Industrial quality inspection method and system integrating real-time defect detection and dynamic updating

Abstract

The invention provides an industrial quality inspection method and system integrating real-time defect detection and dynamic updating. The method comprises the following steps: S1, synchronously acquiring a product image by adopting visible light and an infrared sensor; s2, inputting the image into an edge side improved DETR model for real-time reasoning to obtain a detection result, monitoring the detection result at the same time, and if the detection result does not reach a preset effect, performing online training; s3, multi-station detection model co-evolution is carried out through a production line federal aggregation center; and S4, performing simulation verification and closed-loop feedback on the updated model through a three-level verification system. According to the method, hour-level rapid adaptation to new defect types can be achieved, the performance and robustness of the model are improved, the false detection rate and the omission ratio of the system are effectively and greatly reduced, and rapid and accurate positioning of defect production links is achieved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing and computer vision technology, and in particular to an industrial quality inspection method and system integrating real-time defect detection and dynamic updating. Background Art

[0002] At the intersection of intelligent manufacturing and computer vision, the development of industrial quality inspection systems has undergone a transformation from traditional manual inspection to intelligent, automated quality inspection. Traditional manual inspection methods suffer from low efficiency, high missed detection rates, and high costs. These issues are particularly prominent in precision manufacturing scenarios such as electronics and automotive parts. With intensified market competition and rising customer demands for product quality, traditional quality inspection methods are no longer able to meet the demands of modern manufacturing.

[0003] In recent years, the rapid development of artificial intelligence and machine learning technologies has brought new opportunities to industrial quality inspection. In particular, the application of deep learning technology in image recognition and defect detection has enabled intelligent quality inspection systems to achieve high-precision, high-efficiency, real-time inspections. However, existing intelligent quality inspection systems still face limitations when faced with changing production environments and emerging defect types. For example, model updates often require large amounts of labeled data and lengthy offline training, which cannot meet real-time requirements. Furthermore, collaboration and knowledge sharing between different production lines are insufficient, making it difficult to achieve efficient cross-line collaboration and accurately locate defects. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an industrial quality inspection method and system that integrates real-time defect detection and dynamic updating, so as to overcome the problems of low adaptability of existing defect detection models, long offline training time, insufficient defect localization ability, etc., effectively improve the performance and robustness of the model, and greatly reduce the system's false detection rate and missed detection rate.

[0005] First aspect

[0006] The present invention provides an industrial quality inspection method integrating real-time defect detection and dynamic updating, comprising the following steps:

[0007] Step S1: synchronously capture product images using visible light and infrared sensors;

[0008] Step S2: Input the image into the improved DETR model on the edge side for real-time inference to obtain the detection result. Meanwhile, the detection result is monitored. If the detection result does not meet the preset effect, online training is performed.

[0009] Step S3: Perform collaborative evolution of multi-station detection models through the production line federation aggregation center and regularly issue updated models;

[0010] Step S4: Perform simulation verification and closed-loop feedback on the updated model through a three-level verification system.

[0011] Furthermore, step S1 also includes: achieving pixel-level alignment of visible light and infrared images by setting a circularly polarized light source array, a high frame rate synchronization controller, and a surface reflection feature extractor, and guiding the optical configuration in real time by calculating the BRDF model parameters to adjust the wavelength and incident angle.

[0012] Furthermore, the online training in step S2 specifically includes:

[0013] Automatically close non-relevant detection branches according to material type to achieve dynamic network pruning;

[0014] Generate training samples through feature space interpolation to achieve small sample incremental learning;

[0015] Online knowledge distillation is performed by retaining the three most recent historical models as the teacher network constraint update direction.

[0016] Furthermore, step S3 specifically includes adopting a dynamic weighted average algorithm to realize difference-sensitive aggregation, using the national secret SM9 algorithm for encrypted gradient transmission to achieve end-to-end protection, and constructing a production line knowledge graph to guide model optimization through a defect-process-equipment association database.

[0017] Furthermore, the three-level verification system in step S4 specifically includes:

[0018] The digital twin verification layer generates stress and deformation defect images through ANSYS simulation and inputs them into the updated model for detection. The detection results are obtained to verify the model's detection ability for stress and deformation defects.

[0019] The physical attack test layer uses a robotic arm to perform real scratching / indentation operations and inputs the corresponding images into the updated model for detection. The detection results are obtained to verify the model's detection ability for scratch / indentation defects;

[0020] The mass production closed-loop verification layer compares the correlation between the test results and the final customer return data, and adjusts the model accordingly based on the correlation of the feedback.

[0021] Second aspect

[0022] The present invention provides an industrial quality inspection system integrating real-time defect detection and dynamic updating, comprising:

[0023] Multispectral imaging module, used to synchronously capture product images using visible light and infrared sensors;

[0024] The edge-side dynamic learning engine is used to input images into the improved DETR model on the edge for real-time inference to obtain detection results. The detection results are also monitored. If the detection results do not meet the preset effect, online training is performed.

[0025] The federated learning module is used to collaboratively evolve multi-station inspection models through the production line federation aggregation center and regularly distribute updated models;

[0026] The three-level verification module is used to perform simulation verification and closed-loop feedback on the updated model through the three-level verification system.

[0027] Furthermore, the multispectral imaging module also includes: by setting a circular polarized light source array, a high frame rate synchronization controller and a surface reflection feature extractor, it can achieve pixel-level alignment of visible light and infrared images, and guide the optical configuration in real time by calculating the BRDF model parameters to adjust the wavelength and incident angle.

[0028] Furthermore, the online training in the edge-side dynamic learning engine specifically includes:

[0029] Automatically close non-relevant detection branches according to material type to achieve dynamic network pruning;

[0030] Generate training samples through feature space interpolation to achieve small sample incremental learning;

[0031] Online knowledge distillation is performed by retaining the three most recent historical models as the teacher network constraint update direction.

[0032] Furthermore, the federated learning module specifically includes: using a dynamic weighted average algorithm to achieve difference-sensitive aggregation, using the national secret SM9 algorithm for encrypted gradient transmission to achieve end-to-end protection, and building a production line knowledge graph to guide model optimization through a defect-process-equipment association database.

[0033] Furthermore, the three-level verification system in the three-level verification module specifically includes:

[0034] The digital twin verification layer generates stress and deformation defect images through ANSYS simulation and inputs them into the updated model for detection. The detection results are obtained to verify the model's detection ability for stress and deformation defects.

[0035] The physical attack test layer uses a robotic arm to perform real scratching / indentation operations and inputs the corresponding images into the updated model for detection. The detection results are obtained to verify the model's detection ability for scratch / indentation defects;

[0036] The mass production closed-loop verification layer compares the correlation between the test results and the final customer return data, and adjusts the model accordingly based on the correlation of the feedback.

[0037] The advantages of the present invention are: through the collaboration of multispectral imaging and a dynamic learning engine, rapid adaptation to new defect types can be achieved at the hourly level. By building a process-defect causal reasoning model and a production line federation aggregation center, efficient cross-production line collaboration and precise defect positioning can be achieved, greatly reducing the system's false detection rate and missed detection rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0039] Figure 1 This is a flowchart of an industrial quality inspection method that integrates real-time defect detection and dynamic updating according to the present invention.

[0040] Figure 2 This is a schematic diagram of the framework of an industrial quality inspection system that integrates real-time defect detection and dynamic updating according to the present invention. DETAILED DESCRIPTION

[0041] The embodiments of the present application provide an industrial quality inspection method and system that integrates real-time defect detection and dynamic updates to address some environmental changes and cross-production line collaboration issues that existing AOI systems cannot cope with, and can effectively reduce the system's false detection rate and missed detection rate, and quickly locate the generation link where the defect is located.

[0042] The technical solution in the application embodiment has the following overall idea: The present invention proposes an online evolutionary detection system for industrial quality inspection. Through the collaboration of multispectral imaging and dynamic learning engine, multi-physical field fusion perception is utilized to develop visible light-infrared-polarization multimodal fusion (fusing image data from different sources), reduce detection errors, combine causal-driven learning, construct a causal graph model of process parameters → defect characteristics, and achieve explainable positioning of quality problems. Through online-offline hybrid training, an innovative two-stage update mechanism is designed. In the online stage: obtain the model issued by the federal aggregation center to update the model parameters. In the offline stage: daily fine-tuning of the entire network is performed to achieve the effect of rapid on-site switching without retraining.

[0043] To make the present invention more clearly understood, preferred embodiments are now described in detail below with reference to the accompanying drawings.

[0044] like Figure 1 As shown, an industrial quality inspection method integrating real-time defect detection and dynamic update of the present invention includes the following steps:

[0045] Step S1: synchronously capture product images using visible light and infrared sensors;

[0046] Step S2: Input the image into the improved DETR (DEtection TRansformer) model on the edge side for real-time inference to obtain the detection result. At the same time, the detection result is monitored. If the detection result does not meet the preset effect (for example, when the confidence falls within the preset range), online training is performed;

[0047] Step S3: Perform collaborative evolution of multi-station detection models through the production line federation aggregation center and regularly issue updated models;

[0048] Step S4: Perform simulation verification and closed-loop feedback on the updated model through a three-level verification system.

[0049] Preferably, step S1 includes: setting a circularly polarized light source array, a high frame rate synchronization controller and a surface reflection feature extractor, wherein the circularly polarized light source array is programmable to adjust the wavelength (450-950nm) and the incident angle, setting a high frame rate synchronization controller to achieve pixel-level alignment of visible light and infrared images (error <0.1px), and setting a surface reflection feature extractor, which can guide the optical configuration (such as light source wavelength, incident angle, etc.) in real time by calculating the BRDF model parameters.

[0050] Step S1 also includes inputting the collected multispectral image into a multispectral fusion module. This module is responsible for fusing the collected visible light and infrared images, extracting features from the visible light processing channel and the infrared feature extraction channel, and generating visible spectrum and infrared thermal distribution feature maps. These feature maps are then input into the model in step 2 for real-time inference.

[0051] Preferably, the online training in step S2 specifically includes:

[0052] Automatically close irrelevant detection branches based on material type to achieve dynamic network pruning, such as closing the scratch detection module in mirror material scenes;

[0053] Small-sample incremental learning is achieved by generating training samples through feature space interpolation, requiring only 5 real samples for each new defect.

[0054] Online knowledge distillation is performed by retaining the three most recent historical models as the teacher network constraint update direction.

[0055] Preferably, step S3 specifically includes adopting a dynamic weighted average algorithm to realize difference-sensitive aggregation (specifically, weight = 1 / (detection error^2+0.1)), using the national secret SM9 algorithm for encrypted gradient transmission to achieve end-to-end protection, and constructing a production line knowledge graph, through a defect-process-equipment association database, so that when defects are found, the corresponding process can be quickly and accurately located to guide model optimization.

[0056] The construction of the production line knowledge graph specifically includes:

[0057] Collect operating data, process parameters, and defect detection results of various equipment on the production line. This data mainly comes from equipment sensors, production management systems, and quality inspection reports;

[0058] Clean the collected data: remove noise and erroneous data, unify data formats and units, and extract key information from unstructured data (including equipment maintenance records and technical documents) using text mining technology;

[0059] Data association and integration: Associating and integrating data from different sources to form an associated data set centered around products, equipment, and processes. For example, associating the defect detection results of a product with the corresponding production process parameters and equipment operating status data;

[0060] Knowledge graph construction: Define entity types and relationship types in the knowledge graph, build the ontology structure of the knowledge graph, extract entities and relationships from the integrated data, fill them into the knowledge graph, and store the knowledge graph data in the graph database. The entity types include defect types, process parameters, and equipment components, and the relationship types include cause, association, and belongs to;

[0061] When the model discovers a product defect, it uses the knowledge graph's query function to quickly locate the process parameters and equipment information related to the defect (including process nodes and equipment nodes directly or indirectly associated with a specific defect type), thereby quickly analyzing the possible causes of the defect, such as unreasonable process parameter settings and equipment failures. Based on the analysis results, the corresponding process is promptly optimized and adjusted to reduce the recurrence of similar defects. In addition, the process and equipment knowledge related to the defect in the knowledge graph is fed back to the model as a basis for model optimization. For example, the associated information in the knowledge graph is used to adjust the model's feature selection and weight distribution, thereby improving the model's ability to identify and generalize defects.

[0062] Preferably, the three-level verification system in step S4 specifically includes:

[0063] The digital twin verification layer generates stress and deformation defect images through ANSYS simulation and inputs them into the updated model for detection. The detection results are obtained to verify the model's detection ability for stress and deformation defects.

[0064] The physical attack test layer uses a robotic arm to perform real scratching / indentation operations and inputs the corresponding images into the updated model for detection. The detection results are obtained to verify the model's detection ability for scratch / indentation defects;

[0065] The mass production closed-loop verification layer compares the correlation between the test results and the final customer return data, and adjusts the model accordingly based on the correlation of the feedback.

[0066] Example 2

[0067] like Figure 2 As shown, the industrial quality inspection system of the present invention that integrates real-time defect detection and dynamic update includes:

[0068] Multispectral imaging module, used to synchronously capture product images using visible light and infrared sensors;

[0069] The edge-side dynamic learning engine is used to input images into the improved DETR model on the edge side for real-time inference to obtain detection results. At the same time, the detection results are monitored. If the detection results do not meet the preset effect (for example, when the confidence falls within the preset range), online training is performed;

[0070] The federated learning module is used to collaboratively evolve multi-station inspection models through the production line federation aggregation center and regularly distribute updated models;

[0071] The three-level verification module is used to perform simulation verification and closed-loop feedback on the updated model through the three-level verification system.

[0072] Preferably, the multispectral imaging module includes: setting a circular polarized light source array, a high frame rate synchronization controller and a surface reflection feature extractor, wherein the circular polarized light source array is programmable to adjust the wavelength (450-950nm) and the incident angle, setting a high frame rate synchronization controller to achieve pixel-level alignment of visible light and infrared images (error <0.1px), and setting a surface reflection feature extractor, which can guide the optical configuration (such as light source wavelength, incident angle, etc.) in real time by calculating the BRDF model parameters.

[0073] The multispectral imaging module also includes inputting the collected multispectral images into a multispectral fusion module, which is responsible for fusing the collected visible light and infrared images, extracting features from the visible light processing channel and the infrared feature extraction channel, and generating visible spectrum and infrared thermal distribution feature maps. These feature maps are then input into the corresponding models for real-time inference.

[0074] Preferably, the online training in the edge-side dynamic learning engine specifically includes:

[0075] Automatically close non-relevant detection branches according to material type to achieve dynamic network pruning;

[0076] Generate training samples through feature space interpolation to achieve small sample incremental learning;

[0077] Online knowledge distillation is performed by retaining the three most recent historical models as the teacher network constraint update direction.

[0078] Preferably, the federated learning module specifically includes: using a dynamic weighted average algorithm to achieve difference-sensitive aggregation (specifically, weight = 1 / (detection error^2+0.1)), using the national secret SM9 algorithm for encrypted gradient transmission to achieve end-to-end protection, and building a production line knowledge graph, through a defect-process-equipment association database, so that when defects are found, the corresponding process can be quickly and accurately located to guide model optimization.

[0079] The construction of the production line knowledge graph specifically includes:

[0080] Collect operating data, process parameters, and defect detection results of various equipment on the production line. This data mainly comes from equipment sensors, production management systems, and quality inspection reports;

[0081] Clean the collected data: remove noise and erroneous data, unify data formats and units, and extract key information from unstructured data (including equipment maintenance records and technical documents) using text mining technology;

[0082] Data association and integration: Associating and integrating data from different sources to form an associated data set centered around products, equipment, and processes. For example, associating the defect detection results of a product with the corresponding production process parameters and equipment operating status data;

[0083] Knowledge graph construction: Define entity types and relationship types in the knowledge graph, build the ontology structure of the knowledge graph, extract entities and relationships from the integrated data, fill them into the knowledge graph, and store the knowledge graph data in the graph database. The entity types include defect types, process parameters, and equipment components, and the relationship types include cause, association, and belongs to;

[0084] When the model discovers a product defect, it uses the knowledge graph's query function to quickly locate the process parameters and equipment information related to the defect (including process nodes and equipment nodes directly or indirectly associated with a specific defect type), thereby quickly analyzing the possible causes of the defect, such as unreasonable process parameter settings and equipment failures. Based on the analysis results, the corresponding process is promptly optimized and adjusted to reduce the recurrence of similar defects. In addition, the process and equipment knowledge related to the defect in the knowledge graph is fed back to the model as a basis for model optimization. For example, the associated information in the knowledge graph is used to adjust the model's feature selection and weight distribution, thereby improving the model's ability to identify and generalize defects.

[0085] Preferably, the three-level verification system in the three-level verification module specifically includes:

[0086] The digital twin verification layer generates stress and deformation defect images through ANSYS simulation and inputs them into the updated model for detection. The detection results are obtained to verify the model's detection ability for stress and deformation defects.

[0087] The physical attack test layer uses a robotic arm to perform real scratching / indentation operations and inputs the corresponding images into the updated model for detection. The detection results are obtained to verify the model's detection ability for scratch / indentation defects;

[0088] The mass production closed-loop verification layer compares the correlation between the test results and the final customer return data, and adjusts the model accordingly based on the correlation of the feedback.

[0089] Taking the SMT patch quality inspection scenario as an example, the above technical solution is further explained:

[0090] 1. Hardware deployment:

[0091] 1.1 Installing a 4K polarization camera (frame rate 120fps) and a short-wave infrared camera (InGaAs sensor)

[0092] 1.2 Deploy edge computing box (NVIDIA JetsonAGX Orin, 32GB video memory)

[0093] 2. Defect detection process:

[0094] 2.1 Collecting multispectral images of solder joints (including reflectivity / thermal distribution characteristics)

[0095] 2.2. Dynamic learning engine performs real-time detection:

[0096] 2.21 Normal solder joint: direct output results (delay <8ms)

[0097] 2.22 Suspicious solder joints (confidence level 0.4-0.7): trigger online learning, which is uploaded to the production line federation aggregation center for learning:

[0098] a. Generate adversarial samples: add virtual bubbles / whiskers (based on GAN)

[0099] b. Update the model detection head parameters (learning rate η = 2e-5, limit ||Δw|| < 0.01)

[0100] c. Store the feature vector in memory (using Faiss index)

[0101] 3. Co-evolution of multi-station inspection models through the production line federation aggregation center:

[0102] 3.1 Upload encrypted gradients every 30 minutes (compression rate 95%, using Top-k filtering)

[0103] 3.2 The federation center performs security aggregation:

[0104] a. Eliminate abnormal gradients that deviate from the mean ±3σ

[0105] b. Apply elastic weight consolidation (λ=1e3) to prevent forgetting

[0106] In federated learning, the gradients calculated by each client may be abnormal due to data differences, calculation errors, or malicious attacks. If these abnormal gradients are directly involved in aggregation, they may have a negative impact on the update of the global model, resulting in degraded model performance or unstable training. In order to ensure the reliability of the aggregation results and the stability of the model, the gradients uploaded by each client are first statistically analyzed. Specifically, the mean and standard deviation (σ) of all gradients are calculated, and then the gradients that deviate from the mean by more than ±3σ are regarded as abnormal gradients and excluded. This can effectively filter out abnormal gradients that may have an adverse effect on model training and ensure that the gradients involved in aggregation are within the normal range. In addition, in federated learning, while the model is constantly learning new tasks, it may forget the knowledge it has learned before. This is the so-called "catastrophic forgetting" problem. In order to alleviate this problem, this application applies Elastic Weight Consolidation (Elastic WeightConsolidation,

[0107] The EWC method adds a regularization term to the loss function when learning a new task to protect the parameters of the model that are important for the previous task, preventing these parameters from being over-modified. The strength of this regularization term is controlled by the parameter λ. A larger λ value means stronger protection for important parameters. In this example, λ is set to 1e3, which means that during the optimization process, updates to important parameters will be subject to a larger penalty.

[0108] This effectively prevents the model from forgetting previously learned knowledge.

[0109] 3.3 After model training is completed, the updated model is released (the differential update package is about 650KB)

[0110] 4. Closed-loop verification:

[0111] 4.1 Digital Twin Layer: Simulating Cold Soldering Defects Caused by Abnormal Wave Soldering Temperature

[0112] 4.2 Physical test layer: The robot arm performs real welding defect production (including 4 forms: missing welding, false welding, welding deviation, and wrong welding)

[0113] 4.3 Mass production verification layer: Correlate AOI inspection results with ICT test data to correct model deviations

[0114] In actual tests in PCB board inspection scenarios, the system's false detection rate dropped to 0.08%, the missed detection rate was <0.05%, and the daily model volume growth was controlled within 1.2MB.

[0115] The comparison between the improved industrial quality inspection system of the present invention and the traditional AOI system before improvement is shown in Table 1 below:

[0116] Table 1

[0117]

[0118] In summary, the advantages of the present invention are as follows:

[0119] The present invention can achieve fast adaptation to new defect types at the hourly level, and only 23 samples are needed to increase mAP@0.5 to 90%. During the training process of federated learning, the federal center has taken two measures to ensure the stability and effectiveness of model training: one is to eliminate abnormal gradients through statistical analysis to ensure the reliability of aggregated gradient data; the other is to apply elastic weight consolidation methods to prevent the model from forgetting previous knowledge when learning new tasks. These measures help to improve the performance and robustness of the federated learning model, so that it can stably converge to a good solution in a complex distributed training environment, and improve the accuracy of the detection model by 41% under parallel training of production lines from different manufacturers; by constructing a process-defect causal reasoning model, the production link with quality problems is located, and the accuracy rate reaches 89% after production and use; the system false detection rate and missed detection rate are greatly reduced, and the daily model volume growth is controllable.

[0120] Although the specific embodiments of the present invention are described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and are not intended to limit the scope of the present invention. Equivalent modifications and changes made by those skilled in the art in accordance with the spirit of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. An industrial quality inspection method integrating real-time defect detection and dynamic updating, characterized by: The steps include: Step S1: synchronously capture product images using visible light and infrared sensors; Step S2: Input the image into the improved DETR model on the edge side for real-time inference to obtain the detection result. Meanwhile, the detection result is monitored. If the detection result does not meet the preset effect, online training is performed. Step S3: Perform collaborative evolution of multi-station detection models through the production line federation aggregation center and regularly issue updated models; Step S4: Perform simulation verification and closed-loop feedback on the updated model through a three-level verification system.

2. The industrial quality inspection method integrating real-time defect detection and dynamic updating according to claim 1, characterized in that: The step S1 also includes: achieving pixel-level alignment of visible light and infrared images by setting a circularly polarized light source array, a high frame rate synchronization controller, and a surface reflection feature extractor, and guiding the optical configuration in real time by calculating the BRDF model parameters to adjust the wavelength and incident angle.

3. The industrial quality inspection method integrating real-time defect detection and dynamic updating according to claim 1, characterized in that: The online training in step S2 specifically includes: Automatically close non-relevant detection branches according to material type to achieve dynamic network pruning; Generate training samples through feature space interpolation to achieve small sample incremental learning; Online knowledge distillation is performed by retaining the three most recent historical models as the teacher network constraint update direction.

4. The industrial quality inspection method integrating real-time defect detection and dynamic updating according to claim 1, characterized in that: The step S3 specifically includes adopting a dynamic weighted average algorithm to realize difference-sensitive aggregation, using the national secret SM9 algorithm for encrypted gradient transmission to achieve end-to-end protection, and constructing a production line knowledge graph to guide model optimization through a defect-process-equipment association database.

5. The industrial quality inspection method integrating real-time defect detection and dynamic updating according to claim 1, characterized in that: The three-level verification system in step S4 specifically includes: The digital twin verification layer generates stress and deformation defect images through ANSYS simulation and inputs them into the updated model for detection. The detection results are obtained to verify the model's detection ability for stress and deformation defects. The physical attack test layer uses a robotic arm to perform real scratching / indentation operations and inputs the corresponding images into the updated model for detection. The detection results are obtained to verify the model's detection ability for scratch / indentation defects; The mass production closed-loop verification layer compares the correlation between the test results and the final customer return data, and adjusts the model accordingly based on the correlation of the feedback.

6. An industrial quality inspection system integrating real-time defect detection and dynamic updating, characterized by: include: Multispectral imaging module, used to synchronously capture product images using visible light and infrared sensors; The edge-side dynamic learning engine is used to input images into the improved DETR model on the edge for real-time inference to obtain detection results. The detection results are also monitored. If the detection results do not meet the preset effect, online training is performed. The federated learning module is used to collaboratively evolve multi-station inspection models through the production line federation aggregation center and regularly distribute updated models; The three-level verification module is used to perform simulation verification and closed-loop feedback on the updated model through the three-level verification system.

7. The industrial quality inspection system integrating real-time defect detection and dynamic updating according to claim 6, characterized in that: The multispectral imaging module also includes: by setting up a circular polarized light source array, a high frame rate synchronization controller and a surface reflection feature extractor, it can achieve pixel-level alignment of visible light and infrared images, and guide the optical configuration in real time by calculating the BRDF model parameters to adjust the wavelength and incident angle.

8. The industrial quality inspection system integrating real-time defect detection and dynamic updating according to claim 6, characterized in that: The online training in the edge-side dynamic learning engine specifically includes: Automatically close non-relevant detection branches according to material type to achieve dynamic network pruning; Generate training samples through feature space interpolation to achieve small sample incremental learning; Online knowledge distillation is performed by retaining the three most recent historical models as the teacher network constraint update direction.

9. The industrial quality inspection system integrating real-time defect detection and dynamic updating according to claim 6, characterized in that: The federated learning module specifically includes: using a dynamic weighted average algorithm to achieve difference-sensitive aggregation, using the national secret SM9 algorithm for encrypted gradient transmission to achieve end-to-end protection, and building a production line knowledge graph to guide model optimization through a defect-process-equipment association database.

10. The industrial quality inspection system integrating real-time defect detection and dynamic updating according to claim 6, characterized in that: The three-level verification system in the three-level verification module specifically includes: The digital twin verification layer generates stress and deformation defect images through ANSYS simulation and inputs them into the updated model for detection. The detection results are obtained to verify the model's detection ability for stress and deformation defects. The physical attack test layer uses a robotic arm to perform real scratching / indentation operations and inputs the corresponding images into the updated model for detection. The detection results are obtained to verify the model's detection ability for scratch / indentation defects; The mass production closed-loop verification layer compares the correlation between the test results and the final customer return data, and adjusts the model accordingly based on the correlation of the feedback.

Images