An artificial intelligence-based experimental process state data automatic processing method

Abstract

The application provides an artificial intelligence-based automatic processing method for experimental process state data, and relates to the technical field of artificial intelligence. The method comprises the following steps: performing feature extraction and fusion on multi-source heterogeneous data acquired in an experimental process to output a real-time state data set reflecting the experimental progress; performing dynamic priority scheduling and resource allocation on a plurality of concurrent decision-making calculation tasks according to the real-time state data set and experimental process characteristics to obtain an allocation result; and executing the scheduled tasks according to the allocation result to output a series of decision-making instructions suitable for the current experimental process. The application can provide experimental stage identification and decision-making support based on the rich knowledge base and interdisciplinary knowledge of artificial intelligence, improve the automation level of experimental intervention and experimental safety, reduce the risk of personnel participation, improve the accuracy of experimental records and the depth of experimental result analysis, and significantly improve the experimental efficiency, experimental safety and the probability of discovering new experimental paths and new experimental results.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an automated processing method for experimental process state data based on artificial intelligence. Background Technology

[0002] In scientific research experiments, industrial trials, and process verification, the experimental process usually has the following characteristics: the state of the experimental object changes dynamically over time, the experimental path depends on real-time observation results, and intervention operations are often required based on the phased results during the experiment.

[0003] Current experimental methods primarily rely on intermittent observation and manual decision-making by experimenters, resulting in the following technical drawbacks: Discontinuous perception of the experimental process; experimenters typically observe the experimental status on a minute-by-minute or even lower frequency, making it difficult to capture short-term, low-probability, but crucial experimental changes; Experimental decision-making is highly dependent on personal experience; the selection of experimental paths, intervention timing, and intervention parameters are mainly based on the experimenter's personal knowledge and experience, leading to significant differences between individuals and insufficient stability and repeatability of experimental results; Low degree of automation in experimental intervention; most experiments still require manual execution of key operations, making truly unattended experiments difficult; High experimental safety risks; in high-risk experiments, experimenters need to participate directly or closely, posing safety hazards if the experiment goes out of control; Limited granularity of experimental recording; traditional experimental recording methods struggle to achieve high temporal resolution for the entire process, hindering experimental review, comparative analysis, and the discovery of new patterns. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an automated processing method for experimental process status data based on artificial intelligence, so as to solve the problems of discontinuous monitoring of experimental process, reliance on human experience for experimental decision-making, low degree of automation of experimental intervention and insufficient experimental safety in the prior art, thereby significantly improving experimental efficiency, experimental safety and the probability of discovering new experimental paths and new experimental results.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, an automated processing method for experimental process state data based on artificial intelligence, the method comprising: Step 1: Extract and fuse features from the multi-source heterogeneous data acquired during the experiment to output a real-time status dataset that reflects the experimental progress. Step 2: Based on the real-time status dataset and experimental process characteristics, perform dynamic priority scheduling and resource allocation for multiple concurrent decision computing tasks to obtain allocation results; according to the allocation results, execute the scheduled tasks to output a series of decision instructions adapted to the current experimental process. Step 3: Assign version identifiers to the real-time status dataset and bind the series of decision instructions to the status versions that triggered their generation; after the decision instructions are executed, collect execution feedback data and associate and store it with the corresponding decision instructions and status versions to form a decision-execution link record; Step 4: Based on the decision-execution link record, dynamically load, switch, or update the strategy during the experiment.

[0006] Furthermore, feature extraction and fusion are performed on the multi-source heterogeneous data acquired during the experiment to output a real-time status dataset reflecting the experimental progress, including: Step 1.1: Acquire multimodal experimental process data from different experimental devices or information sources in real time, and classify the experimental process data according to the data source and data type to obtain the classification results; Step 1.2: Based on the classification results, feature extraction processing is performed on different types of experimental process data to obtain processing results; the processing results are mapped to a unified state feature space to obtain the mapped state features. Step 1.3: Based on the mapped state features, perform temporal alignment and feature fusion to construct the experimental state calculation; Step 1.4: When a new experimental process data state change is received, the experimental state calculation representation is incrementally updated for the partial state features corresponding to the new data to form the updated real-time state dataset.

[0007] Furthermore, based on the real-time state dataset and experimental process characteristics, multiple concurrent decision computation tasks are dynamically prioritized and resource-allocated to obtain allocation results, including: Step 2.1: Abstract the intelligent analysis process of state analysis, parameter optimization, or safety assessment during the experiment into multiple independent decision computing tasks; configure task attributes for each decision computing task. Step 2.2: Obtain experimental process feature information related to the current experiment, and dynamically generate a scheduling strategy based on the task attributes and the process feature information; Step 2.3: According to the scheduling strategy, sort the multiple decision calculation tasks to be executed to obtain the task execution sequence and resource allocation results.

[0008] Furthermore, based on the allocation results, the scheduled tasks are executed to output a series of decision instructions adapted to the current experimental progress, including: Step 2.4: Based on the task execution sequence and resource allocation results, instantiate the decision calculation tasks to be executed in sequence; use the real-time status dataset as input and bind it to the corresponding task instance; Step 2.5: Run the task instance on the allocated computing resources, execute state analysis, path planning or parameter optimization calculation logic, and generate preliminary decision results; Step 2.6: The preliminary decision results are formatted and synthesized to transform them into specific control instructions or parameter adjustment instructions that drive the experimental equipment to execute, and output as a series of decision instructions; Step 2.7: During task execution, monitor the usage of computing resources and the task execution status; if the task execution times out, fails, or produces abnormal results, trigger the exception handling process and update the subsequent scheduling strategy.

[0009] Furthermore, a version identifier is assigned to the real-time state dataset, and a series of decision instructions are bound to the state version that triggered their generation, including: Step 3.1: Whenever the construction and update of the real-time state dataset is completed, a unique and incremental state version identifier is generated and assigned to the current dataset to obtain a series of decision instructions; Step 3.2: When issuing a series of decision instructions, record the status version identifier corresponding to the real-time status dataset on which the decision instruction is based for each decision instruction, so as to establish an association between the decision instruction and its source status version. Step 3.3: Encapsulate the decision instruction and its associated status version identifier to form an execution decision unit with version traceability information.

[0010] Furthermore, after the decision instruction is executed, execution feedback data is collected and associated with the corresponding decision instruction and status version for storage, forming a decision-execution link record, including: Step 3.4: After the decision instruction in the decision unit to be executed is sent to the experimental device for execution, capture the execution feedback data returned by the device. The execution feedback data includes at least one or more of the following: execution success flag, execution result data, or execution exception information. Step 3.5: Extract the encapsulated decision instructions and status version identifiers from the decision unit to be executed; associate the execution feedback data with the extracted decision instructions and status version identifiers to form a decision-execution link record containing complete information of status version, decision instructions, and execution feedback.

[0011] Furthermore, in step 4, based on the decision-execution link record, strategies are dynamically loaded, switched, or updated during the experiment to automate the processing of experimental process status data, including: Step 4.1: Obtain the stored decision-execution link record, analyze the execution feedback data and associated experimental status; based on preset evaluation rules or performance indicators, determine whether the strategy currently supporting the decision computing task needs to be updated, replaced or rolled back, and generate a strategy change decision. Step 4.2: Based on the strategy change decision, dynamically load the configuration file and calculation model of the target strategy from the decoupled management strategy library; without interrupting the core experimental process, instantiate the loaded target strategy into an executable strategy object; Step 4.3: Control the timing of the activation of the strategy object so that it replaces the original old strategy object at the beginning of the next experimental decision cycle or during the gap when no critical decision tasks are being executed, so as to complete the hot update and switching of the strategy.

[0012] Secondly, an automated processing system for experimental process state data based on artificial intelligence includes: The multimodal perception module is used to extract and fuse features from multi-source heterogeneous data acquired during the experiment, so as to output a real-time status dataset that reflects the experimental progress. The artificial intelligence analysis and decision-making module is used to dynamically prioritize and allocate resources for multiple concurrent decision computing tasks based on the real-time status dataset and experimental process characteristics to obtain allocation results; and to execute the scheduled tasks according to the allocation results to output a series of decision instructions adapted to the current experimental process. The experiment execution module is used to assign version identifiers to the real-time status dataset and bind a series of decision instructions to the status versions that trigger their generation. After the decision instructions are executed, execution feedback data is collected and associated with the corresponding decision instructions and status versions to form a decision-execution link record. The experimental data processing module is used to dynamically load, switch, or update strategies during the experiment based on the decision-execution link record, so as to realize the automated processing of experimental process status data.

[0013] Thirdly, a computing device, comprising: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0014] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0015] The above-described solution of the present invention has at least the following beneficial effects: This invention enables high-frequency, continuous sensing of the experimental process, provides AI-based experimental stage identification and decision support, significantly improves the automation level of experimental intervention, enhances experimental safety and reduces the risk of human involvement, improves the accuracy of experimental recording and the depth of experimental result analysis, enhances the reliability of experimental data, and reduces the possibility of experimental data being affected by various interference factors. It also accelerates the discovery of new experimental paths, new experimental results, and expected experimental results.

[0016] This invention transforms multi-source heterogeneous data generated during experiments into a unified experimental state computation representation, and dynamically maintains the experimental state during the experiment. It maps experimental process data from different experimental devices, different acquisition frequencies, and different data structures into a unified experimental state computation representation. This enables unified modeling of experimental process data from different experimental devices and information sources, improving the completeness and consistency of experimental state descriptions, supporting dynamic updates of experimental states, reducing unnecessary full-scale computations, improving system response speed and computational efficiency, providing stable and continuous state inputs for intelligent experimental analysis, decision generation, and automatic execution, enhancing the overall performance of the intelligent experimental system, facilitating retrospective analysis of the experimental process, state comparison, and cross-stage processing, allowing for complete reproduction based on historical experimental data, and enhancing the maintainability and scalability of the experimental system.

[0017] In the process of automated experimentation, this invention provides unified management, dynamic scheduling, and resource allocation for different types of decision computing tasks. The experimental intelligent system treats the intelligent analysis and decision processing generated during the experiment as multiple independent computing tasks and schedules and manages these tasks based on at least the following information: it can dynamically adjust the execution priority of decision computing tasks according to changes in the experimental stage or experimental state, improving the real-time performance of key decisions; it avoids disorderly competition for computing resources among multiple decision computing tasks, improving the overall computing efficiency of the experimental intelligent system; it reduces the risk of decision results becoming invalid due to computing delays, improving the reliability of experimental process control; and it helps the experimental intelligent system maintain stable and scalable operation in complex, multi-task scenarios.

[0018] This invention provides unified management and consistency maintenance of the correspondence between experimental state information, intelligent decision results, and experimental execution receipts during automated experiments. Addressing the multiple decision-making, execution, and feedback loops inherent in experiments, it proposes a consistency maintenance mechanism for the decision-execution-feedback computation chain. This mechanism manages the association between experimental state information, decision result versions, and execution receipts during the experiment, avoiding decision deviations caused by data asynchrony or state mismatches. This improves the stability and reliability of the experimental intelligent system in continuous or concurrent operation scenarios. It ensures consistency between computational state, decision version, and execution receipts across multiple decision-making and execution loops; avoids erroneous decisions due to asynchronous experimental state updates during closed-loop experimental control; ensures accurate traceability of each execution receipt to the corresponding decision result and experimental state in parallel or asynchronous execution scenarios; improves the stability and reliability of the experimental intelligent system under long-term operation, multi-round iteration, or complex scheduling conditions; and provides a consistent data foundation for backtracking analysis, problem localization, and system optimization during the experimental process.

[0019] This invention achieves decoupling management of decision-making strategies or models from core computational processes in an experimental intelligent system during automated experiments. The experimental intelligent system includes at least a core computational process for performing experimental state perception, decision generation, and automatic execution, as well as strategy or model units for providing experimental decision-making capabilities. These strategy or model units are managed and invoked in the following ways: they enable updates or replacements of strategies or models within the experimental intelligent system without interrupting the experimental process, ensuring continuous experimental operation; they improve the adaptability of the experimental intelligent system to different experimental types or stages; they reduce risks during system upgrades and maintenance, enhancing system stability; and they facilitate the long-term evolution and functional expansion of the experimental intelligent system.

[0020] This invention allows historical experimental records to be stored in a database, which can be shared via local area network or the internet. Experiments with high similarity can be viewed based on historical records, and if necessary, can be automatically reproduced, with improvements made to certain steps during the reproduction process. This invention provides unified management, refined indexing, and backtracking analysis of various computational events generated during automated experiments. The experimental intelligent system abstracts key computational behaviors during the experiment into computational events and processes these events through at least the following methods, enabling accurate recording and differentiation of densely generated computational events during the experiment, improving the traceability of the experimental process. In the event of experimental anomalies or system failures, it can quickly locate the corresponding computational event or computational stage. It facilitates the analysis, debugging, and optimization of the experimental intelligent system's operational behavior, and improves the reliability and maintainability of the experimental intelligent system in complex, long-term operating scenarios. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating an automated processing method for experimental process state data based on artificial intelligence, provided by an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of an automated processing system for experimental process status data based on artificial intelligence, provided by an embodiment of the present invention. Detailed Implementation

[0023] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, embodiments are provided to provide a more thorough understanding of the present disclosure and to fully convey its scope to those skilled in the art.

[0024] like Figure 1 As shown, an embodiment of the present invention proposes an automated processing method for experimental process state data based on artificial intelligence, the method comprising the following steps: Step 1: Extract and fuse features from the multi-source heterogeneous data acquired during the experiment to output a real-time status dataset that reflects the experimental progress. Step 2: Based on the real-time status dataset and experimental process characteristics, perform dynamic priority scheduling and resource allocation for multiple concurrent decision computing tasks to obtain allocation results; according to the allocation results, execute the scheduled tasks to output a series of decision instructions adapted to the current experimental process. Step 3: Assign version identifiers to the real-time status dataset and bind the series of decision instructions to the status versions that triggered their generation; after the decision instructions are executed, collect execution feedback data and associate and store it with the corresponding decision instructions and status versions to form a decision-execution link record; Step 4: Based on the decision-execution link record, dynamically load, switch, or update the strategy during the experiment.

[0025] In a preferred embodiment of the present invention, feature extraction and fusion are performed on the multi-source heterogeneous data acquired during the experiment to output a real-time status dataset reflecting the experimental progress, including: Step 1.1: Acquire multimodal experimental process data from different experimental devices or information sources in real time, and classify the experimental process data according to the data source and data type to obtain classification results. Among them, the status data includes at least image or video data of the experimental object, microscopic imaging data, sensor data of the experimental environment or experimental object, and information on structural, morphological, color, position or state changes during the experimental process. Specifically, before the experiment begins, the experimenter needs to input or inform the intelligent automated experimental system of all experimental plans or objectives, such as experimental purpose, experimental background, and expected experimental effects, in the form of unstructured or structured data. The intelligent automated experimental system will sort out and output one, multiple, or control experimental steps and required materials based on the information provided by the experimenter. The experimenter prepares the required materials according to the output results and places them in the designated location. During preparation, materials can be prepared manually according to the innovative experimental plan output by artificial intelligence, or the automated experimental platform can automatically select the required materials from the material library according to the current experimental plan. After the materials are prepared and confirmed by AI to be sufficient, the experimenter confirms the start of the experiment. During the experiment, the experimental process sensing module collected multimodal experimental process data in real time from different experimental equipment and information sources. The data covered multiple dimensions, including salt weighing readings output by the electronic balance, solution acidity and alkalinity data detected by the pH meter, environmental and solution temperature data captured by the temperature probe, solution conductivity values ​​measured by the conductivity sensor, liquid level data and weighing change data reflecting the amount of liquid in the container, force and torque data at the end of the robotic arm, workstation in-place detection data to check whether the experimental container was placed in place, on / off status data of the experimental equipment fence door lock, and gas concentration data collected by the volatile organic compound (VOC) sensor. After collection, the data was classified according to its source and type: by source, it can be divided into electronic balance data source, pH meter data source, temperature probe data source, robotic arm data source, etc.; by type, it can be divided into weight data, acidity and alkalinity data, temperature data, mechanical data, status switch data, and gas concentration data, etc., ultimately yielding a clear classification result.

[0026] Step 1.2: Based on the classification results, feature extraction is performed on different types of experimental process data to obtain processing results; the processing results are then mapped to a unified state feature space to obtain the mapped state features. Specifically, this includes: feature extraction based on the classification results; non-visual data feature extraction: extracting features such as weight, pH deviation, temperature trend, and mechanical parameters according to the original logic, keeping them unchanged; visual data feature extraction: for the visual image stream, multi-dimensional process feature extraction and quantization are achieved through the optical flow pixel tidal change algorithm, outputting structured state quantities. Subject Boundary Displacement and Deformation Trend Judgment: The subject's boundary band is considered as the most stable observation area for motion information. When the subject moves, its boundary band pixels exhibit an orderly translation phenomenon in the direction of motion. This phenomenon is called pixel tidal displacement in this algorithm. The algorithm outputs the motion trend of the subject relative to the acquisition device by performing directional consistency and deformation decomposition on the optical flow and displacement vector set of the boundary band. In experimental scenarios such as shaking and stirring, in addition to removing the net motion component of the subject relative to the acquisition device, the pixels of the subject boundary band are also superimposed with cyclic tidal components caused by periodic disturbances, such as reciprocating oscillation, periodic oscillation, and periodic deformation caused by vortices. After calculating the boundary band optical flow displacement set, this algorithm performs periodic detection and decomposition on the displacement sequence: identifies and models the cyclic component, and then removes the cyclic component from the total displacement, outputting a net motion trend and trend confidence that better meets the actual experimental requirements.

[0027] Example: During the stirring process, the algorithm identifies the periodic vortex deformation (cyclic component) of the liquid boundary zone, and outputs the net displacement of the container caused by the slight shaking of the robotic arm after removing it, such as an X-axis offset of 0.2mm with a confidence level of 0.92.

[0028] Color change trend judgment: The color distribution drift of the contents of the main area, such as test tubes and other experimental carriers (mean, quantile and histogram differences) and the color gradient change of the boundary zone are analyzed. The color depth boundary line (gradient edge) is extracted in the observation area, and the edge density, directionality, contrast change and other parameters are calculated as texture quantification indicators. When the knowledge base provides stage prior or color allowable range, the degree of deviation can be output as an anomaly candidate.

[0029] Example: During pH adjustment, the algorithm detected a shift in solution color from colorless to slightly yellow. The RGB mean deviations were calculated as ΔR=5, ΔG=3, and ΔB=1. The deviation was slight and was judged as a normal reaction trend.

[0030] Texture change trend judgment: Extract the boundary line of color depth (gradient edge) in the observation area, and calculate the edge density, directionality, contrast change, etc. as texture quantification indicators. When the knowledge base provides stage prior or texture allowable range, the degree of deviation can be output as anomaly candidate.

[0031] Example: During the salt dissolution process, the algorithm detected that the texture edge density decreased from 28 points / 100 pixels to 5 points / 100 pixels, and the contrast change rate was -0.7. It was determined that the dissolution process was normal and the texture tended to be uniform.

[0032] Bubble generation judgment: The default method uses visual candidate + physical consistency judgment. Visual candidate: Identifies near-circular and elliptical closed structures or highlighted edge structures within the liquid area, and outputs a candidate set including position, radius, roundness, and edge intensity; Physical consistency verification filters false detections and confirms events based on physical laws, including but not limited to consistency between the buoyancy direction and the opposite direction of gravity, consistency between the trend of size and buoyancy speed, temporal consistency of merging and breaking events (approaching → merging → area change; reaching the liquid surface → disappearing), and reasonable coupling with the main motion trend and local flow field (deflection is allowed but long-term anti-gravity drift is not allowed), etc.

[0033] Example: During the reaction, the algorithm identified 12 bubbles that met the physical characteristics: roundness ≥ 0.8, rising speed 0.5 mm / frame, and bubble density 0.3 bubbles / frame.

[0034] Sedimentation formation judgment: Based on the subject naming to limit the observation area, low brightness particle texture candidates are extracted. Combined with physical consistency verification such as gravity sedimentation, low flicker and bottom-hanging stability, the area growth, aggregation and sedimentation trend indicators are output. Under specific experimental backgrounds, bubble formation features need to be excluded and gravity direction differences should be given priority.

[0035] Example: Simultaneously, low-brightness particles were detected at the bottom of the container, verifying that their settling direction was consistent with gravity, and the output sedimentation area growth rate was 0.02 mm. 2 / frame, determined to be slight sedimentation.

[0036] Feature mapping: The structured visual features output by the optical flow pixel tidal change algorithm, such as net motion trend, color deviation level, texture edge density, bubble density, sediment growth rate, etc., are uniformly mapped with non-visual data features, such as weight and pH value, to a standardized state feature space to form a state feature vector with a unified dimension.

[0037] Step 1.3: Based on the mapped state features, perform time alignment and feature fusion to construct the experimental state calculation. Specifically, this includes: First, perform time alignment based on the mapped state features. Due to differences in data acquisition frequencies of different sensors and devices (e.g., pH meter acquires data once per second, temperature probe acquires data once every 2 seconds, and electronic balance acquires data once every 0.5 seconds), using the highest acquisition frequency among all data as a benchmark, interpolation is used to supplement missing time node data for data with lower acquisition frequencies, and data with higher acquisition frequencies are synchronized and matched according to a unified timestamp to ensure that all state features correspond one-to-one on the same time axis, eliminating time sequence differences. Subsequently, feature fusion is performed. Combining the current stage of the experiment, various state features at the same time node are integrated. For example, in the weighing and feeding stage, weight-related features, robotic arm mechanical features, and workstation in-situ status features are fused; in the pH closed-loop calibration stage, pH-related features, temperature features, and stirring state-related features are fused to construct a comprehensive and accurate experimental state representation reflecting the current experimental state.

[0038] Step 1.4: When a new experimental process data state change is received, the experimental state calculation representation is incrementally updated for the partial state features corresponding to the new data to form the updated real-time state dataset. Specifically, this includes: During the experiment, when a new experimental process data state change is received, the data type and corresponding state feature category of the new data are first determined. For example, if the new data is the weight data generated after an electronic balance completes a salt weighing, the corresponding state features are the real-time weighing value and the cumulative weighing amount in the weight category; if the new data is the latest pH value collected by the pH meter, the corresponding state features are the real-time pH value and pH deviation in the acidity / alkalinity category. After determining the corresponding features, only these partial state features are updated, without recalculating all state features. For example, the cumulative weighing amount in the weight category is updated to the latest value, and the real-time pH value and deviation from the target value in the acidity / alkalinity category are updated. Based on the updated partial state features, the values ​​of the corresponding dimensions in the entire experimental state representation are adjusted to maintain the integrity and timeliness of the experimental state representation, ultimately forming the updated real-time state dataset.

[0039] In a preferred embodiment of the present invention, based on the real-time state dataset and experimental process characteristics, dynamic priority scheduling and resource allocation are performed on multiple concurrent decision computation tasks to obtain allocation results, including: Step 2.1: Abstract the intelligent analysis process of state analysis, parameter optimization, or safety assessment during the experiment into multiple independent decision computing tasks; configure task attributes for each decision computing task, specifically including: in the entire process of automated PBS buffer preparation experiment, decompose and abstract various intelligent analysis processes into multiple independent decision computing tasks one by one, ensuring that the functional boundaries of each task are clear and can be executed independently; State analysis-based decision-making tasks specifically include: stability analysis of electronic balance weighing data to determine whether fluctuations in multiple balance readings are within a preset reasonable range; deviation analysis of pH meter readings from the target value of 7.40, correcting deviations based on temperature data and analyzing deviation trends; temperature change trend analysis to monitor real-time changes in environmental and solution temperatures to ensure they meet experimental requirements; normality analysis of robotic arm end-effector force and torque, comparing real-time data with preset normal ranges to determine if there is a risk of collision or jamming; conductivity value pass / fail assessment to check if conductivity data is within the laboratory's preset pass / fail range; workstation presence detection to verify if experimental containers are accurately placed at designated workstations; fence lock closure monitoring to confirm if equipment fence locks are securely closed; VOC sensor concentration change analysis to track real-time fluctuations in volatile gas concentrations; and liquid level and weighing change matching analysis to verify if liquid level and weighing data match the expected correlation of solution volume changes.

[0040] Parameter optimization decision-making calculation tasks specifically include: calculating the amount of acid and alkali added during the pH closed-loop calibration phase to determine the required acid and alkali volumes for this round of adjustment; optimizing the dosing rate to adjust the acid and alkali dosing speed based on the solution volume and stirring status; optimizing the stirring speed to dynamically adjust the stirring motor speed based on the dissolution status or pH adjustment requirements; setting the mixing waiting time after dosing to determine the required mixing time based on the solution volume and stirring power; adjusting the pH retest frequency to optimize the retest interval based on the magnitude and trend of pH deviation; optimizing the stirring duration during the dissolution phase to adjust the stirring duration based on temperature data and salt dissolution progress; optimizing the robotic arm's pick-and-place path to shorten the action time for picking up and placing weighing containers and reagent bottles and avoid collisions; and allocating deionized water dosing volume to plan the dosing ratio for the dissolution and volume-fixing phases.

[0041] Safety assessment-related decision-making tasks specifically include: VOC sensor concentration exceeding limits assessment task, determining whether the gas concentration exceeds the safety threshold; solution overflow risk assessment task, predicting the possibility of overflow based on liquid level data and container capacity; robotic arm collision or jamming risk detection task, identifying safety hazards in real time based on force and torque data; temperature exceeding limits assessment task, checking whether the solution or ambient temperature exceeds the upper limit set in the experiment; equipment calibration status validity determination task, verifying whether the calibration of equipment such as pH meters, electronic balances, and temperature probes is within the validity period and without drift; emergency stop button status monitoring task, confirming whether the emergency stop function is normal and has not been accidentally triggered; and waste liquid tank level safety assessment task, determining whether the waste liquid tank is about to overflow and needs to be dealt with in a timely manner.

[0042] Each abstracted decision computation task is configured with complete task attributes. The task type attribute clearly indicates whether the task belongs to the state analysis, parameter optimization, or safety assessment category. The task urgency attribute is set according to the task's impact on experimental safety and progress. Safety assessment tasks are all set to the highest urgency, parameter optimization tasks to medium urgency, state analysis tasks involving the determination of key experimental indicators to medium-high urgency, and other routine state analysis tasks to normal urgency. The task computational resource requirement attribute is configured according to the complexity of the task's computational logic. Parameter optimization tasks, due to their need to call AI response models for complex calculations, are configured with more high-performance GPU resources. Safety assessment and state analysis tasks... For tasks involving trend analysis or multi-data comparison, medium computing resources are required; for simple state detection tasks, basic computing resources are sufficient. The task execution time limit attribute specifies the maximum time the task must complete: safety assessment tasks should have an execution time limit of 1 second or less, parameter optimization tasks 3 seconds or less, and routine state analysis tasks 5 seconds or less. The task failure impact level attribute categorizes tasks based on the severity of the consequences of failure: safety assessment tasks may result in equipment damage or safety incidents (highest impact level); parameter optimization tasks may result in experimental result deviations or process interruptions (medium impact level); and routine state analysis tasks only affect data record integrity (lowest impact level).

[0043] Step 2.2: Obtain experimental process feature information related to the current experiment. Based on the task attributes and the process feature information, dynamically generate a scheduling strategy. Specifically, this includes: comprehensively obtaining experimental process feature information related to the current PBS buffer preparation experiment through the experimental process perception module, task distribution module, and equipment status monitoring module to ensure that the information covers all dimensions of the experiment. The experimental process characteristics specifically include: the current stage of the experiment, clearly distinguishing between the pre-experiment self-inspection stage, the weighing and feeding stage, the dissolution and mixing stage, the pH closed-loop calibration stage, the volume adjustment and quality verification stage, and the recording and archiving stage; the real-time status of key experimental indicators, including the deviation of the current solution's pH value from the target value of 7.40, the real-time values ​​of the solution and ambient temperature, the real-time detection results of conductivity, the current values ​​of liquid level and weighing data, and the changing trends of each key indicator; and the operating status of the experimental equipment, including the working stability and zero-point status of the electronic balance, whether the pH meter's calibration validity period has drifted, and whether the signal acquisition of the temperature probe and conductivity sensor is accurate. Normal operation includes checking the robotic arm's response speed, the stirring motor's operating power and speed, and the fault alarm status of various sensors and actuators; experimental constraints, including the target volume set for the experiment, the allowable error range of pH value, the maximum limit of acid and alkali addition step size, the upper and lower limits of stirring speed, the highest threshold of experimental temperature, the initial interval of pH retesting, and the strategy level for handling abnormalities; current abnormal conditions in the experiment, including whether there is sensor data drift or signal interruption, whether the robotic arm has motion jamming or positioning deviation, whether the solution level has abnormal fluctuations, whether the VOC concentration is close to the safety threshold, and whether the equipment calibration status has failed, etc. Based on the configured attributes of each decision computing task and combined with the obtained experimental process feature information, a scheduling strategy that is highly adapted to the current experimental scenario is dynamically generated to ensure that the strategy is targeted and flexible. The scheduling strategy is generated as follows: When the experiment is in the pH closed-loop calibration stage, since this stage directly determines whether the core experimental indicator pH meets the standard, the execution priority of parameter optimization tasks is raised to the highest level, and sufficient computing resources are allocated to ensure the rapid execution of key tasks such as acid and alkali addition calculation and dispensing rate optimization; when the experiment is in the weighing and feeding stage, priority is given to the execution of electronic balance weighing data stability analysis tasks and robotic arm pick-and-place path optimization tasks to ensure weighing accuracy and operational safety; when the experiment is in any stage and safety-related process characteristic information is obtained, regardless of the current experimental stage, the corresponding safety assessment tasks are immediately raised to the highest priority, and the execution of current non-urgent tasks is suspended to ensure that safety risks can be detected and dealt with in a timely manner; for parameter optimization tasks with high computing resource requirements, the system... When computing resources are sufficient, allocate adequate resources according to the preset configuration. If system resources are scarce, prioritize the core computing needs of high-urgency tasks and reduce resource consumption of low-urgency tasks. Based on task execution time limits, set up dedicated priority scheduling channels for tasks with high urgency and short execution time limits to avoid timeouts. Considering the impact level of task failures, reserve 10% of spare computing resources for the highest-impact safety assessment tasks and medium-impact core parameter optimization tasks. When the main resource is abnormally occupied, the spare resources are automatically activated to ensure normal task execution. For abnormal situations that occur during the experiment, adjust the scheduling priority of the corresponding tasks. For example, when pH meter drift is detected, increase the priority of the pH meter calibration status validity determination task, while decreasing the execution priority of parameter optimization tasks that rely on pH data, until the anomaly is resolved.

[0044] Step 2.3: According to the scheduling strategy, sort the multiple decision calculation tasks to be executed to obtain the task execution sequence and resource allocation results. Specifically, this includes: according to the dynamically generated scheduling strategy, systematically prioritizing all decision calculation tasks to be executed in the current PBS buffer preparation experiment to form a clear and orderly task execution sequence, ensuring that the task execution meets the experimental requirements and safety requirements. The sequencing process strictly follows the logic of the scheduling strategy. In the pH closed-loop calibration phase, the task execution sequence is as follows: calculation of trace acid / alkali addition amounts, optimization of dispensing rate, analysis of deviation between pH meter readings and target values, assessment of VOC sensor concentration exceeding limits, analysis of temperature change trends, analysis of the normality of force and torque at the robotic arm's end effector, and judgment of conductivity value compliance. In the weighing and dispensing phase, the task execution sequence is as follows: stability analysis of electronic balance weighing data, calculation of salt weighing deviations, optimization of robotic arm pick-and-place paths, analysis of the normality of force and torque at the robotic arm's end effector, and detection of workstation presence. The tasks include monitoring the closure status of the fence gate and assessing the safety of the waste liquid tank level. During the dissolution and mixing stages, the task execution sequence is as follows: optimizing the stirring speed during dissolution, optimizing the stirring time during dissolution, analyzing the temperature change trend, analyzing the matching of liquid level and weighing changes, analyzing the concentration change of VOC sensor, and making a preliminary judgment on the pass / failability of conductivity values. When safety risk-related characteristics appear during the experiment, the corresponding safety assessment task is immediately inserted at the beginning of the task execution sequence. For example, if the fence gate is detected to be not closed, the fence gate closure status monitoring task is placed first in the sequence, and the currently executing non-safety tasks are suspended until the safety risk is eliminated.

[0045] After prioritizing tasks, based on the computational resource requirements of each task as specified in the scheduling strategy and the current resource availability of the system, a detailed resource allocation plan is formulated to ensure that resource allocation is reasonable, efficient and not wasteful. The resource allocation is as follows: Sufficient computing resources are allocated to the highest priority tasks, such as the micro-acid / alkali addition calculation task and the dispensing rate optimization task in the pH closed-loop calibration phase, each allocated 4 GPU computing resources to ensure rapid completion of complex AI model calculations; Appropriate computing resources are allocated to the next highest priority tasks, such as the pH meter detection value deviation analysis task allocated 2 GPU computing resources, and the robotic arm pick-and-place path optimization task allocated 2 GPU computing resources; Basic computing resources are allocated to tasks of normal priority, such as the workstation presence detection task allocated 1 GPU computing resource, and the fence gate lock closure status monitoring task allocated 1 GPU computing resource; 10% of spare computing resources are reserved for tasks with high failure impact, such as VOC sensor concentration... The over-limit assessment task allocates 2 GPU computing resources and reserves an additional 0.2 GPU computing resources as a backup. At the same time, the possibility of parallel execution of tasks is fully considered. For tasks with no functional conflicts and no data dependencies, independent computing resource channels are allocated to achieve parallel computing. For example, in the dissolution and mixing stage, the temperature change trend analysis task and the liquid level and weighing change matching analysis task are assigned to different GPU cores to improve the overall task processing speed without affecting the execution efficiency of each task. For tasks with data dependencies, resources are allocated strictly according to the execution sequence to ensure that the subsequent tasks are started and executed only after the preceding tasks are completed and output results. For example, after the pH value deviation analysis task is completed, resources are allocated and the pH retest frequency adjustment task is started.

[0046] In a preferred embodiment of the present invention, the scheduled task is executed according to the allocation result to output a series of decision instructions adapted to the current experimental process, including: Step 2.4: Based on the task execution sequence and resource allocation results, instantiate the decision calculation tasks to be executed in sequence; use the real-time status dataset as input and bind it to the corresponding task instance. Specifically, according to the determined task execution sequence, starting with the highest priority task, create corresponding task instances for each decision calculation task to be executed in sequence. For example, first create an instance for the pH parameter optimization task, clarifying that the task objective of this instance is to calculate parameters such as the amount of acid and alkali added and the injection rate required in the current stage. Then create an instance for the pH value deviation analysis task, clarifying that its task objective is to analyze the magnitude and trend of the deviation between the current pH value and the target value. Then create instances for safety assessment tasks and other status analysis tasks in sequence. Each instance contains the specific execution logic and expected output results of the task. During the creation of task instances, the generated dataset reflecting the real-time status of PBS buffer preparation is completely input and bound to the corresponding task instance. For pH parameter optimization task instances, the bound real-time status data includes the current solution's pH value, temperature, total amount of salts added, amount of acid and alkali added, stirring speed, and liquid level data. For electronic balance weighing data stability analysis task instances, the bound real-time status data includes the electronic balance's real-time reading, fluctuation values ​​from multiple weighings, balance zero-point status, and the type of salt currently being weighed. For safety assessment task instances, the bound real-time status data includes the VOC sensor's real-time concentration value, fence lock status, robotic arm end effector force and torque data, temperature values, liquid level height, and emergency stop button status, ensuring that each task instance can obtain all the real-time status information required to complete its own calculations.

[0047] Step 2.5: Run the task instances on the allocated computing resources, execute state analysis, path planning, or parameter optimization calculation logic, and generate preliminary decision results. Specifically, this includes: deploying each task instance to the allocated corresponding computing resources and starting the task instance to run; for state analysis task instances, performing calculations according to preset analysis logic, such as the electronic balance weighing data stability analysis task instance, calculating the standard deviation and fluctuation amplitude of multiple weighing readings to determine whether the balance is in a stable working state; if the fluctuation amplitude is within a preset threshold, it is determined to be stable; otherwise, it is determined to be unstable; the pH value deviation analysis task instance calculates the difference between the current pH value and the target value of 7.40, corrects the deviation value in combination with temperature data, and analyzes whether the trend of the deviation change is gradually decreasing or increasing; the robotic arm operation status analysis task instance compares the real-time collected robotic arm end force and torque data with the preset normal range to determine whether there is a risk of collision or jamming. For path planning tasks, the main application is to the motion planning of robotic arms. Combining the workstation's on-site detection data, container position information, door lock status, etc., the optimal path for the robotic arm to pick up and put down weighing containers and reagent bottles is planned to ensure accurate and collision-free movements. For example, in the weighing and feeding stage, the robotic arm needs to pick up and put down the weighing boat from the material warehouse and then add the weighed salts to the reaction container. The path planning needs to avoid other equipment and obstacles while shortening the action time. For parameter optimization tasks, the corresponding optimization calculation logic is executed. For example, in the pH closed-loop calibration phase, the AI's pH response model is called based on the bound real-time pH value, temperature, and amount of acid or alkali added. The addition rate is determined by combining the solution's stirring state, and the waiting time for mixing after addition is set according to historical mixing time data and the current solution volume. The retest frequency is set according to the magnitude and trend of pH deviation. In the dissolution phase, the stirring speed and duration are optimized based on temperature data, salt dissolution status, and stirring power changes to ensure rapid and complete dissolution of salts. Through the calculation logic, each task instance generates corresponding preliminary decision results.

[0048] Step 2.6 involves formatting and synthesizing the preliminary decision results, transforming them into specific control commands or parameter adjustment commands to drive the experimental equipment, and outputting a series of decision commands. Specifically, this includes: collecting the preliminary decision results generated by all task instances and performing unified formatting on these results; organizing the preliminary results from different task instances into a unified standard format, clearly defining the unit identifier and numerical precision of each parameter in the results. For example, the acid addition result generated by the pH parameter optimization task is uniformly formatted as microliters, retaining two decimal places; the stirring speed result is uniformly formatted as revolutions per minute, retaining integer places; the judgment results generated by the state analysis task are uniformly formatted as a clear statement of normal or abnormal, accompanied by relevant data support; and the path results generated by the path planning task are uniformly formatted as the motion angles and motion time sequences of each joint of the robotic arm. After formatting, all preliminary decision results are synthesized. Based on the current stage and task priority of the experiment, related decision results are integrated, and duplicate or conflicting content is removed. For example, in the pH closed-loop calibration stage, the results of pH parameter optimization tasks, such as acid and alkali addition amount, dosing rate, waiting time, and retest frequency, are synthesized with the speed optimization results of the stirring motor to ensure that all parameters are compatible. In the weighing and feeding stage, the results of robotic arm path planning are synthesized with the stability analysis results of electronic balance weighing data. If the balance is unstable, the picking and placing rhythm of the robotic arm is adjusted synchronously. The synthesized decision results are transformed into specific control commands or parameter adjustment commands to drive the experimental equipment, forming a series of decision commands. For different actuators in the automatic execution module, the commands are specific and clear. For example, control commands for the peristaltic pump specify the type of reagent being added (acid or alkali), the amount added, the rate of addition, and the duration of addition; control commands for the stirring motor specify the specific value of the stirring speed and the adjustment time; control commands for the robotic arm specify the movement angle, speed, and dwell time of each joint along the motion path; control commands for the pH meter specify the time interval for retesting and the data acquisition frequency; and control commands for safety equipment specify the shutdown location and alarm method if an anomaly is detected. These specific commands are organized according to the experimental procedure and execution sequence, outputting a series of decision commands.

[0049] Step 2.7: During task execution, monitor the usage of computing resources and the task execution status. If the task execution times out, fails, or produces abnormal results, trigger the exception handling process and update the subsequent scheduling strategy. Specifically, during the execution of all scheduled task instances, the system monitors the usage of computing resources allocated to each task in real time, including the CPU usage, memory usage, and data transmission bandwidth of each task. At the same time, it continuously monitors the execution status of each task, including whether the task has started, whether it is running normally, whether the execution progress meets expectations, and whether the execution has been completed. It also monitors whether the results of task execution output are within a reasonable range and whether there are logical conflicts or abnormal values. If a task timeout is detected, such as a safety assessment task failing to complete within 1 second, or a task failing, such as a parameter optimization task failing to generate valid results due to data anomalies, or a task producing abnormal results, such as a state analysis task determining that the pH value exceeds the safety threshold but there are obvious errors in data acquisition, the anomaly handling process is immediately triggered. The anomaly handling process executes corresponding actions based on the level of impact of the task failure and the type of anomaly. If an anomaly occurs in a safety assessment task, an instruction is immediately sent to stop the operation of the relevant experimental equipment, including stopping liquid addition, stopping the robotic arm movement, stopping stirring, and closing relevant valves. At the same time, the access control is locked, the pumping equipment or alarm device is activated, and the experimenter is prompted to manually confirm and handle the situation. If an anomaly occurs in a parameter optimization task, the execution of the current experimental stage is paused, a backup parameter scheme is activated, or the previous stage is returned to re-execute the relevant steps. If an anomaly occurs in a general state analysis task, the task result is marked as invalid, a redundant state analysis task is activated, or historical data is used for supplementary judgment. After triggering the exception handling process, the system automatically records information such as the time of the exception, task name, exception type, and handling method. Based on this information, the system updates the subsequent scheduling strategy. For example, if a certain type of parameter optimization task times out multiple times, the subsequent scheduling strategy allocates more computing resources to this type of task and extends its execution time limit. If a security assessment task frequently detects anomalies, the subsequent scheduling strategy increases the priority and execution frequency of this type of task and increases the availability of backup computing resources. If a certain type of status analysis task produces abnormal results multiple times, the subsequent scheduling strategy adjusts the execution logic or data input source of this task to ensure that the execution of subsequent tasks is more stable and reliable.

[0050] In a preferred embodiment of the present invention, a version identifier is assigned to the real-time state dataset, and a series of decision instructions are respectively bound to the state version that triggered their generation, including: Step 3.1: Each time the construction and update of the real-time status dataset is completed, a unique and incremental status version identifier is generated and assigned to the current dataset to obtain a series of decision instructions. Specifically, in the entire process of the automated PBS buffer preparation experiment, whenever the experimental sensing module completes the collection, integration, and update of real-time status data to form a new real-time status dataset, the system will automatically generate and assign a status version identifier to the current complete dataset. This status version identifier has a unique attribute and increments sequentially according to the order of dataset updates, ensuring that each real-time status dataset has a unique identifier. Specifically, the construction and update of the real-time status dataset will be triggered at each stage after the task is issued. In the pre-experiment self-check and calibration judgment stage, after the sensing module has collected and integrated all the preset initial status data such as the balance zero point and stability, pH meter calibration status, temperature probe status, and fence lock status, the first complete dataset will be formed. The system generates a real-time status dataset, at which point it assigns the first status version identifier. During the weighing and feeding phase, each time the robotic arm completes a salt weighing and feeds the data back to the system, the sensing module integrates the current weighing data, container position status, balance stability, and other data to form an updated dataset. The system then assigns a new status version identifier with a larger value than the previous one. During the dissolving and mixing phase, when the temperature and stirring status data collected by the sensing module changes and is integrated and updated to form a real-time status dataset reflecting the current dissolving progress, the system continues to generate incremental status version identifiers. During the pH closed-loop calibration phase, after each addition of acid or alkali solution, the sensing module sends back new pH data, solution volume data, temperature data, etc., and integrates them into a new dataset, assigning a unique and incremental status version identifier. During the volume adjustment and quality verification phase, after the sensing module completes the verification and integration of data such as pH, temperature, conductivity, and liquid level, the generation and allocation of status version identifiers are also triggered. Regardless of the experimental phase, once the construction and update of the real-time state dataset is completed, the corresponding state version identifier will be generated strictly in accordance with the above rules to ensure that the version identifiers of all datasets are unique and ordered.

[0051] Step 3.2: During the series of decision-making instructions, record the state version identifier corresponding to the real-time state dataset on which each decision-making instruction is based, so as to establish a correlation between the decision-making instruction and its source state version. Specifically, when the AI ​​analysis and decision-making module generates a series of decision-making instructions based on the current real-time state dataset of the experiment, it will synchronously record the state version identifier corresponding to the real-time state dataset on which each decision-making instruction is based, thereby establishing a clear correlation between the decision-making instruction and the source state version that generated the instruction. This process will continue to be executed at different stages of the experiment: In the weighing and feeding stage, when the AI ​​unit analyzes the current real-time state dataset and determines that the weighing deviation is close to the upper limit and generates a decision-making instruction to switch to a small-step feeding and compensation strategy, it will automatically retrieve the state version identifier corresponding to the real-time state dataset that triggered the decision and associate the identifier with the decision-making instruction to switch to a small-step feeding and compensation strategy. If the AI ​​unit generates a decision-making instruction to extend the stabilization waiting time or reset the weighing due to the fluctuation of the balance reading exceeding the threshold, it will also retrieve the state version identifier corresponding to the real-time state dataset that triggered the decision. The status version identifier of the corresponding real-time status dataset is associated with the new decision instruction. During the dissolution and mixing stage, when the AI ​​unit determines that dissolution is incomplete based on data such as temperature and stirring power changes in the real-time status dataset and generates a decision instruction to increase the stirring speed and extend the stirring time, it will simultaneously record the status version identifier corresponding to the decision instruction. During the pH closed-loop calibration stage, when the AI ​​unit generates a decision instruction based on the current pH value and temperature data in the real-time status dataset to add a specific amount of dilute sodium hydroxide solution, set the addition rate, specify the addition location, determine the mixing waiting time and retest frequency, it will accurately record the status version identifier corresponding to this dataset. If a pH overshoot trend occurs, when the AI ​​unit generates a decision instruction to reduce the alkali addition step size and extend the mixing waiting time, or when it generates a suggestion-type decision instruction to clean the probe, re-insert it, or increase stirring due to pH change lag, it will record the status version identifier of the real-time status dataset from which it originates. Each decision instruction is bound to the version identifier of the real-time status dataset that triggered it from the moment it is generated.

[0052] Step 3.3 encapsulates the decision instruction and its associated state version identifier to form an execution decision unit with version traceability information. Specifically, after completing the association record of each decision instruction and its corresponding state version identifier, the decision instruction and the associated state version identifier are integrated and encapsulated to form an execution decision unit containing complete version traceability information, providing complete data support for subsequent execution module call instructions and subsequent traceability work. During the encapsulation process, the execution decision unit will fully contain two core pieces of information: one is the specific decision instruction content generated by the AI ​​analysis and decision module, covering the operational requirements required for each stage of the experiment, such as the specific amount of material added in the small-step material addition stage, the material addition interval time, the specific duration of the extended stabilization waiting time, and the operation trigger for resetting the weighing. The system includes: adjustments to the stirring speed and extended stirring time during the dissolution and mixing stages; specific volumes, addition rates, coordinates or area descriptions of the addition location, waiting time for mixing after addition, time intervals for retesting, probe cleaning procedures, re-insertion location requirements, and increased stirring speed values ​​during the pH closed-loop calibration stage; and various control parameters or operational guidelines required for each stage. On the other hand, it includes a status version identifier strictly associated with the decision instruction, ensuring a one-to-one correspondence between the identifier and the decision instruction, and that the identifier is indivisible. Through encapsulation, the resulting decision unit to be executed clearly defines the specific operation to be performed by the execution module and also includes the status version information of the source of the operation instruction, providing a complete traceability basis before the decision instruction is transmitted to the automatic execution module.

[0053] In a preferred embodiment of the present invention, after the decision instruction is executed, execution feedback data is collected and associated with the corresponding decision instruction and status version for storage to form a decision-execution link record, including: Step 3.4: After the decision instruction in the decision unit to be executed is sent to the experimental equipment for execution, the system captures the execution feedback data returned by the equipment. The execution feedback data includes at least one or more of the following: execution success flag, execution result data, or execution exception information. Specifically, after the decision instruction encapsulated in the decision unit to be executed is sent to the corresponding experimental equipment, the system uses the experimental process sensing module as a feedback module to capture various types of execution feedback data returned by the experimental equipment after executing the instruction in real time. These feedback data include at least one or more of the following: execution success flag, execution result data, or execution exception information. The specific details are as follows, combined with the various experimental stages of automated PBS buffer preparation. During the weighing and feeding stage, if the decision instruction in the decision-making unit to be executed is to switch to a small-step feeding strategy, extend the stabilization waiting time, or reset the weighing, the execution equipment is a robotic arm and an electronic balance. When the robotic arm executes the small-step feeding instruction, the feedback module collects execution result data such as the actual feeding amount and whether the balance reading is stable in real time through the electronic balance, and generates an execution success flag to confirm whether the feeding action has been completed. If the robotic arm fails to successfully grasp the weighing container or fails to accurately add the salt to the reaction container, the feedback module will capture abnormal force or torque data through the force and torque sensor at the end of the robotic arm, and then generate execution anomaly information, specifying the anomaly type as grasping failure or feeding position deviation. If the balance reading fluctuation continues to exceed the threshold, the feedback module will record the anomaly information and set the execution success flag to the unsuccessful state. During the dissolving and mixing stage, the decision-making instructions are mostly to increase the stirring speed and extend the stirring time, and the executing device is the stirring motor. After the stirring motor executes the adjusted stirring instructions, the feedback module collects real-time solution temperature data through a temperature probe, and collects execution result data such as whether the stirring speed has reached the set value and whether the stirring is stable through a stirring status sensor, and generates an execution success flag; if the stirring motor fails to reach the set value or stops during operation, the feedback module will capture the abnormal situation, generate execution abnormality information, and specify the fault type as insufficient speed or motor stoppage. During the pH closed-loop calibration phase, decision commands include the volume of micro-dosage of acid / alkali solution, dosing rate, dosing location, mixing waiting time, retest frequency, as well as probe cleaning, re-insertion, and increasing stirring speed. The executing devices include a peristaltic pump, stirring motor, and robotic arm. After the peristaltic pump executes the dosing command, the feedback module collects the actual dosing volume via a level sensor or weighing device, the pH value of the solution after dosing via a pH meter, and the solution temperature via a temperature probe. These data constitute the execution result data, and a success flag is generated to confirm the completion of the dosing action. If the peristaltic pump dosing volume exceeds the set range, the dosing rate is abnormal, or the robotic arm fails to accurately reach the cleaning position when executing the probe cleaning command, the feedback module will capture these anomalies through the corresponding sensors, generating execution anomaly information to identify whether the anomaly is due to excessive dosing volume, abnormal dosing rate, or probe cleaning position deviation. If the pH value change is sluggish, the feedback module will record the change data collected by the pH meter as part of the execution result data. If the sluggishness is determined to be caused by insufficient stirring, relevant anomaly prompts will also be generated. During the volume adjustment and quality verification stage, the decision instruction is to adjust the solution to the target volume, and the execution device is the liquid dispensing module. After the module executes the volume adjustment instruction, the feedback module collects the actual volume after volume adjustment through a liquid level sensor or electronic balance, and collects the pH value and conductivity of the solution through a pH meter and conductivity sensor, respectively. These are all execution result data, and a success flag is generated. If the volume after volume adjustment exceeds the allowable error range, or the pH value and conductivity deviate from the qualified range, the feedback module will record these abnormal data and generate execution anomaly information, indicating that the volume adjustment deviation or the indicator is unqualified. During the anomaly handling phase, decision commands include stopping liquid addition, stopping the robotic arm, and stopping agitation. The executing equipment includes the liquid dispensing module, the robotic arm, and the agitator motor. After the equipment executes these commands, the feedback module confirms whether the equipment has stopped operating via the corresponding sensors and generates a success flag. If the equipment fails to stop operating as instructed, the feedback module will capture the anomaly, generate execution anomaly information, and indicate whether the equipment is unresponsive or the access control system is malfunctioning.

[0054] Step 3.5: Extract the encapsulated decision instruction and status version identifier from the decision unit to be executed; associate the execution feedback data with the extracted decision instruction and status version identifier to form a decision-execution link record containing complete information of status version, decision instruction, and execution feedback. Specifically, this includes: firstly, extracting the specific decision instruction content and the status version identifier strictly bound to the decision instruction from the previously encapsulated decision unit to be executed. During the extraction process, the system's built-in verification mechanism ensures that the extracted decision instruction and status version identifier are accurate and that no information is omitted or mismatched; subsequently, performing a one-to-one association and matching between the captured execution feedback data corresponding to the decision instruction and the extracted decision instruction and status version identifier to ensure that the execution feedback data corresponding to each decision instruction can accurately correspond to the status version identifier on which it was generated. Finally, a decision-execution link record containing complete information of status version identifier, decision instruction, and execution feedback data is formed. The specific details are as follows, combined with each experimental stage. During the weighing and feeding stage, the decision instructions extracted from the decision-making unit to be executed may be the specific amount of material to be fed in small increments, the feeding interval, the specific duration of the stabilization waiting time, or the operation trigger conditions for resetting the weighing. The corresponding status version identifier is a unique incrementing identifier of the real-time status dataset on which the decision instruction is based. This identifier was assigned when the real-time status dataset was previously built and updated. The system will associate these extracted decision instructions and status version identifiers with the captured actual feeding amount, stable balance reading, execution success flag, or execution feedback data such as robotic arm grasping failure and balance reading fluctuation exceeding limits. For example, if the decision instruction is to add 0.1 grams of sodium chloride each time with a feeding interval of 5 seconds, the corresponding status version identifier is a specific incrementing value, and the execution feedback data is that 0.102 grams were actually added, the balance reading is stable, and the execution is successful, the system will closely associate these three parts of information to form a complete decision and execution link record, clearly presenting the specific result after the execution of the decision instruction under this status version. During the dissolution and mixing stage, the decision instructions extracted from the decision-making unit to be executed are the adjusted stirring speed and the extended stirring time. The corresponding status version identifier is the identifier of the real-time status dataset on which the decision instruction was based. The extracted information is associated with the collected real-time solution temperature, actual stirring speed, execution success flag, or execution feedback data such as insufficient stirring motor speed or shutdown. For example, if the decision instruction is to increase the stirring speed to 500 rpm and extend the stirring time by 10 minutes, the corresponding status version identifier is a specific value. If the execution feedback data is an actual stirring speed of 500 rpm, a solution temperature of 25 degrees Celsius, and execution success, the system will associate the status version identifier, the decision instruction, and the feedback data to form a link record. If the execution feedback data is that the stirring speed only reaches 400 rpm and execution is abnormal, the system will also accurately associate and record the status version identifier, the decision instruction, and the abnormal feedback data. During the pH closed-loop calibration phase, the decision instructions extracted from the decision unit to be executed include the volume of micro-dosage of acid or alkali solution, dosing rate, dosing location, mixing waiting time, retest frequency, or probe cleaning procedure, re-insertion location requirements, and increased stirring speed. The corresponding status version identifier is the identifier of the real-time status dataset on which the decision instruction is based. The system will associate these extracted contents with the execution feedback data obtained in step 3.4, such as the actual dosing volume, pH value after dosing, solution temperature, execution success flag, or dosing volume exceeding limits, probe cleaning position deviation, etc. For example, if the decision instruction is to add 0.2 ml of dilute sodium hydroxide solution at a rate of 0.05 ml per second, to add the solution near the vortex area, and to wait 30 seconds after mixing, the corresponding status version identifier is a certain value. The execution feedback data is that 0.21 ml was actually added, the pH value increased from 7.2 to 7.3, the temperature reached 25 degrees Celsius, and the execution was successful. The system will associate these three pieces of information to form a link record. If the execution feedback data is that the adding rate is abnormal, the actual rate is 0.08 ml per second, and the execution is abnormal, then the status version identifier, the decision instruction, and the abnormal feedback data will be associated and combined. During the volume adjustment and quality verification stage, the decision instruction extracted from the decision unit to be executed is to adjust the solution to a target volume of 1 liter. The corresponding status version identifier is the identifier of the real-time status dataset on which the decision instruction was generated. This information is associated with the collected actual volume adjustment volume, pH value, conductivity, execution success flag, or execution feedback data such as volume adjustment volume exceeding the allowable error or pH value deviating from the qualified range. For example, if the decision instruction is to adjust the volume to 1 liter, the corresponding status version identifier is a certain value, and the execution feedback data is an actual volume of 1.002 liters, a pH value of 7.42, conductivity within the preset qualified range, and execution success, the three are associated to form a link record. If the execution feedback data is an actual volume of 1.01 liters, exceeding the allowable error, or execution anomaly, the status version identifier, the decision instruction, and the anomaly feedback data are associated and recorded. During the anomaly handling phase, the decision instructions extracted from the decision-making unit to be executed are such as stopping liquid dispensing or stopping the robotic arm operation. The corresponding status version identifier is the identifier of the real-time status dataset on which the decision instruction was based when it was generated. The extracted information is associated with the captured execution feedback data, such as whether the liquid dispensing module has stopped, whether the robotic arm has stopped, execution success flags, or equipment not responding or not stopping operation. For example, if the decision instruction is to immediately stop dilute hydrochloric acid dispensing and stop the robotic arm operation, the corresponding status version identifier is a certain value, and the execution feedback data is that the liquid dispensing module has stopped, the robotic arm has stopped, and execution was successful. These three are associated to form a link record. If the execution feedback data is that the robotic arm has not stopped running or execution is abnormal, then the status version identifier, the decision instruction, and the abnormal feedback data are associated and combined. All decision-making and execution chain records formed by the associated combinations are transmitted in real time to the high temporal resolution recording module for structured storage. Each record fully retains all information such as status version identifier, decision instruction details, and execution feedback data, ensuring that the relationship between each decision and the corresponding execution action is traceable throughout the experiment.

[0055] In a preferred embodiment of the present invention, based on the decision-execution link record, strategies are dynamically loaded, switched, or updated during the experiment to automate the processing of experimental process status data, including: Step 4.1: Obtain the stored decision-execution link records, analyze the execution feedback data and associated experimental states; based on preset evaluation rules or performance indicators, determine whether the strategy currently supporting the decision calculation task needs to be updated, replaced, or rolled back, and generate a strategy change decision. Specifically, this includes: obtaining the full-link decision-execution link records stored in the recording module, which contain raw sensor data, execution command data, AI decision context, timestamps, and abnormal handling results during the experiment; extracting execution feedback data from the records, including execution result data such as the amount of liquid added, stirring speed, and waiting time completed by each actuator, as well as associated experimental state data, covering information such as pH value, temperature, conductivity, solution level, balance reading, and equipment operating status; conducting analysis based on preset evaluation rules and performance indicators, where evaluation rules include whether pH is within the allowable error range, whether the number of adjustment steps exceeds the set threshold, and whether the data fluctuation amplitude meets the requirements, etc., and performance indicators include the response efficiency of pH adjustment, the smoothness of the experimental process, and the occurrence rate of abnormal events, etc. By comparing the results with preset standards, it is determined whether the current strategy supporting the decision-making calculation task is insufficient. If there are situations such as frequent overshooting of pH adjustment, slow response, low efficiency in achieving the target, or an increase in abnormal events, it is determined that the strategy needs to be updated or replaced. If the new strategy does not work well after application, or even causes the experimental state to deviate from expectations, it is determined that the strategy needs to be rolled back. Finally, a clear strategy change decision is generated based on the judgment results.

[0056] Step 4.2: Based on the strategy change decision, dynamically load the configuration file and calculation model of the target strategy from the decoupled management strategy library. Without interrupting the core experimental process, instantiate the loaded target strategy into an executable strategy object. Specifically, this includes: determining the target strategy to be loaded based on the generated strategy change decision. The target strategy comes from the decoupled management strategy library, which contains various strategies adapted to different experimental scenarios, such as different pH adjustment step size strategies, weighing compensation strategies, and dissolution judgment strategies. Accurately retrieve the configuration file and calculation model corresponding to the target strategy from the strategy library. The configuration file contains specific parameter settings for the target strategy, such as the standard micro-dosing volume for acid and alkali addition, dosing rate range, dosing location requirements, mixing waiting time threshold, and retesting frequency settings. The calculation model includes a pH response model learned from historical data and a temperature-dissolution rate relationship model. During the loading process, ensure that the core experimental process is not interrupted. The core experimental process includes ongoing dissolution and stirring, volume adjustment, and other critical steps, avoiding any impact on the experimental progress and status. Simultaneously, convert the loaded target strategy into a directly executable strategy object.

[0057] Step 4.3: Control the activation timing of the strategy object so that it replaces the old strategy object at the start of the next experimental decision cycle or during a gap when no critical decision tasks are being executed, thus completing the hot update and switching of the strategy. Specifically, this includes: accurately controlling the activation timing of the target strategy object, prioritizing activation at the start of the next experimental decision cycle. For example, if the current pH adjustment cycle has been completed and the next pH retest and adjustment decision node is about to begin, activation at this time ensures smooth strategy continuity. If the current experiment is in a gap when no critical decision tasks are being executed, such as the solution dissolution stage, without critical decision operations such as adding acid or alkali, or weighing, activation can also be activated during this gap. At the determined activation timing, the target strategy object replaces the old strategy object, maintaining the continuity of the core experimental process during the replacement process and not interrupting the normal operation of the experiment. After the replacement is completed, all AI decision calculation tasks are carried out based on the new strategy object, thereby achieving hot update and switching of the strategy and ensuring that the experimental process can continue to advance efficiently based on the optimized strategy.

[0058] like Figure 2 As shown, an automated processing system for experimental process state data based on artificial intelligence includes: The multimodal perception module is used to extract and fuse features from multi-source heterogeneous data acquired during the experiment, so as to output a real-time status dataset that reflects the experimental progress. The artificial intelligence analysis and decision-making module is used to dynamically prioritize and allocate resources for multiple concurrent decision computing tasks based on the real-time status dataset and experimental process characteristics to obtain allocation results; and to execute the scheduled tasks according to the allocation results to output a series of decision instructions adapted to the current experimental process. The experiment execution module is used to assign version identifiers to the real-time status dataset and bind a series of decision instructions to the status versions that trigger their generation. After the decision instructions are executed, execution feedback data is collected and associated with the corresponding decision instructions and status versions to form a decision-execution link record. The experimental data processing module is used to dynamically load, switch, or update strategies during the experiment based on the decision-execution link record, so as to realize the automated processing of experimental process status data.

[0059] In this invention, an automated processing system for experimental process state data based on artificial intelligence is provided. Specifically, a multimodal perception module acquires state data during the experimental process. This state data includes at least one of the following: image or video data of the experimental object; microscopic imaging data; sensor data of the experimental environment or the experimental object; and information on structural, morphological, color, positional, or state changes during the experimental process. The multimodal perception module can employ any combination of various sensing devices, and the sensing frequency can be at the second, sub-second, or continuous acquisition level. An artificial intelligence analysis and decision-making module receives the experimental objectives, constraints, or expected results input by the experimenter. Based on the experimental state data acquired by the multimodal perception module, it understands the experimental process. Through an artificial intelligence model, it automatically identifies the stage of the experiment and generates experimental intervention decisions based on the experimental objectives, stages, and historical data. During the experiment, the decisions are dynamically corrected and optimized. The artificial intelligence model can be a large-scale language model, a multimodal model, or an intelligent model with reasoning capabilities; its specific implementation is not limited.

[0060] The experiment execution module is used to execute corresponding experimental intervention actions on the experimental process based on the experimental intervention decisions generated by the artificial intelligence analysis and decision-making module. These intervention actions include, but are not limited to: introducing or removing substances, adjusting experimental parameters, switching experimental conditions, controlling experimental procedures, and terminating or pausing the experimental process. This can be achieved using automated mechanical structures, electronic control devices, or other actuators.

[0061] The experimental data processing module records the state data, decision data, and execution data of the entire experimental process with high temporal resolution, forming a complete data chain of the experimental process. Based on the recorded data, it automatically generates a structured description of experimental results or an experimental report, supporting experimental process backtracking, multi-experiment comparative analysis, and parallel experimental management.

[0062] In embodiments of the present invention, the specific application of an artificial intelligence-based automated processing method for experimental process status data in the automated preparation of PBS buffer and pH closed-loop calibration is presented, specifically including: Phase 1: Feature Extraction and Fusion of Multi-Source Heterogeneous Data. The experimenter, using an AI-input PBS buffer automated preparation experiment, collected real-time data from various sources, including an electronic balance (0.001g accuracy), pH meter (1Hz sampling frequency), conductivity sensor, temperature probe, robotic arm end effector force / torque sensor, level sensor, workstation detection sensor, door lock status sensor, and VOC sensor. Data was categorized by source and type, such as material addition, property monitoring, execution feedback, and safety monitoring, and features were extracted from each category, including quality stability, pH trend, temperature deviation, and motion accuracy. After mapping to a unified state space, temporal alignment and fusion were used to construct an experimental state representation. New data was incrementally updated only for the corresponding features, forming a real-time state dataset.

[0063] The second stage involves dynamic task scheduling and decision command generation. The AI ​​analysis and decision module, based on real-time state datasets and preset experimental characteristics (such as target pH 7.40±0.05, maximum liquid addition step size 50μL, and upper temperature limit 30℃), abstracts the experimental process into multiple concurrent decision-making computation tasks, including state analysis, parameter optimization, safety assessment, quality judgment, and anomaly handling. Each task is configured with attributes such as priority, execution frequency, and resource requirements. Scheduling strategies are dynamically generated based on task attributes and experimental characteristics. For example, in the pH adjustment stage, the priority of the parameter optimization task is increased. Tasks are instantiated sequentially, computational logic is run, and preliminary decision results are generated, which are then formatted into specific control commands, such as "add 10μL NaOH, stir for 60 seconds." During execution, task status and resource usage are monitored, and in case of anomalies, a processing flow is triggered and the scheduling strategy is updated.

[0064] The third stage is version binding and link recording. A unique incremental version identifier, such as State_v0012, is assigned to each updated real-time status dataset. When generating decision instructions, the status version on which they are based is recorded and encapsulated as a decision unit. After the instruction is executed, feedback data is collected, such as the actual liquid volume added being 10.2 μL and the time taken being 8.2 seconds. This data is associated with the corresponding instruction and status version to form a decision-execution link record, which completely saves the full link information of "status → decision → execution → feedback → new status".

[0065] Phase 4: Dynamic strategy update and optimization. Based on the accumulated link records, analyze and adjust performance indicators such as adjustment steps, overshoot rate, and convergence time. When the indicators exceed the threshold, such as overshoot rate > 10%, load a better strategy from the decoupling strategy library, such as reducing the adjustment step size or extending the mixing time. Instantiate the new strategy object without interrupting the experiment, and complete the hot switching in the next decision cycle to achieve dynamic optimization of the strategy.

[0066] Phase 5: Experiment recording and report generation. The system automatically records raw data, instruction flow, decision context and abnormal events, and generates structured reports and traceable "evidence packages" based on timestamps and link records, supporting high-resolution backtracking and reproduction verification of the entire experimental process.

[0067] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An automated processing method for experimental process state data based on artificial intelligence, characterized in that, The method includes: Step 1: Extract and fuse features from the multi-source heterogeneous data acquired during the experiment to output a real-time status dataset that reflects the experimental progress. Step 2: Based on the real-time status dataset and experimental process characteristics, perform dynamic priority scheduling and resource allocation for multiple concurrent decision computing tasks to obtain allocation results; according to the allocation results, execute the scheduled tasks to output a series of decision instructions adapted to the current experimental process. Step 3: Assign version identifiers to the real-time status dataset and bind the series of decision instructions to the status versions that triggered their generation; after the decision instructions are executed, collect execution feedback data and associate and store it with the corresponding decision instructions and status versions to form a decision-execution link record; Step 4: Based on the decision-execution link record, dynamically load, switch, or update the strategy during the experiment to achieve automated processing of the experimental process status data.

2. The automated processing method for experimental process state data based on artificial intelligence according to claim 1, characterized in that, Step 1 involves feature extraction and fusion of the multi-source heterogeneous data acquired during the experiment to output a real-time status dataset reflecting the experimental progress, including: Step 1.1: Acquire multimodal experimental process data from different experimental devices or information sources in real time, and classify the experimental process data according to the data source and data type to obtain the classification results; Step 1.2: Based on the classification results, feature extraction processing is performed on different types of experimental process data to obtain processing results; the processing results are mapped to a unified state feature space to obtain the mapped state features. Step 1.3: Based on the mapped state features, perform temporal alignment and feature fusion to construct the experimental state calculation; Step 1.4: When a new experimental process data state change is received, the experimental state calculation representation is incrementally updated for the partial state features corresponding to the new data to form the updated real-time state dataset.

3. The automated processing method for experimental process state data based on artificial intelligence according to claim 2, characterized in that, Step 2: Based on the real-time state dataset and experimental process characteristics, dynamically prioritize and allocate resources for multiple concurrent decision computation tasks to obtain allocation results, including: Step 2.1: Abstract the intelligent analysis process of state analysis, parameter optimization, or safety assessment during the experiment into multiple independent decision computing tasks; configure task attributes for each decision computing task. Step 2.2: Obtain experimental process feature information related to the current experiment, and dynamically generate a scheduling strategy based on the task attributes and the process feature information; Step 2.3: According to the scheduling strategy, sort the multiple decision calculation tasks to be executed to obtain the task execution sequence and resource allocation results.

4. The automated processing method for experimental process state data based on artificial intelligence according to claim 3, characterized in that, Based on the allocation results, the scheduled tasks are executed to output a series of decision instructions adapted to the current experimental progress, including: Step 2.4: Based on the task execution sequence and resource allocation results, instantiate the decision calculation tasks to be executed in sequence; use the real-time status dataset as input and bind it to the corresponding task instance; Step 2.5: Run the task instance on the allocated computing resources, execute state analysis, path planning or parameter optimization calculation logic, and generate preliminary decision results; Step 2.6: The preliminary decision results are formatted and synthesized to transform them into specific control instructions or parameter adjustment instructions that drive the experimental equipment to execute, and output as a series of decision instructions; Step 2.7: During task execution, monitor the usage of computing resources and the task execution status; if the task execution times out, fails, or produces abnormal results, trigger the exception handling process and update the subsequent scheduling strategy.

5. The automated processing method for experimental process state data based on artificial intelligence according to claim 4, characterized in that, Step 3: Assign version identifiers to the real-time state dataset, and bind a series of decision instructions to the state version that triggered their generation, including: Step 3.1: Whenever the construction and update of the real-time state dataset is completed, a unique and incremental state version identifier is generated and assigned to the current dataset to obtain a series of decision instructions; Step 3.2: When issuing a series of decision instructions, record the status version identifier corresponding to the real-time status dataset on which the decision instruction is based for each decision instruction, so as to establish an association between the decision instruction and its source status version. Step 3.3: Encapsulate the decision instruction and its associated status version identifier to form an execution decision unit with version traceability information.

6. The automated processing method for experimental process state data based on artificial intelligence according to claim 5, characterized in that, After the decision instruction is executed, execution feedback data is collected and associated with the corresponding decision instruction and status version for storage, forming a decision-execution link record, including: Step 3.4: After the decision instruction in the decision unit to be executed is sent to the experimental device for execution, capture the execution feedback data returned by the device. The execution feedback data includes at least one or more of the following: execution success flag, execution result data, or execution exception information. Step 3.5: Extract the encapsulated decision instructions and status version identifiers from the decision unit to be executed; associate the execution feedback data with the extracted decision instructions and status version identifiers to form a decision-execution link record containing complete information of status version, decision instructions, and execution feedback.

7. The automated processing method for experimental process state data based on artificial intelligence according to claim 6, characterized in that, Step 4: Based on the decision-execution link record, dynamically load, switch, or update the strategy during the experiment, including: Step 4.1: Obtain the stored decision-execution link record, analyze the execution feedback data and associated experimental status; based on preset evaluation rules or performance indicators, determine whether the strategy currently supporting the decision computing task needs to be updated, replaced or rolled back, and generate a strategy change decision. Step 4.2: Based on the strategy change decision, dynamically load the configuration file and calculation model of the target strategy from the decoupled management strategy library; without interrupting the core experimental process, instantiate the loaded target strategy into an executable strategy object; Step 4.3: Control the timing of the activation of the strategy object so that it replaces the original old strategy object at the beginning of the next experimental decision cycle or during the gap when no critical decision tasks are being executed, so as to complete the hot update and switching of the strategy.

8. An automated processing system for experimental process state data based on artificial intelligence, the system implementing the method as described in any one of claims 1 to 7, characterized in that, include: The multimodal perception module is used to extract and fuse features from multi-source heterogeneous data acquired during the experiment, so as to output a real-time status dataset that reflects the experimental progress. The artificial intelligence analysis and decision-making module is used to dynamically prioritize and allocate resources for multiple concurrent decision computing tasks based on the real-time status dataset and experimental process characteristics, so as to obtain the allocation result. Based on the allocation results, the scheduled tasks are executed to output a series of decision instructions adapted to the current experimental progress; The experiment execution module is used to assign version identifiers to the real-time state dataset and bind a series of decision instructions to the state version that triggered their generation. After the decision instruction is executed, execution feedback data is collected and associated with the corresponding decision instruction and status version for storage to form a decision-execution link record; The experimental data processing module is used to dynamically load, switch, or update strategies during the experiment based on the decision-execution link record, so as to realize the automated processing of experimental process status data.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images