology vector set is extracted through hierarchical analysis of the model; then the defect morphology vector set is matched with the pre-constructed correlation logic database, and the corresponding internal implicit defect type is inversely deduced; then the time sequence change rule is extracted from the inter-frame feature correlation information, and the distribution trend of the internal implicit defect is mined combining the internal implicit defect type; finally, the internal implicit defect type, the distribution trend, and the related features are integrated into a defect detection result, which contains defect type, distribution range, evolution trend, and influence degree.

[0078] Step 161: Input the surface defect features into a pre-set hierarchical reasoning mining model for hierarchical feature analysis, and extract a defect morphology vector set corresponding to the surface defect features.

[0079] In detail, first, the surface defect features are standardized, and the numerical features are converted into standardized values with a mean of 0 and a variance of 1. The standardized surface defect features are input into a layered reasoning mining model, which includes a feature input layer, a first feature analysis layer, an intermediate feature analysis layer, and a second feature analysis layer. The standardized surface defect features are input into the first feature analysis layer through the feature input layer, and the first layer of convolution kernels is used to preliminarily extract features from the standardized surface defect features to obtain first layer defect features. The first layer defect features are input into the intermediate feature analysis layer, and the feature fusion algorithm is used to splice and fuse the feature components with a correlation relationship in the first layer defect features, while filtering out redundant feature components to obtain middle layer fusion defect features. The middle layer fusion defect features are input into the second feature analysis layer, and the deep neural network is used to analyze the middle layer fusion defect features to mine defect morphology correlation information and output multi-dimensional defect morphology feature components. The multi-dimensional defect morphology feature components output by the second feature analysis layer are subjected to vector quantization processing, and each feature component is converted into a vector form of a unified dimension. Based on the correlation relationship of each feature component, the defect morphology vector set corresponding to the surface defect features is extracted by sorting and integrating, and the vector set is a set of multi-dimensional vectors, each vector corresponding to a defect morphology feature dimension.

[0080] Step 1611: input the standardized surface defect features corresponding to the surface defect features into a preset layered reasoning mining model; wherein the layered reasoning mining model includes a feature input layer, a first feature analysis layer, an intermediate feature analysis layer, and a second feature analysis layer, and the standardized surface defect features are input into the first feature analysis layer through the feature input layer, and the first layer of convolution kernels is used to preliminarily extract features from the standardized surface defect features to obtain first layer defect features.

[0081] For example, the numerical parameters in the surface defect features are processed using the Min-Max standardization rule to convert them into values between 0 and 1, and the non-numerical semantic labels are processed using the one-hot encoding rule to convert them into binary vectors. The standardized features are input into the feature input layer of the layered reasoning mining model, and the feature input layer is responsible for converting the input features into a tensor form that can be processed by the model. The first feature analysis layer includes multiple convolution kernels, each corresponding to a feature extraction dimension. The convolution operation is used to extract features from the standardized surface defect features, and the first layer defect features are output, which are feature map tensors containing multiple feature channels.

[0082] Step 1612: input the first layer defect features into the intermediate feature analysis layer, and use the feature fusion algorithm to fuse the feature components with a correlation relationship in the first layer defect features and filter out redundant feature components to obtain middle layer fusion defect features.

[0083] The intermediate feature analysis layer includes a feature correlation analysis module and a feature filtering module. The feature correlation analysis module determines the correlation between feature components in the first layer of defect features by calculating mutual information values of the feature components, and the higher the mutual information value, the stronger the correlation. The feature components with strong correlation are spliced and fused to obtain preliminary fusion features. The feature filtering module calculates the importance score of each feature component based on a feature importance evaluation rule, filters out redundant feature components with a score lower than a preset threshold, and obtains a middle layer fusion defect feature, which is a tensor set containing key feature components.

[0084] Step 1613: input the middle layer fusion defect feature into a second feature analysis layer, analyze and mine defect morphology correlation information of the middle layer fusion defect feature through a deep neural network, and output multi-dimensional defect morphology feature components.

[0085] It can be understood that the second feature analysis layer adopts a Transformer encoder architecture, including multiple stacked encoder layers, each encoder layer including a multi-head self-attention sublayer and a feedforward neural network sublayer. The middle layer fusion defect feature is calculated by the multi-head self-attention sublayer of the encoder layer to obtain the self-attention weight between the feature components, mine the global correlation information between the features, and then the feature is transformed by the feedforward neural network sublayer to output multi-dimensional defect morphology feature components, each component corresponding to a specific dimension of a defect morphology.

[0086] Step 1614: vector quantization processing is performed on the multi-dimensional defect morphology feature components output by the second feature analysis layer, each feature component is converted into a uniform dimension vector form, and the defect morphology vector set corresponding to the surface defect feature is extracted based on the correlation between the feature components.

[0087] For example, each defect morphology feature component is converted into a fixed length vector using a vector quantization rule, and the dimension of the vector is uniformly set to a preset fixed value. Based on the self-attention weight output by the second feature analysis layer, the correlation between the feature components is determined, and the vectors are sorted in order from high to low according to the correlation strength. The sorted vectors are integrated into a set to obtain the defect morphology vector set corresponding to the surface defect feature, and each vector in the set contains feature information of the corresponding defect morphology dimension.

[0088] Step 162: according to the defect morphology vector set and the association logic of the internal implicit defect, the internal implicit defect type corresponding to the suspected surface defect region is deduced reversely.

[0089] It can be understood that firstly, a correlation logic database of defect morphology vector set and internal implicit defect is pre-constructed, the database stores standard defect morphology vector sets corresponding to different internal implicit defect types and vector matching rules, the vector matching rules are used to define a similarity calculation method and a matching threshold of the defect morphology vector set to be matched and the standard defect morphology vector set; the extracted defect morphology vector set is input into the correlation logic database, and the similarity of the vector set and each standard defect morphology vector set in the database is calculated according to the vector matching rule; the standard defect morphology vector set with the highest similarity and meeting the matching threshold is screened out, and the internal implicit defect type corresponding to the standard defect morphology vector set is determined as a candidate internal implicit defect type; the candidate internal implicit defect type is verified based on the feature level correlation information in the surface defect feature, and whether the defect formation mechanism corresponding to the candidate internal implicit defect type is consistent with the feature evolution law of the surface defect feature is judged; if consistent, the candidate internal implicit defect type is determined as the internal implicit defect type corresponding to the suspected surface defect region; if inconsistent, the matching threshold in the vector matching rule is adjusted again, and the similarity calculation and screening are performed again until the internal implicit defect type consistent with the feature evolution law of the surface defect feature is determined.

[0090] Step 1621: A correlation logic database of defect morphology vector set and internal implicit defect is pre-constructed, the correlation logic database stores standard defect morphology vector sets corresponding to different internal implicit defect types and vector matching rules, the vector matching rules are used to define a similarity calculation method and a matching threshold of the defect morphology vector set to be matched and the standard defect morphology vector set.

[0091] In detail, the correlation logic database is a structured database, including three core tables of defect type table, standard vector set table and matching rule table; the defect type table stores all possible internal implicit defect types and corresponding information such as defect formation mechanism and evolution law; the standard vector set table stores the standard defect morphology vector set corresponding to each internal implicit defect type, and the vector set is trained by a large number of morphology feature vectors of corresponding defect samples collected in advance; the matching rule table stores the similarity calculation method and matching threshold of vector matching, the similarity calculation method adopts cosine similarity calculation rule, and the matching threshold is the optimal threshold obtained based on sample data statistics.

[0092] Step 1622: The extracted defect morphology vector set is input into the correlation logic database, and the similarity of the defect morphology vector set and each standard defect morphology vector set in the correlation logic database is calculated according to the vector matching rule.

[0093] In the step, the defect morphology vector set to be matched and the standard defect morphology vector set corresponding to each internal implicit defect type in the database are matched one by one, the cosine similarity calculation rule is adopted to calculate the similarity value of the two, and the closer to 1 the similarity value is, the higher the matching degree is.

[0094] Step 1623: screening the standard defect morphology vector set with the highest similarity and meeting the matching threshold, and determining the internal implicit defect type corresponding to the standard defect morphology vector set as a candidate internal implicit defect type.

[0095] It can be understood that all the calculated similarity values are sorted, and the standard defect morphology vector set with the highest similarity value is screened out; if the similarity value is greater than or equal to the matching threshold in the matching rule table, the internal implicit defect type corresponding to the vector set is determined as a candidate internal implicit defect type; if all the similarity values are less than the matching threshold, it is marked as an abnormal situation that cannot be matched, and enters the subsequent threshold adjustment and re-matching process.

[0096] Step 1624: verifying the candidate internal implicit defect type based on the feature level association information in the surface defect feature, and determining whether the defect formation mechanism corresponding to the candidate internal implicit defect type is consistent with the feature evolution law of the surface defect feature.

[0097] For example, the feature evolution law is extracted from the feature level association graph of the surface defect feature, including the time sequence change trend of the feature, the change of the association strength between the features, etc.; the defect formation mechanism corresponding to the candidate internal implicit defect type is extracted from the association logic database, including the cause of the defect and the law of development, etc.; the feature evolution law and the defect formation mechanism are compared to determine whether they are consistent, for example, if the defect formation mechanism indicates that the defect will gradually expand over time, the feature evolution law of the surface defect feature should include the trend of gradually increasing area.

[0098] Step 1625: if consistent, the candidate internal implicit defect type is determined as the internal implicit defect type corresponding to the suspected surface defect region; if not consistent, the matching threshold in the vector matching rule is adjusted again, and similarity calculation and screening are performed again until the internal implicit defect type consistent with the feature evolution law of the surface defect feature is determined.

[0099] In the embodiment of the application, if the feature evolution law and the defect formation mechanism are consistent, the candidate internal implicit defect type is directly determined as the final internal implicit defect type; if not consistent, the matching threshold is reduced by a preset adjustment amplitude, and similarity calculation and screening are performed again to obtain a new candidate internal implicit defect type, which is verified again; the process is repeated until the internal implicit defect type consistent with the feature evolution law is found, or after the preset number of adjustments, it is still not found, and then it is marked as a defect type that cannot be determined.

[0100] Step 163: combining the time sequence change law in the inter-frame feature association information to mine the internal implicit defect distribution trend.

[0101] In the embodiment of the present application, first, the feature parameters related to the time sequence change rule are extracted from the inter-frame feature association information, the parameters include the overall motion trend of the PVC compound in the continuous multi-frame visual imaging data, the position offset rate of the defect area, and the morphological expansion rate, and a time sequence change rule model is created based on the above parameters; the feature parameters corresponding to the internal implicit defect type are input into the time sequence change rule model, and the mapping relationship between the evolution of the internal implicit defect in the time sequence dimension and the time sequence change rule in the inter-frame feature association information is output; based on the mapping relationship, the position change, morphological evolution and range expansion of the internal implicit defect in the subsequent frame are predicted based on the internal implicit defect features in the continuous multi-frame visual imaging data using a time sequence prediction algorithm; the spatio-temporal distribution map of the internal implicit defect is generated by combining the historical distribution information of the internal implicit defect in the continuous multi-frame visual imaging data and the subsequent evolution information obtained by prediction; the internal implicit defect distribution trend is mined by analyzing the density change, range expansion direction and evolution rate of the defect distribution in the spatio-temporal distribution map, and the trend includes the main diffusion direction, diffusion rate and affected PVC compound area range of the defect.

[0102] Step 1631: Extracting feature parameters related to the time sequence change rule from the inter-frame feature association information, the feature parameters include the overall motion trend of the PVC compound in the continuous multi-frame visual imaging data, the position offset rate of the defect area, and the morphological expansion rate, and creating a time sequence change rule model based on the feature parameters.

[0103] It can be understood that the fitting direction of the position offset trajectory of the PVC compound extracted from the inter-frame feature association information is taken as the overall motion trend, the position coordinate change value of the suspected surface defect area in the adjacent frames is taken as the position offset rate, and the area change value of the suspected surface defect area in the adjacent frames is taken as the morphological expansion rate; the three types of parameters are taken as input features, and a long short-term memory network (LSTM) is used to construct a time sequence change rule model, the input of the model is a time sequence feature sequence, and the output is an evolution prediction result of the internal implicit defect.

[0104] Step 1632: Inputting the feature parameters corresponding to the internal implicit defect type into the time sequence change rule model, and outputting the mapping relationship between the evolution of the internal implicit defect in the time sequence dimension and the time sequence change rule in the inter-frame feature association information.

[0105] The feature parameters corresponding to the internal implicit defect type are extracted from the associated logic database, including the typical expansion rate of the defect, the diffusion direction, etc.; the above parameters are integrated with the input features of the time sequence change rule model and input into the model; the model learns the time sequence correlation between the features through the internal gating mechanism, and outputs the mapping relationship between the evolution parameters of the internal implicit defect and the time sequence change parameters in the inter-frame feature correlation information, such as the correlation degree between the defect expansion rate and the overall movement trend of the PVC compound, the corresponding relationship between the defect diffusion direction and the position offset trajectory, etc.

[0106] Step 1633: Based on the mapping relationship, the position change, morphological evolution and range expansion of the internal implicit defect in the subsequent frames are predicted based on the internal implicit defect features in the continuous multi-frame visual imaging data using a time sequence prediction algorithm.

[0107] For example, using the ARIMA time sequence prediction algorithm, the time sequence of the position coordinates, area, edge roughness, etc. of the internal implicit defect in the continuous multi-frames is input, and the parameter changes in the subsequent preset frames are predicted in combination with the association rules in the mapping relationship; the prediction result includes the predicted position coordinates, predicted area, predicted edge roughness, etc. of the internal implicit defect in the subsequent frames.

[0108] Step 1634: The spatio-temporal distribution map of the internal implicit defect is generated by combining the historical distribution information of the internal implicit defect in the continuous multi-frame visual imaging data and the subsequent evolution information obtained by prediction.

[0109] It can be understood that the historical position coordinates, area range, etc. of the internal implicit defect in the continuous multi-frames are integrated with the subsequent position coordinates, area range, etc. obtained by prediction to form a spatio-temporal distribution dataset; the dataset is converted into a spatio-temporal distribution map using a heat map construction rule, the horizontal axis of the map is the frame number (representing the time dimension), and the vertical axis is the position coordinates of the PVC compound (representing the space dimension), and the color depth of the heat map represents the density of the defect.

[0110] Step 1635: The internal implicit defect distribution trend is mined by analyzing the density change, range expansion direction and evolution rate of the defect distribution in the spatio-temporal distribution map, and the internal implicit defect distribution trend includes the main diffusion direction of the defect, the diffusion rate and the affected PVC compound area range.

[0111] The image analysis rule is used to analyze the spatio-temporal distribution map, the direction of the increase in the defect density is extracted as the main diffusion direction, the average rate of the defect range expansion is calculated as the diffusion rate, and the maximum area range covered by the defect in the spatio-temporal distribution map is counted as the affected PVC compound area range; the above contents are integrated into the internal implicit defect distribution trend, which includes qualitative description and quantitative parameters.

[0112] Step 164: generating a defect detection result of the PVC compound based on the internal latent defect type and the internal latent defect distribution trend.

[0113] In this step, the information of the internal latent defect type, defect formation mechanism, evolution law, etc. is integrated with the information of the diffusion direction, diffusion rate, influence range, etc. in the internal latent defect distribution trend, and a defect detection result is generated in accordance with a preset standardized format; the result contains three core parts of defect basic information, distribution trend information, and influence evaluation information, the defect basic information contains defect type, position coordinates of suspected surface defect area, morphological parameters, etc., the distribution trend information contains diffusion direction, diffusion rate, predicted evolution, etc., and the influence evaluation information contains potential influence degree of the defect on the performance of the PVC compound and suggestions for measures to be taken, etc.

[0114] Optionally, the embodiments of the present application further include:

[0115] Step 170: decomposing the defect detection result into core feature information containing the internal latent defect type, defect distribution range, defect evolution rate, and defect influence level, and performing feature correlation analysis on the core feature information and key process parameters in the PVC compound production process to obtain defect process parameter correlation information.

[0116] First, the core feature information of the internal latent defect type, defect distribution range, defect evolution rate, and defect influence level, etc. is extracted from the defect detection result; the key process parameters are extracted from the production process data collection system, which include real-time parameters of production links such as raw material ratio, extrusion temperature, extrusion pressure, and cooling rate; the mutual information value between the core feature information and the key process parameters is calculated using mutual information analysis rules, and the higher the mutual information value, the stronger the correlation between the two; the core feature information, the key process parameters, and the corresponding mutual information value are integrated into defect process parameter correlation information, which contains the correlation strength and direction of each defect type and each key process parameter.

[0117] Step 171: based on the defect process parameter correlation information, using feature attribution analysis method to mine the core process parameter influence factors corresponding to different internal latent defect types and distribution trends, determining the action logic of each core process parameter influence factor on defect formation and evolution, and generating process parameter regulation direction and regulation constraint logic data based on the action logic, which is used to define the range of parameter adjustment.

[0118] The causal inference algorithm is used to analyze the defect process parameter correlation information, exclude false correlations, and determine the core process parameter impact factors that truly cause the formation and evolution of defects; the causal relationship between each core process parameter impact factor and the defect is analyzed, for example, too high extrusion temperature can cause uneven melting of PVC compound, and then internal hidden defects are generated, and the action logic is determined; the process parameter control direction is generated based on the action logic, for example, if the extrusion temperature is too high, the control direction is to reduce the extrusion temperature; at the same time, combining factors such as performance limitations of production equipment and product quality standards, generate control constraint logic data to define the range of parameter adjustment, for example, the adjustment range of extrusion temperature cannot be lower than the preset minimum melting temperature, and cannot be higher than the maximum tolerance temperature of the equipment.

[0119] Step 172: Import the process parameter control direction and control constraint logic data into the preset production parameter control model, decompose the control requirements through the hierarchical analysis module of the model, and combine the real-time running data of the current production link to obtain the dynamic parameter control instruction adapted to the current defect state.

[0120] The production parameter control model is a mixed control model of rules and models, including three core modules: hierarchical analysis module, parameter matching module and instruction generation module; the hierarchical analysis module decomposes the process parameter control direction and control constraint logic data into specific parameter adjustment requirements, for example, the control direction of reducing the extrusion temperature is decomposed into the amplitude and timing of temperature adjustment requirements; the parameter matching module matches the adjustment requirements with the real-time running data of the current production link to ensure that the adjustment requirements meet the current production state; the instruction generation module generates dynamic parameter control instructions based on the matching results, and the instructions include the process parameter name to be adjusted, the adjustment amplitude, the adjustment timing, the target value after adjustment and the like.

[0121] Step 173: Based on the dynamic parameter control instruction, the process parameters of the PVC compound production link are adaptively adjusted, the PVC compound production state data after parameter adjustment and new defect detection data are synchronously collected, the production state data and the new defect detection data are fed back to the construction task of the defect process parameter correlation information, and the correlation relationship coefficient is iteratively updated to optimize the correlation logic.

[0122] The dynamic parameter regulation instruction is sent to a control system of the production equipment, and the control system adjusts corresponding process parameters according to the instruction; the parameter-adjusted PVC sizing material production state data, including the adjusted process parameter values, the form parameters of the PVC sizing material and the like, are collected synchronously through the data collection system, and new defect detection data are collected through the PVC sizing material defect detection system; the production state data and the new defect detection data are input into a construction task of defect-process parameter association information, the mutual information values of the core feature information and the key process parameters are recalculated, the association relationship coefficients are iteratively updated, the association logic between the defects and the process parameters is optimized, and the accuracy of subsequent regulation is improved.

[0123] Optionally, the embodiments of the application further include:

[0124] Step 180: The defect detection result is converted into standardized defect information containing defect feature codes, defect timing information and defect severity according to a preset standardization rule, a time alignment algorithm is used to match the standardized defect information with defect records and corresponding production full-process data of historical production batches in the time dimension, and a matching relationship between defect timing evolution and each link of the production process is established.

[0125] First, the defect detection result is coded according to a preset standardization rule, and information such as internal implicit defect type, distribution range and evolution rate is converted into defect feature codes in a unified format; timing information in the defect detection result is extracted, including time points of defect occurrence, duration and the like, and is integrated into defect timing information; based on parameters such as defect type, distribution range and evolution rate, a preset scoring rule is used to calculate a defect severity score; the defect feature codes, the defect timing information and the defect severity score are integrated into standardized defect information; a dynamic time warping algorithm is used to match the timing information in the standardized defect information with the time dimension of defect records and production full-process data of historical production batches, find corresponding production link data in time, and establish a matching relationship between defect timing evolution and each link of the production process, for example, a time point of defect occurrence corresponds to an extrusion link in the production process, and process parameter changes in this link may be related to the generation of defects.

[0126] Step 181: Based on the matching relationship, a production process key node corresponding to the current defect detection result is traced in the timing direction of the production process, the association between process running parameter deviation features and defect evolution features at each production process key node is mined through feature difference analysis, and the driving logic of different deviation features on the generation and development of defects is identified.

[0127] According to the time sequence of the production process, all possible production process key nodes that may affect the generation of defects are traced back from the matched production link, including raw material mixing, extrusion, cooling, traction and other links; the process running parameters at each key node are extracted and compared with the standard running parameters of the node to obtain parameter deviation characteristics; the defect evolution characteristics in the current defect detection results are extracted, including the generation time, development rate and morphological change of the defect; the correlation between the parameter deviation characteristics and the defect evolution characteristics is calculated by using correlation analysis rules to mine the correlation between the two; based on the correlation result, the driving logic of different deviation characteristics on the generation and development of defects is identified, for example, the raw material mixing ratio deviation of the raw material mixing link will lead to uneven composition of PVC compound, and then internal hidden defects will be generated in the extrusion link, and this deviation characteristic is the driving factor of defect generation.

[0128] Step 182: Based on the internal correlation, the defect traceability core node is located, the process deviation propagation path at the defect traceability core node is identified, and defect traceability logic information including traceability node positioning result, deviation propagation logic and defect cause correlation analysis is generated.

[0129] The internal correlation between the parameter deviation characteristics and the defect evolution characteristics is analyzed by using the causal diagram analysis method, the most likely production process key node that leads to the generation of defects, i.e. the defect traceability core node, is determined; the process deviation at the defect traceability core node is analyzed to determine how it propagates to the subsequent production link, for example, the mixing ratio deviation of the raw material mixing link will lead to uneven composition of the melted PVC compound, which will be further amplified in the extrusion link, and finally form internal hidden defects, and the deviation propagation path is identified; the traceability node positioning result, deviation propagation path, defect cause correlation analysis and other contents are integrated into defect traceability logic information, which includes the name of the core node, the specific content of the deviation, the process of deviation propagation, detailed analysis of the cause of the defect, etc.

[0130] Step 183: The defect traceability logic information is data interfaced with the production process control strategy, and a defect traceability report containing defect traceability conclusion, process optimization direction and control focus is output, and based on the defect traceability report, process adjustment logic instructions are generated, which are injected into the corresponding process nodes of the PVC compound production process.

[0131] The defect traceability logic information is input into the production process management system and is connected with the production process management strategy stored in the system, for example, the deviation information of the core node is matched with the management strategy of the node to find the corresponding optimization direction; a defect traceability report is generated based on the connection result, which contains defect traceability conclusion, process optimization direction, management focus and the like, for example, the conclusion is that the deviation of raw material mixing link causes defects, the optimization direction is to improve the accuracy of raw material ratio, and the management focus is the parameter monitoring of the raw material mixing link; a process adjustment logic instruction is generated based on the process optimization direction in the defect traceability report, which contains the process node to be adjusted, the content to be adjusted, the standard to be adjusted and the like; the process adjustment logic instruction is injected into the corresponding process node of the PVC compound production process, for example, the parameter monitoring rule of the raw material mixing link is updated, the control algorithm of the ratio is adjusted and the like.

[0132] Optionally, the embodiments of the present application further comprise:

[0133] Step 190: The defect detection result is multi-dimensionally disassembled to obtain multi-dimension defect characterization data covering defect level information, defect spatial distribution range information, defect time evolution trend information and defect influence range prediction information.

[0134] The defect severity score is extracted from the defect detection result, which is mapped to the preset defect level information, for example, a high score corresponds to a high defect level; the coordinate information and coverage area size of the defect distribution range and the like are extracted and integrated into the defect spatial distribution range information; the evolution rate, diffusion direction and predicted evolution of the defect and the like are extracted and integrated into the defect time evolution trend information; the potential influence assessment of the defect on the subsequent production link and product quality is extracted and integrated into the defect influence range prediction information; the four types of information are integrated into multi-dimension defect characterization data, which is stored in a structured JSON format.

[0135] Step 191: An information interaction logic between the multi-dimension defect characterization data and the process characteristic requirement label of the PVC compound processing link and the storage condition information of the storage management link is established.

[0136] The process characteristic requirement label of the PVC compound processing link is extracted from the processing system, including the applicable processing technology type, processing parameter range and the like of the compound of different defect levels; the storage condition information of the storage management link is extracted from the storage management system, including the storage area, storage temperature and storage humidity requirement of the compound of different defect levels; an information interaction logic between the multi-dimension defect characterization data and the process characteristic requirement label and the storage condition information is established, for example, when the defect level is high, the corresponding processing technology type is limited to a specific process, and the storage area is a specific isolation area, and the logic is stored in the form of a rule engine.

[0137] Step 192: According to the information interaction logic, the multi-dimensional defect characterization data is classified and pushed to the processing system and the warehouse management system, and the defect adaptability evaluation results based on the job requirements fed back by the two systems are received. The adaptability evaluation results include the adaptability of the defect rubber compound to the processing technology and the requirement adaptability to the storage conditions.

[0138] According to the rules in the information interaction logic, the corresponding part of the multi-dimensional defect characterization data is pushed to the processing system and the warehouse management system, such as the defect level information and the defect spatial distribution range information are pushed to the processing system, and the defect level information and the defect influence range prediction information are pushed to the warehouse management system. Based on the received information, the processing system evaluates the adaptability of the defect rubber compound to the processing technology in combination with the current processing task requirements, such as the current processing task requires the use of low defect level rubber compound, if the pushed defect level is high, the adaptability is low. Based on the received information, the warehouse management system evaluates the requirement adaptability of the defect rubber compound to the storage conditions in combination with the current storage resource situation, such as the current storage isolation area is full, if the pushed defect level is high, the adaptability is low. The adaptability evaluation results fed back by the two systems are received, and the results are expressed in the form of standardized scores to represent the adaptability.

[0139] Step 193: Based on the defect adaptability evaluation results, the process adjustment logic for adapting the characteristics of the defect rubber compound is generated for the processing link, and the warehouse adjustment logic for distinguishing the storage area, storage method and storage priority of the defect rubber compound is generated for the warehouse management link.

[0140] If the adaptability feedback by the processing system is low, the process adjustment logic is generated, including adjusting the processing process parameters, replacing the processing equipment, adjusting the processing sequence, etc. to adapt to the characteristics of the defect rubber compound. If the adaptability is high, the logic to keep the current processing technology is generated. For the warehouse management link, if the adaptability feedback by the warehouse management system is low, the warehouse adjustment logic is generated, including adjusting the storage area, storage method and storage priority of the defect rubber compound, such as adjusting the high defect level rubber compound to the standby isolation area for storage, using sealed storage method, and reducing the storage priority. If the adaptability is high, the logic to keep the current storage method is generated.

[0141] Step 194: The process adjustment logic and the warehouse adjustment logic are respectively imported into the corresponding systems and executed, and the product processing quality data and the warehouse management operation data after the execution of the adjustment by the systems are collected and recorded.

[0142] The process adjustment logic is imported into the processing system, and the processing system adjusts the corresponding processing process or flow according to the logic; the storage adjustment logic is imported into the warehouse management system, and the warehouse management system adjusts the corresponding storage mode according to the logic; the product processing quality data after the adjustment of the processing system is synchronously collected, including the size precision and mechanical properties of the product after processing; the storage management operation data after the adjustment of the warehouse management system is collected, including the storage state of the defective rubber compound and the utilization rate of the storage resources; and the above data is recorded and stored for subsequent process optimization and storage management optimization.

[0143] In combination with the above content of the embodiments of the present application, the following is described through a complete application scenario example. In a large PVC pipe production base, a high-speed extrusion production line produces a batch of national standard PVC pipes with a diameter of 110 mm for drainage at a conveying rate of 55 meters per minute. A PVC compound defect detection system is embedded in the production process to realize real-time monitoring. The system first collects continuous multiple frames of visual imaging data through three industrial line array cameras deployed above the outlet of the heating and extrusion section of the assembly line. The camera sampling frame rate is set to 50 frames per second to ensure that each frame of image covers a 1.1-meter-long PVC compound area. The single-frame image is stored in an uncompressed BMP format, and the pixel dimension is a two-dimensional matrix of 4096x2048, with each pixel point corresponding to an RGB three-channel grayscale value. The real-time data of the production conveying state monitoring system are synchronously accessed, including the conveying roller speed of 260 revolutions per minute, the conveying belt tension of 16 kilonewtons, and the PVC compound outlet thickness of 2.3 millimeters.

[0144] Based on the collected continuous visual imaging data and production state monitoring information, the system segments the PVC compound area frame by frame: the background difference method is used to remove the background pixels of the assembly line stainless steel conveying belt, guide roller, etc., and only the pixel area corresponding to the PVC compound is retained. Then the top coordinates of the minimum bounding rectangle of each area are extracted as the morphological contour features. The displacement vectors of 200 feature points in the compound area in adjacent frames are calculated by the optical flow method, and the position offset trajectory of the compound between frames is fitted. The time scale deviation of the trajectory is corrected in combination with the conveying roller speed data to avoid trajectory distortion caused by slight fluctuations in the conveying rate. Finally, the inter-frame feature correlation information is integrated, including the frame pair serial number, the position offset vector set, the morphological contour overlap ratio, the contour deviation feature vector, and the conveying state correction coefficient of each group of adjacent frames.

[0145] Subsequently, the system determines the time sequence association logic of the continuous frames based on the inter-frame feature association information, and takes the previous frame's rubber movement features as the input constraint of the next frame's defect tracking: first, the movement direction vector, movement acceleration, movement trajectory curvature and other features of the previous frame's rubber are extracted, and after Z-Score standardization processing, they are taken as input constraints to limit the area range and movement trend prediction direction of the next frame's defect tracking; an inter-frame jump relationship model containing a Siamese network feature matching module and a regularization screening module is constructed, the constraint parameters are imported into the corresponding modules, the cosine similarity of the adjacent frame feature vectors is calculated, and the shape contour similarity is corrected, the defect tracking nodes with qualified feature matching degree are screened out, the nodes are connected in time sequence order in a chain, and the path is smoothed and corrected, to generate a cross-frame defect tracking link covering the continuous frames. Through link tracking, the system locates an initial abnormal shape area in the 120th to 138th frames of images, the area's edge roughness exceeds the preset threshold compared with the standard template, and the feature consistency is maintained in 19 consecutive frames, the shape evolution trend of which is consistent with the development law of the concave defect, and it is determined as a suspected surface defect area, and the morphological evolution parameters such as the center coordinates, area and edge roughness of the area in each frame are recorded synchronously.

[0146] For the suspected surface defect area, the system constructs a feature mining unit with its 19 consecutive frames of images as a time window, uses a multi-scale Canny edge detection algorithm to extract the edge contour point set of each frame, and obtains the fused edge contour corresponding to the unit through feature point matching and weighted fusion. Further mining features based on the fused contour: identifying the nested relationship between the main contour and the internal secondary contour as the contour level feature, obtaining the texture density and texture direction as the texture feature through gray level co-occurrence matrix analysis, and calculating the gray gradient difference between the defect area and the surrounding normal area as the gray gradient feature; performing hierarchical correlation analysis on the three types of features to generate a feature hierarchical correlation graph, and matching it with a preset defect semantic label library to assign semantic mapping labels such as “surface depression”, “uneven texture” and “gray level mutation”, and form a complete surface defect feature. The surface defect feature is input into a hierarchical reasoning mining model based on the Transformer encoder architecture, and after four layers of feature analysis, a defect shape vector set is extracted, which is matched with the associated logic database to reversely deduce that the internal implicit defect corresponding to the area is a “melting uneven bubble defect”.

[0147] The system combines the time sequence change rule in the inter-frame feature association information to mine the distribution trend of the internal implicit defect: the defect diffuses along the PVC compound conveying direction, the diffusion rate is positively correlated with the compound conveying rate, and the influence range covers the current detection area and the compound of about 3.2 meters in length. Finally, the defect detection result containing the defect type, distribution range, evolution trend and influence degree is generated, and the result is associated with the key process parameters such as the raw material ratio 6.2:2.8:1 (PVC resin: plasticizer: stabilizer), the extrusion temperature 187 degrees Celsius, and the screw rotation speed 120 revolutions per minute. It is found that the fluctuation of ±2 degrees Celsius of the extrusion temperature is strongly associated with the defect formation, and then the process control direction of optimizing the extrusion temperature control precision is generated, and the dynamic parameter control instruction is output to adjust the temperature closed-loop control algorithm of the extrusion link. At the same time, the system matches the defect detection result with the historical production data, traces back to the defect traceability core node of the extrusion section heating module, generates a defect traceability report containing the traceability conclusion and process optimization direction, and synchronously updates the defect compound storage strategy of the warehouse management system, realizes the whole-process closed-loop management from defect detection to production optimization, and effectively improves the production stability and product qualification rate of PVC compound.

[0148] The embodiments of the present application realize the full-dimensional accurate identification of PVC compound defects from the surface to the interior. In detail, the cooperative analysis of production state monitoring information and visual imaging data effectively eliminates the interference of production conveying rate fluctuation on feature extraction, ensures the time sequence consistency and accuracy of inter-frame feature association information; the constraint mechanism of the motion feature of the previous frame compound to the defect tracking of the next frame greatly improves the spatio-temporal coherence of the cross-frame defect tracking link, and avoids defect omission and misjudgment caused by independent frame analysis; hierarchical analysis and semantic mapping of surface defect features provide multi-dimensional feature support for reverse deduction of internal implicit defects, breaking through the limitations of traditional defect judgment relying only on surface morphology; internal implicit defect distribution trend is mined by combining inter-frame time sequence change rule, realizing the upgrade from "static recognition" to "dynamic prediction" of defects, which can perceive the diffusion range and evolution direction of defects in advance. The embodiments of the present application deeply integrate visual perception, time sequence association and hierarchical reasoning as a whole, and improve the accuracy and foresight of PVC compound defect detection.

[0149] Please refer to Figure 2 , which is a schematic diagram of the basic structure of a PVC compound defect detection system 200 provided by the embodiments of the present application. The PVC compound defect detection system 200 comprises: a processor 201; a storage device 202, which stores a computer program 2020; a network interface 203, which is used to provide network communication function; when the computer program 2020 is executed by the processor 201, the processor 201 realizes the AI visual recognition based PVC compound defect detection method.

[0150] Please refer toFigure 3 The function module block diagram of the PVC compound defect detection device is provided in the embodiments of the present application. The PVC compound defect detection device comprises:

[0151] An imaging data acquisition module is configured to acquire continuous multiple frames of visual imaging data of the PVC compound in a production conveying process.

[0152] A feature correlation extraction module is configured to extract inter-frame feature correlation information by analyzing position offset trajectories and morphological profile changes of the PVC compound in the continuous multiple frames of visual imaging data in combination with production conveying state monitoring information of the PVC compound.

[0153] A tracking link generation module is configured to determine time sequence correlation logical information of the continuous multiple frames of visual imaging data based on the inter-frame feature correlation information, take motion features of a previous frame of compound as input constraints of a defect tracking task of a next frame, and generate a cross-frame defect tracking link in combination with the time sequence correlation logical information.

[0154] A defect region determination module is configured to track morphological abnormal regions of the PVC compound in the continuous multiple frames of visual imaging data by using the cross-frame defect tracking link, and determine a suspected surface defect region from the morphological abnormal regions by time sequence feature correlation analysis.

[0155] A defect feature mining module is configured to perform feature mining processing on the suspected surface defect region by edge profile identification based on time sequence distribution information of the suspected surface defect region, and obtain surface defect features comprising feature hierarchical correlation and semantic mapping labels.

[0156] A defect detection generation module is configured to input the surface defect features into a preset hierarchical reasoning mining model for hierarchical feature analysis, extract a defect morphological vector set corresponding to the surface defect features, inversely deduce an internal implicit defect type corresponding to the suspected surface defect region according to an association logic between the defect morphological vector set and the internal implicit defect, mine an internal implicit defect distribution trend in combination with a time sequence change rule in the inter-frame feature correlation information, and generate a defect detection result of the PVC compound based on the internal implicit defect type and the internal implicit defect distribution trend.

[0157] Please refer to Figure 4 In the intelligent sorting interaction scenario, an intelligent sorting system is provided, which is in communication connection with the PVC compound defect detection system, and is configured to locate and sort out a PVC compound section with defects through the PVC compound defect detection system.

[0158] On the basis described above, a readable storage medium is provided, wherein a program or instructions are stored on the readable storage medium, and the program or instructions are executed by a processor to implement the steps of the above method.

[0159] In addition, it should be noted that the embodiments of the present application also provide a computer program product, which can include a computer program stored in a computer readable storage medium. The processor of the PVC adhesive defect detection system reads the computer program from the computer readable storage medium, and the processor can execute the computer program, so that the PVC adhesive defect detection system executes the foregoing Figure 1 The description of the method in the corresponding embodiment is not repeated here. In addition, the beneficial effects of using the same method are not repeated. For technical details not disclosed in the computer program product embodiments involved in the present application, please refer to the description of the method embodiments of the present application.

[0160] It should be noted that the embodiments in the present specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system or device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

Claims

1. An AI vision recognition-based PVC compound defect detection method, characterized in that, The method comprises: Collecting continuous multiple frames of visual imaging data of the PVC compound in the production conveying process; Extracting inter-frame feature association information by analyzing the position offset trajectory and the shape profile change of the PVC compound in the continuous multiple frames of visual imaging data in combination with the production conveying state monitoring information of the PVC compound: integrating the position offset trajectory, the shape profile change parameter and the corrected conveying state information into the inter-frame feature association information; Determining the time sequence association logic information of the continuous multiple frames of visual imaging data based on the inter-frame feature association information, taking the motion feature of the compound in the previous frame as the input constraint of the defect tracking task in the next frame, generating a cross-frame defect tracking link in combination with the time sequence association logic information: performing feature recognition in the time sequence dimension on the inter-frame feature association information, extracting the position offset amount, the shape profile similarity and the motion rate change parameter of the PVC compound between different frames contained in the inter-frame feature association information; quantifying the association strength between adjacent frames in the continuous multiple frames of visual imaging data through a time sequence association feature matrix and determining the time sequence association logic information of the continuous multiple frames of visual imaging data; creating an inter-frame jump relationship model based on the time sequence association logic information, importing the input constraint into the inter-frame jump relationship model, determining the corresponding relationship of the defect tracking target in adjacent frames through the inter-frame jump relationship model, connecting the defect tracking nodes in each frame based on the corresponding relationship, and generating a cross-frame defect tracking link covering the continuous multiple frames of visual imaging data; Tracking the shape abnormal area of the PVC compound in the continuous multiple frames of visual imaging data using the cross-frame defect tracking link, and determining a suspected surface defect area from the shape abnormal area through time sequence feature association analysis; Based on the time sequence distribution information of the suspected surface defect area, performing feature mining processing on the suspected surface defect area through edge profile recognition to obtain surface defect features containing feature hierarchical association and semantic mapping labels: performing hierarchical association analysis on the extracted profile hierarchical features, texture features and gray gradient features and generating a feature hierarchical association graph, and assigning corresponding semantic mapping labels to each feature based on a pre-set defect semantic label library, fusing the feature hierarchical association graph and the semantic mapping labels to obtain surface defect features containing feature hierarchical association and semantic mapping labels; Inputting the surface defect features into a pre-set hierarchical reasoning mining model for hierarchical feature analysis, extracting a defect shape vector set corresponding to the surface defect features, inversely deducing the internal implicit defect type corresponding to the suspected surface defect area according to the defect shape vector set and the association logic of the internal implicit defect, mining the internal implicit defect distribution trend in combination with the time sequence change law in the inter-frame feature association information, and generating a defect detection result of the PVC compound based on the internal implicit defect type and the internal implicit defect distribution trend.

2. The method of claim 1, wherein, The time sequence association logic information of the continuous multi-frame visual imaging data is determined based on the inter-frame feature association information, the rubber movement feature of the previous frame is taken as an input constraint of a defect tracking task of the next frame, a cross-frame defect tracking link covering the continuous multi-frame visual imaging data is generated based on the time sequence association logic information, and the cross-frame defect tracking link includes: The feature recognition in the time sequence dimension is performed on the inter-frame feature association information, the position offset, the shape contour similarity, and the movement rate change parameter of the PVC rubber in different frames included in the inter-frame feature association information are extracted; A time sequence association feature matrix is generated based on the position offset, the shape contour similarity, and the movement rate change parameter, the correlation strength between adjacent frames in the continuous multi-frame visual imaging data is quantified through the time sequence association feature matrix, and the time sequence association logic information of the continuous multi-frame visual imaging data is determined; The rubber movement feature corresponding to the previous frame is separated from the inter-frame feature association information, and the rubber movement feature includes a movement direction vector, a movement acceleration, and a movement trajectory curvature; The rubber movement feature is standardized to obtain a standardized movement feature parameter, the standardized movement feature parameter is taken as an input constraint condition of the defect tracking task of the next frame, and the input constraint condition is used to limit the region range and the movement trend prediction direction of the defect tracking of the next frame; An inter-frame jump relationship model is created based on the time sequence association logic information, the input constraint condition is introduced into the inter-frame jump relationship model, the corresponding relationship of the defect tracking targets in adjacent frames is determined through the inter-frame jump relationship model, the defect tracking nodes in each frame are connected based on the corresponding relationship, and a cross-frame defect tracking link covering the continuous multi-frame visual imaging data is generated.

3. The method of claim 2, wherein, The time sequence association logic information of the continuous multi-frame visual imaging data is determined based on the inter-frame feature association information, the rubber movement feature of the previous frame is taken as an input constraint of a defect tracking task of the next frame, a cross-frame defect tracking link covering the continuous multi-frame visual imaging data is generated based on the time sequence association logic information, and the cross-frame defect tracking link includes: An initial model network of the inter-frame jump relationship model is constructed based on the inter-frame correlation strength and the time sequence arrangement rule in the time sequence association logic information, an inter-frame feature matching module and a tracking target screening module are embedded in the initial model network, the inter-frame feature matching module is used to realize the matching of the related features of the PVC rubber in adjacent frames, and the tracking target screening module is used to screen out a target region meeting the defect tracking condition from the PVC rubber region; The input constraint condition corresponding to the standardized movement feature parameter is disassembled into a region range constraint parameter and a movement trend constraint parameter; wherein the region range constraint parameter is used to define the effective region boundary of the defect tracking in the next frame, and the movement trend constraint parameter is used to limit the potential movement direction and the movement rate range of the defect tracking target, and the disassembled constraint parameters are introduced into the corresponding modules of the inter-frame jump relationship model. The inter-frame feature matching module extracts a first motion feature corresponding to a previous frame defect tracking node and a second motion feature of PVC material in an effective area of a next frame, calculates a similarity of the first motion feature and the second motion feature, and generates a feature similarity matrix; the feature similarity matrix is corrected in combination with a shape contour similarity in the time sequence correlation logic information, and a corrected feature matching result is obtained; Based on the corrected feature matching result, the tracking target screening module screens out a region in the next frame that has a feature matching degree with the previous frame defect tracking node satisfying a preset requirement as a defect tracking node of the next frame, and synchronously records a position mapping relationship and a feature inheritance relationship between adjacent frame defect tracking nodes; Taking a time sequence order of the continuous multiple frames of visual imaging data as an axis, according to the position mapping relationship and the feature inheritance relationship, the defect tracking nodes in each frame of visual imaging data are connected in turn in a chain connection mode, and in the connection process, the connection path is smoothed and corrected in combination with a position offset trajectory in the inter-frame feature correlation information to eliminate connection deviation caused by inter-frame offset, and a cross-frame defect tracking link covering the continuous multiple frames of visual imaging data and having time sequence continuity is generated.

4. The method of claim 1, wherein, The cross-frame defect tracking link is used to track the abnormal shape region of the PVC material in the continuous multiple frames of visual imaging data, and a suspected surface defect region is determined from the abnormal shape region through time sequence feature correlation analysis, including: According to an inter-frame jump order of the cross-frame defect tracking link, the PVC material regions corresponding to each frame in the continuous multiple frames of visual imaging data are positioned in turn, the PVC material regions in each frame are morphologically compared based on a preset standard material shape template, and regions having a morphological deviation from the standard material shape template are marked as initial abnormal shape regions; Morphological feature parameters of each initial abnormal shape region are extracted, including region area, edge roughness, shape irregularity, and gray scale distribution uniformity; A time sequence correlation relationship of the initial abnormal shape regions between frames is established based on the cross-frame defect tracking link; The morphological feature parameter similarity of the initial abnormal shape regions in adjacent frames is determined through the time sequence correlation relationship, and initial abnormal shape regions that are continuous in time sequence and have a morphological feature parameter similarity satisfying a preset threshold are extracted; The extracted initial abnormal shape regions are subjected to time sequence trajectory fitting, the morphological evolution trend of the extracted initial abnormal shape regions in the continuous multiple frames is obtained, the initial abnormal shape regions having the morphological evolution trend matching a preset defect development rule are determined as suspected surface defect regions, and the position coordinates and morphological evolution parameters of the suspected surface defect regions in each frame of visual imaging data are synchronously recorded.

5. The method of claim 1, wherein, Based on the time sequence distribution information of the suspected surface defect regions, the suspected surface defect regions are subjected to feature mining processing through edge contour recognition, and surface defect features including feature level correlation and semantic mapping labels are obtained, including: determine the occurrence time, duration and inter-frame position offset of the suspected surface defect region in the continuous multi-frame visual imaging data based on the time sequence distribution information of the suspected surface defect region, and divide the time window for feature mining according to the time sequence distribution information, taking the suspected surface defect regions in the same time window as a feature mining unit; perform edge contour recognition on the suspected surface defect regions in each feature mining unit by using a multi-scale edge detection algorithm, extract edge contour point sets of the suspected surface defect regions in each frame, and obtain a fused edge contour corresponding to the feature mining unit through matching and fusion of the edge contour point sets between frames; perform feature mining processing based on the fused edge contour, and extract contour level features, texture features and gray gradient features; wherein the contour level features include the nesting relationship between the main contour and the secondary contour, the texture features include the texture density and texture direction of the defect region, and the gray gradient features include the gray change rate of the defect region and the surrounding normal region; perform hierarchical correlation analysis on the extracted contour level features, texture features and gray gradient features, and generate a feature hierarchical correlation graph, assign corresponding semantic mapping labels to each feature based on a pre-set defect semantic label library, fuse the feature hierarchical correlation graph and the semantic mapping labels, and obtain a surface defect feature including feature hierarchical correlation and semantic mapping labels.

6. The method according to any one of claims 1 to 5, characterized in that, The surface defect feature is input into a pre-set hierarchical reasoning mining model for hierarchical feature analysis, and a defect morphology vector set corresponding to the surface defect feature is extracted, including: The surface defect feature is input into a pre-set hierarchical reasoning mining model; wherein the hierarchical reasoning mining model includes a feature input layer, a first feature analysis layer, an intermediate feature analysis layer, and a second feature analysis layer, the standardized surface defect feature enters the first feature analysis layer through the feature input layer, and a first layer defect feature is obtained through preliminary feature extraction of the standardized surface defect feature by a first layer convolution kernel; The first layer defect feature is input into the intermediate feature analysis layer, and the feature components with correlation in the first layer defect feature are fused and the redundant feature components are filtered out through a feature fusion algorithm, and a middle layer fusion defect feature is obtained; The middle layer fusion defect feature is input into the second feature analysis layer, and the middle layer fusion defect feature is analyzed and defect morphology correlation information is mined through a deep neural network, and a multi-dimensional defect morphology feature component is output; The multi-dimensional defect morphology feature component output by the second feature analysis layer is subjected to vector quantization processing, each feature component is converted into a unified dimensional vector form, and the correlation relationship between each feature component is sorted and integrated, and a defect morphology vector set corresponding to the surface defect feature is extracted.

7. The method according to any one of claims 1 to 5, characterized in that, The defect morphology vector set and the associated logic of the internal implicit defect are used to infer the internal implicit defect type corresponding to the suspected surface defect region, including: A pre-constructed defect morphology vector set and internal implicit defect association logic database stores standard defect morphology vector sets corresponding to different internal implicit defect types and vector matching rules for defining similarity calculation methods and matching thresholds of the defect morphology vector set to be matched and the standard defect morphology vector set; The extracted defect morphology vector set is input into the association logic database, and the similarity of the defect morphology vector set and each standard defect morphology vector set in the association logic database is calculated according to the vector matching rule; The standard defect morphology vector set with the highest similarity and meeting the matching threshold is screened out, and the internal implicit defect type corresponding to the standard defect morphology vector set is determined as the candidate internal implicit defect type; Based on the feature hierarchy association information in the surface defect feature, the candidate internal implicit defect type is verified to determine whether the defect formation mechanism corresponding to the candidate internal implicit defect type is consistent with the feature evolution law of the surface defect feature; If consistent, the candidate internal implicit defect type is determined as the internal implicit defect type corresponding to the suspected surface defect region; if not consistent, the matching threshold in the vector matching rule is adjusted, and similarity calculation and screening are performed again until an internal implicit defect type consistent with the feature evolution law of the surface defect feature is determined; The internal implicit defect distribution trend is mined by combining the time sequence change law in the inter-frame feature association information, including: Feature parameters related to the time sequence change law are extracted from the inter-frame feature association information, the feature parameters include the overall motion trend of the PVC compound in the continuous multiple frames of visual imaging data, the position offset rate of the defect region, and the morphology expansion rate, and a time sequence change law model is created based on the feature parameters; The feature parameters corresponding to the internal implicit defect type are input into the time sequence change law model, and the mapping relationship between the evolution of the internal implicit defect in the time sequence dimension and the time sequence change law in the inter-frame feature association information is output: the feature parameters corresponding to the internal implicit defect type are extracted from the association logic database, including the typical expansion rate and diffusion direction of the defect; the feature parameters are integrated with the input features of the time sequence change law model and input into the time sequence change law model; the time sequence change law model learns the time sequence association relationship between the features through an internal gating mechanism, and outputs the mapping relationship between the evolution parameters of the internal implicit defect and the time sequence change parameters in the inter-frame feature association information, including the association degree between the defect expansion rate and the overall motion trend of the PVC compound, and the corresponding relationship between the defect diffusion direction and the position offset trajectory; Based on the mapping relationship, the position change, morphology evolution, and range expansion of the internal implicit defect in the subsequent frames are predicted using a time sequence prediction algorithm based on the internal implicit defect features in the continuous multiple frames of visual imaging data; The spatiotemporal distribution map of the internal implicit defect is generated by combining the historical distribution information of the internal implicit defect in the continuous multiple frames of visual imaging data and the predicted subsequent evolution information. By analyzing the density change, range expansion direction and evolution rate of the defect distribution in the space-time distribution map, an internal implicit defect distribution trend is mined, which includes the main diffusion direction, diffusion rate and affected PVC compound area range of the defect.

8. An intelligent sorting system characterized in that, The intelligent sorting system is in communication connection with the PVC compound defect detection system, and is used for positioning and sorting out the PVC compound section with defects through the PVC compound defect detection system. The PVC compound defect detection system comprises: The computer program is executed by the processor, so that the processor implements the AI vision recognition based PVC compound defect detection method in any one of claims 1-7.

9. A readable storage medium, characterized by, The program or instruction stored on the readable storage medium is executed by the processor to implement the AI vision recognition based PVC compound defect detection method in any one of claims 1-7.

Images