Line image defect detection method and device based on agent arrangement

Abstract

The application discloses a production line image defect detection method and device based on an agent arrangement, which comprises the following steps: receiving a production line image detection task request, converting the task request into a standardized task description; generating a sub-task directed acyclic graph and a tool calling plan according to the standardized task description and tool information in a tool register; calling tools according to the steps defined in the tool calling plan based on the sub-task directed acyclic graph to execute the task in turn; recording the hash value of the input and output data, the version number and the key parameters of the model of each step into an event log, performing playback based on the event log, obtaining the difference between the actual result and the expected result in real time, and outputting a difference report. The scheme can improve the consistency, efficiency and traceability of task execution.

Description

Technical Field

[0001] This application relates to the field of industrial visual inspection, and more specifically, to a method and apparatus for detecting defects in production line images based on intelligent agent orchestration. Background Technology

[0002] Production line image inspection typically involves multiple stages, including image preprocessing, target detection, defect classification, and result annotation, each potentially employing different models and algorithms. Currently, most common inspection systems adopt a fixed "end-to-end" architecture, lacking flexibility and struggling to adapt to changing tasks or tool upgrades. In actual deployment, image acquisition, inference, review, and strategy adjustment operate independently, making it difficult to form a reusable closed loop. Furthermore, anomaly handling is weak, and consistency and replayability are difficult to guarantee in online timeouts, failures, and repeated calls. Therefore, an intelligent agent orchestration system is needed to improve the overall performance and controllability of the inspection system. Summary of the Invention

[0003] The main purpose of this application is to provide a production line image defect detection method and device based on intelligent agent orchestration. With the help of intelligent agent framework and tool registration mechanism, the execution path and dependencies of the task can be automatically inferred, which significantly improves the overall detection efficiency.

[0004] To achieve the above objectives, the first aspect of this application proposes a production line image defect detection method based on intelligent agent orchestration, comprising: receiving a production line image detection task request and converting the task request into a standardized task description; generating a subtask directed acyclic graph and a tool invocation plan based on the standardized task description and tool information in the tool registry; invoking tools sequentially to execute tasks according to the steps defined in the tool invocation plan based on the subtask directed acyclic graph; recording the hash values ​​of the input and output data of each step, the model version number, and key parameters in an event log; performing playback based on the event log, obtaining the difference between the result and the expected result in real time, and outputting a difference report.

[0005] Furthermore, the standardized task description includes the task scope, service level objectives, and expected outputs. The task scope includes product batches, inspection panels, and time windows for performing inspections. The service level objectives include latency requirements, throughput requirements, and upper limits for false positives or false negatives. The expected outputs include classification labels for target objects, segmentation masks, bounding box information, structured inspection reports, and strategy drafts generated based on the inspection results.

[0006] Furthermore, the tool registry is used to manage available tools and their related information, including the tool name, the input and output data formats for each tool, the timeout or retry policy set for each tool, and the estimated resource consumption required for tool execution.

[0007] Furthermore, each node in the directed acyclic graph of subtasks represents a subtask, each edge represents the dependency between tasks, and the tool invocation plan is used to specify the tools and execution order required for each subtask. Its key fields include a unique identifier for each step, the dependency between steps, the data source for each step, the idempotent key, and the output data format.

[0008] Furthermore, based on the directed acyclic graph of subtasks, the tool is invoked sequentially to execute tasks according to the steps defined in the tool invocation plan. This includes: generating a unique idempotent key for each subtask step, where the idempotent key includes the task ID, step ID, key input, and version number; checking whether the idempotent key already exists in the result cache before executing the subtask, and directly returning the existing output if it does; defining multiple checkpoints during task execution, persistently storing the current state and intermediate results at preset intervals; restarting execution from the nearest checkpoint when the task is abnormally interrupted; determining whether to retry based on the error type, and configuring the maximum number of retries and backoff strategy.

[0009] Furthermore, the key fields recorded in the event log include the tool name and its calling order for each step, the hash values ​​of the input and output for each step, the model version for each operation, the execution time, the specific configuration parameters and settings, context metadata, resource consumption, and exception information when errors occur.

[0010] Furthermore, based on the event logs, the replay is executed to obtain the differences between the results and the expected results in real time, and the difference report is output, including: the online main process continues to use the old version to output the formal results, the control test is run in parallel with the same input, and the execution results and control difference indicators are recorded; and / or historical task data samples are selected, and offline batch replay is performed using the new strategy version to output the difference report between the results and the old strategy version.

[0011] A second aspect of this application provides a production line image defect detection device based on intelligent agent orchestration, comprising: The standardization module is used to convert a production line image inspection request into a standardized task description when it is received. The task planning module is used to generate a directed acyclic graph of subtasks and a tool call plan based on standardized task descriptions and tool information in the tool registry. The execution module is used to invoke tools to execute tasks sequentially according to the steps defined in the tool invocation plan, based on the subtask directed acyclic graph. The log replay module records the hash values ​​of the input and output data, the model version number, and key parameters of each step into the event log. Based on the event log, it performs replay, obtains the differences between the results and the expected results in real time, and outputs a difference report.

[0012] 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 method based on intelligent agent orchestration from the first aspect of this disclosure.

[0013] 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 method based on agent orchestration according to a first aspect of this disclosure.

[0014] As can be seen from the above scheme, the production line image defect detection method and device based on intelligent agent orchestration provided in this application can automate task execution by formalizing the detection process into an executable tool invocation plan, ensuring task consistency and efficiency. It also supports subsequent auditing and result playback, making the entire process highly traceable. During execution, generating idempotent keys for each subtask and deduplicating them, resuming execution from breakpoints based on checkpoints, and implementing differentiated retry strategies according to error codes can improve the overall performance and controllability of the detection system. Attached Figure Description

[0015] 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 A flowchart illustrating the production line image defect detection method based on intelligent agent orchestration provided in this application; Figure 2 A schematic diagram of the production line image defect detection device based on intelligent agent orchestration 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

[0016] 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.

[0017] 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.

[0018] This disclosure transforms the traditional fixed process of production line image defect detection into a closed-loop system with dynamic scheduling, intelligent interpretation, and strategy governance capabilities by introducing an intelligent agent architecture. The method, based on the intelligent agent's "planning-execution-summarizing-reuse" framework, covers core aspects such as task orchestration, expert model selection, evidence generation, and strategy evolution, achieving intelligent, standardized, and highly available detection processes.

[0019] Specifically, firstly, image data and its metadata are aggregated according to multiple dimensions such as batch number, product model, and process parameters. A context-aware task entry point is constructed, defining the task context by combining image metadata, detection targets, and the contextual environment, providing an information foundation for subsequent task planning and model invocation. Then, the overall detection process is deconstructed into standardized task chains, and the task execution order and toolchain plan are formulated. Scheduling includes sub-tasks such as image preprocessing, model invocation, result merging, and interpretation generation, ensuring that the process is reasonable, efficient, and monitorable. 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, which are packaged into an evidence package in a unified format, including: original and processed images, model output results, confidence assessment and contextual association information, model invocation chain and version information, etc. 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 improve the consistency, efficiency, and traceability of task execution, this disclosure provides a method and apparatus for detecting defects in production line images based on intelligent agent orchestration. Figure 1 This is a flowchart illustrating the production line image defect detection method based on intelligent agent orchestration provided in this application. (Refer to...) Figure 1 As shown, in step S102, a production line image detection task request is received and converted into a standardized task description.

[0021] The standardized task description includes the task scope, service level objectives, and expected outputs. The task scope includes product batches, inspection panels, and time windows for performing inspections. The service level objectives include latency requirements, throughput requirements, and upper limits for false positives or false negatives. The expected outputs include classification labels of target objects, segmentation masks, bounding box information, structured inspection reports, and strategy drafts generated based on the inspection results.

[0022] Subsequently, in step S104, a subtask directed acyclic graph and a tool invocation plan are generated based on the standardized task description and the tool information in the tool registry.

[0023] The tool registry is used to manage available tools and their related information, including the tool name, the input and output data format of each tool, the timeout or retry policy set for each tool, and the estimated resource consumption required for tool execution.

[0024] In a Directed Acyclic Graph (DAG) for subtasks, each node represents a subtask, and each edge represents the dependency between tasks. The tool invocation plan can specify the tools and execution order required for each subtask. Its key fields include a unique identifier for each step, the dependency between steps, the data source for each step, the idempotent key, and the output data format.

[0025] Next, in step S106, based on the subtask directed acyclic graph, the tool is invoked to execute the tasks sequentially according to the steps defined in the tool invocation plan.

[0026] For example, the execution of each step includes the following: Image capture: Obtain the image to be detected from the specified image source.

[0027] Preprocessing: The acquired images are processed, such as denoising, scaling, and enhancement, to prepare them for subsequent model inference.

[0028] Routing: Assigning the processed image to the appropriate detection model as needed.

[0029] Model inference: Call machine learning / deep learning models to perform defect detection and generate detection results.

[0030] Post-processing: Thresholding, labeling, etc., of the model output to make the results easier to understand and use.

[0031] Report generation: Generates a final report based on the detection results, including defect location, type, and statistical information.

[0032] In online task systems, issues such as timeouts, network jitter, unresponsive interfaces, and node crashes are extremely common, often leading to repeated requests after timeouts. When intermediate steps in task execution fail, recovery from the intermediate state is required, resulting in inconsistent results, resource waste, and duplicate writes. This disclosure, through idempotent keys and breakpoint resumption mechanisms, ensures consistent results or reuse of completed output when the same subtask is executed repeatedly. Combined with configurable retry strategies (retryable / non-retryable, exponential backoff), it improves stability and reduces uncertainty.

[0033] In one embodiment of this disclosure, a unique idempotent key is generated for each subtask. This idempotent key is composed of the task ID, step ID, key input, and version number. This ensures that when the same subtask is executed repeatedly, it can be identified as the same request, avoiding duplicate calculations. Before executing a subtask, the database or cache is checked to see if the result corresponding to the idempotent key exists. If it exists, the existing output is returned directly, thus avoiding resource waste and inconsistencies that may result from repeated execution. Multiple checkpoints are defined during task execution, and the current state and intermediate results are persistently stored at preset intervals. For time-consuming steps, the current state or intermediate results can be written to the checkpoint when execution reaches a certain stage. If a failure is encountered during execution, execution can be restarted from the most recent checkpoint based on the checkpoint information, that is, execution continues from the last successful state, skipping successful steps, avoiding full duplicate calculations, and improving overall efficiency.

[0034] Error types can be categorized based on different error codes, such as I / O errors, insufficient GPU resources, and missing data, to determine the appropriate handling method. For retryable errors, an exponential backoff mechanism can be used, increasing the waiting time with each retry to reduce the impact on the system, increase the success rate, and improve system robustness.

[0035] Finally, in step S108, the hash values ​​of the input and output data of each step, the model version number, and key parameters are recorded in the event log. Based on the event log, playback is performed to obtain the difference between the result and the expected result in real time and output a difference report.

[0036] Key fields recorded in the event log include the tool name and its call order for each step, the hash values ​​of the inputs and outputs for each step, the model version for each operation, the execution time, specific configuration parameters and settings, context metadata, resource consumption, and exception information when errors occur, forming a deterministically replayable execution trajectory. The same task can be executed online or replayed offline based on logs, supporting: shadow mode (online results are not changed, only the strategy draft and comparison metrics are recorded), A / B comparison, regression testing, and acceptance testing.

[0037] For example, without affecting the current online output results, conduct real-time testing of the strategy update draft or the new model version. The main online workflow continues to use the old version to produce official results, running parallel control tests with the same input, recording execution results and control difference indicators, such as precision difference, recall change, and processing time. Select historical task data samples, perform offline batch replay using the new strategy version, and output a report on the difference between the results and the old strategy version, such as TPR / FPR / latency / Top-K recall, etc.

[0038] As can be seen from the above scheme, the production line image defect detection method based on intelligent agent orchestration provided in this application can automate task execution by formalizing the detection process into an executable tool invocation plan, ensuring task consistency and efficiency. It also supports subsequent auditing and result playback, making the entire process highly traceable. During execution, generating idempotent keys for each subtask and deduplicating them, resuming execution from breakpoints based on checkpoints, and implementing differentiated retry strategies according to error codes can improve the overall performance and controllability of the detection system.

[0039] Figure 2 A schematic diagram of the production line image defect detection device based on intelligent agent orchestration provided in this application. (Refer to...) Figure 2 As shown, the device 200 includes: a standardization module 210, a task planning module 220, an execution module 230, and a log playback module 240.

[0040] The standardization module 210 can convert a production line image inspection request into a standardized task description upon receiving it. The standardized task description includes the task scope, service level objectives, and expected outputs. The task scope includes the product batch, inspection panel, and time window for performing the inspection. The service level objectives include latency requirements, throughput requirements, and upper limits for false positives or false negatives. The expected outputs include the target object's classification label, segmentation mask, bounding box information, structured inspection report, and a draft strategy generated based on the inspection results.

[0041] The task planning module 220 can generate a subtask directed acyclic graph (DAG) and a tool invocation plan based on standardized task descriptions and tool information in the tool registry. The tool registry manages available tools and their related information, including tool names, input and output data formats for each tool, timeout or retry policies for each tool, and estimated resource consumption for tool execution. Each node in the subtask DAG represents a subtask, and each edge represents a dependency between tasks. The tool invocation plan specifies the tools and execution order required for each subtask, with key fields including a unique identifier for each step, dependencies between steps, the data source for each step, the idempotent key, and the output data format.

[0042] Execution module 230 can invoke the tool to execute tasks sequentially according to the steps defined in the tool invocation plan, based on the subtask directed acyclic graph. For example, the execution of each step includes the following: Image capture: Obtain the image to be detected from the specified image source.

[0043] Preprocessing: The acquired images are processed, such as denoising, scaling, and enhancement, to prepare them for subsequent model inference.

[0044] Routing: Assigning the processed image to the appropriate detection model as needed.

[0045] Model inference: Call machine learning / deep learning models to perform defect detection and generate detection results.

[0046] Post-processing: Thresholding, labeling, etc., of the model output to make the results easier to understand and use.

[0047] Report generation: Generates a final report based on the detection results, including defect location, type, and statistical information.

[0048] In online task systems, issues such as timeouts, network jitter, unresponsive interfaces, and node crashes are extremely common, often leading to repeated requests after timeouts. Failures in intermediate steps during task execution necessitate recovery from the intermediate state, resulting in inconsistent results, resource waste, and duplicate writes. This disclosure addresses these issues by employing idempotent keys and breakpoint-based continuation mechanisms to ensure consistent results or reuse of completed output when the same subtask is executed repeatedly. Combined with configurable retry strategies (retryable / non-retryable, exponential backoff), it enhances stability and reduces uncertainty.

[0049] In one embodiment of this disclosure, a unique idempotent key is generated for each subtask. This idempotent key is composed of the task ID, step ID, key input, and version number. This ensures that when the same subtask is executed repeatedly, it can be identified as the same request, avoiding duplicate calculations. Before executing a subtask, the database or cache is checked to see if the result corresponding to the idempotent key exists. If it exists, the existing output is returned directly, thus avoiding resource waste and inconsistencies that may result from repeated execution. Multiple checkpoints are defined during task execution, and the current state and intermediate results are persistently stored at preset intervals. For time-consuming steps, the current state or intermediate results can be written to the checkpoint when execution reaches a certain stage. If a failure is encountered during execution, execution can be restarted from the most recent checkpoint based on the checkpoint information, that is, execution continues from the last successful state, skipping successful steps, avoiding full duplicate calculations, and improving overall efficiency.

[0050] Error types can be categorized based on different error codes, such as I / O errors, insufficient GPU resources, and missing data, to determine the appropriate handling method. For retryable errors, an exponential backoff mechanism can be used, increasing the waiting time with each retry to reduce the impact on the system, increase the success rate, and improve system robustness.

[0051] The log replay module 240 can record the hash values ​​of the input and output data of each step, the model version number, and key parameters into the event log. Based on the event log, it can perform replay, obtain the difference between the result and the expected result in real time, and output a difference report.

[0052] Key fields recorded in the event log include the tool name and its call order for each step, the hash values ​​of the inputs and outputs for each step, the model version for each operation, the execution time, specific configuration parameters and settings, context metadata, resource consumption, and exception information when errors occur, forming a deterministically replayable execution trajectory. The same task can be executed online or replayed offline based on logs, supporting: shadow mode (online results are not changed, only the strategy draft and comparison metrics are recorded), A / B comparison, regression testing, and acceptance testing.

[0053] For example, without affecting the current online output results, conduct real-time testing of the strategy update draft or the new model version. The main online workflow continues to use the old version to produce official results, running parallel control tests with the same input, recording execution results and control difference indicators, such as precision difference, recall change, and processing time. Select historical task data samples, perform offline batch replay using the new strategy version, and output a report on the difference between the results and the old strategy version, such as TPR / FPR / latency / Top-K recall, etc.

[0054] As can be seen from the above scheme, the production line image defect detection device based on intelligent agent orchestration provided in this application can automate task execution by formalizing the detection process into an executable tool invocation plan, ensuring task consistency and efficiency. It also supports subsequent auditing and result playback, making the entire process highly traceable. During execution, generating idempotent keys for each subtask and deduplicating them, resuming execution from breakpoints based on checkpoints, and implementing differentiated retry strategies according to error codes can improve the overall performance and controllability of the detection system.

[0055] 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.

[0056] 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.

[0057] Furthermore, in embodiments of this disclosure, the computing device 300 may also include an input device 330, such as a keyboard or mouse. Additionally, the computing device 300 may also include an output device 340, such as a display.

[0058] 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 method based on agent orchestration.

[0059] In summary, the production line image defect detection method and apparatus based on intelligent agent orchestration provided in this application, by formalizing the detection process into an executable tool invocation plan, can achieve automated task execution, ensuring task consistency and efficiency, while supporting subsequent auditing and result playback, making the entire process highly traceable. During execution, generating idempotent keys for each subtask and deduplicating them, resuming execution from breakpoints based on checkpoints, and implementing differentiated retry strategies according to error codes can improve the overall performance and controllability of the detection system.

[0060] 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.

[0061] 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.

[0062] 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 detecting defects in production line images based on intelligent agent orchestration, characterized in that, include: Receive production line image detection task requests and convert the task requests into standardized task descriptions; Based on the standardized task description and tool information in the tool registry, a subtask directed acyclic graph and a tool invocation plan are generated. Based on the subtask directed acyclic graph, the tool is invoked to execute the tasks sequentially according to the steps defined in the tool invocation plan; The hash values ​​of the input and output data for each step, the model version number, and key parameters are recorded in the event log. Based on the event log, playback is performed to obtain the difference between the result and the expected result in real time and output a difference report.

2. The production line image defect detection method based on intelligent agent orchestration according to claim 1, characterized in that, The standardized task description includes the task scope, service level objectives, and expected outputs. The task scope includes product batches, detection panels, and time windows for performing detection. The service level objectives include latency requirements, throughput requirements, and upper limits for false positives or false negatives. The expected outputs include classification labels of target objects, segmentation masks, bounding box information, structured detection reports, and strategy drafts generated based on the detection results.

3. The production line image defect detection method based on intelligent agent orchestration according to claim 1, characterized in that, The tool registry is used to manage available tools and their related information, including tool name, input and output data format for each tool, timeout or retry policy set for each tool, and estimated resource consumption required for tool execution.

4. The production line image defect detection method based on intelligent agent orchestration according to claim 1, characterized in that, Each node in the directed acyclic graph represents a subtask, and each edge represents the dependency between tasks. The tool invocation plan is used to specify the tools and execution order required for each subtask. Its key fields include a unique identifier for each step, the dependency between steps, the data source for each step, the idempotent key, and the output data format.

5. The production line image defect detection method based on intelligent agent orchestration according to claim 1, characterized in that, The step of invoking the tool sequentially to execute tasks based on the directed acyclic graph of the subtasks, according to the steps defined in the tool invocation plan, includes: A unique idempotent key is generated for each subtask step. The idempotent key includes the task ID, step ID, key input, and version number. Before executing the subtask, it is checked whether the idempotent key already exists in the result cache. If it exists, the existing output is returned directly. Multiple checkpoints are defined during task execution, and the current state and intermediate results are persisted at preset intervals. If the task is interrupted abnormally, execution is restarted from the nearest checkpoint. Determine whether to retry based on the error type, and configure the maximum number of retries and backoff strategy.

6. The production line image defect detection method based on intelligent agent orchestration according to claim 1, characterized in that, The key fields recorded in the event log include the tool name and its calling order for each step, the hash values ​​of the input and output for each step, the model version for each operation, the execution time, the specific configuration parameters and settings, context metadata, resource consumption, and exception information when errors occur.

7. The production line image defect detection method based on intelligent agent orchestration according to claim 1, characterized in that, The process of performing replay based on the event log, obtaining the difference between the result and the expected result in real time, and outputting a difference report includes: The online main workflow continues to use the old version for outputting formal results, runs control tests in parallel using the same inputs, and records execution results and control difference metrics; and / or Select historical task data samples, perform offline batch playback using the new strategy version, and output a report showing the difference between the results and the old strategy version.

8. A production line image defect detection device based on intelligent agent orchestration, characterized in that, include: The standardization module is used to convert a production line image detection request into a standardized task description when it receives such a request. The task planning module is used to generate a directed acyclic graph of subtasks and a tool call plan based on the standardized task description and tool information in the tool registry. The execution module is used to invoke tools to execute tasks sequentially according to the steps defined in the tool invocation plan, based on the directed acyclic graph of the subtasks. The log replay module is used to record the hash values ​​of the input and output data of each step, the model version number, and key parameters into the event log. Based on the event log, replay is performed to obtain the difference between the result and the expected result in real time and output a difference report.

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 method based on agent orchestration 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 method based on agent orchestration according to any one of claims 1 to 7.

Images