Plastic injection molding defect real-time detection method and system and storage medium

Abstract

The invention relates to the field of industrial detection, and discloses a plastic injection molding defect real-time detection method and system and a storage medium. The method comprises the following steps: synchronously imaging a defect area of a plastic injection molding part through a multi-angle camera array, and pre-processing to generate an enhanced defect candidate point feature set; carrying out three-dimensional reconstruction through a stereoscopic vision algorithm and combining with convolutional neural network classification to generate a defect type set; curvature analysis and surface texture gradient analysis are carried out on the defect type set, and shadow texture analysis and texture roughness calculation results are fused to generate a defect cause parameter identification set; adopting a dynamic tracking algorithm to generate a defect space position evolution trend information set; and fusing the defect type set to optimize injection molding process parameters to obtain a mold design improvement data set. Through fusion of multiple technologies such as multi-angle imaging, stereoscopic vision, the convolutional neural network and dynamic tracking, the problems of low detection precision, weak interference resistance and no closed-loop optimization in the prior art are solved, and the injection molding defect detection efficiency and the process improvement scientificity are improved.

Description

Technical Field

[0001] This application relates to the field of industrial inspection technology, and in particular to a method, system and storage medium for real-time detection of defects in plastic injection molding. Background Technology

[0002] Thin-walled injection molded parts are indispensable components in modern manufacturing, widely used in automobiles, electronics, and medical devices. Their quality directly affects product performance and reliability. High-quality thin-walled injection molded parts require minimal surface and internal defects, especially micron-level white spot defects, which can lead to reduced part strength or unacceptable appearance, making them a critical issue that urgently needs to be addressed in production. However, existing inspection and optimization technologies have proven insufficient in complex production environments, making it difficult to meet the demands for high precision and intelligent manufacturing.

[0003] Existing injection molding defect detection technologies suffer from several significant shortcomings: First, they are weakly resistant to interference in complex environments. When faced with uneven lighting, traditional imaging methods easily misjudge normal surface textures of injection molded parts as white spot defects, or miss tiny white spots due to light and shadow obscuring the image. Humidity fluctuations can cause camera lens fogging and reduced image contrast, further exacerbating defect identification errors. When parts are slightly misaligned, static inspection equipment cannot dynamically adjust the imaging angle, leading to misjudgments of defect locations. Second, they lack spatial information acquisition capabilities. Relying solely on two-dimensional images cannot distinguish the dimensional attributes of defects. For example, tiny white spots adhering to the surface and internally encapsulated air bubbles have similar characteristics in two-dimensional images, easily leading to confusion of defect types and affecting subsequent causal analysis. Third, the correlation between defects and process parameters is broken. Existing technologies can only identify the existence of defects but cannot trace them back to root process problems such as uneven mold temperature distribution, injection pressure fluctuations, and abnormal material flow through the spatial location and morphological characteristics of the defects. This results in process adjustments relying on manual experience, lacking data support, and making it difficult to reduce defect generation at the source. Fourth, the dynamic tracking and prediction capabilities are insufficient. It is impossible to monitor the positional evolution trend of defects in the injection molding process in real time, nor can it predict the development of defects based on historical data and real-time environmental parameters. This results in control commands lagging behind the defect formation process and the inability to intervene in a timely manner.

[0004] To address the aforementioned deficiencies, this application acquires the original image set through multi-angle imaging and preprocesses it to enhance features. The spatial location of the defect is obtained through three-dimensional reconstruction using a stereo vision algorithm. The defect type is classified by combining a neural network. Then, the cause of the defect is identified through multi-view curvature and texture analysis. The evolution trend of the defect is dynamically tracked. Finally, the process parameters are optimized based on the trend, and mold design improvement data is generated, forming a closed loop of "imaging-analysis-tracing-optimization". This solves the problems of low accuracy, weak anti-interference, and lack of closed-loop optimization in existing detection technologies, improves the efficiency of injection molding defect detection and the scientific nature of process improvement, and meets the quality control requirements of high-precision injection molding production. Summary of the Invention

[0005] This application provides a method, system, and storage medium for real-time detection of defects in plastic injection molding, which solves the problems of low accuracy, weak anti-interference, and lack of closed-loop optimization in existing detection technologies, improves the efficiency of injection molding defect detection and the scientific nature of process improvement, and meets the quality control requirements of high-precision injection molding production.

[0006] In a first aspect, this application provides a method for real-time detection of defects in plastic injection molding, the method comprising: Step S101: Simultaneously image the defect area of ​​the plastic injection molded part using a multi-angle camera array to obtain an original image set; preprocess the original image set to extract a clear feature set and generate an enhanced defect candidate point feature set; Step S102: Perform three-dimensional reconstruction of the defect candidate point feature set using a stereo vision algorithm to generate a defect spatial location information set; use a convolutional neural network to extract features from the defect spatial location information set, determine the distinction criteria between surface white spots and internal bubbles in plastic injection molding defects based on the extracted features, classify defects based on the distinction criteria, and generate a classified defect type set. Step S103: Obtain defect curvature change detection data by performing curvature analysis and surface texture gradient analysis on the defect type set respectively; generate a defect cause parameter identification set by fusing the texture analysis results with the defect curvature change detection data, wherein the texture analysis includes light and shadow texture analysis and texture roughness calculation; Step S104: Based on the defect cause parameter identification set, use a dynamic tracking algorithm to update the spatial location change of the defect and generate a spatial location evolution trend information set of the defect. Step S105: Based on the spatial location evolution trend information set of the defects, fuse the defect type set, optimize the injection molding process parameters, and obtain the mold design improvement dataset.

[0007] Secondly, this application provides a real-time detection system for defects in plastic injection molding, the system comprising: The imaging preprocessing module is used to simultaneously image the defect area of ​​the plastic injection molded part through a multi-angle camera array to obtain the original image set, preprocess the original image set to extract a clear feature set, and generate an enhanced defect candidate point feature set. The defect classification module is used to generate a set of spatial location information of defects by 3D reconstruction of the feature set of defect candidate points using a stereo vision algorithm, and to extract features using a convolutional neural network to determine the distinction criteria and classify and generate a set of defect types. The curvature and texture analysis module is used to perform curvature and surface texture gradient analysis on the defect type set to obtain defect curvature change detection data, and to fuse the texture analysis results to generate a defect cause parameter identification set. The dynamic tracking module is used to update the spatial location changes of defects based on the defect cause parameter identification set and the dynamic tracking algorithm to generate a set of defect spatial location evolution trend information. The parameter optimization module is used to optimize injection molding process parameters by integrating the defect type set with the defect spatial location evolution trend information set, and to obtain a mold design improvement dataset.

[0008] This application proposes a method, system, and storage medium for real-time detection of defects in plastic injection molding, solving the problems of low accuracy, weak anti-interference, and lack of closed-loop optimization in existing detection technologies. It improves the efficiency of injection molding defect detection and the scientific basis of process improvement, meeting the quality control requirements of high-precision injection molding production. Compared with existing technologies, the beneficial effects of this application's technical solution are at least as follows: First, by using multi-angle camera array imaging and image preprocessing to remove interference, combined with stereo vision 3D reconstruction and convolutional neural network extraction of depth and volume features, the surface white spots and internal bubbles can be accurately distinguished, improving the accuracy of plastic injection molding defect type identification and solving the problem that existing technologies are unable to accurately classify defect categories.

[0009] Secondly, by fusing diffuse reflection and direct light imaging, activating camera mode, and combining Kalman filtering and particle filtering to correct for temperature and humidity interference, the detection stability in complex environments is improved, and the problem of distorted detection results caused by uneven illumination and humidity fluctuations is solved.

[0010] Third, by generating a set of defect cause parameters through curvature and texture analysis, and associating them with process parameters such as mold temperature and injection pressure, the scientific nature of process optimization is improved, solving the problem that existing technologies cannot trace the root cause of defects and have no basis for process adjustment.

[0011] Fourth, by optimizing process parameters through genetic optimization algorithms and updating the database in conjunction with execution feedback, a closed loop of "collection-identification-optimization-feedback" is constructed, which improves the continuous improvement capability of injection molding production quality and solves the problems of lagging control and lack of closed-loop optimization in existing technologies. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating the real-time detection method for plastic injection molding defects in this application. Figure 2 The image shows the imaging and feature analysis results of plastic injection molding defect detection in this application. Figure 3This is a verification diagram of the results of plastic injection molding defect detection and optimization in this application. Figure 4 This is a schematic diagram of the structure of the real-time detection system for plastic injection molding defects in this application; Detailed Implementation

[0014] This application provides a method, system, and storage medium for real-time detection of defects in plastic injection molding. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0015] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of a real-time detection method for plastic injection molding defects in this application includes: Step S101: Simultaneously image the defect area of ​​the plastic injection molded part using a multi-angle camera array to obtain an original image set; preprocess the original image set to extract a clear feature set and generate an enhanced defect candidate point feature set.

[0016] In one specific embodiment, step S101, obtaining the original image set, may specifically include the following steps: The multi-angle camera array is deployed at the plastic injection molding part inspection station; The image data captured by the multi-angle camera array is preliminarily processed to obtain a processed image set. All images in the processed image set containing micron-level white spot defects are used as the original image set. The positional uncertainty of defects in the original image set is calculated by calculating the pixel deviation of the same defect in images from different viewpoints, wherein the pixel deviation is used to characterize the positional uncertainty; The size variability of defects in the original image set is calculated by calculating the size difference of the same defect in images from different viewpoints, wherein the size difference is used to characterize the size variability; The image data containing the location uncertainty and the size variability are integrated to generate the original image set.

[0017] Specifically, a multi-angle camera array is deployed at the plastic injection molding part inspection station. This array contains at least three cameras with different viewing angles to simultaneously image micron-sized white spot areas on and inside the plastic injection molding part. Synchronous imaging is triggered by a central clock signal to capture images simultaneously from all cameras, avoiding deviations caused by time differences. During imaging, the effects of uneven lighting and humidity must be considered. For example, when the humidity exceeds 60%, the camera lens automatically activates a defogging mode, and the intensity of the auxiliary LED light source array is adjusted based on real-time light sensor feedback to compensate for uneven lighting. The image data captured by the multi-angle camera array undergoes preliminary processing to filter out all images containing micron-sized white spot defects. These images are used as the original image set. Images without defects or with unclear defects are excluded during the preliminary processing to ensure that the original image set contains only valid inspection objects.

[0018] When calculating the location uncertainty of defects in the original image set, white point feature points are first extracted from images from each viewpoint by thresholding (above 150) grayscale values. Then, the SIFT feature point matching algorithm is used to extract key points from each image, and the corresponding points are fitted using the least squares method. The pixel deviation of the same defect in images from different viewpoints is calculated, and this pixel deviation represents the location uncertainty. For example, an average deviation of 3.5 pixels is obtained. The pixel deviation calculation is based on the coordinate difference of the same defect in images from different viewpoints, and the magnitude of the deviation is measured by Euclidean distance. This process can quantify the fluctuation of defect location caused by viewpoint differences, camera shake, or small changes in the position of the injection molded part, solving the technical problem in the background technology that traditional detection methods cannot accurately locate defects due to the non-fixed position of the part. When calculating the size variability of defects in the original image set, the Canny edge detection algorithm is used to delineate the white point boundaries. The bounding box area of ​​the same defect under different viewpoints is calculated, and the size difference under different viewpoints represents the size variability. For example, the defect area is 100 pixels from the front view and 80 pixels from the side view, a relative difference of 20%. The size difference is calculated using the variance formula to obtain a variability index of 0.15. This index can reflect the projection distortion caused by the tilt of the viewpoint or the size difference of the defect itself in different directions, addressing the problem that traditional detection relies on only a single viewpoint and cannot fully grasp the size characteristics of defects. The image data containing positional uncertainty and size variability are integrated to generate a complete original image set. The integration process ensures that the two types of data are associated with the corresponding images one by one. Each image is accompanied by its corresponding positional uncertainty deviation value and size variability index, forming a structured original image set.

[0019] In one specific embodiment, step S101, generating the enhanced defect candidate point feature set, may specifically include the following steps: The original image set is subjected to grayscale conversion and filtering to remove dust interference and background noise, while enhancing image contrast, to generate a processed image set. Pixels with gray values ​​higher than a preset gray threshold in the processed image are marked as candidate points for white spot defects using a threshold segmentation method. Calculate the standard deviation of pixel intensity in the region where the candidate white spot defects are located to determine the severity fluctuation; Spatial clustering analysis was used to analyze the spatial distribution of the candidate white spot defects and determine the uneven distribution. Extract a clear feature set under the severity fluctuation and the uneven distribution, the clear feature set including the edge contour and gray-scale distribution features of the region where the white spot defect candidate point is located; The clear feature set is associated with the corresponding white spot defect candidate points to form the enhanced defect candidate point feature set.

[0020] Specifically, the original image set is converted to grayscale by assigning different weight values ​​to the red (R), green (G), and blue (B) channels, thus converting the color images in the original image set into grayscale images. For example, the formula is used. This reduces the interference of color information on subsequent defect identification. After grayscale conversion, the grayscale image is filtered using a 3×3 window median filtering algorithm. The grayscale value of the center pixel within the window is replaced with the median of all pixel grayscale values ​​within the window. This operation removes dust interference and background noise from the original image set. For example, when isolated dust points exist in the image, median filtering can effectively adjust their grayscale values ​​to the median of surrounding pixels, preventing dust points from being misjudged as defects, while preserving edge details of the defect area. This solves the technical problem of misjudging or missing defects due to dust interference and background noise in traditional background detection techniques. After filtering, a histogram equalization algorithm is used to enhance image contrast. The cumulative distribution function of pixel grayscale values ​​in the grayscale image is calculated, and the cumulative distribution function value is mapped to a grayscale range of 0~255. This expands the grayscale difference in low-contrast areas, enhances the distinction between white point defects and the background, and addresses the problem of low image contrast caused by uneven illumination. Finally, the processed image set is generated.

[0021] The processed image set is processed using a threshold segmentation method. A grayscale threshold is preset, for example, set to 150. The pixels of each image in the processed image set are traversed, and pixels with grayscale values ​​higher than 150 are marked as candidate white spot defects. This process uses the characteristic that white spot defects have higher grayscale values ​​than the background area to filter and can initially identify potential defect areas. After marking the candidate white spot defects, the standard deviation of the pixel intensity in the region where each candidate white spot defect is located is calculated. Specifically, the region range corresponding to each candidate white spot defect is first determined. For example, a 5×5 pixel region is selected centered on the candidate point as the calculation range, and the mean grayscale value of all pixels in this region is calculated. Then according to the formula Calculate the standard deviation, where Here, represents the grayscale value of a single pixel within the region, and N represents the total number of pixels within the region. This standard deviation characterizes the severity fluctuation; the larger the standard deviation, the greater the difference in pixel grayscale values ​​within the region, and the more pronounced the fluctuation in defect severity. This parameter quantifies the changes in defect severity, solving the problem that traditional detection methods struggle to quantify defect severity fluctuations.

[0022] K-means clustering was used to analyze the spatial distribution of candidate white spot defects. The initial number of clusters was set to 3. Candidate points were assigned to the nearest cluster centers using Euclidean distance, and the cluster centers were iteratively updated until convergence. After clustering, the intra-cluster compactness and inter-cluster separation of each cluster were calculated. The Davis-Boulding index (DBI) was used to assess the uneven distribution. The DBI is calculated as the average ratio of the intra-cluster average distance to the inter-cluster center distance. A higher DBI value indicates a more uneven spatial distribution of candidate white spot defects. This analysis process can clearly define the spatial distribution characteristics of defects on the surface of injection molded parts, solving the problem that traditional inspection methods cannot accurately determine the uniformity of defect distribution in the background technology.

[0023] When extracting clear feature sets under varying severity and uneven distribution, the process first extracts the edge contours of the region where each white point defect candidate is located in the processed image set with labeled white point defect candidate points. Using the Canny edge detection algorithm, the image is first calculated in terms of edge contours... x and y The gradients in the directions are obtained using the Sobel operator, specifically the horizontal gradient Gx and the vertical gradient Gy, through the formula... Calculate the gradient magnitude; then use the formula The gradient direction is calculated, and then two thresholds are set, for example, a high threshold of 200 and a low threshold of 100. Pixels with gradient magnitudes greater than the high threshold are marked as strong edges, and pixels with gradient magnitudes between the high and low thresholds and connected to the strong edges are marked as weak edges. Finally, the strong and weak edges are integrated to form the edge contour of the region where the white spot defect candidate point is located. This contour can clearly reflect the shape and boundary range of the defect. Even with fluctuations in severity, the continuity and integrity of the edges can distinguish the defect from the background, solving the problem of inaccurate feature extraction caused by blurred defect edges in traditional detection.

[0024] After edge contour extraction, grayscale distribution features are extracted for the regions containing candidate white-dot defects. Using the edge contour as the boundary, the region corresponding to each candidate white-dot defect is defined. The grayscale values ​​of all pixels within this region are statistically analyzed to construct a grayscale histogram. The mean and variance of the grayscale histogram are calculated; the mean reflects the overall grayscale level of the region, while the variance reflects the dispersion of grayscale values ​​within the region. Simultaneously, the frequency of grayscale values ​​is statistically analyzed to determine the peak range of grayscale values. This range reflects the concentration of grayscale within the defect region. These parameters form the grayscale distribution features. This feature can quantify the grayscale changes in the defect region, addressing the problem of grayscale differences caused by uneven distribution and solving the technical problem that traditional detection methods cannot quantify the internal grayscale features of defects.

[0025] In the process of integrating edge contours and grayscale distribution features to form a clear feature set, it is ensured that each white-dot defect candidate point corresponds one-to-one with its corresponding edge contour and grayscale distribution features. For example, the association is established through image pixel coordinates, and the coordinate information of each white-dot defect candidate point is stored in the same data structure along with the set of edge contour coordinates and grayscale distribution parameters of the area where the point is located. Through this association method, each white-dot defect candidate point is accompanied by corresponding shape and grayscale features, forming an enhanced defect candidate point feature set. This feature set contains information on the location, shape, and grayscale variation of the defect, which can effectively support the subsequent 3D reconstruction operation of stereo vision algorithms, solving the problem that traditional detection can only provide defect location information and lacks shape and grayscale features, resulting in insufficient accuracy in subsequent analysis.

[0026] Step S102: Perform three-dimensional reconstruction of the defect candidate point feature set using a stereo vision algorithm to generate a defect spatial location information set; use a convolutional neural network to extract features from the defect spatial location information set, determine the distinction criteria between surface white spots and internal bubbles in plastic injection molding defects based on the extracted features, classify defects based on the distinction criteria, and generate a classified defect type set.

[0027] In one specific embodiment, step S102, generating the defect spatial location information set, may specifically include the following steps: If the density of the enhanced defect candidate point feature set is higher than the preset density threshold, the stereo vision algorithm is used to perform feature point correspondence matching on the multi-view data. The disparity is calculated based on the matching results of the feature points, and the three-dimensional coordinates of the white point defect candidate points are calculated in combination with the multi-view geometric constraints. The multi-view geometric constraints are determined by the calibration parameters of the multi-angle camera array. The three-dimensional coordinates are corrected by integrating the effects of temperature changes and camera shake. The temperature changes are corrected for offsets using the thermal expansion coefficient of the material, and the camera shake is corrected for deviations using a dynamic compensation algorithm. The three-dimensional coordinates of all the corrected white spot defect candidate points are integrated to generate a defect spatial location information set.

[0028] Specifically, the density of the enhanced defect candidate point feature set is first calculated, and the number of white point defect candidate points per unit area is counted, for example, per square millimeter. The preset density threshold is 5 points / square millimeter. If the calculated density is higher than 5 points / square millimeter, a stereo vision algorithm is used to perform feature point correspondence matching on the multi-view data. During feature point correspondence matching, the edge contour and gray-level distribution features of each white point defect candidate point are extracted from the enhanced defect candidate point feature set as the matching basis. The SIFT algorithm is used to extract the key point descriptors of the feature points. By calculating the Euclidean distance of the feature point descriptors under different views, feature points with a distance less than a preset threshold (such as 30) are determined as matching pairs, thus completing the feature point correspondence matching of multi-view data. This process can solve the problem that a single view cannot obtain spatial information of defects in traditional detection.

[0029] The disparity is calculated based on the matching results of feature points. For each pair of matched feature points, its pixel coordinates in images from different viewpoints are obtained. The difference in the horizontal pixel coordinates of the same feature point in the left and right viewpoints is calculated, and this difference is the disparity. The three-dimensional coordinates of white point defect candidate points are calculated by combining multi-view geometric constraints. The multi-view geometric constraints are determined by the calibration parameters of the multi-angle camera array, including camera intrinsic parameters such as focal length. f Principal point coordinates Extrinsic parameters: rotation matrix R, translation vector T, based on the principle of triangulation, using parallax. d ,focal length f and the distance between the optical centers of the two cameras B Through formula Calculate the depth information Z of the feature points, and then combine it with the pixel coordinates. Through formula , The three-dimensional coordinates (X,Y,Z) of the feature points are calculated. This process can accurately obtain the spatial location of the defect, solving the technical problem that traditional detection can only obtain two-dimensional images and cannot determine the three-dimensional location of the defect.

[0030] The 3D coordinates are corrected by incorporating the effects of temperature changes and camera shake. Temperature changes are corrected for offsets using the material's coefficient of thermal expansion; for example, when the injection-molded part is made of polypropylene, its coefficient of thermal expansion is... for / ℃, if the ambient temperature change ∆T is 5℃, then according to the formula Calculate the expansion offset in each axis direction in three-dimensional coordinates. Offset compensation is performed on the initially calculated three-dimensional coordinates, where The initial position coordinates are provided. Camera shake is addressed through a dynamic compensation algorithm, employing a Kalman filter. Shake data collected by the camera accelerometer is used as observations to establish state and observation equations. Coordinate deviations caused by camera shake are estimated through prediction and update steps, correcting the 3D coordinates. This correction process reduces the impact of environmental factors on coordinate accuracy, resolving the inaccuracy issues caused by temperature changes and equipment shake in traditional detection methods. The corrected 3D coordinates of all white spot defect candidate points are integrated and indexed by viewpoint and defect number, linking each candidate point to its corresponding corrected 3D coordinates, forming a structured dataset—the defect spatial location information set.

[0031] In one specific embodiment, step S102, generating the classified defect type set, may specifically include the following steps: The three-dimensional coordinates of the defect spatial location information set are normalized. Depth and volume features are extracted from the normalized spatial location information of the defects using a convolutional neural network. The convolutional neural network includes multiple convolutional layers and pooling layers. The depth features are extracted through the convolutional kernels of the convolutional layers, and the volume features are extracted through three-dimensional spatial point cloud analysis. The depth features and volume features are fused using a convolutional neural network to form a comprehensive feature vector. Based on the comprehensive feature vector, the criteria for distinguishing between surface white spots and internal bubbles in plastic injection molding defects are determined. The classification layer of the convolutional neural network classifies the comprehensive feature vector and outputs the category probability corresponding to each defect. Based on the category probability, the defect is determined to be either a surface white spot or an internal bubble. The classification results of all defects are integrated to generate a set of classified defect types.

[0032] Specifically, the three-dimensional coordinates in the defect spatial location information set are normalized. The defect spatial location information set contains the three-dimensional coordinates (X, Y, Z) of white point defect candidate points. During the normalization process, each coordinate value is mapped to the [0,1] interval, using the formula... , and ,in and In the three-dimensional coordinates of all defects x Minimum and maximum values ​​in the axial direction, , , and Similarly, this process eliminates the impact of differences in coordinate magnitudes between injection molded parts of different sizes or under different detection scenarios on subsequent feature extraction, ensuring the consistency of input data for the convolutional neural network and solving the problem of feature extraction deviation caused by inconsistent coordinate magnitudes in traditional detection.

[0033] A convolutional neural network (CNN) is used to extract depth and volume features from a normalized set of defect spatial location information. The CNN contains multiple convolutional and pooling layers; for example, three convolutional layers and two pooling layers are used. The first convolutional layer uses 32 3×3×3 kernels with a stride of 1 and identical padding. Convolution operations are performed on the normalized 3D coordinate data using the formula... Local features are extracted, where W is the convolution kernel weight, b is the bias term, and X is the input data. This layer outputs 32 feature maps. A max-pooling layer is then connected, with a 2×2×2 kernel size and a stride of 2, preserving key information in the feature maps and reducing data dimensionality. The second convolutional layer uses 64 3×3×3 kernels, repeating the convolution and pooling operations. The third convolutional layer uses 128 3×3×3 kernels, performing only convolution operations without connecting to pooling layers. Finally, depth features are extracted through the convolution kernels of multiple convolutional layers. Depth features reflect the extension of defects in the thickness direction of the injection molded part. For example, the depth feature of surface white spots has a smaller value in the Z-axis direction, while the depth feature of internal bubbles has a larger value in the Z-axis direction. Meanwhile, the normalized three-dimensional coordinates are analyzed in three-dimensional space to convert the three-dimensional coordinates into point cloud data. The point cloud is divided into a 16×16×16 voxel grid using the voxelization method. The number of points in each voxel is counted to form a volume feature matrix. The volume features reflect the overall three-dimensional size of the defect. For example, the volume feature matrix of white spots on the surface has fewer non-zero elements, while the volume feature matrix of internal bubbles has more non-zero elements.

[0034] A fully connected layer in a convolutional neural network is used to fuse depth and volume features. The depth features are transformed into a 128-dimensional feature vector using a global average pooling layer, and the volume feature matrix is ​​flattened into a 256-dimensional feature vector. The two feature vectors are then concatenated by the fully connected layer to form a 384-dimensional composite feature vector, which integrates the depth and volume information of the defect. This vector is used to determine the distinction between surface white spots and internal bubbles. For example, if the depth feature dimension is less than 0.3 and the volume feature dimension is less than 0.2 in the composite feature vector, it is identified as a surface white spot; if the depth feature dimension is greater than 0.5 and the volume feature dimension is greater than 0.4, it is identified as an internal bubble. This distinction solves the technical problem that traditional detection methods cannot accurately distinguish between surface white spots and internal bubbles.

[0035] The classification layer of a convolutional neural network classifies the comprehensive feature vector. This classification layer contains a fully connected layer and a softmax activation function. The fully connected layer maps the 384-dimensional comprehensive feature vector into a 2-dimensional vector, corresponding to the category scores of surface white points and internal bubbles, respectively. The softmax activation function is defined by the formula... Convert category scores to category probabilities, where For the first i Class score, Let be the probability of the i-th class. The defect category is determined based on the category probability. If the probability of a surface white spot is greater than 0.5, it is classified as a surface white spot; if the probability of an internal bubble is greater than 0.5, it is classified as an internal bubble. All defect classification results are associated with the defect number and its corresponding three-dimensional coordinates, and integrated to form a classified defect type set.

[0036] Step S103: Obtain defect curvature change detection data by performing curvature analysis and surface texture gradient analysis on the defect type set respectively; generate a defect cause parameter identification set by fusing the texture analysis results with the defect curvature change detection data, wherein the texture analysis includes light and shadow texture analysis and texture roughness calculation.

[0037] In one specific embodiment, step S103, obtaining defect curvature change detection data, may specifically include the following steps: Associate the image data captured by the multi-angle camera array with the defect regions corresponding to the defect type set; Curvature analysis is performed on each defect in the defect type set through multi-view geometric reconstruction, and the reconstruction model is fitted using a surface fitting algorithm to generate the local radius of curvature of the defect surface. The surface texture gradient of the defect region is calculated using a gradient operator to obtain the magnitude and direction information of the texture gradient. The density distribution mapping of the defect region is then constructed by calculating the spatial distribution of pixel intensity. The curvature calculation deviation caused by reflection problem is compensated by the lighting model, and the texture gradient distortion caused by the angle deviation is corrected based on the calibration parameters of the multi-angle camera array. The local radius of curvature, density distribution mapping, compensated curvature, and corrected texture gradient analysis results of each defect are integrated to generate defect curvature change detection data.

[0038] Specifically, the image data captured by the multi-angle camera array is associated with the defect area corresponding to the defect type set. By matching the number of each defect in the defect type set with the image data containing the defect captured by the multi-angle camera array from different perspectives, a mapping relationship between the defect number and the multi-view image data is established. This avoids confusion between different defect data and solves the problem of analysis bias caused by the unclear correspondence between multi-view data and defects in traditional detection.

[0039] Curvature analysis was performed on each defect in the defect type set through multi-view geometric reconstruction. This was achieved using correlated multi-view image data, combined with calibration parameters of a multi-angle camera array (intrinsic parameters including focal length). f Principal point coordinates The extrinsic parameters include the rotation matrix R and the translation vector T), which are used to calculate the 3D point cloud data of the defect region. A surface fitting algorithm is employed to fit the reconstructed model constructed from the 3D point cloud data. The least squares method is used to fit the quadratic surface equation, which has the following form: The equation coefficients are determined by minimizing the sum of squared distances from each point in the point cloud to the surface. Based on the fitted quadratic surface equation, the principal curvatures of the surface are calculated. and The principal curvature is obtained by solving for the eigenvalues ​​of the Weingarten matrix at a certain point on the surface, and then according to the formula... and Calculate the local radius of curvature and The local radius of curvature reflects the degree of curvature of the defect surface. For example, the local radius of curvature of white spots on the surface is usually small, while the local radius of curvature of the defect surface corresponding to internal bubbles is relatively large. This process can quantify the surface morphology of defects and solve the technical problem that traditional detection methods cannot obtain defect curvature information.

[0040] The surface texture gradient of the defect region is calculated using a gradient operator. The Sobe gradient operator is selected and applied to the following areas: x direction and y The direction is used to perform convolution operations on the image data of the defect area. x The Sobel operator template for direction is , y The Sobel operator template for direction is Through formula , calculate x and y The gradient value in the direction, where lmg is the image data of the defect region, and * indicates convolution operation. Then, according to the formula... Calculate the magnitude of the texture gradient according to the formula. The direction of the texture gradient is calculated to obtain its magnitude and direction information. A density distribution map of the defect region is constructed by calculating the spatial distribution of pixel intensity. The grayscale value of each pixel within the defect region is statistically analyzed, and a grayscale histogram is created with pixel coordinates as the horizontal axis and grayscale values ​​as the vertical axis. The proportion of pixels corresponding to each grayscale value interval is calculated, and this proportion data is mapped onto the image pixel coordinate grid to form a density distribution map. This density distribution map reflects the distribution of pixel intensity within the defect region; for example, the density distribution map of an internal bubble defect region typically exhibits non-uniform characteristics.

[0041] To compensate for curvature calculation errors caused by reflection issues, the Lambertian lighting model is selected. The formula for this model is: Where I is the observed pixel intensity, The diffuse reflectance coefficient is... Light source intensity, Let be the angle between the incident ray and the surface normal vector. The diffuse reflectance coefficient is calculated by inversion based on image data captured by a multi-angle camera array under different lighting conditions. ,use Correcting deviations in 3D point cloud data caused by reflection during curvature calculation; for example, when reflection causes abnormal grayscale values ​​for some points in the point cloud data, by... Adjust point cloud coordinates. Based on the calibration parameters of a multi-angle camera array, texture gradient distortion caused by angular deviations is corrected. Using extrinsic parameters (rotation matrix R, translation vector T) from the calibration parameters, texture gradient data from different viewpoints are transformed to a unified coordinate system using coordinate transformation formulas. To correct the angle deviation, among which These are the original texture gradient coordinates. These are the corrected coordinates. This compensation and correction process reduces the impact of ambient lighting and viewing angle deviations on the analysis results, solving the problem of data distortion caused by reflection and angular deviations in traditional detection methods.

[0042] The local curvature radius, density distribution mapping, compensated curvature data, and corrected texture gradient analysis results of each defect are integrated. Data associations are established according to defect numbers so that each defect corresponds to a complete set of curvature and texture gradient data. The resulting dataset is the defect curvature change detection data.

[0043] In one specific embodiment, step S103, generating the defect cause parameter identification set, may specifically include the following steps: Light and shadow texture analysis was performed on the original image set using a lighting model to extract texture features, and texture roughness was calculated on the original image set using surface height distribution statistics. Based on the light and shadow texture analysis results and the texture roughness calculation results, the regional density quantification index of the defect is calculated through the spatial distribution of defect points; The region density quantification index, the size variability of the defect, the severity fluctuation, and the corresponding light and shadow texture and texture roughness data are integrated and combined with the preset defect cause mapping rules to generate a defect cause parameter identification set.

[0044] Specifically, a lighting model, specifically the Lambertian lighting model, is used to perform light and shadow texture analysis on the original image set. Pixel intensity data of the defect area is extracted from the original image set. Combined with different lighting angle information captured by a multi-angle camera array, the diffuse reflectance coefficient is calculated. Texture features are extracted through the distribution differences of the diffuse reflectance coefficient. For example, when the diffuse reflectance coefficient exhibits high-frequency fluctuations in a certain area, it indicates significant changes in light and shadow texture in that area, possibly corresponding to the irregular shape of the defect surface. Simultaneously, texture roughness is calculated on the original image set through surface height distribution statistics. Three-dimensional coordinate data of the defect area is obtained from the defect spatial location information set, and surface height values ​​in the Z-axis direction are extracted to construct a height value dataset. The standard deviation of the height values ​​is calculated; the larger the standard deviation, the rougher the defect surface. This process quantifies the light and shadow and roughness features of the defect surface, solving the problem of insufficient causal analysis caused by the inability of traditional detection methods to accurately extract texture details.

[0045] Based on the results of light and shadow texture analysis and texture roughness calculation, the regional density quantification index of defects is calculated through the spatial distribution of defect points. Coordinate data of white point defect candidate points are obtained from the defect candidate point feature set. The statistical grid size is set to 5×5 pixels, with the bounding rectangle of the defect region as the statistical range. Each grid is traversed, and the number of defect candidate points contained within the grid is counted to calculate the point density of each grid. ,in Let S be the number of candidate points within the grid, and S be the grid area (unit: square pixels). Normalize the point density of all grids to obtain the region density distribution matrix, and select the maximum value in the matrix. As a quantitative indicator of regional density The larger the value, the denser the defect distribution in that area. During the calculation, the diffuse reflectance coefficient distribution from the light and shadow texture analysis results is used to eliminate false candidate points caused by uneven lighting, and the standard deviation of the height value calculated from the texture roughness is used to filter out interference points caused by surface roughness, ensuring that all statistically selected candidate points are genuine defect points. This index reflects the spatial clustering degree of defects, solving the technical problem that traditional detection methods cannot quantify defect distribution density.

[0046] This system integrates regional density quantization indicators, defect size variability, severity fluctuations, and corresponding lighting and texture data. Size variability refers to the size difference of the same defect under different viewpoints; severity fluctuations are the standard deviation of pixel intensity in the defect area; lighting and texture data are the distribution matrix of diffuse reflectance coefficients; and texture roughness data are the height standard deviation. Combined with preset defect cause mapping rules, such as "when the regional density quantization index...", the system... When the parameters are >0.8, size variability >20%, severity fluctuation >15, diffuse reflectance coefficient fluctuation frequency >5Hz, and texture roughness >0.5 micrometers, the corresponding causal parameters are bubble aggregation caused by uneven material flow; when When the size variation is <0.3, the severity fluctuation is <10%, the diffuse reflectance coefficient is uniformly distributed, and the texture roughness is <0.2 micrometers, the corresponding causative parameter is "surface white spots caused by scratches on the mold surface." The integrated data for each defect is traversed, and the mapping rules are used to determine the causative parameter for each defect. All causative parameters are then associated with the defect number, three-dimensional coordinates, and classification results, forming a structured dataset, which is the defect causative parameter identification set. This integration process establishes a direct correlation between defect features and their causes, solving the problem of traditional detection relying on experience to determine causes and lacking data support.

[0047] Step S104: Based on the defect cause parameter identification set, use a dynamic tracking algorithm to update the spatial location change of the defect and generate a spatial location evolution trend information set of the defect.

[0048] In one specific embodiment, step S104 may specifically include the following steps: If the defect cause parameter identification set matches the real-time production data, a dynamic tracking algorithm is used to process multiple frame image sequences to update the spatial location changes of the defects. Environmental factor correction is performed on the changes in the spatial location based on the humidity and temperature values ​​in real-time production data. Kalman filtering and particle filtering are used to predict and optimize the corrected spatial position changes, generating accurate data on the spatial position changes of defects. The precise spatial location change data of the defects are arranged in chronological order to form a trajectory of the defect location change; The influence of material-related randomness and injection molding process parameters on the change trajectory is integrated to generate a set of influencing factors. The material-related randomness is the random variation of polymer particle distribution in the injection molding material, and the injection molding process parameters are injection temperature and pressure data. The change trajectory is optimized using the set of influencing factors to generate a set of information on the spatial location evolution trend of the defect.

[0049] Specifically, the defect cause parameter identification set is first matched with real-time production data. The defect cause parameter identification set includes the variability of regional density quantification index size, the fluctuation of severity, and the corresponding cause parameters. The real-time production data includes parameters such as humidity, temperature, pressure, and injection speed during the injection molding process. During matching, the feature similarity between the two is calculated. For example, the cause of "uneven material flow" in the defect cause parameter identification set is compared with the parameter combination "pressure fluctuation value > 5MPa" and "injection speed deviation > 10%" in the real-time production data. If the similarity is higher than a preset threshold (e.g., 0.8), the match is considered successful. After successful matching, a dynamic tracking algorithm is used to process multi-frame image sequences captured by a multi-angle camera array at a fixed frame rate (e.g., 30fps). The dynamic tracking algorithm calculates the inter-frame defect position change using optical flow. The Lucas-Kanade optical flow algorithm is selected, and for each frame's candidate defect points, the formula is used... and Calculate the optical flow vector ( , ),in , The images are respectively in x , y Gradient of direction, This refers to the grayscale variation between frames. The weights are Gaussian. The spatial position change of the defect between adjacent frames is updated by the optical flow vector. This process can track the defect position in real time, solving the technical problem that traditional detection cannot dynamically grasp the change of defect position.

[0050] Environmental factor corrections are applied to spatial location changes based on humidity and temperature values ​​from real-time production data. Humidity correction is performed according to the formula... Calculate the positional offset caused by humidity ,in Humidity influence coefficient (e.g., for polypropylene materials) =1.2x10 -4 / %RH), h Real-time humidity value (unit: %RH). The initial coordinates of the defect are used; during temperature correction, the coefficient of thermal expansion α of the material is used, through the formula... Calculate the position offset caused by temperature ,in This represents the difference between the real-time temperature and the standard temperature. Spatial location change data is then compared with... , The data are superimposed to obtain the corrected spatial position change. This correction process can eliminate the interference of environmental factors on position data and solve the problem of position tracking deviation caused by changes in humidity and temperature in traditional detection methods.

[0051] When using Kalman filtering and particle filtering to predict and optimize corrected spatial position changes, the corrected spatial position change data is used as the observation value, which includes defects. x , y , z The position increments in the three axes have been calculated, and environmental biases caused by humidity and temperature have been eliminated. When performing Kalman filtering, a state vector and an observation model are first constructed. The state vector contains the real-time position and velocity of the defect, and the observation model establishes the correspondence between the state vector and the observed values. Simultaneously, the covariance matrices of process noise and observation noise are introduced to characterize the uncertainty of the system's dynamic changes and the error of the measurement process, respectively. The Kalman filtering process consists of two steps: prediction and update. In the prediction stage, based on the state estimate from the previous moment and combined with the state transition rules (e.g., the inter-frame time interval is set to 0.03 seconds), the prior state estimate and prior covariance matrix for the current moment are calculated. In the update stage, based on the deviation between the observed values ​​and the prior estimate, the Kalman gain is calculated to adjust the weights, thereby obtaining the corrected posterior state estimate, completing the initial data optimization. Next, particle filtering is performed, generating 800 particles as a representation of the state probability distribution, with each particle corresponding to a possible defect position state. By employing importance sampling, particle weights are calculated based on the degree of agreement between observed values ​​and predicted particle positions. The top 30% of particles by weight are retained and resampled to generate a new particle set, maintaining particle diversity and focusing on high-probability regions. Finally, the mean of all positions in the new particle set is taken as the particle filtering optimization result. The posterior state estimate obtained from Kalman filtering is fused with the mean result of particle filtering, and the position data from both are integrated using an arithmetic mean method to generate accurate defect spatial position change data. This process can effectively suppress position data fluctuations caused by environmental interference (such as humidity fluctuations and temperature changes in the injection molding workshop) and equipment measurement noise (such as camera shake), solving the technical problems of poor stability and large errors in defect position data in traditional detection.

[0052] When arranging precise defect spatial location change data in chronological order, time is used as the horizontal axis (unit: seconds). x , y , z The cumulative position value along the axial direction is the vertical axis (unit: mm). Defect position data is recorded sequentially at time intervals (0.03 seconds). For example, at time 0, the defect position is (2.1, 4.3, 0.6); at time 0.03, the position is updated to (2.12, 4.3, 0.6); at time 0.06, the position is further updated to (2.13, 4.31, 0.61). All position data at all time points are organized into a structured time-location data table in ascending order of time, ensuring that each timestamp uniquely corresponds to a set of data. x , y , zThe position coordinates of the axis form the trajectory of the defect's changing position. This trajectory can intuitively show the movement pattern of the defect in space over time, solving the technical problem that traditional detection can only obtain static position information and cannot dynamically track the evolution of the defect's position.

[0053] Step S105: Based on the spatial location evolution trend information set of the defects, fuse the defect type set, optimize the injection molding process parameters, and obtain the mold design improvement dataset.

[0054] In one specific embodiment, step S105 may specifically include the following steps: Based on the spatial location evolution trend information set of the defects and the defect type set, combined with the thermodynamic model, the genetic optimization algorithm is used to optimize the injection molding process parameters and injection temperature. Based on the optimized injection molding process parameters and injection temperature, the pressure distribution equilibrium state and material flow simulation basis in the mold cavity are obtained through fluid dynamics simulation and finite element analysis. Based on the pressure distribution equilibrium state and the material flow simulation, adjust the cooling control parameters; The optimized injection molding process parameters, injection temperature, pressure distribution balance, material flow simulation basis, and cooling control parameters are integrated to generate the mold design improvement dataset.

[0055] Specifically, based on the spatial location evolution trend information set and defect type set of defects, combined with a thermodynamic model, a genetic optimization algorithm is used to optimize the injection molding process parameters and injection temperature. The spatial location evolution trend information set of defects includes the location trajectory of defects changing over time. For example, an internal bubble defect moves from near the mold gate to the edge of the cavity within 0-10 seconds, and the trajectory data shows that the position increases from 0.5mm to 0.8mm in the z-axis direction. The defect type set clearly identifies the defect as a surface white spot or an internal bubble, such as the defect labeled as an internal bubble. The thermodynamic model is used to simulate the material temperature field distribution during injection molding. The input parameters include the initial injection temperature, mold temperature, and material thermal conductivity. The model calculates the solidification time and temperature gradient of the material under different process parameters. The genetic optimization algorithm uses injection molding process parameters (screw speed, injection pressure) and injection temperature as optimization variables. The screw speed range is set to 100-150 r / min, the injection pressure range to 80-120 MPa, and the injection temperature range to 180-220℃. The fitness function is the defect reduction rate. Fifty parameter combinations are initialized as a population. Through selection, crossover, and mutation operations, the algorithm iterates for 15 generations, retaining the top 30% of parameter combinations in each generation. Finally, the parameter combination that achieves the highest defect reduction rate is selected, such as a screw speed of 130 r / min, an injection pressure of 100 MPa, and an injection temperature of 200℃. This process combines defect evolution and type data to optimize parameters, solving the technical problem of insufficient optimization targeting caused by traditional detection methods that rely solely on experience to adjust the process.

[0056] Based on the optimized injection molding process parameters and injection temperature, fluid dynamics simulation and finite element analysis were used to obtain the pressure distribution equilibrium state and material flow simulation basis within the mold cavity. The fluid dynamics simulation used the Navier-Stokes equations to describe the flow state of the molten material within the mold cavity, inputting optimized injection pressure, screw speed, and material viscosity (e.g., polypropylene viscosity of 1000 Pa·s at 200℃). The mold cavity was divided into 1000 mesh elements, and the pressure value of each element was iteratively calculated. When the pressure deviation between elements was less than 5%, it was considered a pressure distribution equilibrium state. For example, the simulation results showed that the pressure within the mold cavity fluctuated between 95-105 MPa, meeting the equilibrium requirement. The finite element analysis discretized the mold and molten material into finite elements, calculated the velocity field and shear rate of the molten material during the flow process, and generated a material flow path diagram. If the path diagram showed that the molten material uniformly filled the cavity after entering from the gate without any stagnant areas, this path diagram served as the basis for the material flow simulation. The results of fluid dynamics simulation and finite element analysis are interrelated. The pressure distribution equilibrium state is used to verify whether the material flow is stable, and the material flow simulation is used to explain the cause of the pressure distribution. Together, they provide data support for mold design and adjustment, solving the technical problem that traditional testing cannot accurately determine the pressure and flow state inside the mold cavity.

[0057] Based on the pressure distribution equilibrium and material flow simulation, the cooling control parameters are adjusted. These parameters include cooling water temperature, water flow rate, and cooling time. If the pressure distribution equilibrium indicates slightly lower pressure at the mold edge, and considering the slower melt filling speed in the edge area according to the material flow simulation, it is determined that the cooling rate in the edge area needs to be reduced to extend the melt flow time. Therefore, the edge cooling water temperature is adjusted from 25℃ to 30℃, and the water flow rate is reduced from 2m / s to 1.5m / s. If the material flow simulation shows melt accumulation near the gate, and the corresponding pressure distribution shows higher pressure in that area, the cooling time near the gate is shortened from 15 seconds to 12 seconds to ensure sufficient cooling of the accumulated melt without affecting the overall molding. This adjustment process is based on pressure and flow data, avoiding blind adjustments that could lead to defects and addressing the lack of data support in traditional cooling parameter adjustment methods.

[0058] The integrated and optimized injection molding process parameters (screw speed 130r / min, injection pressure 100MPa), injection temperature (200℃), pressure distribution equilibrium state (95-105MPa), material flow simulation basis (flow path diagram), and cooling control parameters (edge ​​water channel temperature 30℃, water flow velocity 1.5m / s, gate cooling time 12 seconds) are categorized and associated according to mold design modules. For example, pressure distribution data is associated with mold cavity structure dimensions, and material flow data is associated with gate location, forming a structured dataset, which is the mold design improvement dataset.

[0059] Figure 2 This is an image showing the imaging and feature analysis results of plastic injection molding defect detection; please refer to [link / reference]. Figure 2 The detection system simultaneously images from three perspectives: 45° from the left, 60° from the top, and 45° from the right. Each perspective image is processed in parallel: first, highlighted defect candidate points are extracted to initially locate suspected areas; then, the Canny and Sobel operators are used to capture the clear outlines of the defects, and the Prewitt operator is used to sensitively extract surface texture gradient changes. This multi-view, multi-feature parallel processing architecture constitutes a multi-layered feature information system. This feature strongly demonstrates that the method can achieve highly reliable detection and accurate spatial localization of surface defects (such as white spots and bubbles) in plastic injection molded parts by fusing geometric information from different perspectives with different types of image features.

[0060] Figure 3 This is a verification chart of the results of plastic injection molding defect detection and optimization; please refer to [link / reference]. Figure 3The defect dynamic tracking trajectory diagram, based on three-dimensional coordinates, marks the spatial change paths of the overall defect trajectory, surface white spot trajectory, and internal bubble trajectory, intuitively presenting the positional movement patterns of different types of defects over time during injection molding (such as the Z-axis height change of internal bubbles from near the mold gate to the cavity edge), achieving visualized tracking of the spatial positional evolution trend of defects. The process parameter optimization comparison diagram, in the form of a bar chart, directly compares the "before optimization" and "after optimization" values ​​of three key injection molding process parameters: screw speed, injection pressure, and injection temperature. It clearly shows the specific adjustment direction and magnitude of the parameters, achieving a quantitative presentation of the process parameter optimization process. This demonstrates that the method can combine thermodynamic models and genetic optimization algorithms to output precise parameter adjustment schemes for defect characteristics, rather than relying on blind adjustments based on human experience, ensuring the scientific and targeted nature of process optimization. The cooling system optimization comparison diagram, also using a bar chart, focuses on cooling control parameters, comparing the differences before and after optimization of cooling water temperature, water flow rate, and cooling time. It clarifies the coordinated adjustments made by the cooling system to adapt to process parameter optimization, achieving verification of the matching between cooling parameters and injection molding process parameters. This demonstrates that the optimization process of this method does not involve adjusting a single parameter in isolation. Instead, it combines pressure distribution equilibrium obtained from fluid dynamics simulation and finite element analysis with material flow simulation data to simultaneously optimize the cooling system, ensuring uniform material solidification within the mold cavity and further reducing defect generation. The optimization effect verification chart visually presents the optimization results through the number of defects and quality scores: the number of defects decreased from 1363 before optimization to 382 after optimization, a reduction rate of 72.0%; the quality score increased from 65 points before optimization to 92 points after optimization, an improvement of 27.0%, achieving quantitative verification of the effects of defect detection and production quality optimization. This indicates that this method can not only accurately identify defects but also fundamentally reduce the defect rate and improve product quality through parameter optimization, effectively solving the core problem of existing detection methods that can only identify defects but lack closed-loop optimization.

[0061] The above describes the real-time detection method for plastic injection molding defects in the embodiments of this application. Please refer to [link / reference]. Figure 4 The following describes the real-time detection system 400 for plastic injection molding defects in the embodiments of this application, including: The imaging preprocessing module 401 is used to acquire an original image set by synchronously imaging the defect area of ​​the plastic injection molded part through a multi-angle camera array, preprocess the original image set to extract a clear feature set, and generate an enhanced defect candidate point feature set.

[0062] The defect classification module 402 is used to generate a defect spatial location information set by three-dimensional reconstruction of the defect candidate point feature set through a stereo vision algorithm, and to extract features using a convolutional neural network to determine the distinction criteria and classify and generate a defect type set.

[0063] The curvature and texture analysis module 403 is used to perform curvature and surface texture gradient analysis on the defect type set to obtain defect curvature change detection data, and to fuse the texture analysis results to generate a defect cause parameter identification set.

[0064] The dynamic tracking module 404 is used to update the spatial location changes of defects based on the defect cause parameter identification set and the dynamic tracking algorithm, thereby generating a set of information on the evolution trend of defect spatial location.

[0065] The parameter optimization module 405 is used to optimize injection molding process parameters by fusing the defect type set with the defect spatial location evolution trend information set, and obtain a mold design improvement dataset.

[0066] Through the collaborative efforts of the aforementioned modules, the system constructs a closed-loop detection and improvement system encompassing "imaging acquisition - defect analysis - cause tracing - dynamic tracking - process optimization." This system achieves intelligent processing throughout the entire process, from synchronously imaging and acquiring raw data of defective areas in plastic injection molded parts to generating mold design improvement datasets. The imaging preprocessing module 401, through multi-angle camera array deployment and synchronous imaging, combined with preprocessing operations such as grayscale conversion and filtering, generates an enhanced defect candidate point feature set, providing high-quality feature data for subsequent 3D reconstruction and defect classification, solving the problems of weak anti-interference and unclear defect feature extraction in traditional imaging. The defect classification module 402, based on the defect candidate point feature set, obtains the spatial location information set of defects through 3D reconstruction using stereo vision algorithms, and then uses a convolutional neural network to extract depth and volume features and classify them, accurately distinguishing surface white spots from internal bubbles, filling the gap in traditional detection methods that cannot obtain defect spatial information and accurate classification. The curvature and texture analysis module 403 associates multi-view image data with the defect type set, performs curvature analysis and surface texture gradient analysis, and integrates light and shadow textures. The calculation results of texture roughness generate a defect cause parameter identification set, establish the correlation between defect features and causes, and solve the problem that traditional detection is difficult to trace the root cause of defects. The dynamic tracking module 404 matches the defect cause parameter identification set with real-time production data, and generates a defect spatial location evolution trend information set through dynamic tracking algorithm combined with filtering optimization, realizing dynamic monitoring and trend prediction of defect location, overcoming the shortcomings of insufficient dynamic tracking capability of traditional detection. The parameter optimization module 405 integrates the defect spatial location evolution trend information set and defect type set, and optimizes injection molding process and cooling parameters by combining thermodynamic models, fluid dynamics simulations, etc., and integrates to generate mold design improvement dataset, forming a closed loop from defect detection to process optimization, solving the problems of traditional detection without closed-loop optimization and lack of data support for process adjustment.

[0067] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for real-time detection of defects in plastic injection molding, characterized in that, The method includes: Step S101: Simultaneously image the defect area of ​​the plastic injection molded part using a multi-angle camera array to obtain an original image set; preprocess the original image set to extract a clear feature set and generate an enhanced defect candidate point feature set; Step S102: Perform three-dimensional reconstruction of the defect candidate point feature set using a stereo vision algorithm to generate a defect spatial location information set; use a convolutional neural network to extract features from the defect spatial location information set, determine the distinction criteria between surface white spots and internal bubbles in plastic injection molding defects based on the extracted features, classify defects based on the distinction criteria, and generate a classified defect type set. Step S103: Obtain defect curvature change detection data by performing curvature analysis and surface texture gradient analysis on the defect type set respectively; generate a defect cause parameter identification set by fusing the texture analysis results with the defect curvature change detection data, wherein the texture analysis includes light and shadow texture analysis and texture roughness calculation; Step S104: Based on the defect cause parameter identification set, use a dynamic tracking algorithm to update the spatial location change of the defect and generate a spatial location evolution trend information set of the defect. Step S105: Based on the spatial location evolution trend information set of the defects, fuse the defect type set, optimize the injection molding process parameters, and obtain the mold design improvement dataset.

2. The method for real-time detection of defects in plastic injection molding according to claim 1, characterized in that, In step S101, obtaining the original image set includes: The multi-angle camera array is deployed at the plastic injection molding part inspection station; The image data captured by the multi-angle camera array is preliminarily processed to obtain a processed image set. All images in the processed image set containing micron-level white spot defects are used as the original image set. The positional uncertainty of defects in the original image set is calculated by calculating the pixel deviation of the same defect in images from different viewpoints, wherein the pixel deviation is used to characterize the positional uncertainty; The size variability of defects in the original image set is calculated by calculating the size difference of the same defect in images from different viewpoints, wherein the size difference is used to characterize the size variability; The image data containing the location uncertainty and the size variability are integrated to generate the original image set.

3. The method for real-time detection of defects in plastic injection molding according to claim 2, characterized in that, In step S101, the enhanced defect candidate point feature set is generated, including: The original image set is subjected to grayscale conversion and filtering to remove dust interference and background noise, while enhancing image contrast, to generate a processed image set. Pixels with gray values ​​higher than a preset gray threshold in the processed image are marked as candidate points for white spot defects using a threshold segmentation method. Calculate the standard deviation of pixel intensity in the region where the candidate white spot defects are located to determine the severity fluctuation; Spatial clustering analysis was used to analyze the spatial distribution of the candidate white spot defects and determine the uneven distribution. Extract a clear feature set under the severity fluctuation and the uneven distribution, the clear feature set including the edge contour and gray-scale distribution features of the region where the white spot defect candidate point is located; The clear feature set is associated with the corresponding white spot defect candidate points to form the enhanced defect candidate point feature set.

4. The method for real-time detection of defects in plastic injection molding according to claim 1, characterized in that, In step S102, generating a defect spatial location information set includes: If the density of the enhanced defect candidate point feature set is higher than the preset density threshold, the stereo vision algorithm is used to perform feature point correspondence matching on the multi-view data. The disparity is calculated based on the matching results of the feature points, and the three-dimensional coordinates of the white point defect candidate points are calculated in combination with the multi-view geometric constraints. The multi-view geometric constraints are determined by the calibration parameters of the multi-angle camera array. The three-dimensional coordinates are corrected by integrating the effects of temperature changes and camera shake. The temperature changes are corrected for offsets using the thermal expansion coefficient of the material, and the camera shake is corrected for deviations using a dynamic compensation algorithm. The three-dimensional coordinates of all the corrected white spot defect candidate points are integrated to generate a defect spatial location information set.

5. The method for real-time detection of defects in plastic injection molding according to claim 4, characterized in that, In step S102, the classified defect type set is generated, including: The three-dimensional coordinates of the defect spatial location information set are normalized. Depth and volume features are extracted from the normalized spatial location information of the defects using a convolutional neural network. The convolutional neural network includes multiple convolutional layers and pooling layers. The depth features are extracted through the convolutional kernels of the convolutional layers, and the volume features are extracted through three-dimensional spatial point cloud analysis. The depth features and volume features are fused using a convolutional neural network to form a comprehensive feature vector. Based on the comprehensive feature vector, the criteria for distinguishing between surface white spots and internal bubbles in plastic injection molding defects are determined. The classification layer of the convolutional neural network classifies the comprehensive feature vector and outputs the category probability corresponding to each defect. Based on the category probability, the defect is determined to be either a surface white spot or an internal bubble. The classification results of all defects are integrated to generate a set of classified defect types.

6. The method for real-time detection of defects in plastic injection molding according to claim 1, characterized in that, In step S103, acquiring defect curvature change detection data includes: Associate the image data captured by the multi-angle camera array with the defect regions corresponding to the defect type set; Curvature analysis is performed on each defect in the defect type set through multi-view geometric reconstruction, and the reconstruction model is fitted using a surface fitting algorithm to generate the local radius of curvature of the defect surface. The surface texture gradient of the defect region is calculated using a gradient operator to obtain the magnitude and direction information of the texture gradient. The density distribution mapping of the defect region is then constructed by calculating the spatial distribution of pixel intensity. The curvature calculation deviation caused by reflection problem is compensated by the lighting model, and the texture gradient distortion caused by the angle deviation is corrected based on the calibration parameters of the multi-angle camera array. The local radius of curvature, density distribution mapping, compensated curvature, and corrected texture gradient analysis results of each defect are integrated to generate defect curvature change detection data.

7. The method for real-time detection of defects in plastic injection molding according to claim 6, characterized in that, In step S103, generating a defect cause parameter identification set includes: Light and shadow texture analysis was performed on the original image set using a lighting model to extract texture features, and texture roughness was calculated on the original image set using surface height distribution statistics. Based on the light and shadow texture analysis results and the texture roughness calculation results, the regional density quantification index of the defect is calculated through the spatial distribution of defect points; The region density quantification index, the size variability of the defect, the severity fluctuation, and the corresponding light and shadow texture and texture roughness data are integrated and combined with the preset defect cause mapping rules to generate a defect cause parameter identification set.

8. The method for real-time detection of defects in plastic injection molding according to claim 1, characterized in that, Step S104 includes: If the defect cause parameter identification set matches the real-time production data, a dynamic tracking algorithm is used to process multiple frame image sequences to update the spatial location changes of the defects. Environmental factor correction is performed on the changes in the spatial location based on the humidity and temperature values ​​in real-time production data. Kalman filtering and particle filtering are used to predict and optimize the corrected spatial position changes, generating accurate data on the spatial position changes of defects. The precise spatial location change data of the defects are arranged in chronological order to form a trajectory of the defect location change; The influence of material-related randomness and injection molding process parameters on the change trajectory is integrated to generate a set of influencing factors. The material-related randomness is the random variation of polymer particle distribution in the injection molding material, and the injection molding process parameters are injection temperature and pressure data. The change trajectory is optimized using the set of influencing factors to generate a set of information on the spatial location evolution trend of the defect.

9. The method for real-time detection of defects in plastic injection molding according to claim 1, characterized in that, Step S105 includes: Based on the spatial location evolution trend information set of the defects and the defect type set, combined with the thermodynamic model, the genetic optimization algorithm is used to optimize the injection molding process parameters and injection temperature. Based on the optimized injection molding process parameters and injection temperature, the pressure distribution equilibrium state and material flow simulation basis in the mold cavity are obtained through fluid dynamics simulation and finite element analysis. Based on the pressure distribution equilibrium state and the material flow simulation, adjust the cooling control parameters; The optimized injection molding process parameters, injection temperature, pressure distribution balance, material flow simulation basis, and cooling control parameters are integrated to generate the mold design improvement dataset.

10. A real-time detection system for defects in plastic injection molding, characterized in that, For implementing the real-time detection method for plastic injection molding defects as described in any one of claims 1 to 9, the real-time detection system for plastic injection molding defects comprises: The imaging preprocessing module is used to simultaneously image the defect area of ​​the plastic injection molded part through a multi-angle camera array to obtain the original image set, preprocess the original image set to extract a clear feature set, and generate an enhanced defect candidate point feature set. The defect classification module is used to generate a set of spatial location information of defects by 3D reconstruction of the feature set of defect candidate points using a stereo vision algorithm, and to extract features using a convolutional neural network to determine the distinction criteria and classify and generate a set of defect types. The curvature and texture analysis module is used to perform curvature and surface texture gradient analysis on the defect type set to obtain defect curvature change detection data, and to fuse the texture analysis results to generate a defect cause parameter identification set. The dynamic tracking module is used to update the spatial location changes of defects based on the defect cause parameter identification set and the dynamic tracking algorithm to generate a set of defect spatial location evolution trend information. The parameter optimization module is used to optimize injection molding process parameters by integrating the defect type set with the defect spatial location evolution trend information set, and to obtain a mold design improvement dataset.

Images