Batch measurement task initiation and result collection laboratory management method and system

Abstract

The present application relates to a kind of batch measurement task initiation and result collection laboratory management method, comprising: knowledge enhancement based on RAG, to provide intelligent decision support for measurement task by knowledge enhancement;Task scheduling step based on hypergraph isomorphism calculation, laboratory resource scheduling is modeled as hypergraph matching problem, realize multi-instrument collaborative scheduling;Measurement task type identification and initiation based on multi-expert hybrid model;Through visual technology, sample position verification and instrument state identification of different types of measurement tasks are carried out;Automatic operation strategy learning based on improved PPO algorithm and measurement task and result collection based on deep reinforcement learning control Automatic operation;Based on graph neural network, multi-source data are fused to carry out multi-modal anomaly detection on the result collected;Federal quality evaluation, including: data quality evaluation corresponding to measurement task based on privacy protection distributed learning.It also discloses the corresponding system.

Description

Technical Field

[0001] This invention relates to the field of laboratory automation and artificial intelligence, and particularly to a laboratory management method and system for initiating and collecting results of batch measurement tasks. More specifically, it relates to a laboratory management method for initiating and collecting results of batch measurement tasks by integrating RAG knowledge base, expert model and hypergraph computation. Background Technology

[0002] Modern water quality testing laboratories are typically equipped with dozens of precision measuring instruments, such as silicate analyzers, ion chromatographs, and total organic carbon analyzers. Traditional operating methods face the following technical bottlenecks:

[0003] (1) Operational complexity: Each instrument has a different software interface and a complex operation process, requiring professional training;

[0004] (2) Inefficient: The laboratory technicians need to operate each instrument one by one and manually input sample information, which takes an average of 5-10 minutes per sample;

[0005] (3) High error rate: When multiple tasks are performed in parallel, sample information and measurement parameters are easily confused;

[0006] (4) Monitoring difficulties: Abnormalities in the measurement process are difficult to detect and handle in real time;

[0007] (5) Knowledge silos: Expert experience and historical data are not effectively utilized.

[0008] Existing solutions, such as simple remote control scripts or basic LIMS systems, cannot achieve true intelligence and adaptive learning. In particular, they lack effective AI technology support in areas such as batch task scheduling, anomaly detection, and decision support. Summary of the Invention

[0009] The purpose of this invention is to provide a laboratory management method and system for initiating and collecting batch measurement tasks. It aims to overcome the limitations of existing technologies by integrating advanced AI technologies such as RAG knowledge base, expert models and hypergraph computing to build an intelligent laboratory management system with self-learning and self-adaptive capabilities, thereby realizing fully automated batch execution and intelligent monitoring of measurement tasks.

[0010] The first aspect of the present invention provides a laboratory management method for initiating batch measurement tasks and collecting results, comprising:

[0011] S1, knowledge enhancement based on RAG, thereby providing intelligent decision support for measurement tasks through the knowledge enhancement; including: knowledge retrieval based on a knowledge vectorization library and a dynamic retrieval mechanism, and converting unstructured data into vector representations, wherein the unstructured data includes one or more of historical operation manuals, expert experience, and fault cases;

[0012] S2, based on hypergraph isomorphic computation, performs task scheduling steps, models laboratory resource scheduling as a hypergraph matching problem, and realizes multi-instrument collaborative scheduling;

[0013] S3, Measurement task type identification and initiation based on multi-expert hybrid model, including: an expert network based on multi-expert hybrid model identifies measurement task types and initiates different types of measurement tasks in different ways;

[0014] S4, Visual recognition and automated control, including: verifying sample position and identifying instrument status for different types of measurement tasks using visual technology;

[0015] S5, Automated operation based on deep reinforcement learning control, including: learning automated operation strategies based on the improved PPO algorithm and automating measurement tasks and result acquisition based on deep reinforcement learning control.

[0016] S6, perform multimodal anomaly detection on the collected results based on graph neural network fusion of multi-source data;

[0017] S7, Federal Quality Assessment, includes: conducting a data quality assessment corresponding to the measurement task based on privacy-preserving distributed learning.

[0018] Preferably, S1 includes:

[0019] S11, the knowledge is vectorized by converting unstructured data into vector representation and embedding it into text using the BERT model. The knowledge vectorization library is represented by equation (1) as follows:

[0020] (1);

[0021] in, The input is unstructured text data, including historical operation manuals, expert experience, and failure cases;

[0022] S12, the knowledge that has undergone the vectorization process is dynamically retrieved using a dynamic retrieval mechanism. This dynamic retrieval mechanism is used to retrieve relevant knowledge fragments based on the current task context, including: retrieving relevant knowledge fragments based on the current task context and calculating a similarity score between the retrieved relevant knowledge fragments and the current task context. :

[0023] (2);

[0024] in: Indicates the weight of knowledge timeliness; These are the weighting coefficients;

[0025] S13, Based on the relevant knowledge fragments, perform generative decision-making to generate an operation instruction sequence. As shown in equation (3):

[0026] (3);

[0027] in, For retrieving relevant knowledge sets, This is a description of the current task. This is the generated sequence of operation instructions.

[0028] Preferably, S2 includes:

[0029] S21, Construct the hypergraph H=(V,E), including:

[0030] (1) Establish a vertex set V = {task node, instrument node, time node, resource node};

[0031] (2) Establish the hyperedge set E = {task constraints, resource dependencies, and temporal relationships};

[0032] S22, Establish a hypergraph isomorphism algorithm to calculate the hypergraph matching degree between task requirements and resource supply, as shown in equation (4):

[0033] (4);

[0034] in, This represents the hypergraph matching degree or hypergraph isomorphic matching value, characterizing the task requirement hypergraph. With resource supply supermap The overall matching degree between them, the larger the value, the better the matching effect between task and resource supply and demand; This represents a task requirement hypergraph, which is a set of task requirements to be assigned modeled in the form of a hypergraph. Hyperedges are used to represent multi-dimensional constraints and relationships within the task. Represents a resource supply hypergraph, which is a set of available resource supplies modeled in the form of a hypergraph. Hyperedges are used to represent resource clusters and resource combination constraints. This represents the mapping relationship between nodes in the hypergraph, defining the hypergraph from the task requirements. For mapping The set of matching hyperedges below; nodes to resource supply hypergraph The one-to-one mapping rule for nodes; This indicates traversing all feasible mappings. The optimal matching result is the one with the highest matching degree. For mapping The set of matching hyperedges under the node mapping At the time of its establishment, the task requirements exceeded the map. With resource supply supermap A set of hyperedges whose structure, attributes, and constraints can be matched one-to-one; This represents the cardinality or number of the matching hyperedge set, i.e., the mapping. The total number of superedges that can be successfully matched; Indicates the task requirement hypergraph The set of all superedges; This represents the total number of all hyperedges in the task hypergraph; Indicates resource supply exceeding the chart The set of all superedges; This represents the total number of all hyperedges in the resource hypergraph.

[0035] S23, establish constraints to satisfy the optimization, in order to minimize the degree of constraint violation, as shown in equation (5):

[0036] (5);

[0037] in, For the first The degree of violation of each constraint is related to the decision variables. The function is used to quantize the first step in the current scheme. The extent to which a constraint is violated; when constraints are fully satisfied. The more severe the violation, the larger the value. The index variable is used to iterate through all constraints, where the constraint number is the index variable. This represents the total number of all constraints, i.e., the total number of constraints that need to be satisfied in the optimization problem. For the first The weight coefficient of a constraint represents its importance, priority, or penalty; the more critical the constraint, the higher its weight. The larger the value, the higher the penalty for violation; These are decision variables / system state variables, representing the variables to be solved in this optimization problem, such as task-resource matching schemes, resource allocation strategies, and hypergraph mapping parameters.

[0038] Preferably, S3 includes:

[0039] S31, construct dedicated expert networks, each optimized for a specific type of measurement task, as shown in Equation (6):

[0040] (6);

[0041] in, Indicates the first A dedicated expert network, with feature vectors as input. The output is the expert network's prediction / decision results for a specific measurement task; each For each specific type of measurement task, network parameters are optimized independently. This indicates the expert network number, used to distinguish various specialized expert subnetworks for different types of measurement tasks; The input feature vector to the network is a multi-dimensional input vector composed of the original data, task features, resource features, or hypergraph matching features corresponding to the measurement task. Indicates the first The first layer of an expert network, i.e., the hidden layer weight matrix, is used to weight the input features. Perform a linear transformation; Indicates the first The bias vector of the first layer of the expert network, together with the weight matrix, completes the linear mapping of the hidden layer; The activation function is represented by nonlinear transformation, which improves the ability of expert networks to fit the features of complex measurement tasks. This includes the use of activation functions such as ReLU, Sigmoid, and Tanh. Indicates the first The second layer of the expert network, i.e., the output layer weight matrix, performs a linear transformation on the output of the hidden layer. Indicates the first The second layer bias vector of the expert network is used for linear compensation of the output layer; Indicates the first Each expert network has its own dedicated output mapping function, customized for the corresponding type of measurement task. It can realize the functions of normalization, classification output, regression prediction, and matching scoring, and adapt to the output requirements of specific tasks.

[0042] S32, assign the gating network, and use the softmax function to assign expert weights, as shown in equation (7):

[0043] (7);

[0044] in, The output of the gating network represents the first... The weights of the nth expert network are in the range of [0,1], and the sum of all expert weights is 1. These weights are used to characterize the nth expert network weight under the current input feature. The contribution percentage of each dedicated expert network; This indicates that it corresponds to the first step in S31 of the previous step. A dedicated network of experts for specific measurement tasks; This represents the traversal index of all expert networks, traversing all available expert subnetworks; The input feature vector is consistent with the input in equation (6) and is a multidimensional feature vector related to the measurement task, resource supply, and hypergraph matching. Indicating the first in the gating network The transpose of the weight vector corresponding to each expert is the trainable parameters of the gating network, used to adjust the input features. Perform linear weighting; This represents the natural exponential function, which, when used with softmax, normalizes the weights and ensures that the output is non-negative. Indicating the first in the gating network Transpose the weight vector corresponding to each expert, and iterate through the gating weight parameters of all experts; This represents summing the exponential terms of all expert networks to achieve softmax normalization, so that the weighted sum of the weights of each expert is 1;

[0045] S33, Task type identification and initiation, including: automatically identifying task types based on task characteristics and selecting the most suitable expert network combination.

[0046] Preferably, S4 includes:

[0047] S41, the sample position verification is achieved through a deep learning-based object detection and pose estimation method, as shown in equation (8):

[0048] (8);

[0049] in, This represents the total loss function for detection and pose estimation, used for deep learning model training. It characterizes the error between the sample target position, pose prediction results, and the true label. The model training aims to minimize this loss to achieve sample position verification. This represents the coordinate loss weight coefficient, a hyperparameter used to adjust the proportion of the target box coordinate position loss in the overall loss, amplifying / weakening the impact of positional deviation on model training; The feature map grid size is used to divide the input image into... The grid, The total number of grid cells. Traverse all grid cells; Represents the grid cell index, traversing all the grid cells after image partitioning. One grid; This indicates the number of predefined bounding boxes or anchor boxes within a single grid cell, meaning the maximum number of predictions per grid cell. One sample target bounding box; This represents the bounding box index within a single grid, and iterates through all preset bounding boxes within a single grid. This indicates that the target has an indicator function or a binary mask, with a value of 0 or 1; when the... The first grid The value is 1 if the bounding box contains the real target of the sample, and 0 otherwise. The position loss is calculated only for the box containing the target. and Represents the true center coordinates of the sample target, the first... The x and y coordinates of the center point of the true target bounding box of the sample within each grid; and The x and y coordinates represent the center coordinates of the sample target predicted by the model, and the corresponding center coordinates of the target box output by the deep learning network. The squared error of the target center point coordinates represents the sample position prediction deviation. The remaining terms of the loss function include target confidence loss, class classification loss, pose angle loss, and bounding box width and height loss, which are used for integrity-constrained sample detection, classification, and pose estimation.

[0050] S42, the instrument state identification is achieved by state comparison based on the Siamese network, as shown in equation (9):

[0051] (9);

[0052] in, This represents the contrastive loss function of the Siamese network, used for model training loss in instrument state comparison and state recognition. Optimizing this loss makes the distance between features of the same instrument state smaller and the distance between features of different instruments state larger. This represents the total number of training sample pairs, i.e., the total number of instrument state feature sample pairs participating in model training; Use sample pair indexes to iterate through all instrument status sample pairs; Represents a sample pair labeled, a binary variable; when the two input samples are in the same instrument state... For different instrument states ; The distance represents the Euclidean distance between two instrument state feature vectors, i.e., the feature similarity distance, which characterizes the degree of difference between the two sets of instrument state features extracted by the Siamese network. The smaller the distance, the more similar the states are. This represents the boundary threshold or interval hyperparameter, a manually set safety distance threshold; used to constrain the minimum feature distance between sample pairs in different states to avoid feature aliasing. This represents the margin loss term, where the feature distance between different sample pairs... Less than the set threshold Loss occurs when the distance is greater than the threshold; the greater the distance, the greater the loss. When this item is 0, no further penalty is applied; The normalization coefficient is the average loss of all sample pairs and scaled to stabilize the training gradient and unify the magnitude of the loss. For positive sample loss terms, for sample pairs with the same instrument state, minimize the feature distance to make the features of the same state as close as possible; This represents the negative sample loss term. For sample pairs with different instrument states, it forces the feature distance to be greater than a set threshold, so that features of different states can be distinguished as much as possible.

[0053] Preferably, S5 includes:

[0054] S51, Automated operation strategy learning is performed based on the improved PPO algorithm, and the improved PPO algorithm for strategy learning is shown in equation (10):

[0055] (10);

[0056] in, Indicates the belt An improved PPO policy loss function with divergence penalty is used for automated operational policy learning. By constraining the difference in distributions between the old and new policies, it stabilizes the policy update process. The network parameters for the current policy to be updated; The expectation operator represents taking the mathematical expectation of the state-action trajectory sample distribution and calculating the average loss based on sampling experience; The network parameters of the current policy to be updated are used to improve the neural network parameters to be iteratively optimized in PPO, which are used to output the probability of automated operation actions. This represents the network parameters of the old strategy. The parameters of the previous strategy are updated and used as a reference to prevent the strategy update from being too large. Indicates the current policy, in the state Below, the parameters are The strategy network output action The probability of; This represents the old strategy, in state Below, the old parameter policy network outputs actions. The probability of; The importance sampling ratio represents the probability ratio of the same action under the old and new strategies, and is used to measure the magnitude of strategy updates. The environmental state refers to the system state that matches the instrument, sample, and task during automated operation. The actions performed by the intelligent agent correspond to various automated operation instructions and control behaviors; The advantage function estimate at time t represents the payoff advantage of the current action compared to the average action, and is used to guide the strategy to update to higher-return actions; for The divergence penalty coefficient is a hyperparameter used to control the strength of the penalty for the distribution difference between the old and new strategies. The larger the value, the stronger the constraint. for Divergence relative entropy, used to measure the old strategy With new strategies In the same state The degree of difference in the distribution of lower action; Representing the given states respectively At that time, the probability distribution of the complete actions corresponding to the old strategy and the new strategy;

[0057] S52 is an automated operation for measurement tasks and result acquisition based on deep reinforcement learning control, including: automatically executing sample information entry, parameter setting, and measurement start operations based on the learning strategy.

[0058] Preferably, S6 includes:

[0059] S61, Multimodal feature fusion is performed based on graph neural network and multimodal anomaly detection algorithm, the multimodal anomaly detection algorithm is shown in equation (11):

[0060] (11);

[0061] in, Indicates the first Nodes in a layered graph neural network The multimodal fusion feature vector has nodes corresponding to entities such as instruments, samples, and tasks. The features fuse multimodal information from images, text, and sensor data for subsequent anomaly detection. Indicates the first Neighbor nodes in a layered graph neural network The original multimodal feature vector is the node feature output from the previous layer; This is the layer index of the graph neural network, indicating the current feature propagation layer number; The central node index represents the target entity node whose feature needs to be updated; Indexing neighbor nodes, traversing nodes All adjacent nodes that have a relationship; For nodes The set of first-order neighbors, representing the set of neighbors of the current node. All connected related nodes; It is a trainable weight parameter matrix, a feature transformation matrix of a graph neural network, used to perform linear mapping and dimension adjustment on the aggregated neighbor features, and adapt to multimodal feature fusion. For nodes The normalized degree of a node in a graph structure. The number of connected neighbors is used for node degree normalization; Neighboring nodes Normalized degree; This is the degree matrix normalization term. It normalizes the feature aggregation by taking the square root of the product of the node degrees, so as to avoid high-connectivity nodes contributing too much to the features and balance the weights of different nodes. For neighbor feature aggregation operations, the normalized features of all neighboring nodes are accumulated to achieve local neighborhood information fusion;

[0062] S62 monitors the measurement process in real time, identifies abnormal patterns, and triggers corresponding processing mechanisms.

[0063] Preferably, S7 includes:

[0064] S71, the federated model is trained using a privacy-preserving distributed learning algorithm, as shown in Equation (12):

[0065] (12);

[0066] S72 evaluates the quality of measurement results based on the trained federated model and provides feedback to optimize the measurement process.

[0067] A second aspect of the present invention provides a laboratory management system for initiating and collecting batch measurement tasks, used to implement the method of the first aspect, comprising:

[0068] The RAG-based knowledge enhancement module (101) is used to perform knowledge enhancement based on RAG, thereby providing intelligent decision support for measurement tasks through the knowledge enhancement.

[0069] The task scheduling module (102) based on hypergraph isomorphic computation is used to perform task scheduling steps based on hypergraph isomorphic computation, modeling laboratory resource scheduling as a hypergraph matching problem, and realizing multi-instrument collaborative scheduling.

[0070] The measurement task type identification and initiation module (103) based on the multi-expert hybrid model is used for measurement task type identification and initiation based on the multi-expert hybrid model, including: the expert network based on the multi-expert hybrid model identifies the measurement task type and initiates different types of measurement tasks in different ways;

[0071] A visual recognition and automation control module (104) is used for visual recognition and automation control, including: verifying the sample position and identifying the instrument status for different types of measurement tasks using visual technology;

[0072] The automated operation module (105) based on deep reinforcement learning control is used to perform automated operations based on deep reinforcement learning control, including: learning automated operation strategies based on the improved PPO algorithm and performing automated operations of measurement tasks and result acquisition based on deep reinforcement learning control.

[0073] A multimodal anomaly detection module (106) is used to perform multimodal anomaly detection on the collected results based on graph neural network fusion of multi-source data;

[0074] The Federal Quality Assessment Module (107) is used for federal quality assessment, including: performing data quality assessment corresponding to the measurement task based on privacy-preserving distributed learning.

[0075] Preferably, the RAG-based knowledge enhancement module (101) consists of two parts: a knowledge retrieval unit and a knowledge enhancement generation unit. The knowledge retrieval unit performs knowledge retrieval based on a knowledge vectorization library and a dynamic retrieval mechanism. The knowledge vectorization library is used to convert unstructured data into vector representations. The unstructured data includes one or more of historical operation manuals, expert experience, and fault cases.

[0076] The laboratory management method and system for batch measurement task initiation and result acquisition of the present invention have the following beneficial effects:

[0077] 1. Improved operational efficiency: Batch task processing time reduced by 85%, manual intervention reduced by 90%;

[0078] 2. Reduced error rate: Knowledge-enhanced decision-making reduces the operational error rate to below 0.05%;

[0079] 3. Adaptive learning: The system can continuously learn and optimize from historical operations;

[0080] 4. Resource optimization: Instrument utilization rate increased from an average of 45% to 85%;

[0081] 5. Quality traceability: full-process data recording and quality analysis. Attached Figure Description

[0082] To more clearly illustrate the technical solutions in the specific embodiments or related technologies of the present invention, the drawings used in the description of the specific embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0083] Figure 1 This is a schematic diagram of a laboratory management method for initiating batch measurement tasks and collecting results, provided in an embodiment of the present invention.

[0084] Figure 2 This is a schematic diagram of step S1 of the laboratory management method for batch measurement task initiation and result acquisition provided in an embodiment of the present invention.

[0085] Figure 3 This is a schematic diagram of step S2 of the laboratory management method for batch measurement task initiation and result acquisition provided in an embodiment of the present invention.

[0086] Figure 4 This is a flowchart of step S3 of the laboratory management method for batch measurement task initiation and result acquisition provided in an embodiment of the present invention.

[0087] Figure 5 This is a flowchart illustrating step S4 of the laboratory management method for initiating batch measurement tasks and collecting results, provided in an embodiment of the present invention.

[0088] Figure 6 This is a flowchart of step S5 of the laboratory management method for batch measurement task initiation and result acquisition provided in an embodiment of the present invention.

[0089] Figure 7 This is a flowchart of step S6 of the laboratory management method for batch measurement task initiation and result acquisition provided in an embodiment of the present invention.

[0090] Figure 8 This is a flowchart of step S7 of the laboratory management method for batch measurement task initiation and result acquisition provided in an embodiment of the present invention.

[0091] Figure 9 This is a schematic diagram of the laboratory management system for batch measurement task initiation and result acquisition provided in an embodiment of the present invention. Detailed Implementation

[0092] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0093] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0094] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0095] like Figure 1 As shown, the first aspect of the present invention provides a laboratory management method for initiating batch measurement tasks and collecting results, comprising:

[0096] S1, knowledge enhancement based on RAG, thereby providing intelligent decision support for measurement tasks through the knowledge enhancement; including: knowledge retrieval based on a knowledge vectorization library and a dynamic retrieval mechanism, and converting unstructured data into vector representations, wherein the unstructured data includes one or more of historical operation manuals, expert experience, and fault cases;

[0097] like Figure 2 As shown, in a preferred embodiment, S1 includes:

[0098] S11, the knowledge is vectorized by converting unstructured data into vector representation and embedding it into text using the BERT model. The knowledge vectorization library is represented by equation (1) as follows:

[0099] (1);

[0100] in, The input consists of unstructured text data, including historical operation manuals, expert experience, and failure cases.

[0101] S12, the knowledge that has undergone the vectorization process is dynamically retrieved using a dynamic retrieval mechanism. This dynamic retrieval mechanism is used to retrieve relevant knowledge fragments based on the current task context, including: retrieving relevant knowledge fragments based on the current task context and calculating a similarity score between the retrieved relevant knowledge fragments and the current task context. :

[0102] (2);

[0103] in: Indicates the weight of knowledge timeliness; These are the weighting coefficients.

[0104] S13, Based on the relevant knowledge fragments, perform generative decision-making to generate an operation instruction sequence. As shown in equation (3):

[0105] (3);

[0106] in, For retrieving relevant knowledge sets, This is a description of the current task. This is the generated sequence of operation instructions.

[0107] In this embodiment, a knowledge graph containing instrument operation specifications and fault handling solutions is constructed, and the knowledge base is updated in real time to ensure the timeliness and accuracy of knowledge. At the same time, knowledge fusion based on the attention mechanism improves the quality of decision-making.

[0108] S2, based on hypergraph isomorphic computation, performs task scheduling steps, models laboratory resource scheduling as a hypergraph matching problem, and realizes multi-instrument collaborative scheduling;

[0109] like Figure 3 As shown, in a preferred embodiment, S2 includes:

[0110] S21, Construct the hypergraph H=(V,E), including:

[0111] (1) Establish a vertex set V = {task node, instrument node, time node, resource node};

[0112] (2) Establish the hyperedge set E = {task constraints, resource dependencies, and temporal relationships};

[0113] S22, Establish a hypergraph isomorphism algorithm to calculate the hypergraph matching degree between task requirements and resource supply, as shown in equation (4):

[0114] (4);

[0115] in, This represents the hypergraph matching degree or hypergraph isomorphic matching value, characterizing the task requirement hypergraph. With resource supply supermap The overall matching degree between them, the larger the value, the better the matching effect between task and resource supply and demand; This represents a task requirement hypergraph, which is a set of task requirements to be assigned modeled in the form of a hypergraph. Hyperedges are used to represent multi-dimensional constraints and relationships within the task. Represents a resource supply hypergraph, which is a set of available resource supplies modeled in the form of a hypergraph. Hyperedges are used to represent resource clusters and resource combination constraints. This represents the mapping relationship between nodes in the hypergraph, defining the hypergraph from the task requirements. For mapping The set of matching hyperedges below; nodes to resource supply hypergraph The one-to-one mapping rule for nodes; This indicates traversing all feasible mappings. The optimal matching result is the one with the highest matching degree. For mapping The set of matching hyperedges under the node mapping At the time of its establishment, the task requirements exceeded the map. With resource supply supermap A set of hyperedges whose structure, attributes, and constraints can be matched one-to-one; The cardinality (number) of the matching hyperedge set, i.e., the mapping. The total number of superedges that can be successfully matched; Indicates the task requirement hypergraph The set of all superedges; This represents the total number of all hyperedges in the task hypergraph; Indicates resource supply exceeding the chart The set of all superedges; This represents the total number of all hyperedges in the resource hypergraph.

[0116] S23, establish constraints to satisfy the optimization, in order to minimize the degree of constraint violation, as shown in equation (5):

[0117] (5);

[0118] in, For the first The degree of violation of each constraint is related to the decision variables. The function is used to quantize the first step in the current scheme. The extent to which a constraint is violated; when constraints are fully satisfied. The more severe the violation, the larger the value. The index variable is used to iterate through all constraints, where the constraint number is the index variable. This represents the total number of all constraints, i.e., the total number of constraints that need to be satisfied in the optimization problem. For the first The weight coefficient of a constraint represents its importance, priority, or penalty; the more critical the constraint, the higher its weight. The larger the value, the higher the penalty for violation; These are decision variables / system state variables, which can represent the variables to be solved in this optimization problem, such as task-resource matching schemes, resource allocation strategies, and hypergraph mapping parameters.

[0119] S3, Measurement task type identification and initiation based on multi-expert hybrid model, including: an expert network based on multi-expert hybrid model identifies measurement task types and initiates different types of measurement tasks in different ways;

[0120] like Figure 4 As shown, in a preferred embodiment, S3 includes:

[0121] S31, construct dedicated expert networks, each optimized for a specific type of measurement task, as shown in Equation (6):

[0122] (6);

[0123] in, Indicates the first A dedicated expert network, with feature vectors as input. The output is the expert network's prediction / decision results for a specific measurement task; each For each specific type of measurement task, network parameters are optimized independently. This indicates the expert network number, used to distinguish various specialized expert subnetworks for different types of measurement tasks; The input feature vector to the network is a multi-dimensional input vector composed of the original data, task features, resource features, or hypergraph matching features corresponding to the measurement task. Indicates the first The weight matrix of the first layer (hidden layer) of the expert network is used to weight the input features. Perform a linear transformation; Indicates the first The bias vector of the first layer of the expert network, together with the weight matrix, completes the linear mapping of the hidden layer; The activation function is represented by a nonlinear transformation, which improves the ability of the expert network to fit the features of complex measurement tasks. Common activation functions such as ReLU, Sigmoid, and Tanh can be used. Indicates the first The weight matrix of the second layer (output layer) of the expert network is used to perform a linear transformation on the output of the hidden layer. Indicates the first The second layer bias vector of the expert network is used for linear compensation of the output layer; Indicates the first Each expert network has its own dedicated output mapping function, customized for the corresponding type of measurement task. It can realize functions such as normalization, classification output, regression prediction, and matching scoring, and adapt to the output requirements of specific tasks.

[0124] S32, assign the gating network, and use the softmax function to assign expert weights, as shown in equation (7):

[0125] (7);

[0126] in, The output of the gating network represents the first... The weights of the nth expert network are in the range of [0,1], and the sum of all expert weights is 1. These weights are used to characterize the nth expert network weight under the current input feature. The contribution percentage of each dedicated expert network; This indicates that it corresponds to the first step in S31 of the previous step. A dedicated network of experts for specific measurement tasks; This represents the traversal index of all expert networks, traversing all available expert subnetworks; The input feature vector is consistent with the input in equation (6) and is a multi-dimensional feature vector related to measurement tasks, resource supply, hypergraph matching, etc. Indicating the first in the gating network The transpose of the weight vector corresponding to each expert is the trainable parameters of the gating network, used to adjust the input features. Perform linear weighting; This represents the natural exponential function, which, when used with softmax, normalizes the weights and ensures that the output is non-negative. Indicating the first in the gating network Transpose the weight vector corresponding to each expert, and iterate through the gating weight parameters of all experts; This represents the summation of the exponents of all expert networks, achieving softmax normalization so that the weighted sum of the weights of each expert is 1.

[0127] S33, Task type identification and initiation, including: automatically identifying task types based on task characteristics and selecting the most suitable expert network combination.

[0128] In this embodiment, the implementation details of step S3 include: establishing a task feature extraction model to automatically identify key task attributes; designing an expert network collaborative working mechanism to improve the identification accuracy; and realizing intelligent configuration of task startup parameters.

[0129] S4, Visual recognition and automated control, including: verifying sample position and identifying instrument status for different types of measurement tasks using visual technology;

[0130] like Figure 5 As shown, in a preferred embodiment, S4 includes:

[0131] S41, the sample position verification is achieved through a deep learning-based object detection and pose estimation method, as shown in equation (8):

[0132] (8);

[0133] in, This represents the total loss function for detection and pose estimation, used for deep learning model training. It characterizes the error between the sample target position, pose prediction results, and the true label. The model training aims to minimize this loss to achieve sample position verification. This represents the coordinate loss weight coefficient, a hyperparameter used to adjust the proportion of the target box coordinate position loss in the overall loss, amplifying / weakening the impact of positional deviation on model training; The feature map grid size is used to divide the input image into... The grid, The total number of grid cells. Traverse all grid cells; Represents the grid cell index, traversing all the grid cells after image partitioning. One grid; This indicates the number of predefined bounding boxes or anchor boxes within a single grid cell, meaning the maximum number of predictions per grid cell. One sample target bounding box; This represents the bounding box index within a single grid, and iterates through all preset bounding boxes within a single grid. This indicates that the target has an indicator function or a binary mask, with a value of 0 or 1; when the... The first grid The value is 1 if the bounding box contains the real target of the sample, and 0 otherwise. The position loss is calculated only for the box containing the target. and Represents the true center coordinates of the sample target, the first... The x and y coordinates of the center point of the true target bounding box of the sample within each grid; and The x and y coordinates represent the center coordinates of the sample target predicted by the model, and the corresponding center coordinates of the target box output by the deep learning network. The squared error of the target center point coordinates represents the sample position prediction deviation. The remaining terms of the loss function include target confidence loss, category classification loss, pose angle loss, bounding box width and height loss, etc., which are used for complete constraint sample detection, classification and pose estimation.

[0134] S42, the instrument state identification is achieved by state comparison based on the Siamese network, as shown in equation (9):

[0135] (9);

[0136] in, This represents the contrastive loss function of the Siamese network, used for model training loss in instrument state comparison and state recognition. Optimizing this loss makes the distance between features of the same instrument state smaller and the distance between features of different instruments state larger. This represents the total number of training sample pairs, i.e., the total number of instrument state feature sample pairs participating in model training; Use sample pair indexes to iterate through all instrument status sample pairs; Represents a sample pair labeled, a binary variable; when the two input samples are in the same instrument state... For different instrument states ; The Euclidean distance (feature similarity distance) between two instrument state feature vectors represents the degree of difference between the two sets of instrument state features extracted by the Siamese network. The smaller the distance, the more similar the states are. This represents the boundary threshold or interval hyperparameter, a manually set safety distance threshold; used to constrain the minimum feature distance between sample pairs in different states to avoid feature aliasing. This represents the margin loss term, where the feature distance between different sample pairs... Less than the set threshold Loss occurs when the distance is greater than the threshold; the greater the distance, the greater the loss. When this item is 0, no further penalty is applied; The normalization coefficient is the average loss of all sample pairs and scaled to stabilize the training gradient and unify the magnitude of the loss. For positive sample loss terms, for sample pairs with the same instrument state, minimize the feature distance to make the features of the same state as close as possible; This represents the negative sample loss term. For sample pairs with different instrument states, it forces the feature distance to be greater than a set threshold, so that features of different states can be distinguished as much as possible.

[0137] In this embodiment, the implementation details of step S4 include: establishing a standard template library for sample placement to achieve rapid position comparison; developing a multi-angle visual inspection system to improve recognition reliability; and designing a real-time status monitoring mechanism to ensure instrument readiness.

[0138] S5, Automated operation based on deep reinforcement learning control, including: learning automated operation strategies based on the improved PPO algorithm and automating measurement tasks and result acquisition based on deep reinforcement learning control.

[0139] like Figure 6 As shown, in a preferred embodiment, S5 includes:

[0140] S51, Automated operation strategy learning is performed based on the improved PPO algorithm, and the improved PPO algorithm for strategy learning is shown in equation (10):

[0141] (10);

[0142] in, Indicates the belt An improved PPO policy loss function with divergence penalty is used for automated operational policy learning. By constraining the difference in distributions between the old and new policies, it stabilizes the policy update process. The network parameters for the current policy to be updated; The expectation operator represents taking the mathematical expectation of the state-action trajectory sample distribution and calculating the average loss based on sampling experience; The network parameters of the current policy to be updated are used to improve the neural network parameters to be iteratively optimized in PPO, which are used to output the probability of automated operation actions. This represents the network parameters of the old strategy. The parameters of the previous strategy are updated and used as a reference to prevent the strategy update from being too large. Indicates the current policy, in the state Below, the parameters are The strategy network output action The probability of; This represents the old strategy, in state Below, the old parameter policy network outputs actions. The probability of; The importance sampling ratio represents the probability ratio of the same action under the old and new strategies, and is used to measure the magnitude of strategy updates. The environmental state refers to the system state that matches the instrument, sample, and task during automated operation. The actions performed by the intelligent agent correspond to various automated operation instructions and control behaviors; The advantage function estimate at time t represents the payoff advantage of the current action compared to the average action, and is used to guide the strategy to update to higher-return actions; for The divergence penalty coefficient is a hyperparameter used to control the strength of the penalty for the distribution difference between the old and new strategies. The larger the value, the stronger the constraint. for Divergence relative entropy, used to measure the old strategy With new strategies In the same state The degree of difference in the distribution of lower action; Representing the given states respectively At that time, the probability distribution of the complete actions corresponding to the old strategy and the new strategy.

[0143] conventional Punishment PPO is only fixed Regarding the coefficient constraint strategy update magnitude, this embodiment of the invention makes two key improvements: ① Introducing time-series advantage weighting, which reduces the single-step advantage... ① Embedding the expected loss to strengthen the reward weighting for batch measurement of long-term operation trajectories; ② Designing dynamic adaptive... Adjustment mechanism, based on old and new strategies Divergence real-time correction penalty coefficient, Automatically amplify when value exceeds the limit Tighten the update interval. When the value is small, it decreases. Improve exploration efficiency and solve fixed problems The automated laboratory operation suffers from slow convergence and strategy oscillation issues. The original PPO lacks adaptive behavior. Coupled with the advantages of timing, it cannot adapt to long-process operations involving multiple instruments.

[0144] In this embodiment, 85 surface water ion chromatography batch detection tasks were selected as the test set, and compared with the standard Clip-PPO and fixed... -PPO-Penalty, Improvements in this embodiment -PPO: 1) Policy convergence rounds: Standard PPO requires 1280 rounds to converge, which is fixed. - PPO requires 960 rounds, while this improved algorithm only requires 420 rounds, increasing the convergence speed by 56.25%; 2) Operational error rate: The standard PPO automated operation has a mismatch rate of 0.82% for instrument parameters and sample misplacement, while the fixed β-PPO rate is 0.36%. This approach reduces the error rate to 0.04%. 3) Batch task time: Standard PPO single batch processing time is 120 minutes, fixed. -PPO took 86 minutes, while this improved algorithm only takes 52 minutes, resulting in a 39.5% improvement in task efficiency; 4) Strategy oscillation range: based on Divergence variance is a measure of variance, with a standard PPO variance of 0.187, fixed. -PPO0.092, while this solution is only 0.021, significantly improving the stability of strategy updates.

[0145] Data demonstrates that the improved PPO adapts to the long-process automated operation of multiple devices in the laboratory, balancing convergence speed, operational accuracy, and operational stability.

[0146] Furthermore, in this embodiment, the intelligent agent is a deep reinforcement learning intelligent agent for laboratory automation scheduling, including three types: instrument control sub-agent, sample introduction intelligent agent, and parameter configuration sub-agent. (Action) It includes three categories: 1) Sample operation actions: sample scanning and entry, sample tray positioning, sample switching; 2) Instrument control actions: instrument start and stop, detection parameter configuration, cleaning / purging commands; 3) Process scheduling actions: task switching, multi-instrument resource reallocation, abnormal interruption retesting.

[0147] S52 is an automated operation for measurement tasks and result acquisition based on deep reinforcement learning control, including: automatically executing sample information entry, parameter setting, and measurement start operations based on the learning strategy.

[0148] In this embodiment, the implementation details of step S5 include: constructing operation state space and action space models; designing a reward function to balance operation efficiency and accuracy; and realizing online learning and optimization of operation strategies.

[0149] S6, perform multimodal anomaly detection on the collected results based on graph neural network fusion of multi-source data;

[0150] like Figure 7 As shown, in a preferred embodiment, S6 includes:

[0151] S61, Multimodal feature fusion is performed based on graph neural network and multimodal anomaly detection algorithm, the multimodal anomaly detection algorithm is shown in equation (11):

[0152] (11);

[0153] in, Indicates the first Nodes in a layered graph neural network The multimodal fusion feature vector has nodes corresponding to entities such as instruments, samples, and tasks. The features fuse multimodal information such as images, text, and sensor data for subsequent anomaly detection. Indicates the first Neighbor nodes in a layered graph neural network The original multimodal feature vector is the node feature output from the previous layer; This is the layer index of the graph neural network, indicating the current feature propagation layer number; The central node index represents the target entity node whose feature needs to be updated; Indexing neighbor nodes, traversing nodes All adjacent nodes that have a relationship; For nodes The set of first-order neighbors, representing the set of neighbors of the current node. All connected related nodes; It is a trainable weight parameter matrix, a feature transformation matrix of a graph neural network, used to perform linear mapping and dimension adjustment on the aggregated neighbor features, and adapt to multimodal feature fusion. For nodes The normalized degree of a node in a graph structure. The number of connected neighbors is used for node degree normalization; Neighboring nodes Normalized degree; This is the degree matrix normalization term. It normalizes the feature aggregation by taking the square root of the product of the node degrees, so as to avoid high-connectivity nodes contributing too much to the features and balance the weights of different nodes. The neighbor feature aggregation operation accumulates the normalized features of all neighboring nodes to achieve local neighborhood information fusion.

[0154] Equation (11) is the normalized graph convolution (GCN) node feature update formula. By aggregating neighborhood multimodal features and combining node degree normalization with trainable weight matrix, it completes the deep fusion of multimodal information and provides structured feature input for multimodal anomaly detection.

[0155] S62 monitors the measurement process in real time, identifies abnormal patterns, and triggers corresponding processing mechanisms.

[0156] In this embodiment, the implementation details of step S6 include: establishing a multimodal data acquisition and feature extraction pipeline; designing a hierarchical anomaly detection model to improve detection sensitivity; and implementing anomaly classification and hierarchical processing mechanisms.

[0157] S7, Federal Quality Assessment, includes: conducting a data quality assessment corresponding to the measurement task based on privacy-preserving distributed learning.

[0158] like Figure 8 As shown, in a preferred embodiment, S7 includes:

[0159] S71, the federated model is trained using a privacy-preserving distributed learning algorithm, as shown in Equation (12):

[0160] (12);

[0161] S72 evaluates the quality of measurement results based on the trained federated model and provides feedback to optimize the measurement process.

[0162] In this embodiment, the implementation details of step S7 are as follows: establish a distributed quality assessment network; design a privacy-preserving data sharing mechanism; and realize real-time feedback of quality assessment results and process optimization.

[0163] like Figure 9 As shown, a second aspect of the present invention provides a laboratory management system for batch measurement task initiation and result acquisition, which is a multi-level, multi-modal intelligent laboratory management framework for implementing the method of the first aspect, including:

[0164] (i) A knowledge enhancement module (101) based on RAG (Retrieval-Augmented Generation) consists of two parts: a knowledge retrieval unit and a knowledge enhancement generation unit. The knowledge retrieval unit performs knowledge retrieval based on a knowledge vectorization library and a dynamic retrieval mechanism. The knowledge vectorization library is used to convert unstructured data into vector representations. The unstructured data includes one or more of historical operation manuals, expert experience, and fault cases.

[0165] In this embodiment, the knowledge vectorization library is represented by equation (1) as follows:

[0166] (1);

[0167] The dynamic retrieval mechanism is used to retrieve relevant knowledge fragments based on the current task context, as shown in equation (2):

[0168] (2);

[0169] in: Indicates the weight of knowledge timeliness; These are the weighting coefficients.

[0170] The knowledge enhancement generation unit is used to generate operation instructions based on generative decision-making. The generative decision is used to generate the sequence of operation instructions by combining the retrieved knowledge. As shown in equation (3):

[0171] (3);

[0172] in, For retrieving relevant knowledge sets, This is a description of the current task. This is the generated sequence of operation instructions.

[0173] (ii) Task scheduling module (102) based on hypergraph isomorphic computation, used to model laboratory resource scheduling as a hypergraph matching problem;

[0174] The method of modeling laboratory resource scheduling as a hypergraph matching problem includes:

[0175] 1. Construct a hypergraph, including:

[0176] (1) Establish a vertex set V = {task node, instrument node, time node, resource node};

[0177] (2) Establish the hyperedge set E = {task constraints, resource dependencies, temporal relationships};

[0178] 2. Establish a hypergraph isomorphism algorithm, as shown in equation (4):

[0179] (4);

[0180] in, For mapping The set of matching superedges below.

[0181] 3. Establish the constraints to satisfy the optimization, as shown in equation (5):

[0182] (5);

[0183] in, For the first The degree of violation of a constraint.

[0184] (iii) Measurement task type identification and initiation module (103) based on multi-expert hybrid model, used to identify measurement task types based on expert network of multi-expert hybrid model, and to initiate different types of measurement tasks in different ways; wherein, the multi-expert hybrid model is as shown in equation (6):

[0185] (6);

[0186] The gating network uses softmax allocation, as shown in equation (7):

[0187] (7);

[0188] (iv) A visual recognition and automated control module (104) is used for verifying sample position and identifying instrument status for different types of measurement tasks; wherein:

[0189] The deep learning-based object detection and pose estimation algorithm is used to verify the sample position. The deep learning-based object detection and pose estimation algorithm is shown in Equation (8):

[0190] (8);

[0191] The instrument state recognition is achieved by state comparison based on the Siamese network, as shown in equation (9):

[0192] (9);

[0193] (v) An automated operation module (105) based on deep reinforcement learning control, used for learning automated operation strategies based on the improved PPO algorithm and for automated operation of measurement tasks and result acquisition based on deep reinforcement learning control; wherein, the improved PPO algorithm is shown in equation (10):

[0194] (10);

[0195] (vi) Multimodal anomaly detection module (106), used to perform multimodal anomaly detection on the collected results based on graph neural network fusion of multi-source data.

[0196] (vii) A federated quality assessment module (107) is used to perform data quality assessment corresponding to the measurement task based on privacy-preserving distributed learning, as shown in Equation (12):

[0197] (12).

[0198] Application Example 1

[0199] This application example 1 is a specific implementation method for initiating and collecting batch ion chromatography measurement tasks in a water quality testing laboratory. The application scenario is that a municipal water quality testing center conducts batch ion chromatography detection of indicators such as fluoride ions, chloride ions, nitrate ions, and sulfate ions on samples of drinking water, surface water, and industrial wastewater. The laboratory is equipped with three ion chromatographs of different brands and one autosampler, and can process an average of 80 to 120 samples per day.

[0200] During implementation, knowledge augmentation is first performed using S1 based on RAG. Unstructured data such as historical operation manuals, expert experience, and failure cases are converted into 768-dimensional vectors using the BERT model to build a knowledge vectorization library, based on the formula. Complete the knowledge vectorization process, based on the current task context and according to the formula Calculate the similarity score, where the weight coefficients are... Taking 0.2, and combining it with the weight of knowledge timeliness to complete the state knowledge retrieval, then according to the formula... A sequence of operation instructions matching the surface water ion chromatography detection is generated to complete intelligent decision support; subsequently, S2 task scheduling based on hypergraph isomorphic computation is executed to construct a hypergraph whose vertex set contains task nodes, instrument nodes, time nodes, and resource nodes, and whose hyperedge set contains task constraints, resource dependencies, and temporal relationships. = (V, E), according to the formula The matching degree between the task requirement hypergraph and the resource supply hypergraph was calculated to be approximately 0.615, which meets the matching requirements. Furthermore, according to the formula... The time, load, and priority constraint weights were set to 0.5, 0.3, and 0.2 respectively to minimize constraint violations and complete multi-instrument collaborative scheduling, rationally allocating 85 samples to three ion chromatographs for parallel detection. Next, S3, based on a multi-expert hybrid model, was executed to identify and initiate the measurement task type, constructing dedicated expert networks adapted for low-concentration ion detection in drinking water, conventional ion detection in surface water, and high-concentration ion detection in wastewater, according to the formula... Build an expert network structure, using formulas Gating weights were assigned, and the calculated expert weight for conventional ions in surface water reached 0.78. The task type was automatically identified, and the corresponding detection parameters were loaded to complete the task initiation. Then, S4 visual recognition and automated control were executed, using deep learning target detection and pose estimation methods according to the formula. Verify sample location, among which , , According to the formula via the Siamese network Perform instrument status identification, among which , After confirming that the sample position is correct and the instrument is ready, execute the S5 automated operation based on deep reinforcement learning control, using the improved PPO algorithm according to the formula. Conduct strategy learning, in which , If the divergence threshold is less than 0.02, the entire process of sample information entry, parameter configuration, and measurement initiation is automatically completed; then, S6 multimodal anomaly detection is performed, using a graph neural network according to the formula. By integrating multi-source data such as chromatograms, pressure, temperature, and timestamps, the system identifies anomalies such as baseline drift and pressure exceeding thresholds in real time and triggers processing mechanisms. Finally, it performs an S7 federal quality assessment according to the formula. Privacy-preserving distributed learning is employed to complete data quality assessment, ensuring that the test results meet the requirements of GB / T5750-2022 standard. After implementation, this method can reduce batch task processing time by 85%, manual intervention by 90%, and the operational error rate to below 0.04%, while increasing instrument utilization from 42% to 86%, achieving fully automated start-up, execution, result acquisition, and quality control for batch ion chromatography measurement tasks.

[0201] Application Example 2

[0202] This application example 2 illustrates a specific implementation method for initiating and collecting results of a batch silicate / total organic carbon (TOC) joint measurement task in a power laboratory. The application scenario involves a provincial-level power research institute conducting batch detection of silicate and TOC on samples from the steam and water systems of thermal power plants and hydropower stations. The laboratory is equipped with two silicate analyzers, two TOC analyzers, and an automated solution preparation system, processing an average of 60 to 90 samples per day. The aim is to address technical issues such as large errors in manual solution preparation, difficulties in coordinating multiple instruments, data dispersion, and missed anomalies. In implementation, firstly, knowledge augmentation based on RAG is performed using S1. Unstructured data such as power plant steam and water detection procedures, abnormal silicate color development, high TOC blank values, and ultrapure water system malfunctions are converted into 768-dimensional vectors using a BERT model, and then processed according to the formula... Construct a dedicated knowledge vectorization library for the power industry, based on the current water vapor sample detection task context and according to the formula Calculate the similarity score, where the weight coefficients are... Using a weight of 0.2 and a timeliness weight of 0.92, relevant knowledge such as low-concentration silicate color development, total organic carbon purging, and ultrapure water load matching is dynamically retrieved, and then applied according to the formula. A joint detection operation instruction is generated, specifying key parameters such as color development time, oxidation temperature, and ultrapure water supply strategy. Subsequently, task scheduling based on hypergraph isomorphism computation (S2) is executed, constructing a vertex set containing task nodes, silicate analyzer nodes, total organic carbon analyzer nodes, ultrapure water resource nodes, and solution preparation system nodes, as well as a hyperedge set containing time-series dependencies, ultrapure water constraints, solution preparation exclusivity, and instrument maintenance cycles, according to the formula... The hypergraph matching degree was calculated to be 0.625, which meets the requirements for collaborative scheduling. Furthermore, according to the formula... The constraints for time, ultrapure water, and solution preparation efficiency were set to weights of 0.4, 0.4, and 0.2, respectively. To minimize the degree of constraint violation, 72 water vapor samples were evenly distributed to various detection instruments, achieving multi-device joint batch scheduling. Next, S3, based on a multi-expert hybrid model, was executed to identify and initiate the measurement task type. Dedicated expert networks were constructed to adapt to low silicate concentration detection, high-temperature combustion detection of total organic carbon, and joint water vapor index detection, according to the formula... Construct an expert network model using formulas Gating weights are assigned, and the calculated expert weight for joint water vapor detection reaches 0.56. The task type is automatically identified, and dual-expert collaborative detection is initiated. Then, S4 visual recognition and automated control are executed, using deep learning target detection and pose estimation methods according to the formula. Completed the verification of the positions of the colorimetric tubes and sample boat, among which , , According to the formula via the Siamese network Perform instrument status identification, among which After confirming accurate sample positioning and normal instrument status, the S5 automated operation based on deep reinforcement learning control is executed, using the improved PPO algorithm according to the formula. Optimize operational strategies, among which , With a divergence threshold less than 0.015, the system automatically completes solution preparation, drug addition, color development, simultaneous startup of multiple instruments, and result acquisition. Subsequently, it performs S6 multimodal anomaly detection, using a graph neural network according to the formula... Integrating multimodal data such as spectrum, temperature, flow rate, and time, the system identifies anomalies in real time, such as incomplete color development, high blank values, and insufficient carrier gas, and automatically triggers retesting; finally, it performs an S7 federal quality assessment according to the formula. The results quality assessment was completed using a distributed federated model, ensuring that the relative standard deviation of silicate was less than 1.5% and the total organic carbon recovery rate was between 94% and 105%, meeting the DL / T502-2021 power industry standard. After implementation, this method can reduce batch processing time by 62.5%, reduce manual intervention by 91%, lower the data error rate to below 0.03%, increase instrument utilization from 46% to 88%, and reduce the anomaly miss rate from 8.2% to 0.4%, achieving intelligent management of the entire process of joint silicate / total organic carbon measurement in the power laboratory.

[0203] Through the above description of the embodiments, those skilled in the art can clearly understand that the above embodiments can be implemented by software, or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of the above embodiments can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.), including several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0204] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A laboratory management method for initiating batch measurement tasks and collecting results, characterized in that, include: S1, knowledge enhancement based on RAG, thereby providing intelligent decision support for measurement tasks through the knowledge enhancement; including: knowledge retrieval based on a knowledge vectorization library and a dynamic retrieval mechanism, and converting unstructured data into vector representations, wherein the unstructured data includes one or more of historical operation manuals, expert experience, and fault cases; S2, based on hypergraph isomorphic computation, performs task scheduling steps, models laboratory resource scheduling as a hypergraph matching problem, and realizes multi-instrument collaborative scheduling; S3, Measurement task type identification and initiation based on multi-expert hybrid model, including: an expert network based on multi-expert hybrid model identifies measurement task types and initiates different types of measurement tasks in different ways; S4, Visual recognition and automated control, including: verifying sample position and identifying instrument status for different types of measurement tasks using visual technology; S5, Automated operation based on deep reinforcement learning control, including: learning automated operation strategies based on the improved PPO algorithm and automating measurement tasks and result acquisition based on deep reinforcement learning control. S6, perform multimodal anomaly detection on the collected results based on graph neural network fusion of multi-source data; S7, Federal Quality Assessment, includes: conducting a data quality assessment corresponding to the measurement task based on privacy-preserving distributed learning.

2. The laboratory management method for batch measurement task initiation and result acquisition according to claim 1, characterized in that, S1 includes: S11, the knowledge is vectorized by converting unstructured data into vector representation and embedding it into text using the BERT model. The knowledge vectorization library is represented by equation (1) as follows: （1）； in, The input is unstructured text data, including historical operation manuals, expert experience, and failure cases; S12, perform dynamic knowledge retrieval on the vectorized knowledge using a dynamic retrieval mechanism. This dynamic retrieval mechanism is used to retrieve relevant knowledge fragments based on the current task context; including: retrieving relevant knowledge fragments based on the current task context and calculating a similarity score between the retrieved relevant knowledge fragments and the current task context. : （2）； in: Indicates the weight of knowledge timeliness; These are the weighting coefficients; S13, Based on the relevant knowledge fragments, perform generative decision-making to generate an operation instruction sequence. As shown in equation (3): （3）； in, For retrieving relevant knowledge sets, This is a description of the current task. This is the generated sequence of operation instructions.

3. The laboratory management method for batch measurement task initiation and result acquisition according to claim 2, characterized in that, S2 includes: S21, Construct the hypergraph H=(V,E), including: (1) Establish a vertex set V = {task node, instrument node, time node, resource node}; (2) Establish the hyperedge set E = {task constraints, resource dependencies, and temporal relationships}; S22, Establish a hypergraph isomorphism algorithm to calculate the hypergraph matching degree between task requirements and resource supply, as shown in equation (4): （4）； in, This represents the hypergraph matching degree or hypergraph isomorphic matching value, characterizing the task requirement hypergraph. With resource supply supermap The overall matching degree between them, the larger the value, the better the matching effect between task and resource supply and demand; This represents a task requirement hypergraph, which is a set of task requirements to be assigned modeled in the form of a hypergraph. Hyperedges are used to represent multi-dimensional constraints and relationships within the task. Represents a resource supply hypergraph, which is a set of available resource supplies modeled in the form of a hypergraph. Hyperedges are used to represent resource clusters and resource combination constraints. This represents the mapping relationship between nodes in the hypergraph, defining the hypergraph from the task requirements. For mapping The set of matching hyperedges below; nodes to resource supply hypergraph The one-to-one mapping rule for nodes; This indicates traversing all feasible mappings. The optimal matching result is the one with the highest matching degree. For mapping The set of matching hyperedges under the node mapping At the time of its establishment, the task requirements exceeded the map. With resource supply supermap A set of hyperedges whose structure, attributes, and constraints can be matched one-to-one; This represents the cardinality or number of the matching hyperedge set, i.e., the mapping. The total number of superedges that can be successfully matched; Indicates the task requirement hypergraph The set of all superedges; This represents the total number of all hyperedges in the task hypergraph; Indicates resource supply exceeding the chart The set of all superedges; This represents the total number of all hyperedges in the resource hypergraph. S23, establish constraints to satisfy the optimization, in order to minimize the degree of constraint violation, as shown in equation (5): （5）； in, For the first The degree of violation of each constraint is related to the decision variables. The function is used to quantize the first step in the current scheme. The extent to which a constraint is violated; when constraints are fully satisfied. The more severe the violation, the larger the value. The index variable is used to iterate through all constraints, where the constraint number is the index variable. This represents the total number of all constraints, i.e., the total number of constraints that need to be satisfied in the optimization problem. For the first The weight coefficient of a constraint represents its importance, priority, or penalty; the more critical the constraint, the higher its weight. The larger the value, the higher the penalty for violation; These are decision variables / system state variables, representing the variables to be solved in this optimization problem, such as task-resource matching schemes, resource allocation strategies, and hypergraph mapping parameters.

4. The laboratory management method for batch measurement task initiation and result acquisition according to claim 3, characterized in that, S3 includes: S31, construct dedicated expert networks, each optimized for a specific type of measurement task, as shown in Equation (6): （6）； in, Indicates the first A dedicated expert network, with feature vectors as input. The output is the expert network's prediction / decision results for a specific measurement task; each For each specific type of measurement task, network parameters are optimized independently. This indicates the expert network number, used to distinguish various specialized expert subnetworks for different types of measurement tasks; The input feature vector to the network is a multi-dimensional input vector composed of the original data, task features, resource features, or hypergraph matching features corresponding to the measurement task. Indicates the first The first layer of an expert network, i.e., the hidden layer weight matrix, is used to weight the input features. Perform a linear transformation; Indicates the first The bias vector of the first layer of the expert network, together with the weight matrix, completes the linear mapping of the hidden layer; The activation function is represented by nonlinear transformation, which improves the ability of expert networks to fit the features of complex measurement tasks. This includes the use of activation functions such as ReLU, Sigmoid, and Tanh. Indicates the first The second layer of the expert network, i.e., the output layer weight matrix, performs a linear transformation on the output of the hidden layer. Indicates the first The second layer bias vector of the expert network is used for linear compensation of the output layer; Indicates the first Each expert network has its own dedicated output mapping function, customized for the corresponding type of measurement task. It can realize the functions of normalization, classification output, regression prediction, and matching scoring, and adapt to the output requirements of specific tasks. S32, assign the gating network, and use the softmax function to assign expert weights, as shown in equation (7): （7）； in, The output of the gating network represents the first... The weights of the nth expert network are in the range of [0,1], and the sum of all expert weights is 1. These weights are used to characterize the nth expert network weight under the current input feature. The contribution percentage of each dedicated expert network; This indicates that it corresponds to the first step in S31 of the previous step. A dedicated network of experts for specific measurement tasks; This represents the traversal index of all expert networks, traversing all available expert subnetworks; The input feature vector is consistent with the input in equation (6) and is a multidimensional feature vector related to the measurement task, resource supply, and hypergraph matching. Indicating the first in the gating network The transpose of the weight vector corresponding to each expert is the trainable parameters of the gating network, used to adjust the input features. Perform linear weighting; This represents the natural exponential function, which, when used with softmax, normalizes the weights and ensures that the output is non-negative. Indicating the first in the gating network Transpose the weight vector corresponding to each expert, and iterate through the gating weight parameters of all experts; This represents summing the exponential terms of all expert networks to achieve softmax normalization, so that the weighted sum of the weights of each expert is 1; S33, Task type identification and initiation, including: automatically identifying task types based on task characteristics and selecting the most suitable expert network combination.

5. The laboratory management method for batch measurement task initiation and result acquisition according to claim 4, characterized in that, S4 includes: S41, the sample position verification is achieved through a deep learning-based object detection and pose estimation method, as shown in equation (8): （8）； in, This represents the total loss function for detection and pose estimation, used for deep learning model training. It characterizes the error between the sample target position, pose prediction results, and the true label. The model training aims to minimize this loss to achieve sample position verification. This represents the coordinate loss weight coefficient, a hyperparameter used to adjust the proportion of the target box coordinate position loss in the overall loss, amplifying / weakening the impact of positional deviation on model training; The feature map grid size is used to divide the input image into... The grid, The total number of grid cells. Traverse all grid cells; Represents the grid cell index, traversing all the grid cells after image partitioning. One grid; This indicates the number of predefined bounding boxes or anchor boxes within a single grid cell, meaning the maximum number of predictions per grid cell. One sample target bounding box; This represents the bounding box index within a single grid, and iterates through all preset bounding boxes within a single grid. This indicates that the target has an indicator function or a binary mask, with a value of 0 or 1; when the... The first grid The value is 1 if the bounding box contains the real target of the sample, and 0 otherwise. The position loss is calculated only for the box containing the target. and Represents the true center coordinates of the sample target, the first... The x and y coordinates of the center point of the true target bounding box of the sample within each grid; and The x and y coordinates represent the center coordinates of the sample target predicted by the model, and the corresponding center coordinates of the target box output by the deep learning network. The squared error of the target center point coordinates represents the sample position prediction deviation. The remaining terms of the loss function include target confidence loss, class classification loss, pose angle loss, and bounding box width and height loss, which are used for integrity-constrained sample detection, classification, and pose estimation. S42, the instrument state identification is achieved by state comparison based on the Siamese network, as shown in equation (9): （9）； in, This represents the contrastive loss function of the Siamese network, used for model training loss in instrument state comparison and state recognition. Optimizing this loss makes the distance between features of the same instrument state smaller and the distance between features of different instruments state larger. This represents the total number of training sample pairs, i.e., the total number of instrument state feature sample pairs participating in model training; Use sample pair indexes to iterate through all instrument status sample pairs; Represents a sample pair labeled, a binary variable; when the two input samples are in the same instrument state... For different instrument states ; The distance represents the Euclidean distance between two instrument state feature vectors, i.e., the feature similarity distance, which characterizes the degree of difference between the two sets of instrument state features extracted by the Siamese network. The smaller the distance, the more similar the states are. This represents the boundary threshold or interval hyperparameter, a manually set safety distance threshold; used to constrain the minimum feature distance between sample pairs in different states to avoid feature aliasing. This represents the margin loss term, where the feature distance between different sample pairs... Less than the set threshold Loss occurs when the distance is greater than the threshold; the greater the distance, the greater the loss. When this item is 0, no further penalty is applied; The normalization coefficient is the average loss of all sample pairs and scaled to stabilize the training gradient and unify the magnitude of the loss. For positive sample loss terms, for sample pairs with the same instrument state, minimize the feature distance to make the features of the same state as close as possible; This represents the negative sample loss term. For sample pairs with different instrument states, it forces the feature distance to be greater than a set threshold, so that features of different states can be distinguished as much as possible.

6. The laboratory management method for batch measurement task initiation and result acquisition according to claim 5, characterized in that, S5 includes: S51, Automated operation strategy learning is performed based on the improved PPO algorithm, and the improved PPO algorithm for strategy learning is shown in equation (10): （10）； in, Indicates the belt An improved PPO policy loss function with divergence penalty is used for automated operational policy learning. By constraining the difference in distributions between the old and new policies, it stabilizes the policy update process. The network parameters for the current policy to be updated; The expectation operator represents taking the mathematical expectation of the state-action trajectory sample distribution and calculating the average loss based on sampling experience; The network parameters of the current policy to be updated are used to improve the neural network parameters to be iteratively optimized in PPO, which are used to output the probability of automated operation actions. This represents the network parameters of the old strategy. The parameters of the previous strategy are updated and used as a reference to prevent the strategy update from being too large. Indicates the current policy, in the state Below, the parameters are The strategy network output action The probability of; This represents the old strategy, in state Below, the old parameter policy network outputs actions. The probability of; The importance sampling ratio represents the probability ratio of the same action under the old and new strategies, and is used to measure the magnitude of strategy updates. The environmental state refers to the system state that matches the instrument, sample, and task during automated operation. The actions performed by the intelligent agent correspond to various automated operation instructions and control behaviors; The advantage function estimate at time t represents the payoff advantage of the current action compared to the average action, and is used to guide the strategy to update to higher-return actions; for The divergence penalty coefficient is a hyperparameter used to control the strength of the penalty for the distribution difference between the old and new strategies. The larger the value, the stronger the constraint. for Divergence relative entropy, used to measure the old strategy With new strategies In the same state The degree of difference in the distribution of lower action; Representing the given states respectively At that time, the probability distribution of the complete actions corresponding to the old strategy and the new strategy; S52 is an automated operation for measurement tasks and result acquisition based on deep reinforcement learning control, including: automatically executing sample information entry, parameter setting, and measurement start operations based on the learning strategy.

7. The laboratory management method for batch measurement task initiation and result acquisition according to claim 6, characterized in that, S6 includes: S61, Multimodal feature fusion is performed based on graph neural network and multimodal anomaly detection algorithm, the multimodal anomaly detection algorithm is shown in equation (11): （11）； in, Indicates the first Nodes in a layered graph neural network The multimodal fusion feature vector has nodes corresponding to entities such as instruments, samples, and tasks. The features fuse multimodal information from images, text, and sensor data for subsequent anomaly detection. Indicates the first Neighbor nodes in a layered graph neural network The original multimodal feature vector is the node feature output from the previous layer; This is the layer index of the graph neural network, indicating the current feature propagation layer number; The central node index represents the target entity node whose feature needs to be updated; Indexing neighbor nodes, traversing nodes All adjacent nodes that have a relationship; For nodes The set of first-order neighbors, representing the set of neighbors of the current node. All connected related nodes; It is a trainable weight parameter matrix, a feature transformation matrix of a graph neural network, used to perform linear mapping and dimension adjustment on the aggregated neighbor features, and adapt to multimodal feature fusion. For nodes The normalized degree of a node in a graph structure. The number of connected neighbors is used for node degree normalization; Neighboring nodes Normalized degree; This is the degree matrix normalization term. It normalizes the feature aggregation by taking the square root of the product of the node degrees, so as to avoid high-connectivity nodes contributing too much to the features and balance the weights of different nodes. For neighbor feature aggregation operations, the normalized features of all neighboring nodes are accumulated to achieve local neighborhood information fusion; S62 monitors the measurement process in real time, identifies abnormal patterns, and triggers corresponding processing mechanisms.

8. The laboratory management method for batch measurement task initiation and result acquisition according to claim 7, characterized in that, S7 includes: S71, the federated model is trained using a privacy-preserving distributed learning algorithm, as shown in Equation (12): （12）； S72 evaluates the quality of measurement results based on the trained federated model and provides feedback to optimize the measurement process.

9. A laboratory management system for batch measurement task initiation and result acquisition, used to implement the method described in any one of claims 1-8, characterized in that, include: The RAG-based knowledge enhancement module (101) is used to perform knowledge enhancement based on RAG, thereby providing intelligent decision support for measurement tasks through the knowledge enhancement. The task scheduling module (102) based on hypergraph isomorphic computation is used to perform task scheduling steps based on hypergraph isomorphic computation, modeling laboratory resource scheduling as a hypergraph matching problem, and realizing multi-instrument collaborative scheduling. The measurement task type identification and initiation module (103) based on the multi-expert hybrid model is used for measurement task type identification and initiation based on the multi-expert hybrid model, including: the expert network based on the multi-expert hybrid model identifies the measurement task type and initiates different types of measurement tasks in different ways; A visual recognition and automation control module (104) is used for visual recognition and automation control, including: verifying the sample position and identifying the instrument status for different types of measurement tasks using visual technology; The automated operation module (105) based on deep reinforcement learning control is used to perform automated operations based on deep reinforcement learning control, including: learning automated operation strategies based on the improved PPO algorithm and performing automated operations of measurement tasks and result acquisition based on deep reinforcement learning control. A multimodal anomaly detection module (106) is used to perform multimodal anomaly detection on the collected results based on graph neural network fusion of multi-source data; The Federal Quality Assessment Module (107) is used for federal quality assessment, including: performing data quality assessment corresponding to the measurement task based on privacy-preserving distributed learning.

10. A laboratory management system for batch measurement task initiation and result acquisition according to claim 9, characterized in that, The RAG-based knowledge enhancement module (101) consists of two parts: a knowledge retrieval unit and a knowledge enhancement generation unit. The knowledge retrieval unit performs knowledge retrieval based on a knowledge vectorization library and a dynamic retrieval mechanism. The knowledge vectorization library is used to convert unstructured data into vector representations. The unstructured data includes one or more of historical operation manuals, expert experience, and fault cases.

Images