A kind of inspection method based on Nanobot multi-agent framework
By optimizing the inspection system using the Nanobot multi-agent framework, configuring intelligent agent units and combining them with a vector memory module for dynamic early warning, the centralized and rigid problems of traditional inspection systems are solved, achieving efficient and flexible automated inspection and proactive early warning, and improving detection accuracy and system adaptability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SUZHOU LEGO MOTORS CO LTD
- Filing Date
- 2026-06-26
- Publication Date
- 2026-07-24
AI Technical Summary
Traditional manual inspection methods are inefficient, costly, and their quality is easily affected by subjective factors. Existing automated inspection systems have insufficient adaptability due to their centralized integrated architecture, fixed functional modules that are difficult to expand flexibly, and large fluctuations in detection accuracy.
An inspection method based on the Nanobot multi-agent framework is adopted, which configures inspection planning, video processing and data aggregation agents, uses standardized input and output interfaces to call existing inspection system modules, and combines vector memory module for dynamic early warning and anomaly detection to achieve distributed operation and autonomous identification of abnormal working conditions.
The system has improved its adaptability to different scenarios and ease of iteration, reduced system modification costs, achieved highly accurate proactive early warning and autonomous anomaly identification, and reduced aquaculture losses.
Smart Images
Figure CN122454652A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence and automated inspection technology, specifically to an inspection method based on the Nanobot multi-agent framework. Background Technology
[0002] Large-scale farming is the mainstream development direction of the poultry farming industry today. In the daily operation of chicken farms, inspection work such as counting the number of poultry is a core link to ensure stable production, reduce the risk of disease, and improve farming efficiency. The traditional manual inspection mode is not only inefficient and costly, but also susceptible to the influence of subjective factors on the quality of inspections, making it difficult to meet the high-frequency and comprehensive inspection needs of modern large-scale chicken farms.
[0003] Currently, automated inspection methods are gradually becoming more common in the industry, with specialized chicken farm inspection robots being widely used to complete on-site inspection tasks. The inspection system used by this type of inspection robot adopts a centralized control architecture. All business operation logic within the system, such as data detection, target recognition, and data uploading, is integrated and deployed within a single running process, which coordinates and executes all inspection-related operations.
[0004] This centralized integrated architecture has significant technical shortcomings in practical applications. Specifically, it suffers from insufficient system adaptability. Limited by a single-process centralized management model, the internal functional modules are rigidly separated, preventing staff from flexibly adding, deleting, or independently upgrading corresponding functional modules according to the actual inspection needs of the chicken farm. This makes expansion and optimization difficult. Furthermore, the aforementioned automated inspection methods have deficiencies in early warning capabilities. For example, detection thresholds typically require manual setting to complete the detection. The lack of automatic threshold adjustment capability leads to significant fluctuations in detection accuracy.
[0005] Therefore, how to overcome the shortcomings of the existing technology is the subject of this invention. Summary of the Invention
[0006] The purpose of this invention is to provide an inspection method based on the Nanobot multi-agent framework.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0008] An inspection method based on the Nanobot multi-agent framework includes:
[0009] Multiple intelligent agents are configured, including an Inspection Planner Agent, a Video Processing Agent, and a Data Aggregator Agent. Each agent is deployed and runs independently within the Nanobot multi-agent framework.
[0010] Configure Nanobot tool modules for at least some agents and configure input / output interfaces for Nanobot tool modules so that agents can call the corresponding Nanobot tool modules through the corresponding input / output interfaces.
[0011] The inspection planning agent sends inspection tasks to the video processing agent.
[0012] After receiving the inspection task, the video processing agent calls the corresponding Nanobot tool module to complete the video detection, obtain the inspection information, and then sends the inspection information to the data aggregation agent.
[0013] The data aggregation agent aggregates the inspection information and uploads the aggregated inspection information to the vector memory module (or memory module) of the Nanobot multi-agent framework.
[0014] The inspection information is collected and summarized in the vector memory module based on the early warning mechanism, and early warnings are issued selectively.
[0015] The early warning mechanism includes:
[0016] Periodically retrieve the inspection information summarized in the vector memory module, and obtain and update the historical data of the target information in the inspection information;
[0017] Dynamic statistical baseline values for target information are derived from historical data;
[0018] The statistical value of the target information in the current inspection information is compared with the corresponding dynamic statistical benchmark value. If the comparison results are all lower than the corresponding dynamic statistical benchmark value in multiple consecutive comparisons, an early warning is issued.
[0019] The management platform is deployed as a remote host computer service. The vector memory module is integrated into the local Nanobot multi-agent framework, and the vector data generated by this module is stored in the device's internal storage.
[0020] The Nanobot multi-agent framework is an existing framework, which can be briefly explained as follows: The Nanobot multi-agent framework is a lightweight framework applicable to embedded and edge devices. It allows the construction of multiple agents with independent functions. These agents interact and collaborate through a shared message bus and tool invocation mechanism to jointly execute and complete complex tasks. This framework is particularly suitable for hardware-restricted environments, enabling efficient inference and task orchestration for multi-agent systems on hardware platforms such as ARM processors and NVIDIA Jetson.
[0021] The Nanobot tool module can directly utilize the internal modules of the inspection system described in the background technology, as detailed below. Specifically, by deploying an inspection planning agent, a video processing agent, and a data aggregation agent within the Nanobot multi-agent framework, modules from existing mature inspection systems can be directly called through the configured input / output interfaces, resulting in lower implementation costs.
[0022] In summary, this inspection method optimizes the existing inspection (system) architecture based on the Nanobot multi-agent framework. By deploying agent units adapted to different inspection tasks within the framework and utilizing the standardized input / output interfaces configured by the framework, it can directly call various functional modules within the existing mature inspection system without requiring a complete reconstruction of the original inspection system. This optimization approach can fully reuse existing technical resources and reduce the implementation costs of system modification and functional iteration. Furthermore, relying on the distributed operation characteristics of multi-agent systems, the addition, removal, and upgrade of corresponding agents can be completed independently according to the actual business needs of aquaculture inspection, with each agent operating independently. This solution fundamentally solves the drawbacks of the centralized architecture of traditional inspection systems, such as functional rigidity and inconvenient expansion, effectively improving the scenario adaptability of this inspection method and the iterative convenience of the adopted inspection architecture.
[0023] Furthermore, the system uses an early warning mechanism to detect and aggregate inspection information in the vector memory module, and selectively issues warnings. The detection of the aggregated inspection information can be continuous, with no specific restrictions. For example, after the inspection information is aggregated in the vector memory module, it's necessary to determine if there are any issues with the number of chickens. If chicken coop area 1 had 5 chickens identified in the previous inspection task, but only 3 were identified in the latest task, there might be chicken deaths. To prevent infectious disease outbreaks, an early warning is needed to alert backend personnel to conduct an on-site inspection.
[0024] And the early warning mechanism includes:
[0025] Periodically retrieve the inspection information summarized in the vector memory module, and obtain and update the historical data of the target information in the inspection information;
[0026] Dynamic statistical baseline values for target information are derived from historical data;
[0027] The statistical value of the target information in the current inspection information is compared with the corresponding dynamic statistical benchmark value. If the comparison results are all lower than the corresponding dynamic statistical benchmark value in multiple consecutive comparisons, an early warning is issued.
[0028] For example, based on the cage number, the cage status data (such as the number of chickens, the number of eggs laid, etc.) obtained from each inspection of each cage area is persistently written into the vector memory module (the data in this module can be subsequently transmitted to the management platform). This cage status data constitutes the historical data of each cage area. After each upload is completed, the historical total egg production of each cage area (this section uses cage area one as an example) is updated. The historical total egg production is divided by the number of uploads to obtain the dynamic statistical baseline value (or mean) of the egg production of cage area one. After each upload, before updating the dynamic statistical baseline value of the egg production of cage area one, the currently uploaded egg production of cage area one is compared with the current corresponding dynamic statistical baseline value (not updated based on the latest uploaded data). If the comparison results are all lower than the corresponding dynamic statistical baseline value for multiple consecutive times, an early warning is issued, and this warning information can be reported to the management platform.
[0029] In summary, this solution also constructs a continuous learning and anomaly early warning system equipped with a vector memory module. This system integrates time-series statistical analysis technology with intelligent agent memory mechanisms, endowing inspection robots with the ability to autonomously identify abnormal operating conditions. This solution eliminates the need for manual pre-configuration of judgment thresholds and rules; the memory data within the module can be continuously accumulated across sessions, enabling long-term time-series trend analysis of monitored objects. It can effectively uncover the periodic fluctuation patterns of business indicators such as egg production rate, facilitating the upgrade from passive detection to proactive early warning. This allows for the early detection of potential problems with high accuracy, reducing losses in aquaculture.
[0030] It's worth noting that the vector memory module, based on a vectorized storage mechanism, can persistently store historical inspection data in the form of semantic vectors. This allows the inspection planning agent to perform semantic similarity retrieval (rather than simple keyword queries), thereby enabling intelligent comparison of historical anomaly patterns and dynamic baseline value calculation—capabilities that are difficult for ordinary relational databases to natively support. Furthermore, the module's ability to be deployed locally ensures that data is not lost during network outages. The combination of these two core functions is a significant advantage of this module (compared to a management platform).
[0031] A further technical solution involves configuring the video processing intelligent agent with a first Nanobot tool module, a second Nanobot tool module, and a third Nanobot tool module;
[0032] The first Nanobot tool module is configured with a first input / output interface. The video processing agent can continuously call the first Nanobot tool module through the first input / output interface to complete the target region image segmentation task for the image frame.
[0033] The second Nanobot tool module is equipped with a second input / output interface. The video processing agent can continuously call the second Nanobot tool module through the second input / output interface to complete the target recognition task for the target area image.
[0034] The third Nanobot tool module is equipped with a third input / output interface. The video processing agent can continuously call the third Nanobot tool module through this third input / output interface to complete the number recognition task for the target area image.
[0035] For example, the existing Chicken Farm Node system (i.e. the existing chicken farm inspection system) includes a cage detection module (cage_model inference), an object detection module (item_model inference), a cage number OCR recognition module (Paddle OCR), and a data upload module. These modules are software algorithm modules in the system and are mature and widely used technologies.
[0036] The chicken cage detection module is encapsulated as the first Nanobot tool module in the Nanobot framework and configured with a first input / output interface (cage_detect_tool). The video processing agent can continuously call the first Nanobot tool module through the first input / output interface to complete the chicken cage region image segmentation task for image frames (or video frames).
[0037] The object detection module is encapsulated as the second Nanobot tool module in the Nanobot framework and configured with a second input / output interface (item_detect_tool). The video processing agent can continuously call the second Nanobot tool module through this second input / output interface to complete the tasks of identifying the number of chickens and the number of eggs in the chicken coop area image.
[0038] The cage number OCR recognition module is encapsulated as the third Nanobot tool module in the Nanobot framework, and is configured with a third input / output interface (ocr_recognize_tool). The video processing agent can continuously call the third Nanobot tool module through this third input / output interface to complete the task of recognizing the cage number in the image of the cage area.
[0039] For the example above, by encapsulating the mature detection modules that come with the chicken farm inspection robot into standardized tools in the Nanobot framework, and pre-defining unified input and output interfaces (which can also be uniformly referred to as input and output interfaces) for the standardized tools, at least some agents in the Nanobot framework can call the corresponding standardized tools through the corresponding standardized interfaces to complete the visual inspection and recognition tasks of the chicken farm.
[0040] By encapsulating the existing deep learning detection modules, such as the object detection module, into tools, the visual detection capabilities of these modules can be invoked by the Nanobot agent. Through encapsulation and isolation, the architecture of the visual detection unit and the agent's decision layer is decoupled. Furthermore, different toolchains can be dynamically combined during system operation (without restricting a module to being invoked by only one agent). Each module is independent and replaceable; system upgrades only require updating the corresponding agent without reconstructing the entire system, significantly reducing development and maintenance costs.
[0041] To further illustrate the existing modules, let's take the chicken coop detection module as an example. It can use the existing YOLO model to segment the chicken coop area images in the image frame, forming several images to be identified that contain complete chicken coop areas. Continuing with the example of the cage number OCR recognition module, it can use existing OCR recognition technology to complete character recognition in the chicken coop area images to obtain the chicken coop number corresponding to the chicken coop area images.
[0042] A further technical solution involves configuring a fourth Nanobot tool module for the data aggregation agent. This fourth Nanobot tool module is equipped with a fourth input / output interface. The data aggregation agent can continuously call the fourth Nanobot tool module through this fourth input / output interface to upload the aggregated inspection information to the management platform or vector memory module.
[0043] For example, the data upload module described above can be encapsulated as the fourth Nanobot tool module in the Nanobot framework and configured with a fourth input / output interface (data_upload_tool). The data aggregation agent can continuously call the fourth Nanobot tool module through this fourth input / output interface to upload the aggregated inspection information to the management platform or vector memory module.
[0044] By encapsulating the existing data upload module into a tool, the data upload capability of this module can be invoked by the Nanobot agent. Through encapsulation and isolation, the architecture of the data upload unit and the agent's decision layer is decoupled, and different tool links can be dynamically combined during system operation.
[0045] A further technical solution involves configuring multiple video processing agents.
[0046] The inspection planning agent is equipped with a large language model. The steps for the inspection planning agent to issue inspection tasks to the video processing agent include:
[0047] The large language model is invoked to perform semantic parsing of the inspection instructions;
[0048] Based on the analysis results, the overall inspection task is broken down into a list of multiple sub-tasks;
[0049] The Nanobot message bus is used to centrally relay and distribute task messages, and the decomposed list of subtasks is distributed to each video processing agent.
[0050] For example, a large language model (such as the locally deployed lightweight model Qwen-1.8B) is invoked to perform semantic parsing and intent understanding on the natural language inspection command input by the user, and the parsing results are obtained. Then, based on the parsing results, the overall inspection task is broken down into a list of sub-tasks that match specific business levels and are limited to specific time windows. The task messages are then uniformly relayed and distributed through the Nanobot message bus, and the decomposed list of sub-tasks is distributed to each (distributed) video processing agent.
[0051] For example, multiple video processing agents are configured. Each video processing agent is responsible for managing the complete lifecycle of one video stream. It calls cage_detect_tool, item_detect_tool, and ocr_recognize_tool to complete the detection according to the instructions of the inspection planning agent, maintains the chicken cage registry and cross-frame tracking status of the corresponding chicken cage area, and reports the results to the data aggregation agent.
[0052] The video processing agents are deployed and run based on the Nanobot framework. The hardware is layered as follows: the inspection robot serves as the front-end hardware carrier, responsible for on-site video data acquisition; edge computing devices handle the computational processing tasks of each video processing agent. Taking the first Nanobot tool module as an example, multiple video processing agents are configured, each equipped with one first Nanobot tool module. These video processing agents operate independently and do not share Nanobot tool modules.
[0053] The inspection planning agent is equipped with a large language model, which enables natural language interactive control. This allows farm managers to flexibly adjust inspection strategies through natural language commands without needing programming knowledge, greatly reducing the barrier to entry for the system and the difficulty of implementing inspection methods in multiple scenarios.
[0054] It is important to note that for the inspection planning agent, the task message is ultimately relayed and distributed uniformly via the Nanobot message bus, distributing the broken-down subtask list to each video processing agent. Each video processing agent independently summarizes the inspection information to the data aggregation agent. Based on this, a three-layer agent architecture is constructed, separating decision-making (planning), execution (perception), and integration (reporting) responsibilities. This conforms to the design principles of a distributed control system, ensuring that the failure of any single video processing agent does not affect the operation of other video processing agents, thus giving the overall system partial fault tolerance.
[0055] A further technical solution is that when the inspection command is a priority command, the parameters of the corresponding video processing agent are dynamically adjusted through the subtask list. These parameters include at least one of the following: video frame sampling rate (frame_skip), OCR task scheduling frequency (ocr_request_interval), and object detection confidence threshold (conf_thresholds).
[0056] Priority instructions are instructions that contain priority detection information or key detection information.
[0057] For example, when a user issues the instruction to "focus on inspecting the second layer of chicken coops", the inspection planning agent recognizes this priority instruction through a large language model and dynamically adjusts the video processing agent corresponding to the second layer of chicken coops through the subtask list. The detection frame rate of the video processing agent is increased, while the detection frame rate of other video processing agents is decreased, so as to concentrate GPU computing resources on the inspection and recognition task of the second layer of chicken coops, thereby improving the detection accuracy and response speed of key areas without changing the overall computing power.
[0058] An adaptive resource allocation mechanism is established based on a large language model. This mechanism combines the natural language understanding capabilities of the large language model with the parameter adjustment of the video processing agent, enabling non-coded reconfiguration of inspection behavior. Furthermore, parameter adjustments are asynchronously sent through the Nanobot message mechanism without interrupting the current detection process.
[0059] A further technical solution involves deploying multiple video processing agents on at least two edge computing devices, with the data aggregation agent acquiring the inspection information from each video processing agent.
[0060] When a single edge computing device (such as NVIDIA Jetson) cannot handle the visual inspection tasks of all cameras (belonging to the inspection robot), the Nanobot framework, through its built-in distributed messaging mechanism, deploys some inspection agents (i.e., video processing agents) on remote edge nodes (i.e., another edge computing device). The inspection agents deployed on different edge nodes can complete cross-node message interaction and data transmission via MQTT or WebSocket protocols, enabling distributed collaboration among multiple edge nodes to complete the visual inspection task of the chicken farm. The data aggregation agent subscribes to the detection results of all nodes, providing a unified view of the entire farm.
[0061] By deploying multiple video processing agents across at least two edge computing devices and leveraging Nanobot distributed agents to build a multi-machine collaborative architecture, the limitations of single-machine four-channel video processing are overcome, enabling comprehensive inspection of several cages in a large-scale farm. The data aggregation agent is responsible for collecting inspection information output by each video processing agent. The edge computing devices support independent operation, and local data is fully preserved even during network outages, effectively preventing data loss.
[0062] A further technical solution involves a data aggregation agent that aggregates inspection information and uploads the aggregated inspection information to the management platform and vector memory module.
[0063] The system periodically retrieves the inspection information aggregated in the management platform to obtain the statistical values or trends of the target information in the inspection information. If the statistical values or trends obtained from N consecutive retrievals do not match the expectations, an early warning is issued, where N is a preset positive integer.
[0064] In this early warning mechanism, N is a preset threshold, such as 3. This threshold is a dynamic value that can be adjusted manually and flexibly according to needs.
[0065] For example, inspection information is uploaded to the management platform. Multiple searches yield the following egg counts for chicken coop area 1: 5, 6, 5, 2, 5, 2, and 3. The fourth inspection result is 2, lower than the expected value of 4, indicating a problem. However, the fifth inspection result is 5, higher than the expected value of 4, so no warning is issued at this point. But the sixth and seventh inspection results are both lower than the expected value of 4 (where N is 2), so a warning is issued. The expected value of 4 can be derived from historical egg production data, assuming that under normal circumstances, chickens in one coop should produce at least four eggs between two inspections. If the number is lower, there is generally a problem, such as illness. However, considering various factors (such as missed inspections due to feathers obscuring eggs), it is possible that some chickens in certain coops will produce fewer eggs than expected between two inspections. Manual on-site verification each time would significantly increase labor costs. Therefore, the specific value of N can be flexibly adjusted according to needs to balance labor costs and farming risks.
[0066] It should be noted that the management platform, as a remote host computer service, enables centralized storage and visualization of inspection data, supports cross-regional coordinated supervision of multiple chicken farms, and can connect to the early warning mechanism to synchronously push remote alarms. Its core function is to collect fragmented inspection data (local inspection data) from various edge terminals, build a unified data view across the entire domain, and form a closed-loop management link that allows for remote manual intervention. This core function cannot be achieved by a pure edge deployment solution that relies solely on local storage.
[0067] By setting up a dual early warning mechanism (early warning based on the management platform and early warning based on the vector memory module), potential problems can be detected in advance, reducing aquaculture losses.
[0068] A further technical solution involves the inspection planning agent periodically retrieving and summarizing the inspection information in the vector memory module, and adjusting the parameters of the corresponding video processing agent based on the retrieval results. These parameters include at least one of the following: video frame sampling rate, OCR task scheduling frequency, and target detection confidence threshold.
[0069] The inspection planning agent can also periodically retrieve and summarize inspection information from the management platform.
[0070] For example, continuing from the above description, in the previous inspection task, 5 chickens were identified in chicken coop area 1, but in the latest inspection task, only 3 chickens were identified, which may indicate that some chickens have died. In order to avoid an outbreak of infection, it is necessary to issue an early warning to remind the back-end personnel to check on-site. This information is retrieved and analyzed by the inspection planning agent, which can adjust the parameters of the video processing agent corresponding to chicken coop area 1 to focus on monitoring chicken coop area 1.
[0071] It should be noted that, compared with fixed threshold rule retrieval, semantic retrieval based on the vector memory module (implemented by the inspection planning agent) can adaptively identify historical abnormal patterns and dynamically adjust the benchmark reference value (such as the dynamic statistical benchmark value mentioned above), avoiding the problem of manually preset rules lagging behind actual working conditions and having stronger adaptive capabilities.
[0072] The terms "first," "second," etc., used in this article do not specifically refer to order or sequence, nor are they intended to limit this case; they are merely used to distinguish components or operations described using the same technical terms.
[0073] The terms "connection" or "positioning" as used in this article can refer to two or more components or devices making direct physical contact with each other, or making indirect physical contact with each other, or to two or more components or devices operating or moving with each other.
[0074] The terms “include,” “including,” and “have” used in this article are all open-ended, meaning they include but are not limited to.
[0075] Unless otherwise specified, the terms used herein generally have their ordinary meaning in the context of the art, the subject matter, and the specific context. Certain terms used to describe this case will be discussed below or elsewhere in this specification to provide additional guidance to those skilled in the art in describing this case.
[0076] The terms “front,” “back,” “up,” “down,” “left,” and “right” used in this article are directional terms. In this case, they are only used to describe the positional relationship between the structures and are not intended to limit the specific direction of the protection scheme or its actual implementation.
[0077] The working principle and advantages of this invention are as follows:
[0078] This inspection method optimizes the existing inspection architecture based on the Nanobot multi-agent framework. By deploying agent units adapted to different inspection tasks within the framework and utilizing the standardized input / output interfaces configured by the framework, it can directly call various functional modules within existing mature inspection systems without requiring a complete reconstruction of the original inspection system. This optimization approach can fully reuse existing technical resources and reduce the implementation costs of system modification and functional iteration. Furthermore, leveraging the distributed operation characteristics of multi-agent systems, it can independently complete the addition, removal, and upgrade of corresponding agents according to the actual business needs of aquaculture inspections, with each agent operating independently. This solution fundamentally solves the drawbacks of the centralized architecture of traditional inspection systems, such as functional rigidity and inconvenient expansion, effectively improving the scenario adaptability of this inspection method and the ease of iteration of the adopted inspection architecture.
[0079] Furthermore, this inspection method is based on an early warning mechanism that detects and aggregates inspection information in the vector memory module, and selectively issues early warnings. The early warning mechanism includes:
[0080] Periodically retrieve the inspection information summarized in the vector memory module, and obtain and update the historical data of the target information in the inspection information;
[0081] Dynamic statistical baseline values for target information are derived from historical data;
[0082] The statistical value of the target information in the current inspection information is compared with the corresponding dynamic statistical benchmark value. If the comparison results are all lower than the corresponding dynamic statistical benchmark value in multiple consecutive comparisons, an early warning is issued.
[0083] In summary, this inspection method also constructs a continuous learning and anomaly early warning system equipped with a vector memory module. It integrates time-series statistical analysis technology with intelligent agent memory mechanisms, endowing the inspection robot with the ability to autonomously identify abnormal working conditions. This solution eliminates the need for manual pre-configuration of judgment thresholds and rules; the memory data within the module can be continuously accumulated across sessions, enabling long-term time-series trend analysis of monitored objects. It can effectively uncover the periodic fluctuation patterns of business indicators such as egg production rate, facilitating the upgrade from passive detection to proactive early warning. This allows for the early detection of potential problems with high accuracy, reducing losses in aquaculture. Attached Figure Description
[0084] Figure 1 This is a flowchart of a specific implementation of the inspection method according to an embodiment of the present invention. Detailed Implementation
[0085] The present invention will be clearly described below with illustrations and detailed description. Any person skilled in the art who understands the embodiments of the present invention can make changes and modifications based on the technology taught in the present invention without departing from the spirit and scope of the present invention.
[0086] The terminology used herein is for the purpose of describing specific embodiments only and is not intended to limit the scope of this work. Singular forms such as “a,” “this,” “this,” “the,” and “the” as used herein also include plural forms.
[0087] See Figure 1 An inspection method based on the Nanobot multi-agent framework includes:
[0088] Multiple intelligent agents are configured, including an Inspection Planner Agent, a Video Processing Agent, and a Data Aggregator Agent. Each agent is deployed and runs independently within the Nanobot multi-agent framework.
[0089] Configure Nanobot tool modules for at least some agents and configure input / output interfaces for Nanobot tool modules so that agents can call the corresponding Nanobot tool modules through the corresponding input / output interfaces.
[0090] The inspection planning agent sends inspection tasks to the video processing agent.
[0091] After receiving the inspection task, the video processing agent calls the corresponding Nanobot tool module to complete the video detection, obtain the inspection information, and then sends the inspection information to the data aggregation agent.
[0092] The data aggregation agent aggregates the inspection information and uploads the aggregated inspection information to the vector memory module (or memory module) of the Nanobot multi-agent framework.
[0093] The inspection information is collected and summarized in the vector memory module based on the early warning mechanism, and early warnings are issued selectively.
[0094] The early warning mechanism includes:
[0095] Periodically retrieve the inspection information summarized in the vector memory module, and obtain and update the historical data of the target information in the inspection information;
[0096] Dynamic statistical baseline values for target information are derived from historical data;
[0097] The statistical value of the target information in the current inspection information is compared with the corresponding dynamic statistical benchmark value. If the comparison results are all lower than the corresponding dynamic statistical benchmark value in multiple consecutive comparisons, an early warning is issued.
[0098] The management platform is deployed as a remote host computer service. The vector memory module is integrated into the local Nanobot multi-agent framework, and the vector data generated by this module is stored in the device's internal storage.
[0099] The Nanobot multi-agent framework is an existing framework, which can be briefly explained as follows: The Nanobot multi-agent framework is a lightweight framework applicable to embedded and edge devices. It allows the construction of multiple agents with independent functions. These agents interact and collaborate through a shared message bus and tool invocation mechanism to jointly execute and complete complex tasks. This framework is particularly suitable for hardware-restricted environments, enabling efficient inference and task orchestration for multi-agent systems on hardware platforms such as ARM processors and NVIDIA Jetson.
[0100] The Nanobot tool module can directly utilize the internal modules of the inspection system described in the background technology, as detailed below. Specifically, by deploying an inspection planning agent, a video processing agent, and a data aggregation agent within the Nanobot multi-agent framework, modules from existing mature inspection systems can be directly called through the configured input / output interfaces, resulting in lower implementation costs.
[0101] In summary, this inspection method optimizes the existing inspection (system) architecture based on the Nanobot multi-agent framework. By deploying agent units adapted to different inspection tasks within the framework and utilizing the standardized input / output interfaces configured by the framework, it directly calls various functional modules within the existing mature inspection system without requiring a complete reconstruction of the original system. This optimization approach fully reuses existing technical resources, reducing the implementation costs of system modification and functional iteration. Furthermore, leveraging the distributed operation characteristics of multi-agent systems, it allows for the independent addition, removal, and upgrading of corresponding agents based on actual aquaculture inspection needs, with each agent operating independently. This solution fundamentally solves the drawbacks of the centralized architecture of traditional inspection systems, such as functional rigidity and inconvenient expansion, effectively improving the scenario adaptability of this inspection method and the ease of iteration of the adopted inspection architecture.
[0102] Furthermore, the system uses an early warning mechanism to detect and aggregate inspection information in the vector memory module, and selectively issues warnings. The detection of the aggregated inspection information can be continuous, with no specific restrictions. For example, after the inspection information is aggregated in the vector memory module, it's necessary to determine if there are any issues with the number of chickens. If chicken coop area 1 had 5 chickens identified in the previous inspection task, but only 3 were identified in the latest task, there might be chicken deaths. To prevent infectious disease outbreaks, an early warning is needed to alert backend personnel to conduct an on-site inspection.
[0103] And the early warning mechanism includes:
[0104] Periodically retrieve the inspection information summarized in the vector memory module, and obtain and update the historical data of the target information in the inspection information;
[0105] Dynamic statistical baseline values for target information are derived from historical data;
[0106] The statistical value of the target information in the current inspection information is compared with the corresponding dynamic statistical benchmark value. If the comparison results are all lower than the corresponding dynamic statistical benchmark value in multiple consecutive comparisons, an early warning is issued.
[0107] For example, based on the cage number, the cage status data (such as the number of chickens, the number of eggs laid, etc.) obtained from each inspection of each cage area is persistently written into the vector memory module (the data in this module can be subsequently transmitted to the management platform). This cage status data constitutes the historical data of each cage area. After each upload is completed, the historical total egg production of each cage area (this section uses cage area one as an example) is updated. The historical total egg production is divided by the number of uploads to obtain the dynamic statistical baseline value (or mean) of the egg production of cage area one. After each upload, before updating the dynamic statistical baseline value of the egg production of cage area one, the currently uploaded egg production of cage area one is compared with the current corresponding dynamic statistical baseline value (not updated based on the latest uploaded data). If the comparison results are all lower than the corresponding dynamic statistical baseline value for multiple consecutive times, an early warning is issued, and this warning information can be reported to the management platform.
[0108] In summary, this solution also constructs a continuous learning and anomaly early warning system equipped with a vector memory module. This system integrates time-series statistical analysis technology with intelligent agent memory mechanisms, endowing inspection robots with the ability to autonomously identify abnormal operating conditions. This solution eliminates the need for manual pre-configuration of judgment thresholds and rules; the memory data within the module can be continuously accumulated across sessions, enabling long-term time-series trend analysis of monitored objects. It can effectively uncover the periodic fluctuation patterns of business indicators such as egg production rate, facilitating the upgrade from passive detection to proactive early warning, enabling the early detection of potential problems and reducing losses in aquaculture.
[0109] It's worth noting that the vector memory module, based on a vectorized storage mechanism, can persistently store historical inspection data in the form of semantic vectors. This allows the inspection planning agent to perform semantic similarity retrieval (rather than simple keyword queries), thereby enabling intelligent comparison of historical anomaly patterns and dynamic baseline value calculation—capabilities that are difficult for ordinary relational databases to natively support. Furthermore, the module's ability to be deployed locally ensures that data is not lost during network outages. The combination of these two core functions is a significant advantage of this module (compared to a management platform).
[0110] In one embodiment of this application, the video processing intelligent agent is configured with a first Nanobot tool module, a second Nanobot tool module, and a third Nanobot tool module;
[0111] The first Nanobot tool module is configured with a first input / output interface. The video processing agent can continuously call the first Nanobot tool module through the first input / output interface to complete the target region image segmentation task for the image frame.
[0112] The second Nanobot tool module is equipped with a second input / output interface. The video processing agent can continuously call the second Nanobot tool module through this second input / output interface to complete the target recognition task for the target area image and obtain some inspection information.
[0113] The third Nanobot tool module is equipped with a third input / output interface. The video processing agent can continuously call the third Nanobot tool module through this third input / output interface to complete the number recognition task for the target area image and obtain some inspection information.
[0114] For example, the existing Chicken Farm Node system (i.e., the existing chicken farm inspection system, such as the Mujilang series of caged layer chicken intelligent inspection robots developed by Fuzhou Mujilang Intelligent Technology Co., Ltd.) includes a cage detection module (cage_model inference), an object detection module (item_model inference), a cage number OCR recognition module (Paddle OCR), and a data upload module. These modules belong to the software algorithm modules of the system and are mature and used technologies.
[0115] The chicken cage detection module is encapsulated as the first Nanobot tool module in the Nanobot framework and configured with a first input / output interface (cage_detect_tool). The video processing agent can continuously call the first Nanobot tool module through the first input / output interface to complete the chicken cage region image segmentation task for image frames (or video frames).
[0116] The object detection module is encapsulated as the second Nanobot tool module in the Nanobot framework and configured with a second input / output interface (item_detect_tool). The video processing agent can continuously call the second Nanobot tool module through this second input / output interface to complete the tasks of identifying the number of chickens and the number of eggs in the chicken coop area image.
[0117] The cage number OCR recognition module is encapsulated as the third Nanobot tool module in the Nanobot framework, and is configured with a third input / output interface (ocr_recognize_tool). The video processing agent can continuously call the third Nanobot tool module through this third input / output interface to complete the task of recognizing the cage number in the image of the cage area.
[0118] For the example above, by encapsulating the mature detection modules that come with the chicken farm inspection robot into standardized tools in the Nanobot framework, and pre-defining unified input and output interfaces (which can also be collectively referred to as the above input and output interfaces) for the standardized tools, at least some agents in the Nanobot framework can call the corresponding standardized tools through the corresponding standardized interfaces to complete the visual inspection and recognition tasks of the chicken farm.
[0119] By encapsulating the existing deep learning detection modules, such as the object detection module, into tools, the visual detection capabilities of these modules can be invoked by the Nanobot agent. Through encapsulation and isolation, the architecture of the visual detection unit and the agent's decision layer is decoupled. Furthermore, different toolchains can be dynamically combined during system operation (without restricting a module to being invoked by only one agent). Each module is independent and replaceable; system upgrades only require updating the corresponding agent without reconstructing the entire system, significantly reducing development and maintenance costs.
[0120] To further illustrate the existing modules, let's take the chicken coop detection module as an example. It can use the existing YOLO model to segment the chicken coop area images in the image frame, forming several images to be identified that contain complete chicken coop areas. Continuing with the example of the cage number OCR recognition module, it can use existing OCR recognition technology to complete character recognition in the chicken coop area images to obtain the chicken coop number corresponding to the chicken coop area images.
[0121] In one embodiment of this application, the data aggregation agent is configured with a fourth Nanobot tool module. The fourth Nanobot tool module is configured with a fourth input / output interface. The data aggregation agent can continuously call the fourth Nanobot tool module through the fourth input / output interface to upload the aggregated inspection information to the management platform or vector memory module.
[0122] For example, the data upload module described above can be encapsulated as the fourth Nanobot tool module in the Nanobot framework and configured with a fourth input / output interface (data_upload_tool). The data aggregation agent can continuously call the fourth Nanobot tool module through this fourth input / output interface to upload the aggregated inspection information to the management platform or vector memory module.
[0123] By encapsulating the existing data upload module into a tool, the data upload capability of this module can be invoked by the Nanobot agent. Through encapsulation and isolation, the architecture of the data upload unit and the agent's decision layer is decoupled, and different tool links can be dynamically combined during system operation.
[0124] In one embodiment of this application, multiple video processing agents are configured.
[0125] The inspection planning agent is equipped with a large language model. The steps for the inspection planning agent to issue inspection tasks to the video processing agent include:
[0126] The large language model is invoked to perform semantic parsing of the inspection instructions;
[0127] Based on the analysis results, the overall inspection task is broken down into a list of multiple sub-tasks;
[0128] The Nanobot message bus is used to centrally relay and distribute task messages, and the decomposed list of subtasks is distributed to each video processing agent.
[0129] For example, a large language model (such as the locally deployed lightweight model Qwen-1.8B) is invoked to perform semantic parsing and intent understanding on the natural language inspection command input by the user, and the parsing results are obtained. Then, based on the parsing results, the overall inspection task is broken down into a list of sub-tasks that match specific business levels and are limited to specific time windows. The task messages are then uniformly relayed and distributed through the Nanobot message bus, and the decomposed list of sub-tasks is distributed to each (distributed) video processing agent.
[0130] For example, multiple video processing agents are configured. Each video processing agent is responsible for managing the complete lifecycle of one video stream. It calls cage_detect_tool, item_detect_tool, and ocr_recognize_tool to complete the detection according to the instructions of the inspection planning agent, maintains the chicken cage registry and cross-frame tracking status of the corresponding chicken cage area, and reports the results to the data aggregation agent.
[0131] The video processing agents are deployed and run based on the Nanobot framework. The hardware is layered as follows: the inspection robot serves as the front-end hardware carrier, responsible for on-site video data acquisition; edge computing devices handle the computational processing tasks of each video processing agent. Taking the first Nanobot tool module as an example, multiple video processing agents are configured, each equipped with one first Nanobot tool module. These video processing agents operate independently and do not share Nanobot tool modules.
[0132] The inspection planning agent is equipped with a large language model, which enables natural language interactive control. This allows farm managers to flexibly adjust inspection strategies through natural language commands without needing programming knowledge, greatly reducing the barrier to entry for the system and the difficulty of implementing inspection methods in multiple scenarios.
[0133] It is important to note that for the inspection planning agent, the task message is ultimately relayed and distributed uniformly via the Nanobot message bus, distributing the broken-down subtask list to each video processing agent. Each video processing agent independently summarizes the inspection information to the data aggregation agent. Based on this, a three-layer agent architecture is constructed, separating decision-making (planning), execution (perception), and integration (reporting) responsibilities. This conforms to the design principles of a distributed control system, ensuring that the failure of any single video processing agent does not affect the operation of other video processing agents, thus giving the overall system partial fault tolerance.
[0134] In one embodiment of this application, when the inspection instruction is a priority instruction, the parameters of the corresponding video processing agent are dynamically adjusted through the subtask list. The parameters include at least one of the following: video frame sampling rate (frame_skip), OCR task scheduling frequency (ocr_request_interval), and target detection confidence threshold (conf_thresholds).
[0135] Priority instructions are instructions that contain priority detection information or key detection information.
[0136] For example, when a user issues the instruction to "focus on inspecting the second layer of chicken coops", the inspection planning agent recognizes this priority instruction through a large language model and dynamically adjusts the video processing agent corresponding to the second layer of chicken coops through the subtask list. The detection frame rate of the video processing agent is increased, while the detection frame rate of other video processing agents is decreased, so as to concentrate GPU computing resources on the inspection and recognition task of the second layer of chicken coops, thereby improving the detection accuracy and response speed of key areas without changing the overall computing power.
[0137] An adaptive resource allocation mechanism is established based on a large language model. This mechanism combines the natural language understanding capabilities of the large language model with the parameter adjustment of the video processing agent, enabling non-coded reconfiguration of inspection behavior. Furthermore, parameter adjustments are asynchronously sent through the Nanobot message mechanism without interrupting the current detection process.
[0138] In one embodiment of this application, multiple video processing agents are deployed on at least two edge computing devices, and a data aggregation agent obtains the inspection information of each video processing agent.
[0139] When a single edge computing device (such as NVIDIA Jetson) cannot handle the visual inspection tasks of all cameras (belonging to the inspection robot), the Nanobot framework, through its built-in distributed messaging mechanism, deploys some inspection agents (i.e., video processing agents) on remote edge nodes (i.e., another edge computing device). The inspection agents deployed on different edge nodes can complete cross-node message interaction and data transmission via MQTT or WebSocket protocols, enabling distributed collaboration among multiple edge nodes to complete the visual inspection task of the chicken farm. The data aggregation agent subscribes to the detection results of all nodes, providing a unified view of the entire farm.
[0140] By deploying multiple video processing agents across at least two edge computing devices and leveraging Nanobot distributed agents to build a multi-machine collaborative architecture, the limitations of single-machine four-channel video processing are overcome, enabling comprehensive inspection of several cages in a large-scale farm. The data aggregation agent is responsible for collecting inspection information output by each video processing agent. The edge computing devices support independent operation, and local data is fully preserved even during network outages, effectively preventing data loss.
[0141] In one embodiment of this application, the data aggregation agent aggregates the inspection information and uploads the aggregated inspection information to the management platform and the vector memory module;
[0142] The system periodically retrieves the inspection information aggregated in the management platform to obtain the statistical values or trends of the target information in the inspection information. If the statistical values or trends obtained from N consecutive retrievals do not match the expectations, an early warning is issued, where N is a preset positive integer.
[0143] In this early warning mechanism, N is a preset threshold, such as 3. This threshold is a dynamic value that can be adjusted manually and flexibly according to needs.
[0144] For example, inspection information is uploaded to the management platform. Multiple searches yield the following egg counts for chicken coop area 1: 5, 6, 5, 2, 5, 2, and 3. The fourth inspection result is 2, lower than the expected value of 4, indicating a problem. However, the fifth inspection result is 5, higher than the expected value of 4, so no warning is issued at this point. But the sixth and seventh inspection results are both lower than the expected value of 4 (where N is 2), so a warning is issued. The expected value of 4 can be derived from historical egg production data, assuming that under normal circumstances, chickens in one coop should produce at least four eggs between two inspections. If the number is lower, there is generally a problem, such as illness. However, considering various factors (such as missed inspections due to feathers obscuring eggs), it is possible that some chickens in certain coops will produce fewer eggs than expected between two inspections. Manual on-site verification each time would significantly increase labor costs. Therefore, the specific value of N can be flexibly adjusted according to needs to balance labor costs and farming risks.
[0145] It should be noted that the management platform, as a remote host computer service, enables centralized storage and visualization of inspection data, supports cross-regional coordinated supervision of multiple chicken farms, and can connect to the early warning mechanism to synchronously push remote alarms. Its core function is to collect fragmented inspection data (local inspection data) from various edge terminals, build a unified data view across the entire domain, and form a closed-loop management link that allows for remote manual intervention. This core function cannot be achieved by a pure edge deployment solution that relies solely on local storage.
[0146] By setting up a dual early warning mechanism (early warning based on the management platform and early warning based on the vector memory module), potential problems can be detected in advance, reducing aquaculture losses.
[0147] In one embodiment of this application, the inspection planning agent periodically retrieves and summarizes the inspection information in the vector memory module, and adjusts the parameters of the corresponding video processing agent based on the retrieval results. The parameters include at least one of the video frame sampling rate, OCR task scheduling frequency, and target detection confidence threshold.
[0148] The inspection planning agent can also periodically retrieve and summarize inspection information from the management platform.
[0149] For example, continuing from the above description, in the previous inspection task, 5 chickens were identified in chicken coop area 1, but in the latest inspection task, only 3 chickens were identified, which may indicate that some chickens have died. In order to avoid an outbreak of infection, it is necessary to issue an early warning to remind the back-end personnel to check on-site. This information is retrieved and analyzed by the inspection planning agent, which can adjust the parameters of the video processing agent corresponding to chicken coop area 1 to focus on monitoring chicken coop area 1.
[0150] It should be noted that, compared with fixed threshold rule retrieval, semantic retrieval based on the vector memory module (implemented by the inspection planning agent) can adaptively identify historical abnormal patterns and dynamically adjust the benchmark reference value (such as the dynamic statistical benchmark value mentioned above), avoiding the problem of manually preset rules lagging behind actual working conditions and having stronger adaptive capabilities.
[0151] The following provides specific examples for illustration.
[0152] Example 1: Application in the inspection scenario of a medium-sized layer chicken farm (approximately 2000 cages)
[0153] Hardware configuration: 2 NVIDIA Jetson Orin NX edge computing devices, each connected to 4 IP cameras, covering a total of 8 video feeds; 1 management server running the host computer management platform and Nanobot message broker service.
[0154] Agent deployment: Each edge computing device runs 4 video processing agents (corresponding to 4 video streams) and 1 local data aggregation agent; the management server runs 1 global inspection and planning agent (calling a large language model in the cloud or locally) and 1 global data aggregation agent.
[0155] Operation process:
[0156] Step 1: Managers input natural language commands via mobile app: This morning, focus on patrolling the first two floors of Zone 1, and patrol the remaining floors as usual. The patrol planning agent calls the local large language model to parse the command, generate a priority scheduling strategy, and send it to the corresponding video processing agent through the Nanobot message bus.
[0157] Step 2: Increase the detection frame rate of the video processing agents in layers 1-2 of Zone 1 to 30 FPS, and reduce the frame rate of the remaining video processing agents to 10 FPS. At the same time, reduce the OCR request frequency and dynamically allocate GPU resources to priority areas.
[0158] Step 3: Each video processing agent continuously calls the corresponding Nanobot tool module to complete the detection, and gradually reports the results to the global data aggregation agent. These data are eventually written into at least one of the vector memory module and the management platform.
[0159] Step 4: The inspection planning intelligent agent retrieves the memory module and finds that the egg production of cage 03 in Zone 1 has been 0 for 3 consecutive inspections (the historical average is 2.3). It automatically triggers an abnormality warning and pushes it to the mobile phone of the management personnel.
[0160] The inspection system used in the inspection method has the ability to synchronously and parallelly analyze and process 8 channels of 1080P video streams. It optimizes latency for business priority areas, and the average response latency is controlled within 50ms. It relies on algorithms to achieve intelligent early warning of anomalies. After manual sampling verification, the early warning accuracy rate reaches 89%. The whole machine can meet the requirements of 24 / 7 uninterrupted automated operation and maintenance, and can run smoothly and continuously without human intervention.
[0161] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.
Claims
1. An inspection method based on the Nanobot multi-agent framework, characterized in that: include: Multiple intelligent agents are configured, including an inspection planning agent, a video processing agent, and a data aggregation agent. Each agent is deployed and runs independently within the Nanobot multi-agent framework. Configure Nanobot tool modules for at least some agents and configure input / output interfaces for Nanobot tool modules so that agents can call the corresponding Nanobot tool modules through the corresponding input / output interfaces. The inspection planning agent sends inspection tasks to the video processing agent. After receiving the inspection task, the video processing agent calls the corresponding Nanobot tool module to complete the video detection, obtain the inspection information, and then sends the inspection information to the data aggregation agent. The data aggregation agent aggregates the inspection information and uploads the aggregated inspection information to the vector memory module of the Nanobot multi-agent framework; The inspection information is collected and summarized in the vector memory module based on the early warning mechanism, and early warnings are issued selectively. The early warning mechanism includes: Periodically retrieve the inspection information summarized in the vector memory module, and obtain and update the historical data of the target information in the inspection information; Dynamic statistical baseline values for target information are derived from historical data; The statistical value of the target information in the current inspection information is compared with the corresponding dynamic statistical benchmark value. If the comparison results are all lower than the corresponding dynamic statistical benchmark value in multiple consecutive comparisons, an early warning is issued.
2. The inspection method based on the Nanobot multi-agent framework according to claim 1, characterized in that: The video processing agent is equipped with a first Nanobot tool module, a second Nanobot tool module, and a third Nanobot tool module; The first Nanobot tool module is configured with a first input / output interface. The video processing agent can continuously call the first Nanobot tool module through the first input / output interface to complete the target region image segmentation task for the image frame. The second Nanobot tool module is equipped with a second input / output interface. The video processing agent can continuously call the second Nanobot tool module through the second input / output interface to complete the target recognition task for the target area image. The third Nanobot tool module is equipped with a third input / output interface. The video processing agent can continuously call the third Nanobot tool module through this third input / output interface to complete the number recognition task for the target area image.
3. The inspection method based on the Nanobot multi-agent framework according to claim 1, characterized in that: The data aggregation agent is equipped with a fourth Nanobot tool module, which has a fourth input / output interface. The data aggregation agent can continuously call the fourth Nanobot tool module through the fourth input / output interface to upload the aggregated inspection information to the management platform or vector memory module.
4. The inspection method based on the Nanobot multi-agent framework according to claim 1, characterized in that: The video processing agent is configured as multiple; The inspection planning agent is equipped with a large language model. The steps for the inspection planning agent to issue inspection tasks to the video processing agent include: The large language model is invoked to perform semantic parsing of the inspection instructions; Based on the analysis results, the overall inspection task is broken down into a list of multiple sub-tasks; The Nanobot message bus is used to centrally relay and distribute task messages, and the decomposed list of subtasks is distributed to each video processing agent.
5. The inspection method based on the Nanobot multi-agent framework according to claim 4, characterized in that: When the inspection command is a priority command, the parameters of the corresponding video processing agent are dynamically adjusted through the subtask list. These parameters include at least one of the following: video frame sampling rate, OCR task scheduling frequency, and target detection confidence threshold.
6. The inspection method based on the Nanobot multi-agent framework according to claim 4, characterized in that: Multiple video processing agents are deployed on at least two edge computing devices, and the data aggregation agent obtains the inspection information of each video processing agent.
7. The inspection method based on the Nanobot multi-agent framework according to claim 1, characterized in that: The data aggregation agent aggregates the inspection information and uploads the aggregated inspection information to the management platform and vector memory module; The system periodically retrieves the inspection information aggregated in the management platform to obtain the statistical values or trends of the target information in the inspection information. If the statistical values or trends obtained from N consecutive retrievals do not match the expectations, an early warning is issued, where N is a preset positive integer.
8. The inspection method based on the Nanobot multi-agent framework according to claim 1, characterized in that: The inspection planning agent periodically retrieves and summarizes the inspection information in the vector memory module, and adjusts the parameters of the corresponding video processing agent based on the retrieval results. These parameters include at least one of the following: video frame sampling rate, OCR task scheduling frequency, and target detection confidence threshold.