Industrial quality inspection-oriented pre-training data system construction method and system, and medium

Abstract

The invention discloses a pre-training data system construction method and system for industrial quality inspection and a medium, and the method comprises the following steps: collecting data according to a preset field-product-mode three-stage classification architecture system in combination with a multi-dimensional demand; organizing and storing the collected data, and ensuring data set availability and subsequent model training efficiency; and performing labeling and label management on the data based on the process information, the defect risk level and the product type information, and constructing a pre-training data system for industrial quality inspection. According to the method, a high-quality data set which meets industrial practical application requirements and can effectively adapt to complex working conditions and detail changes is constructed, a data basis can be provided for subsequent unsupervised learning and anomaly detection, the performance of the model in a real production environment is further improved, and the method is suitable for large-scale popularization and application. The method can be widely applied to a plurality of subdivided technology and product fields such as industrial artificial intelligence, computer vision and intelligent manufacturing.

Description

Technical Field

[0001] This invention relates to the field of data infrastructure design technology for intelligent analysis of industrial images and unsupervised anomaly detection, and in particular to a method, system and medium for constructing a pre-training data system for industrial quality inspection. Background Technology

[0002] Currently, existing industrial visual quality inspection systems typically rely on transfer learning techniques, especially pre-trained models transferred from natural image tasks (such as ImageNet). Models based on ImageNet pre-training are difficult to adapt to complex production line conditions (such as production environment, product materials, diverse types of defects, and multi-source imaging interference), so it is necessary to build pre-trained datasets adapted to the industrial field.

[0003] These technologies generally employ the following technical solutions: Transfer learning: In existing technologies, many industrial visual quality inspection systems rely on pre-trained models on natural image datasets (such as ImageNet). By transferring the learning outcomes of these models on natural images to industrial quality inspection tasks, defect detection and classification can be performed using the pre-trained feature representations of the models.

[0004] Contrastive learning and self-supervised learning: In recent years, some systems have adopted self-supervised learning and contrastive learning techniques. By learning features from a large amount of unlabeled data, the model can autonomously discover patterns from the data and improve the robustness of the model.

[0005] Industrial vision datasets: Some industrial vision datasets (such as MVTec-AD) are widely used for defect detection tasks. They assist models in identifying defects on industrial products by providing labeled images of different defect types. While these datasets provide a foundation for unsupervised learning, they still have some limitations, such as the scarcity of defect samples, idealization and simplification of scenes, etc.

[0006] The currently used technical solution has the following drawbacks: 1. Domain differences: Most existing technologies rely on pre-trained models based on natural image datasets (such as ImageNet), which differs significantly from the imaging methods, texture features, and processes of industrial quality inspection tasks. This makes it difficult for these transfer learning models to fully adapt to the complex working conditions and details of real industrial production lines.

[0007] Domain differences can lead to insufficient model generalization ability: Existing industrial visual quality inspection systems generally rely on pre-trained models based on natural images (such as ImageNet). These models are trained in natural scenes, and the features they learn differ significantly from the geometric structure, surface texture, and manufacturing processes of industrial images. As a result, the models perform poorly in detecting complex conditions in industrial images, especially small defects and subtle texture changes, exhibiting poor generalization ability.

[0008] 2. Dataset Limitations: While existing industrial vision datasets (such as MVTec-AD) possess a certain degree of diversity, they still have limitations in terms of defect types, scale, and morphology, and cannot fully reflect all the complex situations encountered in the industrial quality inspection process. In particular, their ability to capture small defects and subtle texture changes is poor.

[0009] The existing industrial vision datasets (such as MVTec-AD) lack limitations and diversity, failing to capture the complexities of real-world industrial production environments. While they cover a range of defect types, most are based on artificial synthesis or simple scene generation, failing to adequately reflect the intricate background interference, process variations, material properties, and equipment noise inherent in industrial production. In particular, existing datasets struggle to cover scenarios involving small defects, texture disturbances, and the coexistence of multiple defects.

[0010] 3. Ignoring dynamic changes during the production process: Existing industrial quality inspection systems often overlook the constantly changing processes, environments, and equipment conditions during data acquisition and defect detection. This limitation makes the model prone to performance fluctuations when dealing with variable operating conditions, complex backgrounds, and unexpected situations, and it cannot effectively adapt to the dynamic changes of the production line.

[0011] In summary, existing technologies have failed to effectively address the domain differences between industrial and natural images, and lack label-free anomaly modeling methods adapted to industrial quality inspection tasks, resulting in insufficient robustness and generalization ability of existing systems. Summary of the Invention

[0012] The main objective of this invention is to address the problems of domain differences, dataset limitations, and insufficient unlabeled modeling in existing technologies by proposing a method, system, and medium for constructing a pre-training data system for industrial quality inspection. The aim is to build a high-quality dataset that meets the actual needs of industrial applications and can effectively adapt to complex working conditions and detailed changes, providing a data foundation for subsequent unsupervised learning and anomaly detection, and further improving the performance of the model in real production environments. It can be widely applied to multiple sub-technologies and product fields such as industrial artificial intelligence, computer vision, and intelligent manufacturing.

[0013] To achieve the above objectives, this invention provides a method for constructing a pre-training data system for industrial quality inspection, the method comprising the following steps: Step S10: Based on the pre-set three-level classification architecture of domain-product-mode, data is collected in combination with multi-dimensional requirements. The multi-dimensional requirements include at least one or more of the following: device adaptation, image acquisition consistency, process diversity, and data enhancement. Step S20: Organize and store the collected data to ensure dataset availability and subsequent model training efficiency; Step S30: Based on process information, defect risk level and product type information, the data is labeled and tagged to build a pre-training data system for industrial quality inspection.

[0014] A further technical solution of the present invention is that step S10 includes: Step S101, Process Modeling and Defect Definition: Define the location and mode of defects for each type of product at each process stage; Step S102, Imaging Equipment Adaptation and Selection: Select the imaging equipment that matches each process and product. Step S103: Workstation-level data acquisition to ensure data acquisition consistency and reduce external interference; Step S104, Data Acquisition Enhancement: Increase image diversity and simulate disturbances in the real environment; Step S105, Abnormal Interference Retention and Data Cleaning: Retain abnormal noise in the actual industrial environment while avoiding the introduction of unnecessary interference.

[0015] A further technical solution of the present invention is that step S20 includes: Step S201, File Structure and Storage Management: Clearly manage data and ensure easy access and querying; Step S202, Image Slicing and Sample Balancing: Ensure that the dataset can support multi-level and fine-grained training tasks; Step S203, Image cleaning and quality screening: Ensure the quality of the dataset and remove low-quality samples; Step S204, Multi-view and Multi-environment Adaptation: Simulate image data changes under different environments.

[0016] A further technical solution of the present invention is that step S30 includes: Step S301: Use a combination of automated annotation tools and manual quality inspection to annotate the data to ensure annotation accuracy; Step S302: Regularly review the annotations to ensure the quality and consistency of all labeled data and avoid affecting the training results due to annotation errors.

[0017] A further technical solution of the present invention is that the annotation content includes image-level meta-information and risk level. In the annotation of image-level meta-information, each image includes process type, product information, shooting conditions, and defect information. In the process of annotating the risk level, each image is classified according to the severity of the defect to help the subsequent model identify defects of different levels.

[0018] To achieve the above objectives, the present invention also proposes a pre-training data system construction system for industrial quality inspection. The system includes a memory, a processor, and a pre-training data system construction program for industrial quality inspection stored on the processor. When the processor runs the pre-training data system construction program for industrial quality inspection, it executes the following steps: Based on a pre-defined three-level classification architecture of domain-product-mode, and combined with multi-dimensional demand data collection, the multi-dimensional demand includes at least one or more of the following: device adaptation, image acquisition consistency, process diversity, and data enhancement. Organize and store the collected data to ensure dataset availability and subsequent model training efficiency; Data is labeled and tagged based on process information, defect risk level, and product type information to build a pre-training data system for industrial quality inspection.

[0019] A further technical solution of the present invention is that, when the pre-training data system construction program for industrial quality inspection is run by the processor, the following steps are also performed: Process modeling and defect definition: Define the location and pattern of defects for each type of product at each process stage; Equipment compatibility and selection: Select imaging equipment that matches each process and product. Workstation-level data acquisition ensures data consistency and reduces external interference; Enhanced data acquisition: Increases image diversity and simulates disturbances in real-world environments.

[0020] A further technical solution of the present invention is that, when the pre-training data system construction program for industrial quality inspection is run by the processor, the following steps are also performed: File structure and storage management: Clearly manage data and ensure easy access and querying; Image slicing and sample balancing: ensure that the dataset can support multi-level and fine-grained training tasks; Image cleaning and quality screening: Ensure dataset quality and remove low-quality samples; Multi-view and multi-environment adaptation: Simulates changes in image data under different environments.

[0021] A further technical solution of the present invention is that step S30 includes: The data is labeled using a combination of automated labeling tools and manual quality control to ensure labeling accuracy. Regularly review the annotations to ensure the quality and consistency of all labeled data and avoid affecting training results due to annotation errors.

[0022] To achieve the above objectives, the present invention also proposes a computer-readable storage medium storing a pre-training data system construction program for industrial quality inspection, wherein the pre-training data system construction program for industrial quality inspection is executed by a processor to perform the steps of the method described above.

[0023] The present invention provides a pre-training data system construction method, system, and medium for industrial quality inspection, which solves the technical problems in the prior art and achieves the following technical effects: 1. The model's generalization ability and robustness have been improved. By using a data system specifically designed for industrial quality inspection, the problem of domain differences in transfer learning has been solved.

[0024] 2. It enhances the diversity and realism of the data, simulates more complex changes in the production environment, and improves the model's ability to perceive complex backgrounds and minor defects.

[0025] 3. It supports unsupervised learning, enabling training on unlabeled data, which reduces the cost of data labeling and improves the model's ability to detect anomalies.

[0026] 4. It has strong adaptability and can cover industrial production with multiple processes and specifications, improving the generalization of the model and enhancing its application value in actual industrial scenarios.

[0027] The aforementioned beneficial effects make this invention highly valuable in the field of industrial quality inspection, and can significantly improve the efficiency and accuracy of industrial visual quality inspection systems. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0029] Figure 1 This is a flowchart illustrating a preferred embodiment of the pre-training data system construction method for industrial quality inspection of the present invention; Figure 2 This is a detailed flowchart of step S10; Figure 3 This is a detailed flowchart of step S20; Figure 4 This is a detailed flowchart of step S30; Figure 5 This is a schematic diagram of a three-level classification architecture; Figure 6 This is a schematic diagram of the third level of the data system—schema partitioning; Figure 7 This is a schematic diagram of the process for collecting and organizing industrial quality inspection data.

[0030] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0032] This invention proposes a method for constructing a pre-training data system for industrial quality inspection. In the field of industrial quality inspection, designing an efficient and reliable dataset is crucial, as it can provide strong support for both unsupervised and self-supervised learning of models.

[0033] The core objective of this invention is to construct an industrial “normal image” dataset that supports self-supervised learning, providing a transfer foundation for downstream tasks.

[0034] The data structure of this invention adopts a three-level classification system of domain → product → pattern. Each domain contains multiple processes and product types, and the data types include normal images and defective images.

[0035] This invention uses automated acquisition equipment and a robotic arm synchronous control system to ensure high consistency during the data acquisition process; it also introduces industrial environmental interferences such as lighting and background changes.

[0036] The data organization strategy of this invention includes image-level process metadata h and pixel-free labels, which assists in subsequent model training by labeling defect types, risk levels, etc.

[0037] In the self-supervised pre-training adaptation process, this invention designs enhancement strategies to adapt to tasks such as contrastive learning, MAE, and patch-based learning, thereby improving the model's ability to learn from local defects.

[0038] The application goal of this invention is to provide high-quality pre-training data for general encoders and downstream tasks in the industrial field, supporting detection tasks such as PatchCore and GLASS.

[0039] This invention leverages the data acquisition capabilities and automation systems of existing industrial projects to ensure the feasibility and scalability of datasets, providing a guarantee of feasibility for industrial quality inspection.

[0040] In industrial quality inspection scenarios, data acquisition is not merely about obtaining images, but also about ensuring that these images possess sufficient diversity, representativeness, and quality to facilitate subsequent training and model optimization. Therefore, designing a standardized and efficient data acquisition process is crucial for the entire industrial vision pre-training data system.

[0041] The core objective of data acquisition is to collect a sufficient number of samples from real production lines and ensure that these samples reflect the diversity and complexity of process and defect characteristics under different environmental changes, process differences, and product specifications. The detailed data acquisition and organization process is described below.

[0042] Specifically, such as Figures 1 to 7 As shown, a preferred embodiment of the pre-training data system construction method for industrial quality inspection of the present invention includes the following steps: Step S10: Based on the pre-defined three-level classification architecture of domain-product-mode, data is collected in combination with multi-dimensional requirements. The multi-dimensional requirements include at least one or more of the following: device adaptation, image acquisition consistency, process diversity, and data enhancement.

[0043] This invention employs a three-tiered classification architecture of domain, product, and defect pattern to manage and organize industrial quality inspection data from multiple levels, including domain, product, and defect pattern. Each level has clear definitions and standard categories, providing a solid framework for data collection, annotation, and model training.

[0044] This systematic classification method not only helps organize diverse industrial data but also effectively supports detection tasks across multiple domains, products, and defect types, improving the model's generalization ability. In practical applications, suitable self-supervised learning tasks and data augmentation methods can be designed based on the characteristics of each domain, product, and defect pattern to further enhance detection accuracy.

[0045] The following is an introduction to the three-level classification architecture of this field: product, business model.

[0046] 1. Domain A "domain" represents a manufacturing process or industrial process. Different domains have different processes and techniques, resulting in different image characteristics. Data acquisition methods, image backgrounds, lighting, reflections, and material types should all differ within each domain. The purpose of domain design is to provide different perspectives and contexts for subsequent product classification and defect type modeling.

[0047] The domain classifications and descriptions are shown in Table 1: Table 1 Key elements of domain definition include: (1) Process equipment: Different fields use different process equipment (such as punch press, welding machine, polishing machine, spraying equipment, etc.), and the process of these equipment directly affects the quality of the product.

[0048] (2) Environmental control: such as the impact of different environments such as temperature, humidity, and pressure on production.

[0049] (3) Material properties: Different materials are used in each field, such as metals, ceramics, and plastics. Their physical properties have different effects on product quality.

[0050] 2. Product The "product" level describes specific product parts. Different products have different shapes, structures, dimensions, and processing requirements. Each product belongs to a specific field, within which there are multiple different specifications. The purpose of the product level is to clarify specific process applications and the key points of inspection on different production lines.

[0051] Product categories, descriptions, and common specifications are shown in Table 2: Table 2 Key elements of a product hierarchy include: (1) Product structure and size: Each product has different structural features and size, which affect the manifestation of defects.

[0052] (2) Material and processing: Different products use different materials (such as metal, plastic, glass, etc.), and the surface texture, reflectivity and brittleness of these materials also determine the way defects are manifested.

[0053] (3) Production line and process: Each product may go through different processes during production (such as stamping, welding, polishing, etc.), which will have different effects on the quality of the final product.

[0054] 3. Pattern "Defect patterns" refer to both normal images and all possible different types of defects in a product, which have different shapes, sizes, distributions, and manifestations. The hierarchy of defect patterns is the core content that the final detection model needs to learn and recognize.

[0055] The pattern classifications, descriptions, and examples are shown in Table 3: Table 3 The key elements of defect mode include: (1) Size and shape of defects: Defects can be large or small. They may be large-area deformations on the surface or tiny cracks or scratches.

[0056] (2) Distribution and location of defects: Some defects may appear in specific areas or may be unevenly distributed throughout the product.

[0057] (3) Visibility of defects: Different defects react differently to light. Some may be difficult to see with the naked eye, while others may be easy to identify.

[0058] During the data acquisition process, multiple requirements need to be considered, including equipment compatibility, image acquisition consistency, process diversity, and data augmentation.

[0059] Please refer to Figure 2 In this embodiment, step S10 specifically includes: Step S101, Process Modeling and Defect Definition: Define the location and mode of defects for each type of product at each process stage.

[0060] In this embodiment, the goal of process modeling and defect definition is to define the location and pattern of defects for each type of product at each process stage.

[0061] The details of process modeling and defect definition include the following three aspects: (1) Determine the types of defects that may occur in each product in each area (such as stamping, welding, polishing, etc.).

[0062] (2) Based on the product’s usage scenario and process steps, identify the possible defect areas for each product (such as metal surfaces, electronic component soldering points, packaging material seams, etc.).

[0063] (3) Determine the location, shape, and severity of different defects and label them as the risk level of the defect occurrence.

[0064] Step S102, Imaging Equipment Adaptation and Selection: Select an imaging equipment that matches each process and product.

[0065] In this embodiment, the goal of imaging device adaptation and selection is to select an imaging device that matches each process and product.

[0066] The details of imaging equipment adaptation and selection include the following two aspects: (1) For different processes, select different types of imaging equipment, such as area array camera (suitable for large area shooting), line scan camera (suitable for scanning long strip objects), infrared camera (suitable for detecting thermal non-uniformity), 3D structured light camera (suitable for capturing three-dimensional defects), etc.

[0067] (2) For different products, select appropriate resolution, focal length and light source type to ensure that the image quality is clear enough to capture details.

[0068] Step S103: Workstation-level data acquisition to ensure data acquisition consistency and reduce external interference.

[0069] In this embodiment, the goal of workstation-level data acquisition is to ensure the consistency of data acquisition and reduce external interference.

[0070] The details of workstation-level data acquisition include the following two aspects: (1) Use robotic arms or synchronously controlled track systems on the production line to acquire images, ensuring the consistency of each image acquisition position and angle.

[0071] (2) Ensure high consistency of images collected at different workstations and production line segments through an automated control system to avoid image differences caused by human operation.

[0072] Step S104, data acquisition enhancement: increase image diversity and simulate disturbances in the real environment.

[0073] In this embodiment, the goal of data acquisition enhancement is to increase image diversity and simulate disturbances in a real production environment.

[0074] The details of the enhanced data acquisition include the following two aspects: (1) During the data collection process, actively adjust the light source, focal length, and shooting angle to create different perspectives and lighting changes to ensure data diversity.

[0075] (2) By using dynamic backgrounds or complex environmental interference, such as simulating the effects of reflection, dust and reflection in a factory, the difficulty of training data can be increased.

[0076] Step S105, Abnormal Interference Retention and Data Cleaning: Retain abnormal noise in the actual industrial environment while avoiding the introduction of unnecessary interference.

[0077] In this embodiment, the goal of abnormal interference retention and data cleaning is to retain abnormal noise in the actual industrial environment while avoiding the introduction of unnecessary interference.

[0078] Abnormal interference retention and data cleaning include the following two aspects: (1) Do not artificially remove disturbances such as reflections and friction textures caused by equipment failures or changes in operating conditions, because these disturbances can simulate common noises in the production process.

[0079] (2) Use automated image cleaning tools to filter low-quality images (such as blurry images, blank images, etc.) based on indicators such as image clarity, contrast, and edge distribution to ensure the effectiveness of subsequent image annotation and model training.

[0080] Step S20: Organize and store the collected data to ensure dataset availability and subsequent model training efficiency.

[0081] The data organization strategy in this embodiment (for training purposes) is crucial for ensuring dataset availability and subsequent model training efficiency by effectively organizing and storing the data after collection.

[0082] Please refer to Figure 3 In this embodiment, step S20, the strategy for organizing data includes the following steps: Step S201, File Structure and Storage Management: Clearly manage data and ensure easy access and querying.

[0083] In this embodiment, the goal of file structure and storage management is to clearly manage data and ensure easy access and querying.

[0084] The details of file structure and storage management include the following two aspects: (1) The file structure is designed as follows: Domain / Product / Process Name / Image ID.png. Each data category is classified by domain, product, and process for easy management.

[0085] (2) Each image is accompanied by a metadata file (such as JSON or YAML format) that records relevant information during the acquisition, such as product type, process, lighting conditions, focal length, shooting angle, etc.

[0086] Step S202, Image Slicing and Sample Balancing: Ensure that the dataset can support multi-level and fine-grained training tasks.

[0087] In this embodiment, the goal of image slicing and sample balancing is to ensure that the dataset can support multi-level and fine-grained training tasks.

[0088] The details of image slicing and sample equalization include the following two aspects: (1) Slice each image to generate sliding window slices of size 512×512 or 224×224 to adapt to the patch-level pre-training task.

[0089] (2) Ensure a balanced number of images for each type of product to avoid bias in model training due to an excessive number of images for a particular type of product. This can be addressed through acquisition strategies and sample balancing techniques.

[0090] Step S203, Image cleaning and quality screening: Ensure the quality of the dataset and remove low-quality samples.

[0091] In this embodiment, the goal of image cleaning and quality screening is to ensure the quality of the dataset and remove low-quality samples.

[0092] Image cleaning and quality screening involve the following two aspects: (1) Based on indicators such as image clarity, contrast, and edge distribution, use an automated filter to clean the data and remove blurry images, noisy images, and empty images.

[0093] (2) Retain high-quality images that meet industrial quality inspection standards to ensure that subsequent training will not be affected by poor-quality data.

[0094] Step S204, Multi-view and Multi-environment Adaptation: Simulate image data changes under different environments.

[0095] In this embodiment, the goal of multi-view and multi-environment adaptation is to simulate changes in image data under different environments.

[0096] The details of multi-view and multi-environment adaptation include the following aspects: In each domain, ensure that the images cover a variety of lighting conditions, focal lengths, shooting angles, background interference, and other factors. Ensure that the dataset has strong environmental heterogeneity to improve the model's adaptability to changing environments.

[0097] Step S30: Based on process information, defect risk level and product type information, the data is labeled and tagged to build a pre-training data system for industrial quality inspection.

[0098] After data collection is completed, this embodiment requires data annotation and label management. This embodiment does not perform pixel-level annotation of defects, but focuses on information such as process information, defect risk level, and product type.

[0099] The annotation content in this embodiment includes image-level metadata and risk level.

[0100] For image-level metadata, each image includes metadata such as process type, product information, shooting conditions, and defect information (such as cracks, scratches, and other defect types).

[0101] For risk levels, each image is classified according to the severity of the defect to help subsequent models identify defects of different levels.

[0102] Please refer to Figure 4 In this embodiment, step S30, the annotation tool and process, includes the following steps: Step S301: Use a combination of automated annotation tools and manual quality inspection to annotate the data to ensure annotation accuracy.

[0103] Step S302: Regularly review the annotations to ensure the quality and consistency of all labeled data and avoid affecting the training results due to annotation errors.

[0104] The pre-trained data system design for industrial quality inspection provided by this invention has the following significant advantages and positive effects compared with the prior art, mainly reflected in the following aspects: 1. It solves the problem of domain differences and improves the model's generalization ability: Comparison with existing technologies: Existing technologies typically rely on pre-training on natural image datasets such as ImageNet. When transferred to industrial quality inspection tasks, the models are affected by domain differences (such as imaging methods, texture features, and process flows), resulting in insufficient robustness and difficulty in adapting to complex working conditions and subtle defects in real industrial environments.

[0105] Advantages of this invention: This invention constructs a pre-trained data system specifically designed for industrial quality inspection tasks, directly using real industrial production line data and covering multiple manufacturing processes and product categories. This allows the model to learn feature representations specific to the industrial environment, rather than generic features from natural images. Consequently, the model's generalization ability and robustness in industrial environments are significantly improved, effectively handling complex scenarios involving small defects, subtle texture variations, and multiple coexisting defects.

[0106] Beneficial effects: This improves the adaptability of industrial quality inspection systems, enabling models to better adapt to various industrial production environments, enhancing the accuracy and generalization ability of defect detection, thereby reducing quality problems caused by inaccurate detection in industrial production.

[0107] 2. Enhanced data diversity and realism Comparison with existing technologies: Existing datasets, such as MVTec-AD, while covering some types of industrial defects, are mostly created in artificially synthesized or simplified industrial scenarios. They lack sufficient simulation of the complex backgrounds and process variations in real production environments and cannot represent the real production differences and environmental changes in industrial quality inspection.

[0108] Advantages of this invention: This invention creates an industrial dataset encompassing multiple fields, products, processes, and defect patterns, realistically reflecting various processes and product types in industrial production. It offers greater diversity, particularly in defect details, product specification variations, and changes in the imaging environment. The dataset not only includes normal product images but also simulates image disturbances under different processes, materials, and lighting conditions, such as scratches, oil reflections, glare, and equipment friction.

[0109] Beneficial effects: This enhances the diversity and realism of the dataset, enabling model training to move beyond simple, idealized, artificially synthesized data and instead learn within complex and dynamic real-world industrial environments. This not only makes the model more accurate in real-world scenarios but also allows it to maintain high-efficiency anomaly detection capabilities even when faced with uncontrollable factors such as environmental interference and equipment changes.

[0110] 3. Supports unsupervised learning and anomaly detection, improving the effectiveness of model training. Comparison with existing technologies: Existing technologies often rely on large amounts of labeled data to train models. In particular, many industrial quality inspection tasks require precise labeling of each image. This approach is not only costly but also difficult to scale. In particular, when there are insufficient labeled samples, the model performance may be greatly reduced.

[0111] Advantages of this invention: The pre-training data system of this invention adopts an unsupervised learning approach, utilizing comparative learning between normal product images and defective images, and is trained through self-supervised learning tasks (such as contrastive learning, Masked AE, PatchMIM, etc.). This not only reduces the cost of data annotation but also enables the learning of efficient anomaly detection features even without labels.

[0112] By employing various data augmentation strategies (such as occlusion, noise reduction, blurring, and cropping), the model can better understand the differences between normal and defective images, thereby achieving unsupervised anomaly detection.

[0113] Beneficial effects: It reduces labeling costs and alleviates the burden of manual labeling by training models with unlabeled data. Especially in large-scale industrial quality inspection scenarios, it can efficiently handle inspection tasks of different product and process types.

[0114] It improves the model's ability to identify anomalies, especially minor defects and uncommon anomaly patterns, enhancing the model's flexibility and robustness.

[0115] 4. Supports multiple processes and specifications in industrial applications, demonstrating strong adaptability. Comparison with existing technologies: Most existing industrial datasets focus on a single process or product category, meaning that models can only learn from specific types of images during training. This approach leads to a lack of adaptability to products with multiple processes and specifications, and in industrial production, the models often fail to generalize well when faced with diverse process requirements and complex product specifications.

[0116] Advantages of this invention: The data system of this invention covers multiple fields and products, such as: 3C electronics (e.g., mobile phone lenses, linear motors, integrated circuit boards, etc.), metal processing (e.g., connectors, heat sinks), and light industry (e.g., non-woven fabrics, packaging bags, cotton and linen fabrics, etc.). Each product category contains data samples with multiple specifications and different process combinations, fully considering the manufacturing processes and defect characteristics of different products.

[0117] Beneficial effects: Its strong adaptability enables it to handle defect detection tasks under different processes, product specifications, and materials. When faced with complex and ever-changing industrial production lines, the model can quickly adapt to different production environments and needs, and demonstrates excellent detection performance in multi-process and multi-product scenarios.

[0118] In summary, the present invention solves the technical problems in the prior art and achieves the following technical effects: 1. The model's generalization ability and robustness have been improved. By using a data system specifically designed for industrial quality inspection, the problem of domain differences in transfer learning has been solved.

[0119] 2. It enhances the diversity and realism of the data, simulates more complex changes in the production environment, and improves the model's ability to perceive complex backgrounds and minor defects.

[0120] 3. It supports unsupervised learning, enabling training on unlabeled data, which reduces the cost of data labeling and improves the model's ability to detect anomalies.

[0121] 4. It has strong adaptability and can cover industrial production with multiple processes and specifications, improving the generalization of the model and enhancing its application value in actual industrial scenarios.

[0122] The aforementioned beneficial effects make this invention highly valuable in the field of industrial quality inspection, and can significantly improve the efficiency and accuracy of industrial visual quality inspection systems.

[0123] To achieve the above objectives, the present invention also proposes a pre-training data system construction system for industrial quality inspection. The system includes a memory, a processor, and a pre-training data system construction program for industrial quality inspection stored on the processor. When the processor runs the pre-training data system construction program for industrial quality inspection, it executes the following steps: Based on a pre-defined three-level classification architecture of domain-product-mode, and combined with multi-dimensional demand data collection, the multi-dimensional demand includes at least one or more of the following: device adaptation, image acquisition consistency, process diversity, and data enhancement. Organize and store the collected data to ensure dataset availability and subsequent model training efficiency; Data is labeled and tagged based on process information, defect risk level, and product type information to build a pre-training data system for industrial quality inspection.

[0124] Furthermore, the pre-training data system construction program for industrial quality inspection, when run by the processor, also performs the following steps: Process modeling and defect definition: Define the location and pattern of defects for each type of product at each process stage; Equipment compatibility and selection: Select imaging equipment that matches each process and product. Workstation-level data acquisition ensures data consistency and reduces external interference; Enhanced data acquisition: Increases image diversity and simulates disturbances in real-world environments.

[0125] Furthermore, the pre-training data system construction program for industrial quality inspection, when run by the processor, also performs the following steps: File structure and storage management: Clearly manage data and ensure easy access and querying; Image slicing and sample balancing: ensure that the dataset can support multi-level and fine-grained training tasks; Image cleaning and quality screening: Ensure dataset quality and remove low-quality samples; Multi-view and multi-environment adaptation: Simulates changes in image data under different environments.

[0126] Further, step S30 includes: The data is labeled using a combination of automated labeling tools and manual quality control to ensure labeling accuracy. Regularly review the annotations to ensure the quality and consistency of all labeled data and avoid affecting training results due to annotation errors.

[0127] To achieve the above objectives, the present invention also proposes a computer-readable storage medium storing a pre-training data system construction program for industrial quality inspection, wherein the pre-training data system construction program for industrial quality inspection is executed by a processor to perform the steps of the method described above.

[0128] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made under the concept of the present invention using the description and drawings of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A method for constructing a pre-training data system for industrial quality inspection, characterized in that, The method includes the following steps: Step S10: Based on the pre-set three-level classification architecture of domain-product-mode, data is collected in combination with multi-dimensional requirements. The multi-dimensional requirements include at least one or more of the following: device adaptation, image acquisition consistency, process diversity, and data enhancement. Step S20: Organize and store the collected data to ensure dataset availability and subsequent model training efficiency; Step S30: Based on process information, defect risk level and product type information, the data is labeled and tagged to build a pre-training data system for industrial quality inspection.

2. The method for constructing a pre-training data system for industrial quality inspection according to claim 1, characterized in that, Step S10 includes: Step S101, Process Modeling and Defect Definition: Define the location and mode of defects for each type of product at each process stage; Step S102, Imaging Equipment Adaptation and Selection: Select the imaging equipment that matches each process and product. Step S103: Workstation-level data acquisition to ensure data acquisition consistency and reduce external interference; Step S104, Data Acquisition Enhancement: Increase image diversity and simulate disturbances in the real environment; Step S105, Abnormal Interference Retention and Data Cleaning: Retain abnormal noise in the actual industrial environment while avoiding the introduction of unnecessary interference.

3. The method for constructing a pre-training data system for industrial quality inspection according to claim 1, characterized in that, Step S20 includes: Step S201, File Structure and Storage Management: Clearly manage data and ensure easy access and querying; Step S202, Image Slicing and Sample Balancing: Ensure that the dataset can support multi-level and fine-grained training tasks; Step S203, Image cleaning and quality screening: Ensure the quality of the dataset and remove low-quality samples; Step S204, Multi-view and Multi-environment Adaptation: Simulate image data changes under different environments.

4. The method for constructing a pre-training data system for industrial quality inspection according to claim 1, characterized in that, Step S30 includes: Step S301: Use a combination of automated annotation tools and manual quality inspection to annotate the data to ensure annotation accuracy; Step S302: Regularly review the annotations to ensure the quality and consistency of all labeled data and avoid affecting the training results due to annotation errors.

5. The method for constructing a pre-training data system for industrial quality inspection according to claim 4, characterized in that, The annotation content includes image-level metadata and risk level. In the image-level metadata annotation, each image includes process type, product information, shooting conditions, and defect information. In the process of annotating the risk level, each image is classified according to the severity of the defect to help the subsequent model identify defects of different levels.

6. A pre-training data system construction system for industrial quality inspection, characterized in that, The system includes a memory, a processor, and a pre-training data system construction program for industrial quality inspection stored on the processor. The pre-training data system construction program for industrial quality inspection is executed by the processor to perform the following steps: Based on a pre-defined three-level classification architecture of domain-product-mode, and combined with multi-dimensional demand data collection, the multi-dimensional demand includes at least one or more of the following: device adaptation, image acquisition consistency, process diversity, and data enhancement. Organize and store the collected data to ensure dataset availability and subsequent model training efficiency; Data is labeled and tagged based on process information, defect risk level, and product type information to build a pre-training data system for industrial quality inspection.

7. The method for constructing a pre-training data system for industrial quality inspection according to claim 6, characterized in that, When the processor runs the pre-training data system construction program for industrial quality inspection, it also performs the following steps: Process modeling and defect definition: Define the location and pattern of defects for each type of product at each process stage; Equipment compatibility and selection: Select imaging equipment that matches each process and product. Workstation-level data acquisition ensures data consistency and reduces external interference; Enhanced data acquisition: Increases image diversity and simulates disturbances in real-world environments.

8. The method for constructing a pre-training data system for industrial quality inspection according to claim 6, characterized in that, When the pre-training data system construction program for industrial quality inspection is run by the processor, the following steps are also performed: File structure and storage management: Clearly manage data and ensure easy access and querying; Image slicing and sample balancing: ensure that the dataset can support multi-level and fine-grained training tasks; Image cleaning and quality screening: Ensure dataset quality and remove low-quality samples; Multi-view and multi-environment adaptation: Simulates changes in image data under different environments.

9. The method for constructing a pre-training data system for industrial quality inspection according to claim 6, characterized in that, Step S30 includes: The data is labeled using a combination of automated labeling tools and manual quality control to ensure labeling accuracy. Regularly review the annotations to ensure the quality and consistency of all labeled data and avoid affecting training results due to annotation errors.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a pre-training data system construction program for industrial quality inspection, which, when executed by a processor, performs the steps of the method as described in any one of claims 1 to 5.

Images