Supervision system and method for screening laboratory of component traditional Chinese medicine pharmacodynamic substances

Abstract

The invention provides a supervision system and method for a screening laboratory of component traditional Chinese medicine efficacy substances. The system comprises a robot, an RPA module, a large language model decision module, a multi-agent calling module and an ROS module. The RPA module is used for collecting experimental equipment state information, traditional Chinese medicine sample state information, laboratory environment parameter state information and experimenter state information; the large language model decision module is used for analyzing a management rule of the component traditional Chinese medicine efficacy substance screening laboratory, making a decision and generating a task instruction; and the multi-agent calling module is used for receiving the task instruction and carrying out priority scheduling management on the agents. According to the supervision system and method for the laboratory for screening the component traditional Chinese medicine pharmacodynamic substances, standard operation monitoring of laboratory staff, illegal article identification, real-time monitoring of the experimental process, problem early warning and dynamic adjustment can be achieved, and the inspection efficiency and the laboratory safety are improved.

Description

Technical Field

[0001] The invention belongs to the field of laboratory technology, and in particular relates to a supervision system and method for a component traditional Chinese medicine efficacy substance screening laboratory. Background Art

[0002] The screening laboratory for the effective substances of component traditional Chinese medicine is an experimental site for screening the active substances of component traditional Chinese medicine. It is equipped with necessary instruments and equipment (such as washers, incubation systems, and pipetting systems) and technical resources. Its main function is to verify the active substances of component traditional Chinese medicine through experiments. However, the environment of the screening laboratory for the effective substances of component traditional Chinese medicine is complex, and the high-voltage equipment, high-speed equipment, high-temperature equipment, etc. used bring many safety hazards. Although many laboratories are now equipped with intelligent monitoring systems, such as smoke alarms and temperature sensors, which can monitor the laboratory environment in real time, these systems also have some shortcomings. For example, these systems can usually only detect a single type of hazard, and their comprehensive analysis capabilities for multiple potential hazards are weak.

[0003] In addition, the laboratory still relies on a large amount of manpower for daily monitoring and maintenance of whether the laboratory technicians operate in accordance with regulations and whether illegal items enter the laboratory (food, etc.), and cannot fully realize automated processing. These shortcomings show that the existing intelligent monitoring system still has obvious deficiencies when facing daily maintenance work and various safety hazards in the laboratory environment. It is unable to realize patrol inspection and supervision of the laboratory, which further increases the limitations of the system. These problems urgently need to be solved by improving and enhancing the comprehensive monitoring and automated processing capabilities of the system. Summary of the invention

[0004] In view of this, the present invention aims to propose a supervision system and method for component Chinese medicine efficacy substance screening laboratories, so as to solve the problem that the existing laboratory intelligent monitoring system cannot realize inspection and supervision of the laboratory and has great limitations.

[0005] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0006] In a first aspect, the present invention provides a supervision system for a component Chinese medicine efficacy substance screening laboratory, comprising a robot, an RPA module, a large language model decision module, a multi-agent calling module, and a ROS module;

[0007] Robots are used to inspect and supervise the screening laboratories for the efficacy of component Chinese medicines;

[0008] RPA module is used to collect experimental equipment status information, Chinese medicine sample status information, laboratory environmental parameter status information, and experimental personnel status information;

[0009] A large language model decision module is used to analyze the management rules of the component Chinese medicine efficacy substance screening laboratory, and make decisions on the acquisition of experimental equipment status information, Chinese medicine sample status information, laboratory environmental parameter status information, and experimental personnel status information based on the management rules to generate task instructions;

[0010] A multi-agent calling module is used to receive the task instructions, and use the action- and state-based joint attention reinforcement learning technology to implement the priority scheduling management of the agent, and call the agent to perform the task; wherein the agent includes an RPA module, a large language model decision module, and a ROS module, and each of the agents is provided with an Actor-Critic network framework;

[0011] The ROS module is used to drive the robot to work according to the instructions of the multi-agent calling module.

[0012] Furthermore, the RPA module includes a data acquisition unit and an image acquisition unit.

[0013] Furthermore, the large language model decision module includes a semantic parsing unit, a RAG indexing unit, a CNN image recognition unit, and an instruction scheduling unit;

[0014] A semantic parsing unit is used to perform semantic parsing and intent recognition on the status information to obtain data information; wherein, when the data information is text information, the RAG indexing unit is used to index the information according to the management rules to obtain an index result and generate a task instruction according to the index result; when the data information is image information, the CNN image recognition unit is used to identify and parse the information first to obtain text information, and then the RAG indexing unit is used to index the information according to the management rules to obtain an index result and generate a task instruction according to the index result;

[0015] The instruction scheduling unit is used to uniformly schedule, send and receive information at different stages in the large language model decision module.

[0016] Furthermore, the multi-agent calling module includes a task allocation unit and a multi-objective optimization unit;

[0017] A task allocation unit, used for receiving the task instructions and allocating the task instructions according to the agents corresponding to the task instructions to obtain subtask instructions corresponding to the agents;

[0018] A multi-objective optimization unit is used to prioritize the sub-task instructions using the action- and state-based joint attention reinforcement learning technology, and after optimizing the priority ranking based on the MODPO reinforcement learning technology, the intelligent agent is called to perform tasks according to its corresponding sub-task instructions based on the optimized priority ranking.

[0019] Further, the ROS module includes a node topic unit, a service and action library unit, a parameter server unit, and a driver unit;

[0020] A node topic unit is used to read the calling instruction of the multi-objective optimization unit and control the driving unit according to the calling instruction

[0021] Service and action library units for synchronous request-reply communication and management of long-running tasks;

[0022] Parameter server unit, used to store the robot's parameters and data to ensure configuration consistency;

[0023] The driving unit is used to drive the robot to work.

[0024] In a second aspect, the present invention provides a supervision method for a component Chinese medicine efficacy substance screening laboratory, comprising:

[0025] An Actor-Critic network framework is set up in the RPA module, the large language model decision module, and the ROS module respectively;

[0026] A robot with an RPA module is arranged in the laboratory, and after the RPA module is connected to the experimental equipment in the laboratory, the robot with the RPA module is used to collect data and conduct inspection and supervision of the laboratory;

[0027] The data collected by the RPA module is parsed using a large language model decision module, and decisions are made based on the management rules of the component Chinese medicine efficacy substance screening laboratory to generate task instructions;

[0028] Using a multi-agent calling module to receive the task instructions, and using the action- and state-based joint attention reinforcement learning technology to implement the priority scheduling management of the agent, calling the agent to perform the task;

[0029] The ROS module is used to receive instructions from the multi-agent calling module, and the robot is driven to inspect and supervise the laboratory according to the instructions from the multi-agent calling module.

[0030] Furthermore, an Actor-Critic network framework is set in each of the RPA module, the large language model decision module, and the ROS module, including:

[0031] An Actor network is respectively set in the RPA module, the large language model decision module, and the ROS module; wherein the input part of the Actor network includes the state information of each intelligent agent;

[0032] A Critic network is respectively set in the RPA module, the large language model decision module, and the ROS module; wherein the input part of the Critic network is a state-action pair;

[0033] The Actor network receives current state information as input, and the Critic network receives updated state-action pairs, and then completes training updates to guide the agent to learn how to make optimal decisions in a variety of interaction scenarios.

[0034] Furthermore, the Actor network is used to accept current state information as input, and the Critic network accepts updated state-action pairs, and then completes training updates to guide the agent to learn how to make optimal decisions in a variety of interaction scenarios, including:

[0035] Using a dynamic graph attention mechanism to construct a state graph in the Actor network;

[0036] The Critic network is used to further combine the action information on the basis of the state graph of the Actor network, and the state graph of the Actor network is expanded into a state-action graph to learn the complex dependency between the state information and the action information.

[0037] Furthermore, after the Actor network is used to accept current state information as input, the Critic network accepts updated state-action pairs, and training updates are then completed to guide the agent to learn how to make optimal decisions in a variety of interaction scenarios, the method further includes:

[0038] Based on the MODPO reinforcement learning technology, the optimal decision is subjected to multi-objective direct preference optimization. If the agent performs the operation correctly, a positive reward is given; if the agent performs the operation incorrectly, a negative reward is given.

[0039] Compared with the prior art, the monitoring system and method for component Chinese medicine efficacy substance screening laboratory described in the present invention has the following advantages:

[0040] (1) The present invention describes a supervision system and method for a component Chinese medicine efficacy substance screening laboratory, wherein the system integrates a robot, an RPA module, a large language model decision module, a multi-agent calling module, and a ROS module to achieve standardized operation monitoring of laboratory personnel, identification of illegal items, real-time monitoring of the experimental process, problem warning and dynamic adjustment, thereby improving inspection efficiency and laboratory safety. The system has broad application prospects and is suitable for environments where various laboratories require strict safety supervision.

[0041] (2) The monitoring system and method for a component Chinese medicine efficacy substance screening laboratory described in the present invention forms an efficient operating system by utilizing the autonomous decision-making ability of a large language model combined with the automated execution mechanism of the ROS module. Through in-depth analysis of a large amount of laboratory data, the large language model can identify potential problems and risks in real time, issue early warnings in a timely manner, and provide data-driven, precise decision-making support. Subsequently, the ROS module generates specific operating instructions based on these decisions, and automatically dispatches robots in the laboratory to perform tasks, from daily maintenance to emergency response. The fully automated operation brought about by this fusion reduces the need for human intervention and significantly improves the laboratory's response speed and work efficiency.

[0042] (3) The monitoring system and method for a component Chinese medicine effective substance screening laboratory described in the present invention can continuously monitor the laboratory environment through real-time data collection, capture instantaneous changes and potential risks, and ensure the timeliness and effectiveness of safety and quality management. Multimodal data fusion includes sensor data, video monitoring, environmental parameters, etc., providing comprehensive state perception and improving data accuracy and reliability through cross-validation. This fusion not only provides comprehensive data integration by eliminating information islands, which helps to deeply understand the operating mechanism of complex systems, but also enhances the accuracy of problem prediction and decision-making by improving the accuracy and reliability of data.

[0043] (4) The monitoring system and method for the screening laboratory of component Chinese medicine efficacy substances described in the present invention achieves high efficiency in task execution through efficient task allocation and optimized resource utilization; at the same time, it ensures comprehensive and accurate data support through data integration and analysis functions, thereby improving the decision-making ability of the system. In addition, intelligent scheduling and path planning use MODPO reinforcement learning technology to achieve dynamic adaptation and optimization of scheduling strategies, thereby enhancing the flexibility of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings:

[0045] Figure 1 This is a schematic diagram of the structure of a supervision system for a component Chinese medicine efficacy substance screening laboratory according to the first embodiment of the present invention;

[0046] Figure 2 The present invention is a flowchart of a method for supervising a laboratory for screening effective substances of component traditional Chinese medicines as described in the second embodiment of the present invention. DETAILED DESCRIPTION

[0047] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for ease of description, only parts related to the present invention, rather than all structures, are shown in the accompanying drawings.

[0048] Embodiment 1

[0049] Figure 1 This is a schematic diagram of the structure of a supervision system for a component Chinese medicine efficacy substance screening laboratory as described in Example 1 of the present invention. This supervision system can be used for the supervision of a component Chinese medicine efficacy substance screening laboratory, realize the monitoring of laboratory personnel's standardized operations, identification of illegal items, real-time monitoring of the experimental process, problem warning and dynamic adjustment, improve inspection efficiency and laboratory safety, has a wide range of application prospects, and is suitable for various laboratories that require strict safety supervision. Figure 1 ,This supervision system includes robots, RPA modules, large language model decision modules, multi-agent calling modules, and ROS modules.

[0050] The above-mentioned robot is used to inspect and supervise the component Chinese medicine efficacy substance screening laboratory.

[0051] The above RPA module is used to collect experimental equipment status information, traditional Chinese medicine sample status information, laboratory environment parameter status information, and experimenter status information.

[0052] The above-mentioned large language model decision module is used to parse the management rules of the component Chinese medicine efficacy substance screening laboratory, and make decisions on the collection of experimental equipment status information, Chinese medicine sample status information, laboratory environmental parameter status information, and experimenter status information based on the management rules to generate task instructions.

[0053] The above-mentioned multi-agent calling module is used to receive the task instructions, and use the action- and state-based joint attention reinforcement learning technology to realize the priority scheduling management of the agent, and call the agent to perform the task; wherein the agent includes an RPA module, a large language model decision module, and a ROS module, and each of the agents is provided with an Actor-Critic network framework.

[0054] The above-mentioned ROS (Robot Operating System) module is used to drive the robot to work according to the instructions of the multi-agent calling module.

[0055] Preferably, the robot can adopt a Mecanum wheel mobile platform, and a mechanical arm is arranged on the Mecanum wheel mobile platform. Compared with other forms of robot chassis, the Mecanum wheel mobile platform has better omnidirectional mobility and can meet the precise movement needs of the robot in the laboratory, while the mechanical arm can be used to perform detailed operation tasks such as gripping and obstacle removal, which is conducive to improving the robot's adaptability to different tasks and improving the robot's inspection and supervision effect.

[0056] Preferably, the RPA module includes a data acquisition unit and an image acquisition unit. The data acquisition unit is used to collect laboratory management and related data indicators in real time, including the status of experimental equipment (such as a plate washer, incubation system, and liquid transfer system), the status information of Chinese medicine samples between materials, and laboratory temperature data. The image acquisition unit is used to collect laboratory management and related images in real time, including the status of experimental equipment (such as a plate washer, incubation system, and liquid transfer system), the status information of Chinese medicine samples between materials, monitoring of standardized operations by laboratory technicians, and identification of illegal items in the laboratory.

[0057] In actual application, in addition to connecting with the experimental equipment to directly collect data from the experimental equipment, the data acquisition unit may also include data acquisition sensors such as smoke sensors, gas leakage sensors, temperature sensors, and lidars to collect the status of laboratory equipment (such as plate washers, incubation systems, and liquid transfer systems) under different indicators, laboratory safety data (such as temperature and humidity), and information on the status of traditional Chinese medicine samples. The image acquisition unit may include image acquisition sensors such as (depth) cameras to collect the status of laboratory equipment (such as plate washers, incubation systems, and liquid transfer systems) under different indicators, standardized operations of laboratory technicians, and prohibited items in the laboratory. By using the data acquisition unit and the image acquisition unit to collect laboratory data, real-time data can be provided for the large language model decision module.

[0058] Exemplarily, the working steps of the RPA module are as follows:

[0059] (1) Multiple sensors, such as smoke sensors, gas leak sensors, temperature sensors, (depth) cameras, and lidar, are deployed on the robot system to achieve multimodal data collection;

[0060] (2) Each sensor transmits multimodal data to the large language model decision module.

[0061] Preferably, the large language model decision module includes a semantic parsing unit, a RAG indexing unit, a CNN image recognition unit, and an instruction scheduling unit. Among them, the semantic parsing unit is used to perform semantic parsing and intent recognition on the state information to obtain data information; wherein, when the data information is text information, the RAG indexing unit is used to index the information according to the management rules, obtain the index result and generate the task instruction according to the index result; when the data information is image information, the CNN image recognition unit is first used to identify and parse, and after obtaining the text information, the RAG indexing unit is used to index the information according to the management rules to obtain the index result and generate the task instruction according to the index result. Secondly, the instruction scheduling unit is used to uniformly schedule, send and receive information at different stages in the large language model decision module.

[0062] In the actual application process, the large language model decision module parses and understands the management rules of the component Chinese medicine efficacy substance screening laboratory, plans the inspection route according to the management rules and laboratory environment, implements behavioral decisions, and sends the task instructions to the multi-agent call module. Specifically, the semantic parsing unit can be used to perform semantic parsing and intent recognition on the experimenter's instructions. If the data information is image information such as the laboratory environment, it is sent to the CNN image recognition unit for recognition and processing, and then sent to the RAG index unit. If the data information is text, it is sent to the RAG index unit. Among them, the RAG index unit indexes the relevant documents according to the supervision task, generates automatic execution instructions, and returns the results to the instruction scheduling unit.

[0063] In addition, the instruction scheduling unit uniformly schedules, sends and receives information at different stages in the large language model decision module. Exemplarily, the instruction scheduling unit can implement data scheduling between various sub-units, and between the large language model decision module and the RPA module. For example, when the data information is text information, the instruction scheduling unit can be used to uniformly schedule all text information to the RAG index unit for processing, and when the data information is image information, the instruction scheduling unit can be used to schedule all image information to the CNN image recognition unit for recognition and processing, and then uniformly scheduled to the RAG index unit for processing.

[0064] Exemplarily, the working steps of the large language model decision module are as follows:

[0065] (1) The data collected by the RPA module is preprocessed (such as cleaning, standardization, etc.) and sent to the semantic parsing unit. If the data is an image, it is sent to the CNN image recognition unit (and then to the RAG indexing unit); if it is text data, it is sent to the RAG indexing unit.

[0066] (2) The RAG index unit constructs a local knowledge base of the "Laboratory Usage Management Rules" and a knowledge base of robot action instructions, and uses the search engine to retrieve relevant documents for the instructions in the predefined knowledge action base. Among them, the local knowledge base of the "Laboratory Usage Management Rules" is the usage management rules of the existing laboratory.

[0067] (3) The retrieved relevant documents are used as context, integrated with the original question and returned to the instruction scheduling unit.

[0068] (4) The large language model generates an automated task workflow based on specific steps and passes it to the multi-agent calling module.

[0069] Preferably, the multi-agent calling module includes a task allocation unit and a multi-objective optimization unit. The task allocation unit is used to receive the task instructions and allocate the task instructions according to the agents corresponding to each task instruction to obtain the sub-task instructions corresponding to each agent. Secondly, the multi-objective optimization unit is used to prioritize the sub-task instructions using the joint attention reinforcement learning technology based on action and state, and optimize the priority ranking based on the MODPO reinforcement learning technology, and then call the agent to perform the task according to its corresponding sub-task instruction according to the optimized priority ranking.

[0070] In actual application, the task allocation unit can receive and allocate task instructions from the large language model decision module. The multi-objective optimization unit can use the joint attention reinforcement learning technology of action and state to complete the priority management of intelligent agents; MODPO reinforcement learning technology optimizes multi-objective behaviors (path planning, laboratory equipment status detection, and Chinese medicine sample material delivery), thereby ensuring efficient execution of tasks and ensuring efficient communication and collaboration between intelligent agents, avoiding task conflicts and waste of resources, so that the multi-agent calling module can call intelligent agents to perform specific tasks according to the task instructions generated by the large language model.

[0071] By setting up a multi-agent calling module, using the multi-agent calling module to accept instructions, collaboratively calling multiple agents, and using multi-objective optimization technology to train the RPA module, large language model decision module, and ROS module, the accuracy of data analysis, behavioral decision-making, and machine driving can be improved, and the laboratory inspection path, equipment status detection, material transfer between equipment, and operation decisions can be optimized; at the same time, the ROS module can drive the robot to execute operation control instructions in real time, ensuring the efficiency and flexibility of the inspection process.

[0072] Exemplarily, the working steps of the multi-agent calling module are as follows:

[0073] (1) Receive task instructions from the large language model module, synchronize with the instructions generated by the large language model, and determine the state space.

[0074] (2) Use the joint attention reinforcement learning technology of action and state to complete the priority scheduling of intelligent agents; MODPO completes multi-objective optimization to ensure efficient task execution.

[0075] (3) Based on the task instructions generated by the large language model, call the agent to perform specific tasks.

[0076] Preferably, the ROS module includes a node topic unit, a service and action library unit, a parameter server unit, and a drive unit; the node topic unit is used to read the call instruction of the multi-objective optimization unit, and control the drive unit service and action library unit according to the call instruction, for synchronous request-response communication and management of long-term tasks; the parameter server unit is used to store the robot's parameters and data to ensure configuration consistency; the drive unit is used to drive the robot to work.

[0077] In actual application, node topic units can be created according to different types of requirements. In a laboratory environment, one node can be used to read specific instructions of the multi-objective optimization unit, and the other node is used to control the drive unit. The service and action library unit is divided into two parts: the service library and the action library. The service is used for synchronous request-response communication, and the action library is used for the management of long-term tasks. The parameter server unit is used to store the motion parameters of the robot arm, sensor calibration data, etc., to ensure that all nodes use consistent configurations. The drive unit can directly drive the robot's reduction motor (including the McLennan tire) and the robot arm. The reduction motor provides power to control the robot's moving direction and speed. The robot arm performs detailed operations such as gripping and obstacle removal.

[0078] Exemplarily, the working steps of the ROS module are as follows:

[0079] (1) Select appropriate robot modules to execute different task instructions based on the multi-agent module call.

[0080] (2) Call specific services and perform request-response communication.

[0081] (3) Drive the robot and schedule the corresponding supervisory agent to perform tasks according to the workflow.

[0082] This system can supervise the laboratory according to the following steps:

[0083] (1) Through the RPA module, real-time data collection is performed on the status of experimental equipment (such as microplate washers, incubation systems, and pipetting systems), the status of Chinese medicine samples in the material room, laboratory operations, identification of illegal items, laboratory environmental parameters, and usage supervision.

[0084] (2) Through the large language model decision module, the collected data is parsed and understood, supporting advanced data processing and analysis.

[0085] (3) Through the multi-agent calling module, it accepts instructions, realizes multi-objective optimization decision-making results, calls different agents according to actual problems, and improves the efficiency of supervision task execution.

[0086] (4) Through the ROS module, the robot is driven to automatically execute the supervision workflow and problem warning. For example, the ROS module can drive the robot to respond quickly and move to the problem area according to the call of the multi-agent module, adjust the equipment, transfer materials or check abnormal points, thereby improving the efficiency of the supervision task execution.

[0087] The supervision system for a component Chinese medicine efficacy substance screening laboratory described in this embodiment utilizes a robot, an RPA module, a large language model decision module, a multi-agent calling module, and a ROS module to form a system. Based on this system, multimodal data information is collected and analyzed, and specific supervision task instructions are generated through large-model intelligent decision-making. This can enhance the laboratory's real-time supervision and emergency response and rescue capabilities, improve the deficiencies of existing supervision schemes in data collection, intelligent decision-making, optimization of intelligent agent scheduling, ROS-driven robot technology, etc., and realize efficient, intelligent, and automated laboratory smart supervision.

[0088] In addition, by utilizing this system's automation and intelligent processing, we can reduce dependence on human resources, thereby reducing labor costs. It is also conducive to optimizing resource allocation and experimental processes, reducing resource waste, improving resource utilization, facilitating tracking and auditing, and enhancing compliance and transparency. It not only helps the direct operational benefits of the laboratory, but also promotes a profound change in the laboratory's scientific research, technological innovation, and lean production management methods in the long run.

[0089] Embodiment 2

[0090] Figure 2 This is a flow chart of a supervision method for a component Chinese medicine efficacy substance screening laboratory according to the second embodiment of the present invention. This supervision method can be used for laboratory supervision and collaboratively complete laboratory equipment status detection and Chinese medicine sample material delivery based on path planning. Figure 2 , this regulatory method specifically includes the following steps:

[0091] Step 201: Set up an Actor-Critic network framework in the RPA module, the large language model decision module, and the ROS module respectively.

[0092] Step 202: Arrange a robot with an RPA module in the laboratory, establish a data connection between the RPA module and the experimental equipment in the laboratory, and then use the robot with the RPA module to collect data and conduct inspection and supervision of the laboratory.

[0093] Step 203: Analyze the data collected by the RPA module using the large language model decision module, make decisions based on the management rules of the component Chinese medicine efficacy substance screening laboratory, and generate task instructions.

[0094] Step 204: Utilize the multi-agent calling module to receive the task instruction, and utilize the action- and state-based joint attention reinforcement learning technology to implement the priority scheduling management of the agent, and call the agent to perform the task.

[0095] Step 205: Utilize the ROS module to receive instructions from the multi-agent calling module, and drive the robot to inspect and supervise the laboratory according to the instructions from the multi-agent calling module.

[0096] Preferably, step 201, setting an Actor-Critic network framework in the RPA module, the large language model decision module, and the ROS module respectively, specifically includes the following steps:

[0097] Step 2011: Set up an Actor network in the RPA module, the large language model decision module, and the ROS module respectively; wherein the input part of the Actor network includes the status information of each agent.

[0098] Exemplarily, the input part of the Actor network includes the status information of all current intelligent agents (such as the system health status of the RPA module 1, the input data status of the large language model decision module 2, and the node status of the ROS module 4).

[0099] Step 2012: Set a Critic network in the RPA module, the large language model decision module, and the ROS module respectively; wherein the input part of the Critic network is a state-action pair.

[0100] Step 2013: Using the Actor network to accept current state information as input, the Critic network accepts updated state-action pairs, and then completes training updates to guide the agent to learn how to make optimal decisions in a variety of interaction scenarios.

[0101] In actual application, a dynamic graph attention mechanism can be used in the Actor network to construct a state graph. Then, the Critic network is used to further combine the action information on the basis of the state graph of the Actor network to expand the state graph of the Actor network into a State-Action Graph to learn the complex dependency between state information and action information.

[0102] This embodiment provides a new type of dynamic sharing process of actor and critic state information. The actor accepts the current state information as input, and the critic accepts the updated state-action pair. First, the actor uses a dynamic graph attention mechanism to build a state graph. Based on the actor's state graph, the critic further combines the action information and expands it into a state-action graph to learn the complex dependencies between states and actions.

[0103] Specifically, assuming that the state information is s (dimension is state_dim), the action strategy output of the Actor is π(a|s), which is a probability distribution (dimension is action_dim). s is the weight matrix of state feature transformation (dimension is (d, state_dim)), W a and b a are learnable parameters for generating action policies.

[0104] State feature transformation:

[0105] h s =ReLU(W s ·s+b s );

[0106] Dynamic graph adjacency matrix A generation:

[0107] where a T is a learnable weight vector;

[0108] Then update the state feature node (generate h s new ), and generate an action strategy:

[0109] π(a|s)=softmax(W a ·h s new +b a );

[0110] The Critic network further combines the action information based on the Actor's state graph and expands it into a State-Action Graph.

[0111] First, action feature extraction:

[0112] h a =ReLU(W a ′·π(a|s)+b a ′);

[0113] This generation process combines the dynamic features with the node features updated by the Actor to generate a larger extended graph A′:

[0114]

[0115] Then, the state-action feature node is updated according to this feature graph to complete the input of the Critic to achieve the purpose of the dynamic sharing process.

[0116] In actual application, the Actor inputs state information as s, and the Critic inputs the updated state-action pair (s, a). After that, the training update is completed to guide the agent to learn how to make the best decision in a variety of interaction scenarios.

[0117] Preferably, in step 2013, after the Actor network receives the current state information as input and the Critic network receives the updated state-action pair, and then the training update is completed to guide the agent to learn how to make the best decision in a variety of interaction scenarios, the method further includes:

[0118] Step 2014: Perform multi-objective direct preference optimization on the optimal decision based on MODPO reinforcement learning technology. If the agent performs the operation correctly, a positive reward is given; if the agent performs the operation incorrectly, a negative reward is given.

[0119] In actual application, for multi-objective optimization, if the agent performs the operation correctly, a positive reward is given; if the agent performs the operation incorrectly, a negative reward is given. Therefore, the optimization can be performed according to the following example:

[0120] Example 1: The multi-objective optimization of the robot's laboratory inspection through a specific route is divided into two parts. The large language model calculates (on the basis of completing the target instructions) that the inspection route that can detect the most laboratory equipment is the preferred output, and other routes are non-preferred outputs.

[0121] Example 2: The multi-objective optimization of robot equipment status detection is divided into two parts. Correct detection of the experimental equipment status (such as normal operation of the plate washer, incubation system, pipetting system, and good temperature of the incubation system) is the preferred output, otherwise it is the unpreferred output.

[0122] Example 3: The multi-objective optimization of equipment status detection is divided into two parts. Correct detection of the status information of traditional Chinese medicine samples (such as temperature and humidity) is the preferred output, and vice versa.

[0123] Specifically, the loss function formula is as follows:

[0124]

[0125] Among them, in the formula of the loss function: x is the input data, y ω ,y l There are two different outputs corresponding to input x, where y ω is the preferred output, y l is the unpreferred output. is θ ω The model predicts the probability distribution of y given input x. sft (y ω |x) is the probability distribution of the soft label model predicting y when input x. β is usually a hyperparameter used to balance the proportion of terms. ω k is the weight of the kth target. is the reward vector of all targets except the kth target.

[0126] Therefore, the inspection route is optimized by using a multi-objective optimization unit to continuously reduce the loss value. Then the multi-objective optimization unit calculates the gradient and updates the gradient, and finally performs iterative training. After the training is completed, the multi-objective optimization unit can call different intelligent agents according to the training results.

[0127] Specifically, combined with specific examples, the specific steps for supervising laboratories using this supervision method are as follows:

[0128] Step S1: The RPA module collects laboratory environmental parameters such as smoke concentration, leakage concentration of specific gases (such as natural gas, carbon monoxide), equipment temperature, whether the experimental equipment is operating normally, particulate matter concentration, etc., and detects multi-source data such as whether the experimenter has arrived at the designated location, whether prohibited items appear in the laboratory, and whether the experimenter operates in accordance with regulations. The radar sensor accurately determines the inspection location.

[0129] Step S2: The large language model decision module processes the data and instructions (such as location information) collected in step S1.

[0130] For example, the processed data and instructions may be sent to the semantic parsing unit, and the semantic parsing unit encodes the user supervision instruction representation as a sequence a i Then, during the inspection process, if the instruction is image data (such as standardized operation of the laboratory technician, prohibited items in the laboratory, etc.), the CNN image recognition unit uses the Resnet50 model to recognize it. The Resnet50 model extracts the image data and returns it to the RAG index unit. Finally, the RAG index unit retrieves relevant documents from the predefined knowledge base based on the execution instruction sequence and determines whether the relevant data is qualified (such as whether the inspection location is accurate).

[0131] In actual application, the input of the RAG index unit can be set as x, the document in the knowledge base can be set as D, and the goal of the retrieval module is to find the k documents most similar to x. The similarity calculation formula is as follows: Among them, s(x,d) represents the similarity between x and d. Then the index structure is fused and generated. Let the k documents retrieved be d1, d2, …, d k , the goal of the generation module is to generate text y. The context formula fusion formula is as follows: c = Concat(x, d1, d2, …, d k ); where Concat represents a vector concatenation operation.

[0132] Therefore, different data can infer different status information to generate different text instructions. At this time, the RAG index unit will comprehensively analyze the data to generate results. (For example, the text status information generated by the temperature data may be that the equipment temperature is too high, and the text status information generated by the CO concentration data may be that the CO gas cylinder is leaking. RAG comprehensive analysis and fusion may lead to a laboratory fire caused by a CO gas cylinder leak. The generated instruction is to alarm and shut down the CO gas cylinder.) Then the instruction evolution is generated (adding constraints, deepening, concretizing, increasing reasoning steps, and complicating the input) to generate specific instructions {step1, step2,…, step n} and returns the result to the instruction dispatching unit to further control the robot to perform inspection and path planning operations. Finally, the instruction dispatching unit sends the specific instructions to the multi-agent calling module.

[0133] Step S3: The multi-agent calling module accepts specific instructions, optimizes the agent behavior through multi-objective optimization, and calls different multi-agents to complete specific tasks.

[0134] Specifically, the task assignment unit receives specific instructions. The multi-objective optimization unit first completes the priority order of the agent (joint attention reinforcement learning technology of action and state). Specifically, the multi-objective optimization unit first determines the state space and action space to ensure the set of all considered variables of the agent and all actions that can be taken in a certain state.

[0135] Step S4: The ROS module selects different execution action instructions from the service and action library according to the instructions (calculating the inspection position to be reached) through the node topic publish-subscribe mode), and then establishes multiple nodes to call the drive unit.

[0136] In actual application, when the robot moves, the drive unit controls the direction of movement and controls the reduction motor (including the McLennan tire) to provide power, so that the robot can move along the track; when the robot needs to change the direction of movement, the rotation speed of the McLennan tires on both sides is coordinated to achieve steering, so that the robot can turn along the track. When the robot needs to use a robotic arm to transfer Chinese medicine samples, the drive unit controls the motors of the robotic arm to rotate at different angles, accurately clamps the Chinese medicine samples, and cooperates with the drive unit to place the Chinese medicine samples in the laboratory feeding area. Finally, the ROS module stores the motion parameters of the robotic arm, sensor calibration data, etc. through the parameter server to ensure that all nodes use consistent configurations.

[0137] The final inspection behavior of the robot is as follows: using the reduction motor (including the McLennan tire) to execute the inspection route, the RPA module collects the sample status data of laboratory equipment and materials in real time, and completes the delivery of traditional Chinese medicine samples based on the data-setting goals.

[0138] The present embodiment describes a supervision method for a component Chinese medicine efficacy substance screening laboratory, by setting an Actor-Critic network framework in the RPA module, the large language model decision module, and the ROS module respectively;

[0139] A robot with an RPA module is arranged in the laboratory, and after a data connection is established between the RPA module and the experimental equipment in the laboratory, the robot with the RPA module is used to collect data and conduct patrol inspections on the laboratory; the data collected by the RPA module is parsed using a large language model decision module, and decisions are made based on the management rules of the component Chinese medicine efficacy substance screening laboratory to generate task instructions; the task instructions are received using a multi-agent calling module, and the priority scheduling management of the agent is realized by using the action and state-based joint attention reinforcement learning technology, and the agent is called to perform the task; the instructions of the multi-agent calling module are received using a ROS module, and the robot is driven to conduct patrol inspections on the laboratory according to the instructions of the multi-agent calling module.

[0140] Compared with existing technologies, this supervision method not only helps to reduce inspection time and labor costs and improve overall work efficiency, but also uses multi-sensor and multi-modal data acquisition technology to enable robots to more comprehensively and accurately detect laboratory compliance management and collect various types of experimental process data, thereby quickly discovering potential problems or anomalies and achieving convenient and efficient execution of supervision tasks. It has broad application prospects and significant social benefits.

[0141] Note that the above are only preferred embodiments of the present invention and the technical principles used. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A supervision system for a screening laboratory of effective substances of component traditional Chinese medicines, characterized by: Including robots, RPA modules, large language model decision modules, multi-agent calling modules, and ROS modules; Robots are used to inspect and supervise the screening laboratories for the efficacy of component Chinese medicines; RPA module is used to collect experimental equipment status information, Chinese medicine sample status information, laboratory environmental parameter status information, and experimental personnel status information; A large language model decision module is used to analyze the management rules of the component Chinese medicine efficacy substance screening laboratory, and make decisions on the acquisition of experimental equipment status information, Chinese medicine sample status information, laboratory environmental parameter status information, and experimental personnel status information based on the management rules to generate task instructions; A multi-agent calling module is used to receive the task instructions, and use the action- and state-based joint attention reinforcement learning technology to implement the priority scheduling management of the agent, and call the agent to perform the task; wherein the agent includes an RPA module, a large language model decision module, and a ROS module, and each of the agents is provided with an Actor-Critic network framework; The ROS module is used to drive the robot to work according to the instructions of the multi-agent calling module.

2. A monitoring system for a component Chinese medicine efficacy substance screening laboratory according to claim 1, characterized in that: The RPA module includes a data acquisition unit and an image acquisition unit.

3. A monitoring system for a component Chinese medicine efficacy substance screening laboratory according to claim 1, characterized in that: The large language model decision module includes a semantic parsing unit, a RAG indexing unit, a CNN image recognition unit, and an instruction scheduling unit; A semantic parsing unit is used to perform semantic parsing and intent recognition on the status information to obtain data information; wherein, when the data information is text information, the RAG indexing unit is used to index the information according to the management rules to obtain an index result and generate a task instruction according to the index result; when the data information is image information, the CNN image recognition unit is used to identify and parse the information first to obtain text information, and then the RAG indexing unit is used to index the information according to the management rules to obtain an index result and generate a task instruction according to the index result; The instruction scheduling unit is used to uniformly schedule, send and receive information at different stages in the large language model decision module.

4. The monitoring system for component Chinese medicine efficacy substance screening laboratory according to claim 1 is characterized by: The multi-agent calling module includes a task allocation unit and a multi-objective optimization unit; A task allocation unit, used for receiving the task instructions and allocating the task instructions according to the agents corresponding to the task instructions to obtain subtask instructions corresponding to the agents; A multi-objective optimization unit is used to prioritize the sub-task instructions using the action- and state-based joint attention reinforcement learning technology, and after optimizing the priority ranking based on the MODPO reinforcement learning technology, the intelligent agent is called to perform tasks according to its corresponding sub-task instructions based on the optimized priority ranking.

5. The monitoring system for component Chinese medicine efficacy substance screening laboratory according to claim 1 is characterized by: The ROS module includes a node topic unit, a service and action library unit, a parameter server unit, and a driver unit; A node topic unit is used to read the calling instruction of the multi-objective optimization unit and control the driving unit according to the calling instruction Service and action library units for synchronous request-reply communication and management of long-running tasks; Parameter server unit, used to store the robot's parameters and data to ensure configuration consistency; The driving unit is used to drive the robot to work.

6. A supervision method using the supervision system for component Chinese medicine efficacy substance screening laboratory according to any one of claims 1 to 5, characterized in that: include: An Actor-Critic network framework is set up in the RPA module, the large language model decision module, and the ROS module respectively; A robot with an RPA module is arranged in the laboratory, and after the RPA module is connected to the experimental equipment in the laboratory, the robot with the RPA module is used to collect data and conduct inspection and supervision of the laboratory; The data collected by the RPA module is parsed using a large language model decision module, and decisions are made based on the management rules of the component Chinese medicine efficacy substance screening laboratory to generate task instructions; Using a multi-agent calling module to receive the task instructions, and using the action- and state-based joint attention reinforcement learning technology to implement the priority scheduling management of the agent, calling the agent to perform the task; The ROS module is used to receive instructions from the multi-agent calling module, and the robot is driven to inspect and supervise the laboratory according to the instructions from the multi-agent calling module.

7. The method according to claim 6, characterized in that An Actor-Critic network framework is set in each of the RPA module, the large language model decision module, and the ROS module, including: An Actor network is respectively set in the RPA module, the large language model decision module, and the ROS module; wherein the input part of the Actor network includes the state information of each intelligent agent; A Critic network is respectively set in the RPA module, the large language model decision module, and the ROS module; wherein the input part of the Critic network is a state-action pair; The Actor network receives current state information as input, and the Critic network receives updated state-action pairs, and then completes training updates to guide the agent to learn how to make optimal decisions in a variety of interaction scenarios.

8. The method according to claim 7, characterized in that The Actor network receives the current state information as input, and the Critic network receives the updated state-action pair, and then completes the training update to guide the agent to learn how to make the best decision in a variety of interaction scenarios, including: Using a dynamic graph attention mechanism to construct a state graph in the Actor network; The Critic network is used to further combine the action information on the basis of the state graph of the Actor network, and the state graph of the Actor network is expanded into a state-action graph to learn the complex dependency between the state information and the action information.

9. The method according to claim 7 or 8, characterized in that: After the Actor network is used to accept the current state information as input, the Critic network accepts the updated state-action pair, and then the training update is completed to guide the agent to learn how to make the best decision in a variety of interaction scenarios, the method further includes: Based on the MODPO reinforcement learning technology, the optimal decision is subjected to multi-objective direct preference optimization. If the agent performs the operation correctly, a positive reward is given; if the agent performs the operation incorrectly, a negative reward is given.

Images