Production line image defect detection strategy generation method and device
By generating and managing evidence packages with a unified structure, and combining large model analysis and strategy draft verification, the interpretability and traceability issues of production line image defect detection are solved, enabling efficient and reliable generation and release of detection strategies.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHONGJIA MICROVISION (SHENZHEN) SEMICONDUCTOR TECHNOLOGY CO LTD
- Filing Date
- 2026-02-14
- Publication Date
- 2026-05-15
AI Technical Summary
Traditional production line image defect detection methods struggle to meet the requirements of efficiency, accuracy, and scalability when faced with complex and diverse defect types and highly variable production environments. Furthermore, the detection results of deep learning models are difficult to effectively translate into actionable detection strategies.
By acquiring production line images and metadata, expert models are invoked for defect detection, inference results are output and packaged into a unified evidence package, and a large model is used to parse the evidence package to generate natural language explanations, verify and replay the strategy draft for evaluation, and achieve version management.
It improves the interpretability and traceability of detection strategies, reduces maintenance costs, ensures that each strategy release undergoes rigorous testing, reduces production risks, and supports version management and canary deployment.
Smart Images
Figure CN122048918A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of industrial visual inspection, and more specifically, to a method and apparatus for generating a production line image defect detection strategy. Background Technology
[0002] In the industrial manufacturing sector, production line image defect detection is a crucial step in ensuring product quality. Traditional defect detection methods typically employ rule-based image processing techniques, which often struggle to meet requirements for efficiency, accuracy, and scalability when faced with complex and diverse defect types and rapidly changing production environments.
[0003] With the development of deep learning and artificial intelligence technologies, image defect detection based on deep learning has gradually become mainstream. However, although deep learning models can provide high-precision detection results, how to effectively transform these results into actionable detection strategies, and how to dynamically adjust the detection strategies to cope with different changes in production environments, remains a significant challenge. Summary of the Invention
[0004] The main objective of this application is to provide a method, apparatus, computing device, and storage medium for generating a production line image defect detection strategy. By encapsulating the reasoning results and contextual information into a standardized evidence package, and then parsing the evidence package through a large model to generate a natural language interpretation, a strategy draft is generated based on this. Finally, playback evaluation, gray-scale deployment, and version management are performed, thereby reducing the maintenance cost of the detection strategy and improving the interpretability and traceability of the strategy.
[0005] To achieve the above objectives, the first aspect of this application proposes a method for generating a defect detection strategy for production line images, comprising: acquiring production line images and metadata; calling an expert model to perform a defect detection task on the acquired production line images and outputting inference results; encapsulating the inference results and their contextual information into a unified evidence package and assigning a unique version number to each evidence package; parsing the evidence package through a large model to generate a natural language explanation and a corresponding strategy draft; verifying and replaying the strategy draft for evaluation, and releasing the strategy in a versioned manner based on the version number.
[0006] Furthermore, the system acquires production line images and metadata, invokes expert models to perform defect detection tasks on the acquired production line images, and outputs inference results. This includes: connecting to production line image acquisition devices to obtain real-time production line images and their related contextual metadata, or periodically extracting production line images and their related contextual metadata from the production line data pool. The metadata includes panel, batch, production line, and timestamp information. The acquired production line images undergo preprocessing operations such as image enhancement, format standardization, and defect ROI screening to generate an input format acceptable to the model. The system dynamically selects the optimal expert model based on task requirements and system operating status, manages the input, output, and execution order of each expert model through a directed acyclic graph, completes image recognition, defect segmentation, and classification tasks, and outputs and stores the inference results of each expert model.
[0007] Furthermore, the evidence package includes at least the original image, feature vectors, classification results output by the model, defect type, defect region segmentation mask, confidence level, version, content hash value, heatmap, overlay map, generated table, and report.
[0008] Furthermore, the strategy draft is validated and replayed for evaluation, and the strategy is versioned and released based on the version number. This includes: comparing the new strategy draft with historical successful cases to verify whether the generated strategy draft conforms to the logic of the current version; configuring A / B testing and canary deployment, storing successful strategies based on the version number according to the test results, and reverting to the previous version when a strategy fails.
[0009] Furthermore, the method also includes: recording the tool name and its calling order for each step, the hash values of the input and output of each step, the model version of each operation, the execution time, the specific configuration parameters and settings, context metadata, resource consumption, and exception information when errors occur through the event log; performing online or offline playback based on the evidence package and the event log, and outputting a difference report.
[0010] Furthermore, based on the evidence package and event logs, online or offline playback is performed, and a difference report is output. This includes: directly executing the new strategy in a real environment and generating event logs in real time, or reproducing historical operations in an isolated environment based on the recorded event logs, comparing the playback results with the expected output in the evidence package, and outputting a difference report.
[0011] Furthermore, the draft strategy includes: if the defect is repairable, rework is recommended, with details to be noted; if the defect is serious and cannot be repaired, scrapping is recommended, with reasons explained; for certain boundary cases or uncertain defects, manual re-inspection is recommended.
[0012] A second aspect of this application provides a production line image defect detection strategy generation device, comprising: a task execution module for acquiring production line images and metadata, calling an expert model to perform defect detection tasks on the acquired production line images, and outputting inference results; an encapsulation module for encapsulating the inference results and their context information into a unified structured evidence package, assigning a unique version number to each evidence package; a parsing module for parsing the evidence package through a large model, generating natural language explanations and corresponding strategy drafts; and an auditing module for verifying and replaying the strategy drafts for evaluation, and releasing the strategy in a versioned manner based on the version number.
[0013] According to a third aspect of this disclosure, a computing device is provided, including at least one processor and at least one memory storing a computer program. The processor executes the production line image defect detection strategy generation method of the first aspect of this disclosure.
[0014] According to a fourth aspect of this disclosure, a computer-readable storage medium storing a computer program is provided, wherein the computer program, when executed by a processor, implements the steps of the production line image defect detection strategy generation method according to a first aspect of this disclosure.
[0015] As can be seen from the above scheme, the production line image defect detection strategy generation method and apparatus provided in this application, through a unified structured evidence package, can not only effectively organize the reasoning results and contextual information, but also provide a reliable data foundation for subsequent analysis, improving data availability and traceability. Utilizing a large model to parse the evidence package and generate natural language explanations enables production line workers to quickly understand the detection results and related strategies, thereby making rapid responses and improving decision-making efficiency. Verification and playback evaluation of the strategy draft ensure that each strategy release undergoes rigorous testing, effectively reducing production risks. Simultaneously, version management ensures effective comparison and tracking between different versions, facilitating continuous improvement. Attached Figure Description
[0016] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application and to make other features, objects, and advantages of the application more apparent. The illustrative embodiments and descriptions of this application are used to explain the application and do not constitute an undue limitation of the application. In the drawings: Figure 1 This is a flowchart illustrating the production line image defect detection strategy generation method provided in this application; Figure 2 This is a schematic diagram of the production line image defect detection strategy generation device provided in this application; Figure 3 This is a schematic block diagram of a computing device according to embodiments of the present disclosure. Detailed Implementation
[0017] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.
[0018] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0019] This solution combines task orchestration, multi-expert dynamic routing mechanisms, standardized evidence output, and policy governance assurance mechanisms to ensure that the entire system has high scalability (plug-in expert model access), interpretability (unified evidence generation and semantic summarization), and maintainability (policy versioning and replay mechanisms, policy support for canary deployment, version tracking, and rapid rollback). Specifically, firstly, image data and its metadata are aggregated across multiple dimensions, including batch number, product model, process parameters, and panel type. A context-aware task entry point is constructed, defining the task context by combining image metadata, detection targets, and the surrounding environment, providing an information foundation for subsequent task planning and model invocation. Then, the overall detection process is deconstructed into a standardized task chain (DAG), and the task execution sequence and toolchain plan are determined. Scheduling includes sub-tasks such as image preprocessing, model invocation, result merging, and interpretation generation, ensuring a reasonable, efficient, and monitorable process. Based on the constraints of cost (computation / invocation cost) and cycle time (processing latency) for different detection tasks, the optimal combination of expert models is dynamically selected. The expert model outputs structured detection results, packaged in a unified format into an evidence package, including: original and intermediate processed images, model output results, confidence assessment and contextual information, model invocation chain, and version information. By introducing Large Language Models (LLM) or Multimodal Language Models (VLM), semantic interpretations of the model results are generated based on evidence packages to interpret natural language. Combined with historical context, task summaries are generated to initially formulate subsequent processing suggestions or process strategies. Under the constraint mechanism of "evidence + consensus," illusion and reasoning drift are avoided, improving interpretability and credibility. The generated strategy draft undergoes rule verification, conflict detection, and other legality checks, historical replay evaluation (simulated operation compared with historical results), and canary release (gradual online verification) to generate reusable strategy versions. Multi-version management and rollback mechanisms are supported to achieve governance and controllability.
[0020] To ensure the traceability, auditability, reproducibility, and interpretability of model outputs in production line image defect detection tasks, the model inference results and their generation background (input data, model version, parameters, context, etc.) can be uniformly encapsulated into a structured evidence package. This package is used for front-end interface display, subsequent interpretation and strategy generation for large model / rule systems, offline review, quality analysis, and model accuracy tracking, ensuring compliance and standard acceptance processes. In other words, a reasoning or detection result is treated as an acceptable workpiece, and an output package with supporting documentation is generated around that result, just as a product must have an inspection report and traceability code when it leaves the factory.
[0021] Figure 1 This is a flowchart illustrating the production line image defect detection strategy generation method provided in this application. (Refer to...) Figure 1 As shown, in step S102, production line images and metadata are acquired, an expert model is called to perform a defect detection task on the acquired production line images, and the inference results are output.
[0022] On the production line, high-resolution cameras or sensors can capture product images in real time, covering different perspectives, lighting conditions, and various types of defects. Connecting to production line image acquisition equipment allows for the acquisition of real-time production line images and their associated contextual metadata. Alternatively, in some cases where real-time acquisition may not be possible, production line images and their associated contextual metadata can be periodically extracted from the production line data pool. This metadata includes information such as panel information, batch information, production line information, and timestamps. Panel information specifies the specific product type or series being monitored; batch information identifies the current batch number for traceability and classification; production line information indicates the production line number where the image was acquired; and the timestamp records the specific time of image acquisition to ensure data timeliness.
[0023] The acquired production line images undergo preprocessing operations such as image enhancement, format standardization, and defect ROI screening to generate an input format acceptable to the model. For example, filtering techniques (such as Gaussian filtering and median filtering) are used to remove random noise from the images and enhance image edges, making defect outlines more obvious. The images are uniformly resized to a specific size (such as 256x256 or 512x512) to fit the model's input layer. Edge detection algorithms are used to initially screen out regions of interest (ROIs), i.e., locations where defects may exist, to reduce the computational load of subsequent processing.
[0024] The optimal expert model is dynamically selected based on task requirements and system operating status. The input, output, and execution order of each expert model are managed through a directed acyclic graph to complete image recognition, defect segmentation, and classification tasks. The inference results of each expert model are output and stored.
[0025] Subsequently, in step S104, the reasoning results and their context information are encapsulated into evidence packages with a unified structure, and each evidence package is assigned a unique version number.
[0026] In one embodiment of this disclosure, the evidence package is a structured, versioned output container containing information such as the original image, feature vectors, classification results output by the model, defect type, defect region segmentation mask, confidence score, version, content hash value, heatmap, overlay map, generated tables, and reports, all uniformly encapsulated in a JSON structure. The evidence package can be bidirectionally bound to a multimodal report artifact via a reference ID, enabling users to easily export the evidence package and its corresponding report. Each evidence package is assigned a unique version number, which can be generated based on a timestamp or an incrementing method for subsequent traceability and management.
[0027] Next, in step S106, the evidence package is analyzed using a large model to generate a natural language interpretation and a corresponding policy draft.
[0028] Through API or local model calls, the constructed input is passed to the model (such as GPT-4, CLIP, etc.) for inference, to understand and parse images and related data, and generate natural language descriptions based on context. It supports multimodal prompts (mixed input of images, masks, and text), supports Few-shot historical case alignment (similar defects and handling measures previously processed), and can inject rule-based prompts (such as disallowing the scrapping of high-value panels, prioritizing re-inspection, etc.). Example semantic explanations are as follows: "According to the analysis, the circuit board has multiple defects, mainly concentrated in the upper left corner area. These defects mainly manifest as short circuits and open circuits, which may lead to a decrease in product performance." Based on the reasoning results and natural language interpretation, a specific strategy draft is generated: Rework: If the defect is repairable, rework is recommended. Note the details that need attention.
[0029] Scrap: If the defect is severe and cannot be repaired, scrapping is recommended, and the reasons should be explained.
[0030] Manual re-inspection: For certain boundary conditions or uncertain defects, manual re-inspection is recommended to ensure quality.
[0031] Finally, in step S108, the draft strategy is verified and replayed for evaluation, and the strategy is versioned and released based on the version number.
[0032] Specifically, the new strategy draft is compared with historical success stories to verify whether the generated strategy draft conforms to the logic of the current version, ensuring that the strategy meets production line requirements and does not exceed the set rules. A / B testing can be configured to compare the impact of different strategies on production efficiency and quality. Canary deployment can be configured to evaluate the effectiveness of the strategy through user feedback on a small scale. Based on the test results, successful strategies are versioned and stored according to version numbers. If a strategy fails, it is rolled back to the previous version to ensure the stability and reliability of the system.
[0033] In one embodiment of this disclosure, the event log can also record the tool name and its call order for each step, the hash values of the input and output of each step, the model version of each operation, the execution time, specific configuration parameters and settings, context metadata, resource consumption, and exception information when errors occur. Online or offline playback is performed based on the evidence package and the event log, and a difference report is output. Specifically, the new strategy is executed directly in a real environment, and event logs are generated in real time; or, based on the recorded event logs, historical operations are reproduced in an isolated environment, and the playback results are compared with the expected output in the evidence package to output a difference report.
[0034] As can be seen, the production line image defect detection strategy generation method provided in this application, through a unified structured evidence package, not only effectively organizes reasoning results and contextual information, but also provides a reliable data foundation for subsequent analysis, improving data usability and traceability. Utilizing a large model to parse the evidence package and generate natural language explanations enables production line workers to quickly understand detection results and related strategies, thereby making rapid responses and improving decision-making efficiency. Verification and playback evaluation of the strategy draft ensure that each strategy release undergoes rigorous testing, effectively reducing production risks. Simultaneously, version control ensures effective comparison and tracking between different versions, facilitating continuous improvement.
[0035] Figure 2 This is a schematic diagram of the production line image defect detection strategy generation device provided in this application. (Refer to...) Figure 2 As shown, the device includes a task execution module 210, an encapsulation module 220, a parsing module 230, and an auditing module 240.
[0036] The task execution module 210 is used to acquire production line images and metadata, call an expert model to perform defect detection tasks on the acquired production line images, and output inference results. Specifically, the task execution module can connect to a production line image acquisition device to acquire real-time production line images and their related contextual metadata. Alternatively, in some cases where real-time acquisition may not be possible, production line images and their related contextual metadata can be extracted periodically from the production line data pool. The metadata includes information such as panel, batch, production line, and timestamp. Panel information indicates the specific product type or series currently being monitored; batch information identifies the batch number currently being produced for traceability and classification; production line information indicates the production line number where the image acquisition took place; and the timestamp records the specific time of image acquisition to ensure data timeliness.
[0037] The acquired production line images undergo preprocessing operations such as image enhancement, format standardization, and defect ROI screening to generate an input format acceptable to the model. For example, filtering techniques (such as Gaussian filtering and median filtering) are used to remove random noise from the images and enhance image edges, making defect outlines more obvious. The images are uniformly resized to a specific size (such as 256x256 or 512x512) to fit the model's input layer. Edge detection algorithms are used to initially screen out regions of interest (ROIs), i.e., locations where defects may exist, to reduce the computational load of subsequent processing.
[0038] The optimal expert model is dynamically selected based on task requirements and system operating status. The input, output, and execution order of each expert model are managed through a directed acyclic graph to complete image recognition, defect segmentation, and classification tasks. The inference results of each expert model are output and stored.
[0039] The encapsulation module 220 is used to encapsulate the inference results and their context information into a unified structured evidence package, assigning a unique version number to each evidence package. In one embodiment of this disclosure, the evidence package is a structured, versioned output container containing information such as the original image, feature vectors, classification results output by the model, defect type, defect region segmentation mask, confidence level, version, content hash value, heatmap, overlay map, generated tables, and reports, all uniformly encapsulated in a JSON structure. The evidence package can be bidirectionally bound to the multimodal report artifact via a reference ID, enabling users to easily export the evidence package and the corresponding report. Each evidence package is assigned a unique version number, which can be generated based on a timestamp or an incrementing method for subsequent traceability and management.
[0040] The parsing module 230 is used to parse the evidence package using a large model, generating natural language explanations and corresponding policy drafts. The parsing module can be called via API or local model calls, passing pre-constructed input to the model (such as GPT-4, CLIP, etc.) for inference, understanding and parsing images and related data, and generating natural language descriptions based on context. It supports multimodal prompts (mixed input of images, masks, and text), supports Few-shot historical case alignment (similar defects and handling measures previously processed), and can inject rule-based prompts (such as disallowing the scrapping of high-value panels, prioritizing re-inspection, etc.). An example semantic explanation is as follows: "According to the analysis, the circuit board has multiple defects, mainly concentrated in the upper left corner area. These defects mainly manifest as short circuits and open circuits, which may lead to a decrease in product performance." Based on the reasoning results and natural language interpretation, a specific strategy draft is generated: Rework: If the defect is repairable, rework is recommended. Note the details that need attention.
[0041] Scrap: If the defect is severe and cannot be repaired, scrapping is recommended, and the reasons should be explained.
[0042] Manual re-inspection: For certain boundary conditions or uncertain defects, manual re-inspection is recommended to ensure quality.
[0043] The audit module 240 is used to verify and replay the strategy draft for evaluation, and to release the strategy in a versioned manner based on the version number. According to one embodiment of this disclosure, the audit module can compare the new strategy draft with historical successful cases to verify whether the generated strategy draft conforms to the logic of the current version, ensuring that the strategy meets production line requirements and does not exceed the set rule scope. A / B testing can be configured to compare the impact of different strategies on production efficiency and quality. Canary deployment can be configured to evaluate the strategy's effectiveness through user feedback on a small scale. Successful strategies are versioned and stored based on version numbers according to the test results, and in the event of a strategy failure, it reverts to the previous version, ensuring the stability and reliability of the system.
[0044] As can be seen, the production line image defect detection strategy generation device provided in this application, through a unified evidence package, can not only effectively organize reasoning results and contextual information, but also provide a reliable data foundation for subsequent analysis, improving data usability and traceability. Utilizing a large model to parse the evidence package and generate natural language explanations enables production line workers to quickly understand the detection results and related strategies, thereby making rapid responses and improving decision-making efficiency. Verification and playback evaluation of the strategy draft ensure that each strategy release undergoes rigorous testing, effectively reducing production risks. Simultaneously, version control ensures effective comparison and tracking between different versions, facilitating continuous improvement.
[0045] Figure 3 This is a schematic block diagram of a computing device according to embodiments of the present disclosure. Figure 3 As shown, the computing device 300 may include a processor 310 and a memory 320 storing a computer program. When the computer program is executed by the processor 310, the computing device 300 is able to perform the steps of a production line image defect detection strategy generation method.
[0046] In embodiments of this disclosure, processor 310 may be, for example, a central processing unit (CPU), a microprocessor, a digital signal processor (DSP), a processor based on a multi-core processor architecture, etc. Memory 320 may be any type of memory implemented using data storage technologies, including but not limited to random access memory, read-only memory, semiconductor-based memory, flash memory, disk storage, etc.
[0047] Furthermore, in embodiments of this disclosure, the computing device 300 may also include an input device 330, such as a keyboard, mouse, etc. Additionally, the computing device 300 may also include an output device 340, such as a display, for outputting natural language interpretations and policy drafts.
[0048] In other embodiments of this disclosure, a computer-readable storage medium storing a computer program is also provided, wherein the computer program, when executed by a processor, is capable of implementing the steps of a production line image defect detection strategy generation method.
[0049] In summary, the production line image defect detection strategy generation method and apparatus provided in this application, through a unified evidence package, not only effectively organizes reasoning results and contextual information but also provides a reliable data foundation for subsequent analysis, improving data usability and traceability. Utilizing a large model to parse the evidence package and generate natural language explanations enables production line workers to quickly understand detection results and related strategies, thereby making rapid responses and improving decision-making efficiency. Verification and playback evaluation of the strategy draft ensure that each strategy release undergoes rigorous testing, effectively reducing production risks. Furthermore, version control ensures effective comparison and tracking between different versions, facilitating continuous improvement.
[0050] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0051] Obviously, those skilled in the art should understand that the various units or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps into a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.
[0052] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A method for generating a production line image defect detection strategy, characterized in that, include: Acquire production line images and metadata, call the expert model to perform defect detection tasks on the acquired production line images, and output the inference results; The reasoning results and their context information are encapsulated into evidence packages with a unified structure, and each evidence package is assigned a unique version number. The evidence package is analyzed using a large model to generate a natural language interpretation and a corresponding policy draft. The draft strategy is validated and replayed for evaluation, and the strategy is released in a versioned manner based on the version number.
2. The production line image defect detection strategy generation method according to claim 1, characterized in that, The process involves acquiring production line images and metadata, calling an expert model to perform defect detection on the acquired production line images, and outputting inference results including: Connect to the production line image acquisition device to obtain real-time production line images and their related contextual metadata, or periodically extract production line images and their related contextual metadata from the production line data pool. The metadata includes panel, batch, production line, and timestamp information. The acquired production line images undergo preprocessing operations such as image enhancement, format standardization, and defect ROI screening to generate an input format acceptable to the model. The optimal expert model is dynamically selected based on task requirements and system operating status. The input, output, and execution order of each expert model are managed through a directed acyclic graph to complete image recognition, defect segmentation, and classification tasks. The inference results of each expert model are output and stored.
3. The production line image defect detection strategy generation method according to claim 1, characterized in that, The evidence package includes at least the original image, feature vector, classification results output by the model, defect type, defect region segmentation mask, confidence level, version, content hash value, heatmap, overlay map, generated table, and report.
4. The production line image defect detection strategy generation method according to claim 1, characterized in that, The process of verifying and replaying the draft strategy, and then releasing the strategy in a versioned manner based on the version number, includes: The new strategy draft is compared with historical success stories to verify whether the generated strategy draft conforms to the logic of the current version; Configure A / B testing and canary deployment. Store successful strategies based on version numbers according to the test results, and revert to the previous version when a strategy fails.
5. The production line image defect detection strategy generation method according to claim 4, characterized in that, The method further includes: The event log records the name of the tool called in each step and its calling order, the hash value of the input and output of each step, the model version of each operation, the execution time, the specific configuration parameters and settings, the context metadata, resource consumption, and the exception information when an error occurs. Based on the evidence package and event logs, perform online or offline playback and output a difference report.
6. The production line image defect detection strategy generation method according to claim 5, characterized in that, The process of performing online or offline playback based on the evidence package and event logs, and outputting a difference report, includes: The new strategy can be executed directly in a real environment and event logs can be generated in real time. Alternatively, based on the recorded event logs, historical operations can be reproduced in an isolated environment, and the replay results can be compared with the expected output in the evidence package to generate a difference report.
7. The production line image defect detection strategy generation method according to claim 1, characterized in that, The draft strategy includes: if the defect is repairable, rework is recommended, with details noted; if the defect is serious and cannot be repaired, scrapping is recommended, with reasons explained; for certain boundary cases or uncertain defects, manual re-inspection is recommended.
8. A production line image defect detection strategy generation device, characterized in that, include: The task execution module is used to acquire production line images and metadata, call the expert model to perform defect detection tasks on the acquired production line images, and output inference results; The encapsulation module is used to encapsulate the reasoning results and their context information into evidence packages with a unified structure, and assign a unique version number to each evidence package. The parsing module is used to parse the evidence package through a large model, and generate natural language interpretations and corresponding policy drafts; The audit module is used to verify and replay the draft strategy, and to release the strategy in a versioned manner based on the version number.
9. A computing device, characterized in that, include: At least one processor; as well as At least one memory storing a computer program; When the computer program is executed by the at least one processor, the computing device performs the steps of the production line image defect detection strategy generation method according to any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the production line image defect detection strategy generation method according to any one of claims 1 to 7.