Rubber and plastic sealing element quality evaluation method and system based on big data analysis

Abstract

The invention relates to the technical field of quality evaluation, and particularly discloses a rubber and plastic sealing element quality evaluation method and system based on big data analysis, and the method comprises the steps: training a normal appearance model based on an unsupervised model architecture through constructing and dynamically managing a dynamic baseline library only containing good rubber and plastic sealing element images; learning the internal characteristic mode of the image of the good rubber and plastic sealing element; furthermore, in a quality evaluation stage, a to-be-detected rubber and plastic sealing element image is input into the trained normal appearance model, the normal appearance model carries out image reconstruction on the to-be-detected rubber and plastic sealing element image based on a good product image internal characteristic mode learned by the normal appearance model, and an image reconstruction error is calculated to serve as an abnormal score of the detected image. Therefore, the defect evaluation of the rubber and plastic sealing element is realized. The method not only improves the detection precision and efficiency, but also reduces the collection and labeling cost of the training data, and can meet the detection requirements of novel or accidental defects at the same time.

Description

Technical Field

[0001] The present application relates to the technical field of quality assessment, and more specifically, to a method and system for assessing the quality of rubber and plastic seals based on big data analysis. Background Art

[0002] As essential, fundamental components in industrial production, rubber and plastic seals are widely used in numerous fields, including automotive, aerospace, mechanical equipment, and electronic appliances. They provide reliable seals, preventing fluid or gas leakage, ensuring safe and stable equipment operation and ensuring the overall performance and service life of the product. Therefore, the quality of rubber and plastic seals is directly related to the reliability and safety of the final product. Any minor defects, such as cracks, bubbles, missing material, burrs, or dimensional deviations, can lead to seal failure, potentially causing equipment failure, production accidents, and even serious economic losses and safety hazards.

[0003] Traditional quality inspection of rubber and plastic seals relies primarily on manual visual inspection. This method is not only labor-intensive and inefficient, but is also susceptible to subjective factors, experience level, and fatigue of the inspector, resulting in poor consistency in inspection results, high rates of missed and false detections, and difficulty meeting the needs of large-scale, high-precision production. With the development of machine vision technology, deep learning algorithms based on supervised learning have achieved remarkable success in the field of image recognition. However, they face practical challenges in the application of defect detection in rubber and plastic seals: On the one hand, defect recognition methods based on supervised learning rely on large amounts of labeled data, while defect samples in industrial production are usually scarce, diverse, and complex in form. Collecting and accurately labeling large-scale, diverse defect sample data is costly, time-consuming, and labor-intensive. On the other hand, new or occasional defects are difficult to fully cover with pre-defined training sets, making it difficult to achieve comprehensive and effective quality inspection.

[0004] Therefore, an optimized rubber and plastic seal quality assessment method and system based on big data analysis is expected. Summary of the Invention

[0005] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a rubber and plastic seal quality assessment method and system based on big data analysis, which trains a normal appearance model based on an unsupervised model architecture by constructing and dynamically managing a dynamic baseline library containing only images of good rubber and plastic seals, so that it can learn the intrinsic feature patterns of good rubber and plastic seal images; then, in the quality assessment stage, the rubber and plastic seal image to be inspected is input into the trained normal appearance model, and the normal appearance model reconstructs the image of the rubber and plastic seal to be inspected based on the intrinsic feature patterns of the good images it has learned, and calculates the image reconstruction error as the abnormality score of the rubber and plastic seal image to be inspected, thereby realizing the defect assessment of the rubber and plastic seal to be inspected. This method not only improves the quality inspection accuracy and efficiency of rubber and plastic seals, but also does not need to rely on a large amount of defect sample annotation data, reducing the cost of data collection and annotation, and can also meet the detection needs of new or occasional defects, realizing comprehensive and effective quality inspection.

[0006] Accordingly, according to one aspect of the present application, a method for quality assessment of rubber and plastic seals based on big data analysis is provided, which includes a training phase and a quality assessment phase; The training phase includes: constructing a dynamic baseline library of good product images, the dynamic baseline library including a plurality of images of rubber and plastic seals marked as good products; dynamically managing the dynamic baseline library of good product images; and using the dynamic baseline library of good product images to train a normal appearance model based on an unsupervised model architecture to obtain a trained normal appearance model based on an unsupervised model architecture. The quality assessment stage includes: obtaining an image of the rubber or plastic seal to be inspected; inputting the image of the rubber or plastic seal to be inspected into the trained normal appearance model based on the unsupervised model architecture to obtain an abnormality score; and judging whether the rubber or plastic seal to be inspected is an abnormal part based on a comparison between the abnormality score and a preset threshold.

[0007] According to another aspect of the present application, a rubber and plastic seal quality assessment system based on big data analysis is provided, which can implement the rubber and plastic seal quality assessment method based on big data analysis as described above, wherein the rubber and plastic seal quality assessment system based on big data analysis includes: a training module and a quality assessment module; The training module includes: a baseline library construction unit for constructing a dynamic baseline library of good product images, wherein the dynamic baseline library of good product images includes multiple images of rubber and plastic seals marked as good products; a dynamic management unit for dynamically managing the dynamic baseline library of good product images; and a normal appearance model training unit for training a normal appearance model based on an unsupervised model architecture using the dynamic baseline library of good product images to obtain a trained normal appearance model based on an unsupervised model architecture. The quality assessment module includes: a rubber and plastic seal image acquisition unit, used to acquire an image of the rubber and plastic seal to be inspected; an abnormality assessment unit, used to input the image of the rubber and plastic seal to be inspected into the trained normal appearance model based on the unsupervised model architecture to obtain an abnormality score; and an abnormal part determination unit, used to determine whether the rubber and plastic seal to be inspected is an abnormal part based on a comparison between the abnormality score and a preset threshold.

[0008] Compared with the prior art, the rubber and plastic seal quality assessment method and system based on big data analysis provided by this application trains a normal appearance model based on an unsupervised model architecture by constructing and dynamically managing a dynamic baseline library containing only images of good rubber and plastic seals, so that it learns the intrinsic feature patterns of good rubber and plastic seal images. Furthermore, during the quality assessment stage, the rubber and plastic seal image to be inspected is input into the trained normal appearance model. The normal appearance model reconstructs the image of the rubber and plastic seal to be inspected based on the learned intrinsic feature patterns of the good image. The image reconstruction error is calculated as the abnormality score of the rubber and plastic seal image to be inspected, thereby achieving defect assessment of the rubber and plastic seal to be inspected. This method not only improves the quality inspection accuracy and efficiency of rubber and plastic seals, but also does not require reliance on a large amount of defect sample annotation data, reducing the cost of data collection and annotation. At the same time, it can meet the detection needs of new or occasional defects and achieve comprehensive and effective quality inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 Flowchart of a rubber and plastic seal quality assessment method based on big data analysis according to an embodiment of the present application.

[0011] Figure 2 Flowchart of step S2 in the rubber and plastic seal quality assessment method based on big data analysis according to an embodiment of the present application.

[0012] Figure 3 Flowchart of step S3 in the rubber and plastic seal quality assessment method based on big data analysis according to an embodiment of the present application.

[0013] Figure 4 Flowchart of step S31 in the rubber and plastic seal quality assessment method based on big data analysis according to an embodiment of the present application.

[0014] Figure 5 Flowchart of step S32 in the rubber and plastic seal quality assessment method based on big data analysis according to an embodiment of the present application.

[0015] Figure 6 4 is a block diagram of a rubber and plastic seal quality assessment system based on big data analysis according to an embodiment of the present application. DETAILED DESCRIPTION

[0016] Below, an example embodiment according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiment is only a part of the embodiment of the present application, not all of the embodiments of the present application, and it should be understood that the present application is not limited to the example embodiment described herein. It is worth noting that in the present application, all actions to obtain data are carried out in compliance with the relevant data protection laws and policies of the country where they are located, and with the authorization of the corresponding device owner.

[0017] Figure 1 Flowchart of the rubber and plastic seal quality assessment method based on big data analysis according to an embodiment of the present application. Figure 1 As shown, the rubber and plastic seal quality assessment method based on big data analysis according to the embodiment of the present application includes a training stage and a quality assessment stage; wherein the training stage includes the steps of: S1, constructing a dynamic baseline library of good product images, the dynamic baseline library of good product images including a plurality of rubber and plastic seal images marked as good products; S2, dynamically managing the dynamic baseline library of good product images; S3, using the dynamic baseline library of good product images to train a normal appearance model based on an unsupervised model architecture to obtain a trained normal appearance model based on an unsupervised model architecture; the quality assessment stage includes the steps of: S4, acquiring an image of a rubber and plastic seal to be detected; S5, inputting the image of the rubber and plastic seal to be detected into the trained normal appearance model based on the unsupervised model architecture to obtain an abnormality score; S6, judging whether the rubber and plastic seal to be detected is an abnormal part based on a comparison between the abnormality score and a preset threshold.

[0018] In the training phase of the above-mentioned rubber and plastic seal quality assessment method based on big data analysis, the step S1 constructs a dynamic baseline library of good product images, and the dynamic baseline library of good product images includes a plurality of rubber and plastic seal images that are labeled as good products. It should be understood that in the prior art, defect detection methods based on supervised learning rely on a large number of defect samples, while rubber and plastic seal defect samples are difficult to obtain in industrial scenarios, and it is difficult to cover all defect types, resulting in insufficient generalization ability of the model and high deployment costs. Therefore, in order to establish a quality assessment benchmark that does not require defect samples, this application is based on the principle of self-learning of normal samples, and by collecting images of good rubber and plastic seals under multiple batches and multiple working conditions, a dynamic baseline library of good product images is constructed to achieve adaptive modeling of the visual feature distribution of normal rubber and plastic seals.

[0019] In one specific embodiment of the present application, a highly stable image acquisition system is first deployed on a rubber and plastic seal production line. This image acquisition system includes a high-resolution industrial camera (e.g., a CCD or CMOS camera with a resolution exceeding one megapixel to capture minute details), a uniform and shadowless illumination source (e.g., a ring-shaped LED light source, a strip light source, or a dome light source, selected based on the surface characteristics and shape of the seal to avoid reflections and shadows, ensuring stable presentation of image features), and a workpiece positioning and conveying device (to ensure that the seal is in a similar position and posture during each image acquisition, reducing unnecessary image variation). During the production process, rubber and plastic seals produced in each batch or within a specified time period undergo manual inspection by experienced quality inspectors. Acceptable rubber and plastic seals are then sent to an image acquisition area, where the image acquisition system captures images of these acceptable seals. The captured images of these acceptable seals are then uploaded to a server or local storage device for storage. In addition, each image of a qualified rubber and plastic seal is assigned a unique identifier, and relevant production information is recorded, such as metadata such as production batch, production time, production line number, raw material batch number, mold number, process parameter group, etc., for subsequent dynamic management while ensuring the traceability of the model. Before the qualified rubber and plastic seal image data is stored in the qualified image dynamic baseline library, it is required to undergo a series of preprocessing operations, such as image cropping to remove irrelevant background and size normalization to meet the model input requirements. All images of rubber and plastic seals that have been confirmed as qualified and preprocessed together constitute the initial qualified image dynamic baseline library. The data can be kept current by continuously incorporating new qualified samples, thereby providing a dynamically evolving high-quality normal sample input source for model training. In this way, it is possible to ensure that the machine vision model learns the standard appearance of rubber and plastic seals under various normal production conditions, laying a data foundation for the subsequent accurate identification of abnormal parts that deviate from the standard appearance.

[0020] During the training phase of the aforementioned rubber and plastic seal quality assessment method based on big data analysis, step S2 dynamically manages the dynamic baseline library of good product images. It should be understood that factors such as production line parameter adjustments, material batch changes, and equipment wear can cause the distribution of good product characteristics to drift, and the outdated production conditions represented by the outdated samples in the dynamic baseline library of good product images will gradually deviate from the current actual operating conditions. Therefore, in order to maintain the dynamic baseline library of good product images' ability to represent real-time production conditions, this application further introduces a dynamic management mechanism to achieve intelligent updating of the dynamic baseline library of good product images.

[0021] In a specific embodiment of the present application, dynamically managing the good product image dynamic baseline library includes at least one of the following steps: (1) Regularly remove the oldest batch of sample images stored in the dynamic baseline library of good product images. Generally speaking, the older the sample data is, the greater the deviation between the production status it represents and the current status may be. Therefore, based on the first-in-first-out (FIFO) principle, this application sets a sample cleaning time window (such as one month), and automatically identifies and removes the oldest part of the good product images stored in the baseline library by regularly triggering the sample cleaning process. In specific implementation, each good product image stored in the database will be accompanied by its acquisition timestamp and the identification of the production batch to which it belongs. By regularly running maintenance scripts or tasks, the image records belonging to the earliest N batches and their corresponding image files are queried and deleted from the baseline library. In this way, by continuously eliminating old data that may no longer fully represent the current good product standards, space is made for newly collected good product images that better reflect the current production characteristics, thereby ensuring that the learning focus of the model is always on the appearance pattern of recent good products.

[0022] (2) Only sample images within a preset time period are retained or only sample images from a preset batch are retained. That is, based on the important influence of data freshness on model relevance, the content of the dynamic baseline library of good product images is intelligently updated. In specific implementation, a clear retention period is first set. For example, only good product images collected within the past 90 days are retained, or only good product images from the most recent 50 production batches are retained. When new good product images are collected and added to the baseline library, the system checks whether there are old images in the library that exceed this time limit or batch range. If so, they are automatically removed, thereby ensuring that the size of the baseline library is relatively stable and the sample image data therein always revolves around the current production cycle, allowing the model to quickly respond to changes in recent production conditions.

[0023] (3) When it is confirmed that a certain production condition is no longer in use, all sample images related to the production condition are removed. That is, considering that the changes in some production conditions are clear and intermittent, such as the permanent scrapping of a certain model of mold, the replacement of a certain raw material supplier, or the replacement of an old process parameter group by a new process. In this case, the good product images related to the invalid production condition are not only no longer representative, but may even interfere with the model learning of the current good product pattern. Therefore, when the production management system (such as MES) records that a certain production condition (such as mold M001 has been deactivated) has changed, the production condition change information is passed to the baseline library management module. The baseline library management module queries all good product image records associated with the production condition in the baseline library based on the relevant production condition metadata that was annotated when the good product image was stored in the baseline library, such as mold number, raw material batch number, equipment ID, process version number, etc., and completely deletes the corresponding image files from the baseline library. Through this removal method driven by specific production events, data that has lost its reference value can be accurately eliminated, preventing the model from making misjudgments due to learning outdated normal rubber and plastic seal visual feature patterns.

[0024] By implementing at least one of the above dynamic management strategies, it is possible to effectively cope with the natural evolution of the appearance of qualified rubber and plastic seals during the production process, and ensure that the dynamic baseline library of qualified product images always contains high-quality qualified product samples that are most relevant to the current production status.

[0025] Furthermore, if the dynamic update of the good product image dynamic baseline library only relies on fixed cycles or triggering of specific process change events, it may not be able to respond in time to certain occasional, irregular and sudden changes in the appearance of good rubber and plastic seals. To this end, in one embodiment of the present application, an intelligent update decision mechanism for the baseline library based on cluster analysis is introduced. By modeling the good product image feature space in the good product image dynamic baseline library as a dynamically evolving clustering structure, each time a new good product image is added, by quantifying the deviation between the new sample and the cluster center, active perception and accurate response to data distribution drift can be achieved. Figure 2 As shown, the step S2 includes: S21, performing cluster analysis on the rubber and plastic seal images marked as good products in the good product image dynamic baseline library to obtain the good product image cluster center; S22, acquiring a new rubber and plastic seal image marked as good products; S23, calculating the sample offset coefficient between the new rubber and plastic seal image marked as good products and the good product image cluster center; S24, determining whether to trigger the dynamic update instruction of the good product image dynamic baseline library based on the comparison between the sample offset coefficient and a preset threshold.

[0026] Specifically, step S21 performs cluster analysis on the images of rubber and plastic seals labeled as good in the dynamic baseline library of good product images to obtain cluster centers of good product images. It should be understood that the good product images stored in the dynamic baseline library are diverse, including images of good products of different models, sizes, and production batches. Furthermore, good products of the same type may also exhibit certain legal variations in appearance (e.g., reasonable color differences and allowable texture fluctuations). Therefore, to establish an effective reference standard for newly added good product images, this application performs cluster analysis on the good product images in the baseline library to discover the essential visual feature patterns of various types of good products.

[0027] In an embodiment of the present application, an image encoder is first used to extract visual features from each rubber and plastic seal image in the dynamic baseline library of good product images to obtain a visual feature vector representation of each rubber and plastic seal image. Here, the image encoder can utilize the encoder portion of the normal appearance model based on the unsupervised model architecture, or employ other convolutional neural network models, such as ResNet and VGG, to effectively map the good product rubber and plastic seal images from a high-dimensional pixel space to a low-dimensional feature space, thereby preserving key visual information in the image while removing redundancy and noise. The K-Means clustering algorithm is then used to cluster the visual feature vectors of each rubber and plastic seal image. The appropriate number of clusters, K, is determined by presetting or using methods such as the elbow method and the silhouette coefficient. The visual feature vector of each image is iteratively assigned to one of the K clusters, and the center of each cluster (i.e., the cluster center) is updated to maximize the similarity (minimize the distance) between samples within a cluster and minimize the similarity (maximize the distance) between samples between clusters. This allows the identification of different categories of non-defective rubber and plastic seals and the calculation of the cluster center for each category. After clustering, the resulting K cluster centers (each representing the typical visual features of all non-defective rubber and plastic seal images within that cluster) serve as the essential visual feature representation of the current non-defective standard in feature space, thereby clarifying several major non-defective appearance patterns and their representative representatives known in the dynamic baseline library of non-defective images.

[0028] Specifically, step S22 involves acquiring a new image of a rubber or plastic sealant labeled as a good product. Specifically, the new image of the rubber or plastic sealant labeled as a good product is obtained by capturing images of newly produced rubber or plastic sealants using an image acquisition system on the production line and manually or automatically labeling their good product attributes. This acquisition process is consistent with the aforementioned image acquisition process, including capture with an industrial camera and necessary image preprocessing (such as cropping and normalization). The image represents the appearance of good rubber or plastic sealants within the current production cycle and serves as an important basis for dynamically updating the baseline library.

[0029] Specifically, step S23 calculates the sample offset coefficient between the new rubber and plastic seal image marked as a good product and the cluster center of the good product image. That is, in order to determine whether the new rubber and plastic seal image marked as a good product conforms to the known good product rubber and plastic seal appearance pattern, the present application quantifies the degree of deviation from the known good product appearance pattern in the good product image dynamic baseline library by calculating the sample offset coefficient between it and the good product image cluster center, thereby determining whether to include it in the baseline library. In an embodiment of the present application, first, the same image encoder as described above is used to extract visual features of the new rubber and plastic seal image marked as a good product to obtain its visual feature vector representation. Then, the Euclidean distance between the new image visual feature vector and each good product image cluster center is calculated, and the minimum Euclidean distance value is selected as the sample offset coefficient of the new image to determine the degree of deviation between the new image and the closest good product appearance typical pattern in the baseline library.

[0030] Specifically, step S24 determines whether to trigger a dynamic update instruction for the dynamic baseline library of good product images based on a comparison between the sample offset coefficient and a preset threshold. It should be understood that the sample offset coefficient reveals the offset of the new image of the rubber and plastic seal, labeled as good, relative to the current state of the dynamic baseline library of good product images. Comparing the sample offset coefficient with the preset threshold effectively determines whether the new good product image falls within the known good product appearance pattern, thereby determining whether learning and updating the baseline library is necessary.

[0031] In an embodiment of the present application, first, a reasonable preset threshold value, such as 0.1, is set based on the historical data of the dynamic baseline library of good product images, the tolerance of the appearance variation of good product rubber and plastic seals, and the actual needs of the production line. Then, the sample offset coefficient is compared with the preset threshold value. When the sample offset coefficient is less than or equal to the preset threshold value, it is considered that the new image is consistent with the good product appearance pattern in the baseline library and belongs to the category of known good products, and it can be added to the baseline library; if the sample offset coefficient is greater than the threshold value, it means that the new rubber and plastic seal image marked as good product may be a falsely detected defective product, the production condition is abnormal, or there is a new good product appearance pattern. At this time, the new sample is marked as "to be observed". If multiple good product samples with high offset coefficients appear continuously and these new good product samples form a new cluster, it indicates that there is a new good product appearance pattern. At this time, a dynamic update instruction is triggered to incorporate these new representative good product images into the dynamic baseline library of good product images to enrich the good product appearance pattern representation of the baseline library. At the same time, some old and no longer representative samples are removed as needed, and the model retraining or incremental learning process is started. If no new clustering trends are found during continuous observation, they are fed back to manual review for re-inspection and confirmation by professionals to avoid mistakenly including defective products in the baseline library and ensure the accuracy and timeliness of the dynamic baseline library of good product images. In this way, the dynamic baseline library of good product images can continuously optimize and update itself, ensuring that it always contains the latest and most representative good product samples. This provides high-quality data support for training normal appearance models that accurately reflect the current production status, improving the adaptability and long-term reliability of the quality assessment system.

[0032] During the training phase of the aforementioned rubber and plastic seal quality assessment method based on big data analysis, step S3 utilizes the dynamic baseline library of good product images to train a normal appearance model based on an unsupervised model architecture to obtain a trained normal appearance model based on the unsupervised model architecture. Specifically, to autonomously mine normal appearance patterns from good product rubber and plastic seal image data, the present application trains the normal appearance model based on the unsupervised model architecture using codec image reconstruction learning technology, enabling it to learn the intrinsic visual feature patterns of good product rubber and plastic seal images, and reconstruct a normal appearance image similar to the original input good product image based on the learned visual feature patterns of the good product rubber and plastic seal images.

[0033] In a specific example of the present application, the normal appearance model based on the unsupervised model architecture includes an encoder, a decoder and a reconstruction error calculation module. Figure 3As shown, the step S3 includes: S31, inputting each rubber and plastic seal image marked as good in the good image dynamic baseline library into the encoder to obtain a rubber and plastic seal visual feature latent vector; S32, inputting the rubber and plastic seal visual feature latent vector into the decoder to obtain a reconstructed rubber and plastic seal image; S33, inputting the reconstructed rubber and plastic seal image and the rubber and plastic seal image marked as good into the reconstruction error calculation module to obtain a reconstruction error; S34, constructing a reconstruction loss function value based on the reconstruction error; S35, adjusting the weight parameters of the encoder, the decoder and the reconstruction error calculation module by minimizing the reconstruction loss function value.

[0034] More specifically, in step S31, each rubber and plastic seal image marked as good in the good image dynamic baseline library is input into the encoder to obtain a rubber and plastic seal visual feature latent vector. Here, the encoder is used to extract the key visual features of each good rubber and plastic seal image, and convert the pixel-level data of the good rubber and plastic seal image into an abstract, low-dimensional visual feature latent vector to characterize the intrinsic properties and appearance patterns of the good rubber and plastic seal, providing basic data support for the subsequent image reconstruction process. Figure 4 As shown, the step S31 includes: S311, performing convolution encoding on the rubber and plastic seal image marked as good product through a convolution layer to obtain a rubber and plastic seal visual feature map; S312, performing feature distribution offset constraint correction on the rubber and plastic seal visual feature map to obtain a corrected rubber and plastic seal visual feature map; S313, flattening the corrected rubber and plastic seal visual feature map into a rubber and plastic seal visual feature vector, and inputting it into a fully connected layer to obtain the rubber and plastic seal visual feature latent vector.

[0035] In a specific embodiment of the present application, step S311 performs convolutional encoding on the image of the rubber or plastic seal marked as good through a convolutional layer to obtain a visual feature map of the rubber or plastic seal. Specifically, the surface of the rubber or plastic seal has complex microstructures such as vulcanization marks and mold seams. To extract spatially hierarchical visual features, the present application uses multi-level convolution operations based on local receptive field theory to achieve a multi-scale abstract representation of the visual features of the rubber or plastic seal image. The first convolutional layer uses a large 7×7 convolution kernel (stride 2, padding 3) to capture the overall contours of the seal and large defect features with a large receptive field. The second layer uses a 3×3 convolution kernel (stride 1, padding 1) for deep feature extraction, introducing nonlinear transformations through the ReLU activation function, and superimposing a batch normalization layer (BatchNorm) to accelerate convergence. The third to fifth layers use a depthwise separable convolution structure, each consisting of 3×3 depthwise convolutions and 1×1 pointwise convolutions, reducing the number of parameters while enhancing feature expression. Each convolution layer is followed by a 2×2 max pooling layer (stride 2) to gradually compress the feature map size and deeply explore the detailed features of high-quality rubber and plastic seals. Through multi-level convolutional encoding, the raw pixel data is converted into a visual feature map with a spatial hierarchical structure, fully preserving the multi-granular quality information of high-quality rubber and plastic seals, from macroscopic shape to microscopic texture.

[0036] In a specific embodiment of the present application, step S312 performs feature distribution offset constraint correction on the rubber and plastic seal visual feature map to obtain a corrected rubber and plastic seal visual feature map. It should be understood that due to the sample clustering offset-based update method of the good product image dynamic baseline library, the rubber and plastic seal images in the good product image dynamic baseline library will also have certain dynamic boundary conditions of image semantic distribution. This means that when the rubber and plastic seal images marked as good products are convolutionally encoded through the convolution layer to extract local image semantic features, the image semantic high-dimensional feature distribution of the rubber and plastic seal visual feature map will also show a specific evolutionary tendency, thereby causing a systematic distribution offset and affecting the accuracy of the image semantic expression of the reconstructed rubber and plastic seal image.

[0037] Therefore, for the visual characteristic diagram of the rubber and plastic seal, for example, , each eigenvalue of which is expressed as First, the parameter drift diffusion efficiency quantization coefficient of each characteristic value in the visual characteristic graph of the rubber and plastic seal is calculated, that is: ;in, is the offset diffusion angle hyperparameter, for The corresponding parameter drift diffusion efficiency quantization coefficient. In this way, the implicit evolution tendency based on the fuzzy dynamic boundary condition corresponding to the parameter drift of the eigenvalue can be converted into a quantization coding coefficient composed of the diffusion quantization map.

[0038] Then, based on the parameter drift diffusion efficiency quantification coefficient of each characteristic value in the visual characteristic graph of the rubber and plastic seal, the macroscopic distribution density constraint coefficient of each characteristic value is calculated. As the core of the nonlinear transformation inverse mapping of the dimension representation preservation mapping, the macroscopic distribution density constraint coefficient is constructed, namely: ;in, is a natural constant, for The corresponding macroscopic distribution density constraint coefficient.

[0039] In this way, by calculating the response of the micro-level relative to the macro-level nonlinear transformation inverse mapping space, while performing the implicit evolution tendency representation based on fuzzy dynamic boundary conditions, it is ensured that the macroscopic characteristics of the overall feature distribution conform to the prior assumptions, so as to achieve the gradual stabilization of the feature micro-macro interaction relationship.

[0040] Finally, based on the parameter drift diffusion efficiency quantification coefficient and macroscopic distribution density constraint coefficient of each characteristic value in the visual characteristic graph of the rubber and plastic seal, the characteristic value granularity offset correction is performed on the visual characteristic graph of the rubber and plastic seal to obtain the corrected visual characteristic graph of the rubber and plastic seal. More specifically, the parameter drift diffusion efficiency quantification coefficient is used. and the macroscopic distribution density constraint coefficient The weighted sum of Correction is performed to obtain the corrected visual feature map of the rubber and plastic seal. ,in , that is: ,in, and is a trainable weight parameter, for The corresponding corrected eigenvalues.

[0041] In this way, by defining the correction space as a composite structure of micro-macro dimensional mapping-inverse mapping, and by quantifying the implicit evolution tendency based on fuzzy dynamic boundary conditions, the boundary condition interpretability of the evolution results can be improved, and by constraining the macroscopic structure distribution density under fuzzy boundary conditions, the systematic distribution offset based on the quantified evolution path can be corrected, thereby improving the image semantic expression accuracy of the reconstructed rubber and plastic seal image obtained based on the rubber and plastic seal visual feature map.

[0042] In a specific embodiment of the present application, in step S313, after flattening the corrected rubber and plastic seal visual feature map into a rubber and plastic seal visual feature vector, it is input into a fully connected layer to obtain the rubber and plastic seal visual feature latent vector. It should be understood that since the corrected rubber and plastic seal visual feature map has some information redundancy and dimensionality disaster, directly using it for image reconstruction will lead to low computational efficiency and easily cause overfitting. Therefore, in order to construct a compact potential feature representation, the present application obtains the rubber and plastic seal visual feature latent vector by performing feature dimensionality reduction on the corrected rubber and plastic seal visual feature map to achieve distillation and compression of the essential visual features of the rubber and plastic seal. Specifically, first, the corrected rubber and plastic seal visual feature map is flattened into a one-dimensional feature vector along the spatial dimension, that is, a row-column-first scanning strategy is adopted to traverse the feature matrix of each channel in the corrected rubber and plastic seal visual feature map by row, and its elements are arranged in sequence into a one-dimensional vector to obtain the rubber and plastic seal visual feature vector. The visual feature vector of the rubber and plastic seal is then fed into a fully connected layer. The first fully connected layer uses a compression strategy with a dimensionality reduction ratio of 4:1, with weights initialized to a He normal distribution and bias terms set to zero. The second fully connected layer further compresses the feature dimensions with a dimensionality reduction ratio of 2:1 and a LeakyReLU activation function (negative slope 0.01). The final layer implements key dimension compression, outputting the rubber and plastic seal visual feature latent vector via 128 neurons. Its activation function uses the Tanh function to constrain eigenvalues to the range [-1, 1]. This refines the original high-dimensional feature map into a 128-dimensional latent vector, effectively avoiding the curse of dimensionality and significantly improving the robustness and generalization of the feature representation. This latent vector of the rubber and plastic seal visual feature serves as the core input for subsequent image reconstruction tasks, ensuring the complete preservation of core information during the image reconstruction process while reducing computational complexity and improving reconstruction efficiency.

[0043] More specifically, in step S32, the visual feature latent vector of the rubber and plastic seal is input into the decoder to obtain a reconstructed rubber and plastic seal image. It should be understood that the visual feature latent vector of the rubber and plastic seal contains highly compressed essential visual feature information of the rubber and plastic seal, and its spatial structure information and detail texture need to be restored when reconstructing the image. Therefore, the present application further designs a decoder to achieve the mapping of low-dimensional features to high-dimensional images. Here, the decoder is used to perform the opposite operation of the encoder, by gradually expanding the feature dimension of the visual feature latent vector of the rubber and plastic seal, converting it back to a high-dimensional image space, thereby reconstructing a rubber and plastic seal image similar to the original good image. As Figure 5As shown, the step S32 includes: S321, inputting the rubber and plastic seal visual feature latent vector into the fully connected layer for dimension expansion to obtain the rubber and plastic seal visual feature dimension expansion latent vector; S322, inputting the rubber and plastic seal visual feature dimension expansion latent vector into the feature shape reshaping layer to obtain the rubber and plastic seal visual feature reshaping coding map; S323, inputting the rubber and plastic seal visual feature reshaping coding map into the deconvolution layer for upsampling and spatial detail reconstruction to obtain the reconstructed rubber and plastic seal image.

[0044] In a specific embodiment of the present application, the rubber and plastic seal visual feature latent vector is first input into a fully connected layer for dimensional expansion. The fully connected layer maps the low-dimensional features of the rubber and plastic seal visual feature latent vector to a high-dimensional feature space using a weight matrix. Initial weights are initialized to follow a Xavier normal distribution, and the bias term is set to zero to ensure information integrity during the feature mapping process. LeakyReLU activation (α=0.03) is used to increase the nonlinear expressiveness of the features, thereby achieving dimensional expansion and obtaining a dimensionally expanded latent vector for the rubber and plastic seal visual feature.

[0045] Subsequently, since the one-dimensional vector output by the fully connected layer loses its spatial correlation, the image reconstruction needs to restore the two-dimensional topological structure. Therefore, in order to realize the conversion from feature to spatial coding, this application constructs a three-dimensional feature map through dimensional reorganization based on the principle of tensor reconstruction to obtain a rubber and plastic seal visual feature remodeling coding map. Specifically, first, the rubber and plastic seal visual feature dimension expansion latent vector is tensor reconstructed according to the preset spatial dimension (64×64) and the number of channels (8) to generate a 64×64×8 three-dimensional feature map; then geometric calibration enhancement is performed, and spatial coordinate information is injected through learnable position encoding (generating a 64×64 XY coordinate grid), and 3×3 convolution is used to fuse the coordinate features and content features, and then the calibrated feature map is activated by Sigmoid linear unit (SiLU) to output, that is, the rubber and plastic seal visual feature remodeling coding map. In this way, the rubber and plastic seal visual feature remodeling coding map not only retains the original feature semantics, but also embeds the spatial position prior, so that the subsequent reconstruction has geometric fidelity.

[0046] Next, in order to achieve pixel-level accurate reconstruction, this application is based on the principle of progressive upsampling, and restores the spatial details of the coding image by cascading multiple deconvolution layers to restore the visual features of the rubber and plastic seals. Among them, the deconvolution layer upsamples the low-resolution feature map to a higher resolution by learning trainable weights, and gradually restores the spatial details of the image. Similar to the encoder, the deconvolution layer of the decoder also contains batch normalization layers and activation functions. Finally, through the gradual processing of multiple layers of deconvolution layers, the last layer of the decoder outputs a reconstructed rubber and plastic seal image with the same size and number of channels as the original input image, thereby completing the image reconstruction task of the rubber and plastic seal.

[0047] More specifically, in step S33, the reconstructed rubber-plastic seal image and the image of the rubber-plastic seal labeled as a good product are input into the reconstruction error calculation module to obtain a reconstruction error. It should be understood that the reconstructed rubber-plastic seal image represents the learning effect of the normal appearance model based on the unsupervised model architecture on the normal appearance feature pattern of the original input rubber-plastic seal image. Therefore, in order to quantitatively evaluate the degree to which the normal appearance model grasps the visual feature pattern of normal rubber-plastic seals, the present application further evaluates model performance by calculating the reconstruction error between the reconstructed rubber-plastic seal image and the image of the rubber-plastic seal labeled as a good product.

[0048] In one specific embodiment of the present application, the reconstruction error is calculated as the pixel-by-pixel mean squared error between the reconstructed rubber-plastic seal image and the image of the rubber-plastic seal labeled as a qualified product. Specifically, the average of the squared differences between the corresponding pixel values between the reconstructed rubber-plastic seal image and the image of the rubber-plastic seal labeled as a qualified product is calculated. In other embodiments of the present application, other image similarity evaluation metrics, such as the Mean Absolute Error (MAE) or the Structural Similarity Index (SSIM), can also be used to evaluate the similarity between the reconstructed image and the original qualified product image, thereby revealing the effectiveness of the normal appearance model in learning the normal appearance features of the rubber-plastic seal.

[0049] More specifically, step S34 constructs a reconstruction loss function value based on the reconstruction error. It should be understood that the training process of the normal appearance model based on the unsupervised model architecture is essentially an optimization problem, requiring the definition of an objective function that can guide the update of the model's weight parameters. Therefore, in order to enable the model to continuously adjust its network parameters to minimize reconstruction error during training, the present application uses reconstruction error as the primary optimization objective and constructs a reconstruction loss function. In a specific embodiment of the present application, the reconstruction loss function is the average of the reconstruction errors of all training samples. Specifically, during training, for each batch of training datasets, the reconstruction error of each training sample in the batch is first calculated, and the reconstruction error values for each training sample are stored in an array. Then, after traversing the batch of training datasets, all reconstruction error values in the array are summed and divided by the number of samples N in the batch of training datasets to obtain the final reconstruction loss function value. The reconstruction loss function value comprehensively reflects the reconstruction performance of the model on the batch of training datasets. At the same time, by averaging the reconstruction errors of each sample, it prevents individual sample errors from interfering with the model optimization direction due to excessively large or small errors. In this way, the training effect of the normal appearance model based on the unsupervised model architecture can be fully reflected, providing a unified optimization target for the subsequent adjustment of model parameters based on the loss function, so that the model can continuously optimize the learning and reconstruction capabilities of normal images as a whole.

[0050] More specifically, step S35 adjusts the weight parameters of the encoder, decoder, and reconstruction error calculation module by minimizing the reconstruction loss function. Specifically, to enable the model to more accurately learn and reconstruct images of normal rubber and plastic seals, improve the model's ability to recognize normal appearance patterns, and thus achieve accurate quality assessment, this application uses a gradient descent algorithm to calculate the gradient of the loss function with respect to the weight parameters, then updates the weight parameters in the opposite direction of the gradient to minimize the loss function, thereby iteratively adjusting the weight parameters to optimize the model. In one specific embodiment of this application, the normal appearance model based on the unsupervised model architecture is initialized using the Xavier method to set the convolutional layer weights, with the bias terms initially set to 0. Subsequently, the stochastic gradient descent (SGD) algorithm is used to adjust the weight parameters. First, the model's optimizer is defined. For example, in deep learning frameworks such as TensorFlow or PyTorch, the SGD optimizer can be selected, and hyperparameters such as the learning rate are set. The learning rate determines the step size for each parameter update; for example, the learning rate is set to 0.001. Then, during the training process, for each batch of training samples, the reconstruction loss function value is calculated using automatic differentiation technology (such as TensorFlow's GradientTape or PyTorch's autograd) to calculate the gradient of the reconstruction loss function value with respect to all weight parameters in the encoder, decoder, and reconstruction error calculation module. The gradient represents the direction of change of the loss function under the current weight parameters. Updating the weight parameters in the opposite direction of the gradient can gradually reduce the loss function value. For example, for a certain weight parameter in the encoder , and its update formula is ,in, is the learning rate, is the loss function For weight parameters The above process is repeated, training and updating the weight parameters for each batch of samples until the reconstruction loss function converges to a smaller value or reaches a preset number of iterations, completing the model training process. In this way, the weight parameters of each module in the normal appearance model based on the unsupervised model architecture are gradually optimized, enabling the model to better learn the visual feature patterns of the normal appearance of rubber and plastic seals, thereby improving the model's accuracy and generalization ability, and providing a reliable model foundation for efficient and accurate rubber and plastic seal quality assessment.

[0051] After the training is completed, the normal appearance model based on the unsupervised model architecture has the ability to learn and reconstruct normal rubber and plastic seal images, and can be used to perform quality assessment on new rubber and plastic seal images.

[0052] During the quality assessment phase of the aforementioned rubber and plastic seal quality assessment method based on big data analysis, step S4 involves acquiring images of the rubber and plastic seal to be inspected. Specifically, the acquisition process for the images of the rubber and plastic seal to be inspected is consistent with the aforementioned image acquisition process. The images are captured by the image acquisition system on the production line and subjected to necessary image preprocessing operations such as cropping and normalization to ensure that the image quality meets the model input requirements.

[0053] During the quality assessment phase of the aforementioned rubber and plastic seal quality assessment method based on big data analysis, step S5 involves inputting the image of the rubber and plastic seal to be inspected into the trained normal appearance model based on the unsupervised model architecture to obtain an anomaly score. It should be understood that the trained normal appearance model based on the unsupervised model architecture has already learned a large number of image features of normal rubber and plastic seals and possesses the ability to reconstruct their normal appearance. Therefore, when the image of the rubber and plastic seal to be inspected is input into the trained normal appearance model based on the unsupervised model architecture, the model reconstructs the image of the rubber and plastic seal to be inspected based on the normal appearance feature patterns learned during training. Since the images of the rubber and plastic seals to be inspected may be images of good or defective rubber and plastic seals, and the normal appearance model only learns the image features of normal rubber and plastic seals during the training process, for good rubber and plastic seal images, their normal appearance feature patterns are relatively consistent with those learned by the model, and the model can well reconstruct their normal appearance, with a relatively small reconstruction error and a relatively low anomaly score. However, for defective rubber and plastic seal images, since the defective portion differs from the normal appearance, the model has difficulty accurately restoring the defective portion during the reconstruction process, resulting in a relatively large reconstruction error and a relatively high anomaly score. In this way, the presence of defects in the rubber and plastic seal images to be inspected can be determined based on their anomaly scores, thereby achieving quality assessment of the rubber and plastic seals.

[0054] In one specific embodiment of the present application, an image of a rubber or plastic seal to be inspected is input into a trained normal appearance model. Based on the image feature extraction and reconstruction process described above, the encoder of the normal appearance model first performs convolutional encoding, feature distribution offset constraint correction, feature flattening, and fully connected encoding on the image according to a preset convolutional neural network structure and trained network parameters, extracting a latent vector of visual features for the image. This latent vector is then fed into a decoder. Based on the trained decoding network parameters, the decoder performs dimensionality expansion, feature reshaping, and deconvolution in fully connected layers to achieve upsampling and spatial detail reconstruction, gradually recovering a reconstructed image of the rubber or plastic seal that matches the size of the image. Furthermore, the reconstruction error (i.e., pixel-by-pixel mean squared error) between the image and the reconstructed image is calculated and used as an anomaly score for the image to reveal the degree of deviation between the image and the appearance pattern of a qualified rubber or plastic seal. The lower the anomaly score, the closer the appearance of the rubber or plastic seal under inspection is to that of a qualified product, indicating a higher likelihood that the seal is a qualified product. Conversely, the higher the anomaly score, the more likely the seal is defective. This accurately quantifies the difference between the appearance of the rubber or plastic seal under inspection and a qualified product, providing an objective, quantitative basis for determining whether the seal is an abnormal part.

[0055] In the quality assessment stage of the above-mentioned rubber and plastic seal quality assessment method based on big data analysis, the step S6 judges whether the rubber and plastic seal to be detected is an abnormal part based on the comparison between the abnormal score and the preset threshold. That is, in order to achieve automated and accurate judgment of the quality of rubber and plastic seals, this application is based on the threshold judgment principle, and achieves quality judgment by comparing the abnormal score of the rubber and plastic seal to be detected with a pre-set reasonable threshold. In a specific embodiment of the present application, the preset threshold is determined by using statistical methods combined with actual production experience. First, 1,000 images are randomly selected from a large number of good rubber and plastic seal images, input into the trained normal appearance model, and 1,000 abnormal scores are calculated. By performing statistical analysis on these 1,000 abnormal scores, their mean is calculated. and standard deviation According to the actual production requirements for product quality and the acceptable error rate, the preset threshold is usually set to in the form of is an empirical coefficient, which has been verified through many experiments and actual production. When the system is running, it can effectively control the false positive rate while ensuring the accuracy of detection. In the actual quality assessment process, by comparing the abnormality score of the image of the rubber and plastic seal to be detected with the preset threshold, if the abnormality score is greater than the preset threshold, the rubber and plastic seal to be detected is judged to be an abnormal part, and the system automatically issues an alarm and highlights the image of the seal and related information, such as production batch, acquisition time, etc., on the display interface. At the same time, the information of the abnormal part is recorded in the abnormal product database to facilitate subsequent traceability and analysis; if the abnormality score is less than or equal to the preset threshold, the rubber and plastic seal to be detected is judged to be a good product, its information is recorded in the qualified product database, and the product is allowed to enter the next production process or packaging link. In this way, the quality of rubber and plastic seals is quickly and accurately judged, production efficiency and product quality stability are improved, and the subjectivity and uncertainty of manual judgment are effectively avoided.

[0056] In summary, the rubber and plastic seal quality assessment method based on big data analysis according to the embodiment of the present application is explained. It trains a normal appearance model based on an unsupervised model architecture by constructing and dynamically managing a dynamic baseline library containing only images of good rubber and plastic seals, so that it learns the intrinsic feature patterns of good rubber and plastic seal images. Then, in the quality assessment stage, the rubber and plastic seal image to be inspected is input into the trained normal appearance model. The normal appearance model reconstructs the image of the rubber and plastic seal to be inspected based on the intrinsic feature patterns of the good images it has learned. The image reconstruction error is calculated as the abnormality score of the rubber and plastic seal image to be inspected, thereby achieving defect assessment of the rubber and plastic seal to be inspected. This method not only improves the quality inspection accuracy and efficiency of rubber and plastic seals, but also does not require reliance on a large amount of defect sample annotation data, reducing the cost of data collection and annotation. At the same time, it can meet the detection needs of new or occasional defects and achieve comprehensive and effective quality inspection.

[0057] Furthermore, the present application also provides a rubber and plastic seal quality assessment system based on big data analysis.

[0058] Figure 6 FIG is a block diagram of a rubber and plastic seal quality assessment system based on big data analysis according to an embodiment of the present application. Figure 6As shown, the rubber and plastic seal quality assessment system 100 based on big data analysis according to the embodiment of the present application includes: a training module 110 and a quality assessment module 120; wherein, the training module 110 includes: a baseline library construction unit 111 for constructing a dynamic baseline library of good product images, wherein the dynamic baseline library of good product images includes a plurality of rubber and plastic seal images marked as good products; a dynamic management unit 112 for dynamically managing the dynamic baseline library of good product images; a normal appearance model training unit 113 for using the dynamic baseline library of good product images to train a normal appearance model based on the normal appearance model. The normal appearance model of the supervised model architecture is used to obtain the trained normal appearance model based on the unsupervised model architecture; the quality assessment module 120 includes: a rubber and plastic seal image acquisition unit 121, used to obtain the image of the rubber and plastic seal to be detected; an abnormality assessment unit 122, used to input the image of the rubber and plastic seal to be detected into the trained normal appearance model based on the unsupervised model architecture to obtain an abnormality score; an abnormal part determination unit 123, used to determine whether the rubber and plastic seal to be detected is an abnormal part based on a comparison between the abnormality score and a preset threshold.

[0059] Here, those skilled in the art will understand that the specific operations of each module in the rubber and plastic seal quality assessment system based on big data analysis have been described in the above. Figures 1 to 5 The description of the rubber and plastic seal quality assessment method based on big data analysis has been introduced in detail, and therefore, its repeated description will be omitted.

[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A rubber and plastic seal quality assessment method based on big data analysis, characterized in that: include: Training phase and quality assessment phase; The training phase includes: Constructing a dynamic baseline library of good product images, wherein the dynamic baseline library includes a plurality of images of rubber and plastic seals marked as good products; Dynamically managing the good product image dynamic baseline library; Using the good product image dynamic baseline library to train a normal appearance model based on an unsupervised model architecture to obtain a trained normal appearance model based on an unsupervised model architecture; The quality assessment stage includes: Acquire the image of the rubber and plastic seal to be inspected; Inputting the rubber and plastic seal image to be detected into the trained normal appearance model based on the unsupervised model architecture to obtain an abnormality score; Based on the comparison between the abnormality score and a preset threshold, it is determined whether the rubber-plastic seal to be inspected is an abnormal part.

2. The rubber and plastic seal quality assessment method based on big data analysis according to claim 1, characterized in that: The normal appearance model based on the unsupervised model architecture includes an encoder, a decoder, and a reconstruction error calculation module.

3. The rubber and plastic seal quality assessment method based on big data analysis according to claim 2, characterized in that: The normal appearance model based on the unsupervised model architecture is trained using the dynamic baseline library of good product images to obtain a trained normal appearance model based on the unsupervised model architecture, including: Inputting each rubber and plastic seal image marked as good in the good image dynamic baseline library into the encoder to obtain a rubber and plastic seal visual feature latent vector; Inputting the rubber and plastic seal visual feature latent vector into the decoder to obtain a reconstructed rubber and plastic seal image; Inputting the reconstructed rubber and plastic seal image and the rubber and plastic seal image marked as good product into the reconstruction error calculation module to obtain a reconstruction error; Constructing a reconstruction loss function value based on the reconstruction error; The weight parameters of the encoder, the decoder and the reconstruction error calculation module are adjusted by minimizing the reconstruction loss function value.

4. The rubber and plastic seal quality assessment method based on big data analysis according to claim 3 is characterized in that: Inputting each rubber and plastic seal image marked as good in the good image dynamic baseline library into the encoder to obtain a rubber and plastic seal visual feature latent vector, including: Performing convolution encoding on the rubber and plastic seal image marked as good product through a convolution layer to obtain a visual feature map of the rubber and plastic seal; Performing feature distribution offset constraint correction on the visual feature map of the rubber and plastic seal to obtain a corrected visual feature map of the rubber and plastic seal; After the corrected rubber and plastic seal visual feature map is flattened into a rubber and plastic seal visual feature vector, it is input into a fully connected layer to obtain the rubber and plastic seal visual feature latent vector.

5. The rubber and plastic seal quality assessment method based on big data analysis according to claim 4 is characterized in that: Performing feature distribution offset constraint correction on the visual feature map of the rubber and plastic seal to obtain a corrected visual feature map of the rubber and plastic seal, including: Calculate the parameter drift diffusion efficiency quantization coefficient of each characteristic value in the visual characteristic graph of the rubber and plastic seal; Calculating the macroscopic distribution density constraint coefficient of each eigenvalue based on the parameter drift diffusion efficiency quantification coefficient of each eigenvalue in the visual characteristic graph of the rubber and plastic seal; Based on the parameter drift diffusion efficiency quantization coefficient and the macroscopic distribution density constraint coefficient of each characteristic value in the rubber and plastic seal visual characteristic map, the characteristic value granularity offset correction is performed on the rubber and plastic seal visual characteristic map to obtain the corrected rubber and plastic seal visual characteristic map.

6. The rubber and plastic seal quality assessment method based on big data analysis according to claim 3 is characterized in that: Inputting the rubber and plastic seal visual feature latent vector into the decoder to obtain a reconstructed rubber and plastic seal image, comprising: Inputting the rubber and plastic seal visual feature latent vector into a fully connected layer for dimension expansion to obtain a rubber and plastic seal visual feature dimension expansion latent vector; Inputting the dimension-expanded latent vector of the rubber and plastic seal visual feature into the feature shape reshaping layer to obtain a rubber and plastic seal visual feature reshaping coding map; The rubber and plastic seal visual feature reshaping coding map is input into the deconvolution layer for upsampling and spatial detail reconstruction to obtain the reconstructed rubber and plastic seal image.

7. The rubber and plastic seal quality assessment method based on big data analysis according to claim 3 is characterized in that: Inputting the reconstructed rubber and plastic seal image and the rubber and plastic seal image marked as a good product into the reconstruction error calculation module to obtain a reconstruction error, including: calculating a pixel-by-pixel mean square error between the reconstructed rubber and plastic seal image and the rubber and plastic seal image marked as a good product as the reconstruction error.

8. The rubber and plastic seal quality assessment method based on big data analysis according to claim 1, characterized in that: Dynamically managing the good product image dynamic baseline library includes at least one of the following steps: Regularly remove the earliest batch of sample images stored in the dynamic baseline library of good product images; Only retain sample images within a preset time period or only retain sample images of a preset batch; When it is confirmed that a certain production condition is no longer used, all sample images related to the production condition are removed.

9. The rubber and plastic seal quality assessment method based on big data analysis according to claim 1, characterized in that: Dynamic management of the good product image dynamic baseline library includes: Performing cluster analysis on the rubber and plastic seal images marked as good products in the good product image dynamic baseline library to obtain good product image cluster centers; Acquire a new image of a rubber or plastic seal marked as good quality; Calculating a sample offset coefficient between the new rubber and plastic seal image marked as a good product and the cluster center of the good product image; Based on the comparison between the sample offset coefficient and a preset threshold, it is determined whether to trigger a dynamic update instruction of the good product image dynamic baseline library.

10. A rubber and plastic seal quality assessment system based on big data analysis, which can implement the rubber and plastic seal quality assessment method based on big data analysis according to any one of claims 1 to 9, characterized in that: The rubber and plastic seal quality assessment system based on big data analysis includes: a training module and a quality assessment module; Wherein, the training module includes: A baseline library construction unit, configured to construct a dynamic baseline library of good product images, wherein the dynamic baseline library includes a plurality of images of rubber and plastic seals marked as good products; A dynamic management unit, configured to dynamically manage the dynamic baseline library of good product images; A normal appearance model training unit, configured to train a normal appearance model based on an unsupervised model architecture using the good product image dynamic baseline library to obtain a trained normal appearance model based on an unsupervised model architecture; The quality assessment module includes: A rubber and plastic seal image acquisition unit, used to acquire an image of the rubber and plastic seal to be inspected; an abnormality assessment unit, configured to input the image of the rubber and plastic seal to be detected into the trained normal appearance model based on the unsupervised model architecture to obtain an abnormality score; The abnormal part determination unit is used to determine whether the rubber-plastic sealing part to be detected is an abnormal part based on the comparison between the abnormality score and a preset threshold.

Images