Laboratory risk control method based on graph neural network

Abstract

The invention discloses a laboratory risk control method based on a graph neural network, and the method comprises the following steps: S1, constructing a dynamically updated interaction event graph based on an interaction event sequence of laboratory equipment, personnel and environment; s2, dynamically updating the state space of each node according to the event preorder relation, and generating a synchronous node state sequence; s3, generating an early warning feature set of a node state change trend by adopting a graph convolution operation; s4, constructing an inter-node risk incidence matrix according to the early warning feature set, and generating an updated risk propagation path; s5, evaluating the probability of transition from the node state to the abnormal state, and generating a state risk prediction value; and S6, inputting a wolf pack optimization algorithm, and optimizing a parameter combination to obtain a laboratory risk prediction result. According to the invention, the real-time performance and accuracy of laboratory risk prediction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of risk management, and particularly relates to a laboratory risk control method based on a graph neural network. BACKGROUND

[0002] Laboratory risk control technology is widely used to ensure the safety of laboratory equipment, environment and personnel. Traditional risk control methods mainly rely on physical monitoring and simple rule systems. These methods usually judge laboratory safety through equipment state monitoring, environmental parameter detection and personnel behavior management. Although these methods can play a role in early warning to some extent, they mostly rely on single-dimensional information such as temperature, humidity and other parameters, and are difficult to effectively handle complex laboratory environments with multiple factor interactions.

[0003] With the development of artificial intelligence, especially deep learning technology, data-driven laboratory risk control methods have gradually attracted attention. Graph neural networks, as a powerful model, can effectively handle complex relationships and dependencies between nodes. In laboratory risk control, equipment, personnel, environment and other factors can be regarded as nodes in the graph, and the interaction between nodes is connected through the edges of the graph. GNN can capture these complex relationships and is considered an effective tool to improve the accuracy of laboratory risk assessment and prediction.

[0004] However, the main problem of existing technology when applying GNN is its poor adaptability to dynamic laboratory environments. Traditional GNN models usually rely on static graph structures and are difficult to respond to environmental changes in real time. In addition, existing methods often ignore the complexity of interactions between factors, limiting the predictive ability of the system. Therefore, how to combine GNN with optimization algorithms to improve the real-time performance and accuracy of laboratory risk control is a problem that needs to be solved in the current technical field.

[0005] Therefore, how to provide a laboratory risk control method based on a graph neural network is a problem that needs to be solved by those skilled in the art. SUMMARY

[0006] One object of the present application is to provide a laboratory risk control method based on a graph neural network combined with a wolf optimization algorithm. The present application adopts a technical means combining a graph neural network and a wolf optimization algorithm, constructs a dynamically updated interaction event graph, captures the complex relationships between laboratory equipment, environment and personnel in real time, and optimizes the parameter combination of graph convolution operation using WOA, thereby improving the accuracy and adaptability of laboratory risk prediction. This method effectively solves the problem of poor adaptability of traditional methods to changes in laboratory environments by evaluating the probability of node state transition in real time and dynamically generating risk prediction values.

[0007] According to an embodiment of the present application, a laboratory risk control method based on a graph neural network comprises the following steps:

[0008] S1, based on the interaction event sequence between laboratory equipment, personnel and environment, a dynamically updated interaction event graph is constructed to form a node connection relationship in the time sequence of event occurrence;

[0009] S2, define a state space for each node in the interaction event graph, dynamically update the state space of the node according to the antecedent relationship of event occurrence, and generate a node state sequence synchronized with time;

[0010] S3, based on the node state sequence, a graph convolution operation is used to generate a warning feature set for the trend of node state change within a fixed time window in the future;

[0011] S4, a time-evolving node risk correlation matrix is constructed according to the warning feature set, and the risk correlation strength between nodes is dynamically updated to generate a real-time updated node risk propagation path;

[0012] S5, according to the real-time updated node risk propagation path, the probability of the state of each node changing to an abnormal state is gradually evaluated through the warning feature set, and the state risk prediction value of each node is dynamically generated;

[0013] S6, input the state risk prediction value of each node into the wolf swarm optimization algorithm, take the minimization of the deviation between the risk prediction value and the real node state as the optimization objective, continuously correct the warning feature generation mechanism, dynamically optimize the parameter combination of the graph convolution operation, and obtain the laboratory risk prediction result.

[0014] Optionally, the S1 specifically comprises:

[0015] S11, collect the running event data of all laboratory equipment, dynamically time sequence cluster according to the periodic characteristics of event repetition, and form the periodic characteristic sequence of equipment running events;

[0016] S12, for laboratory environment data, a multi-dimensional change trend synchronous detection mechanism is used to determine the multi-dimensional feature intersection point at which the environment index simultaneously reaches the threshold in dynamic change in real time, and the multi-dimensional feature intersection point is marked as an environment state mutation event;

[0017] S13, according to the correlation analysis of personnel operation behavior type and real-time influence on equipment running state, according to the operation action leading to equipment state change amplitude and response time, the grading rules of operation event influence on equipment state are constructed;

[0018] S14, according to the device cycle characteristic sequence, the environmental state mutation event mark and the grading rule of the influence of the operation event on the device state, the dynamic event chain fusion is carried out, and the event fusion sequence which can reflect the interaction of the device, the environment and the personnel and has time sequence is formed;

[0019] S15, each event in the event fusion sequence is defined as a graph node with an influence factor weight, and a quantitative correlation connection relationship between the nodes is constructed in turn according to the size of the influence factor weight, forming an interactive event graph with a quantitative influence factor weight, and the influence factor weight is determined according to the historical contribution rate of the event to the risk occurrence and the time sequence closeness between events;

[0020] S16, using a real-time dynamic sliding window adaptive mechanism, taking the change rate of the influence factor weight of each node event in the window exceeding a preset threshold as a criterion, the quantitative correlation connection relationship between the nodes of the interactive event graph is updated in real time.

[0021] Optionally, the S2 specifically comprises:

[0022] S21, the basic state space and the derived state space are simultaneously defined for each node state space in the interactive event graph, the basic state space includes event type identification, event duration value, event intensity level, and the derived state space is dynamically generated according to the real-time calculated correlation strength value between the current node and the adjacent node;

[0023] S22, according to the number of the previous nodes of the node in the event chain and the influence factor weight between the nodes, the state dependence depth value of each node is calculated as a supplementary feature of the node state space;

[0024] S23, the state diffusion coefficient of each node is calculated by monitoring the propagation speed of the node state feature to the adjacent node in real time, and the state diffusion coefficient value reflects the immediate influence degree of the node state change on the subsequent node;

[0025] S24, an independent state evolution track is constructed for each event type in the node basic state space, and the state evolution trend of the node state is predicted in real time by using the historical change data of the state space feature, and then the event intensity level and the event duration value of the node state space are dynamically adjusted;

[0026] S25, based on the real-time autocorrelation analysis of the change of the node state space feature, the state stability index value of each node is calculated, and when the state stability index value is lower than the predetermined stability threshold, the node state space is immediately updated;

[0027] S26, according to the state dependence depth value, the state diffusion coefficient value and the state stability index value of each node, a node state sequence strictly synchronized with time is generated in sequence according to the corresponding event accurate time stamp of the node.

[0028] Optionally, the S3 specifically comprises:

[0029] S31, based on the node state sequence, generating a node state trend curve with the numerical change of each node state feature in a time window, and obtaining the continuous evolution feature of the node state by using a curve fitting method;

[0030] S32, calculating the slope, first derivative and second derivative values of each node state trend curve, respectively representing the change rate, change acceleration and change trend stability of the node state, forming a trend dynamics feature set of the node state;

[0031] S33, calculating the similarity value of the node state change between the cosine similarity of the node state trend curve, and constructing a node state trend correlation matrix according to the similarity value, wherein each element value in the correlation matrix represents the similarity of the state trend between nodes;

[0032] S34, determining the aggregation weight of the graph convolution operation by using the node state trend correlation matrix, taking the trend dynamics feature set of the node with high similarity as the weight, performing a weighted graph convolution operation on the target node state feature set, and updating the node state space;

[0033] S35, calculating the trend sensitivity value of the node state by using the difference value of the trend dynamics feature before and after the node state update, and screening the node features after the graph convolution operation by using the trend sensitivity value, and extracting the node state with the trend sensitivity value exceeding a preset sensitivity threshold as a key node early warning feature;

[0034] S36, weighting and aggregating the key node early warning features according to the trend sensitivity value to form an early warning feature set containing the node individual and node group state change sensitivity information.

[0035] Optionally, the S31 specifically comprises:

[0036] S311, dividing the state feature values of the node state sequence into a plurality of micro trend segments in a fixed time window at a set time interval, each micro trend segment being composed of two adjacent continuous state feature values, and recording the feature change direction and change amplitude value of each micro trend segment;

[0037] S312, defining a micro trend mode category according to the change direction and amplitude value of the micro trend segment, and clustering and dividing all micro trend segments according to the category to form a micro trend mode set of the node state feature value;

[0038] S313, calculate the transition probability values between the micro-trend patterns according to the frequency and time sequence of the micro-trend pattern in the time window, and construct a micro-trend transition probability matrix with the micro-trend pattern as the node and the transition probability value as the edge weight;

[0039] S314, define the continuous combination of each possible micro-trend pattern as a node state trend evolution path based on the micro-trend transition probability matrix, and calculate the entropy value of the transition probability value of each path using the information entropy calculation method, determine the uncertainty of the path according to the information entropy value, and select the trend path with the smallest information entropy and the highest stability;

[0040] S315, according to the trend path information entropy calculation result, filter out the trend evolution path with the lowest information entropy value and the highest certainty, and define it as the optimal stable trend path of the node state characteristics;

[0041] S316, take the feature change amplitude value of each micro-trend pattern in the optimal stable trend path as the fitting reference point, and generate the continuous evolution characteristics of the node state by weighted interpolation smoothing fitting.

[0042] Optionally, the S4 specifically comprises:

[0043] S41, by adaptively segmenting the time sequence of the node state characteristic value, a dynamic time window model is constructed, the time lag effect of the state change between nodes is calculated in each time window, the lag effect factor is generated in real time, and the lag effect factor is used as a correction factor of the node state correlation strength to adjust the risk correlation strength between nodes;

[0044] S42, for the risk correlation of each pair of nodes, a multi-level weighted matrix modeling is performed to preliminarily define the direct risk propagation path between nodes, and a reverse propagation strategy is used to adaptively adjust the path weight in the matrix to form a multi-level, multi-stage complex risk propagation path;

[0045] S43, the nonlinear relationship among the laboratory equipment, environment and personnel is comprehensively considered, and the heterogeneous influence factor of the upstream and downstream nodes is introduced in the construction process of the risk correlation matrix, so that the cross-layer influence of each node on the adjacent nodes is dynamically calculated, and the risk propagation direction between nodes is adjusted;

[0046] S44, based on the dynamically adjusted risk correlation matrix, the risk propagation path between nodes is geometrically modeled by using the space embedding technology, and an adaptive path mapping strategy is introduced, so that the weight and direction of the path can be real-time corrected according to the dynamic change of the laboratory environment;

[0047] S45, through multiple rounds of iterative optimization calculation, real-time update of the propagation ability of each node and the correlation strength between nodes, driven by changes in the laboratory environment, equipment and personnel state, dynamically generating global risk propagation path.

[0048] Optionally, the S5 specifically includes:

[0049] S51, according to the real-time updated node risk propagation path, calculate the potential influence of the state change of each node on the adjacent node, and generate the preliminary probability estimate of the state transition of the node by weighting and accumulating the state propagation coefficient between adjacent nodes;

[0050] S52, based on the state change characteristics of each node, the propagation path between nodes and the time sequence relationship between node states, the weighted average method is used to smooth the preliminary transition probability of the node, and the instantaneous fluctuation is eliminated;

[0051] S53, using historical node state transition data, fitting the state change of each node through regression analysis model, dynamically adjusting the prediction function of node state transition, and adjusting the update frequency and update strategy of transition probability in real time according to the change of environment;

[0052] S54, according to the updated transition probability, using Bayesian inference method to correct the state risk prediction value of each node according to the real-time data of the laboratory, forming the risk prediction model based on the current laboratory environment and equipment state;

[0053] S55, according to the prediction value, calculate the state change amplitude of each node, and determine the abnormal state risk index of the node according to the comprehensive evaluation value of the state change amplitude and the transition probability of the node;

[0054] S56, according to the abnormal state risk index of the node, dynamically generating the state risk prediction value of each node in the laboratory.

[0055] Optionally, the S55 specifically includes:

[0056] S551, based on the real-time prediction value of each node, calculate the change amplitude of the node state between the current time window and the last time window;

[0057] S552, according to the node state change amplitude, combined with the transition probability of the node, define the node state risk index as the weighted product of the node state change amplitude and the node transition probability;

[0058] S553, according to the node state risk index, mark the node whose risk index exceeds the predetermined threshold as a high-risk node, and assign a corresponding risk level to each node, the risk level is divided into low risk, medium risk and high risk, and the nodes are managed according to the risk level.

[0059] Optionally, the S6 specifically comprises:

[0060] S61, calculating a deviation value of each node according to the deviation between the state risk prediction value of each node and the actual node state;

[0061] S62, inputting the wolf swarm optimization algorithm for global optimization with the deviation value of each node as a target function, and calculating an optimal weight factor of each node;

[0062] S63, updating the weight parameter in the graph convolution operation according to the optimal weight factor, adjusting the information propagation strength between nodes, and correcting the graph convolution operation with the optimized weight parameter;

[0063] S64, adopting an adaptive mechanism of the wolf swarm optimization algorithm to dynamically adjust the learning rate and the exploration range in each optimization process;

[0064] S65, gradually correcting the early warning feature generation mechanism through multiple rounds of iterative optimization, and generating a new laboratory risk prediction result based on the new optimized parameters after each update.

[0065] Optionally, the S64 specifically comprises:

[0066] S641, in each optimization process, calculating an adjustment factor of the current learning rate according to the error value of the current optimization result and the convergence condition of the historical result, and adjusting the learning rate with the adjustment factor;

[0067] S642, in each optimization process, dynamically adjusting the exploration range according to the amplitude and stability of the node state change, if the amplitude of the node state change is greater than a preset threshold, expanding the exploration range, and if the amplitude of the node state change is smaller, reducing the exploration range;

[0068] S643, dynamically selecting different exploration strategies according to the current optimization stage and the error convergence speed;

[0069] S644, monitoring the convergence speed of each node in the optimization process in real time, and adjusting the exploration range according to the convergence speed, if the convergence speed of the node in the current optimization step is slower, enhancing the global search and increasing the exploration range, and vice versa.

[0070] The beneficial effects of the present application are:

[0071] (1) The present application realizes real-time modeling and risk prediction of complex relationships between laboratory equipment, environment and personnel by combining graph neural networks and wolf optimization algorithm. The method significantly improves the accuracy and real-time performance of laboratory risk prediction through dynamically updated interaction event graph and adaptive parameter optimization mechanism, breaking through the limitations of traditional methods in dealing with dynamic changes in the laboratory, and optimizing the effect of laboratory safety monitoring.

[0072] (2) The present application effectively integrates multi-dimensional laboratory data through graph convolution operation and node state update mechanism, and optimizes it by combining wolf optimization algorithm, so that the risk assessment model can continuously adapt to changes in different laboratory environments. This method shows stronger adaptability and accuracy in the face of complex and variable laboratory environment, ensuring more accurate and efficient risk control in the laboratory.

[0073] (3) The present application effectively solves the problem of existing technology that cannot respond to changes in laboratory environment and multi-factor interaction in real time by dynamically adjusting node state characteristics and risk propagation path, breaking through the bottleneck of traditional risk control methods in predicting accuracy under dynamic changes in the environment, and greatly improving the intelligent level of laboratory safety management. BRIEF DESCRIPTION OF DRAWINGS

[0074] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. In the drawings:

[0075] Fig. 1 An interaction event graph construction schematic diagram of a laboratory risk control method based on a graph neural network is provided for the present application;

[0076] Fig. 2 A node state update and graph convolution operation flowchart of a laboratory risk control method based on a graph neural network is provided for the present application;

[0077] Fig. 3 A risk propagation path generation and optimization flowchart of a laboratory risk control method based on a graph neural network is provided for the present application. DETAILED DESCRIPTION

[0078] The present application will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, and only illustrate the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0079] REFERENCE Figs. 1-3 A laboratory risk control method based on a graph neural network, comprising the following steps:

[0080] S1, based on the interaction event sequence between laboratory equipment, personnel and environment, a dynamically updated interaction event graph is constructed, and the node connection relationship is formed in the time sequence of event occurrence;

[0081] S2, define the state space of each node in the interaction event graph, dynamically update the state space of the node according to the antecedent relationship of the event occurrence, and generate a node state sequence synchronized with time;

[0082] S3, based on the node state sequence, a graph convolution operation is used to generate a warning feature set for the trend of node state change in a fixed time window in the future;

[0083] S4, a time-evolving node risk correlation matrix is constructed according to the warning feature set, and the risk correlation strength between nodes is dynamically updated to generate a real-time updated node risk propagation path;

[0084] S5, according to the real-time updated node risk propagation path, the probability of the state of each node changing to an abnormal state is gradually evaluated through the warning feature set, and the state risk prediction value of each node is dynamically generated;

[0085] S6, input the state risk prediction value of each node into the wolf swarm optimization algorithm, take the minimization of the deviation between the risk prediction value and the real node state as the optimization target, continuously correct the warning feature generation mechanism, dynamically optimize the parameter combination of the graph convolution operation, and obtain the laboratory risk prediction result.

[0086] By combining the graph neural network and the laboratory risk prediction mechanism, the risk propagation path of each node in the laboratory can be evaluated in real time, and the weight value of the node state feature is dynamically adjusted to provide accurate risk prediction. Compared with the prior art, the present application dynamically constructs and updates the interaction event graph, accurately captures the complex interaction relationship between the laboratory environment, equipment and personnel, and effectively improves the prediction accuracy. In addition, by optimizing the risk propagation path between nodes, the present application can realize real-time response to the change of laboratory state, overcome the problems of poor adaptability and prediction lag in traditional technology in dynamic environment, and improve the real-time and accuracy of the risk prediction result.

[0087] In the embodiment, the S1 specifically comprises:

[0088] S11, collect the running event data of all equipment in the laboratory, dynamically time sequence cluster according to the periodic characteristics of repeated event occurrence, and form the periodic characteristic sequence of equipment running event;

[0089] S12, for laboratory environment data, a multi-dimensional change trend synchronous detection mechanism is used to determine the multi-dimensional feature intersection point at which the environment index simultaneously reaches the threshold in dynamic change in real time, and the multi-dimensional feature intersection point is marked as an environment state mutation event;

[0090] S13, constructing a hierarchical rule of the influence of the operation event on the device state according to the operation behavior type of the personnel and the real-time influence on the device running state, and according to the change amplitude of the device state caused by the operation action and the response time;

[0091] S14, dynamically fusing the event chain according to the device periodic characteristic sequence, the environmental state mutation event marker, and the hierarchical rule of the influence of the operation event on the device state, to form an event fusion sequence that can reflect the interaction of the device, the environment, and the personnel and has a time sequence;

[0092] S15, defining each event in the event fusion sequence as a graph node with an influence factor weight, and sequentially constructing a quantitative correlation connection relationship between the nodes according to the size of the influence factor weight, to form an interactive event graph with a quantitative influence factor weight, wherein the influence factor weight is calculated and determined according to the historical contribution rate of the event to the risk occurrence and the time sequence closeness between events;

[0093] S16, using a real-time dynamic sliding window adaptive mechanism to update the quantitative correlation connection relationship between the nodes of the interactive event graph in real time, according to the criterion that the change rate of the influence factor weight of each node event in the window exceeds a preset threshold.

[0094] Through the laboratory risk control method based on the graph neural network, combined with the dynamically updated event graph and the multi-dimensional risk assessment mechanism, the real-time performance and accuracy of laboratory risk prediction are significantly improved. Compared with the traditional method, the present application can timely reflect the changes of the laboratory environment and the complex interaction between nodes by real-time collection of the interaction data between the laboratory equipment, personnel, and environment, and dynamic construction and update of the interactive event graph. This technical solution effectively overcomes the problems of poor adaptability in dynamic environment and inability to comprehensively consider the interaction of multiple risk factors in the prior art, and enhances the accurate prediction and timely response ability of laboratory risks.

[0095] In the embodiment, the S2 specifically comprises:

[0096] S21, simultaneously defining a basic state space and a derived state space for each node state space in the interactive event graph, wherein the basic state space includes an event type identifier, an event duration value, and an event intensity level, and the derived state space is dynamically generated according to the real-time calculated correlation strength value between the current node and the adjacent node;

[0097] S22, calculating a state dependence depth value of each node as a supplementary feature of the node state space according to the number of previous nodes of the node in the event chain and the influence factor weight between the nodes;

[0098] S23, calculate the state diffusion coefficient of each node by monitoring the propagation speed of the node state feature to the adjacent node in real time, and the state diffusion coefficient value reflects the degree of immediate influence of the node state change on the subsequent node;

[0099] S24, respectively construct independent state evolution trajectories in the node basic state space for each event type, and use the state space feature historical change data to predict the node state evolution trend in real time, and then dynamically adjust the event intensity level and event duration value of the node state space;

[0100] S25, based on the real-time autocorrelation analysis of the node state space feature change, calculate the state stability index value of each node, and when the state stability index value is lower than the predetermined stability threshold, trigger the node state space update immediately;

[0101] S26, according to the state dependence depth value, the state diffusion coefficient value and the state stability index value of each node, and in the order of the event accurate time stamp corresponding to the node, generate a node state sequence strictly synchronized with time.

[0102] By constructing an interactive event graph based on laboratory equipment, personnel and environment, dynamic monitoring and risk prediction of laboratory node state are realized. Compared with the prior art, the present application can update the node state in real time and generate accurate state change trend by accurately calculating the state correlation degree between nodes, combining the state frequency characteristics of the nodes and the real-time collected data. This method effectively overcomes the defect that the traditional technology cannot accurately capture the complex interaction relationship between nodes in the laboratory environment, provides more accurate and real-time risk prediction value, can more comprehensively reflect the changes of the laboratory environment, and enhances the adaptability and accuracy of the laboratory risk control.

[0103] In the embodiment, the S3 specifically includes:

[0104] S31, based on the node state sequence, generate a node state trend curve with the numerical change of each node state feature within a time window, and obtain the continuous evolution characteristics of the node state by using a curve fitting method;

[0105] S32, calculate the slope, first derivative and second derivative values of each node state trend curve, respectively representing the change rate, change acceleration and change trend stability of the node state, forming a trend dynamics feature set of the node state;

[0106] S33, calculate the similarity value of the node state change by the cosine similarity between the node state trend curves, and construct a node state trend correlation matrix according to the similarity value, each element value in the correlation matrix representing the similarity of the state trend between nodes;

[0107] S34, determine the aggregation weight of the graph convolution operation by using the node state trend correlation matrix, perform a weighted graph convolution operation on the target node state feature set by using the trend dynamics feature set of the nodes with high similarity as the weight, and update the node state space;

[0108] S35, calculate the trend sensitivity value of the node state by using the difference value of the trend dynamics features before and after the node state is updated, filter the node features after the graph convolution operation by using the trend sensitivity value, and extract the node state with a trend sensitivity value exceeding a preset sensitivity threshold as a key node early warning feature;

[0109] S36, aggregate the key node early warning features according to the trend sensitivity values to form an early warning feature set containing the state change sensitivity information of the individual nodes and the node groups.

[0110] Through state trend analysis based on the node state sequence, accurate prediction and risk assessment of the node state are realized. Compared with the prior art, the application constructs a more detailed state trend feature set by calculating the acceleration, change speed and stability of the change trend of the node state, and overcomes the limitations of traditional methods in dealing with complex change trends. By dynamically adjusting the learning strategy and weighting the node features, the application can more accurately predict the abnormal change of the node, improve the accuracy and real-time performance of risk prediction, and has stronger adaptability in dynamic laboratory environments, and provides a more comprehensive risk management solution.

[0111] In the embodiment, the S31 specifically includes:

[0112] S311, divide the state feature values of the node state sequence into a plurality of micro trend segments at a set time interval within a fixed time window, each micro trend segment is composed of two adjacent continuous state feature values, and the feature change direction and change amplitude value of each micro trend segment are recorded;

[0113] S312, define the micro trend mode category according to the change direction and amplitude value of the micro trend segment, cluster and divide all micro trend segments according to the category, and form a micro trend mode set of the node state feature values;

[0114] S313, calculate the transition probability value between the micro trend modes by using the appearance frequency and time sequence before and after of the micro trend modes of the node state sequence within the time window, and construct a micro trend transition probability matrix taking the micro trend mode as a node and the transition probability value as an edge weight;

[0115] S314, based on the micro-trend transition probability matrix, each possible micro-trend mode is continuously combined to define a node state trend evolution path, and an information entropy calculation method is used to calculate the entropy value of the transition probability value of each path, the uncertainty of the path is judged by the information entropy value, and the trend path with the minimum information entropy and the highest stability is screened out as the optimal stable trend path of the node state characteristics;

[0116] S315, according to the information entropy calculation result of the trend path, the trend evolution path with the lowest information entropy value and the highest certainty is screened out, and is defined as the optimal stable trend path of the node state characteristics;

[0117] S316, taking the characteristic change amplitude value of each micro-trend mode in the optimal stable trend path as the fitting reference point, a weighted interpolation method is used for smoothing fitting to generate the continuous evolution characteristics of the node state.

[0118] By accurately dividing the micro-trend segment in the node state sequence, the trend of the node state change is calculated by using the micro-trend mode and the migration rate, and the state prediction accuracy in the laboratory risk control is significantly improved. Compared with the prior art, the present application can capture the micro-dynamics of the node state change by analyzing the micro-trend segment, so as to accurately evaluate the transition trend of the node state. By combining the optimization of the state migration rate, the adaptability and prediction ability of the model to the change of the laboratory environment are further improved. In addition, the global optimization method based on the micro-trend mode adopted in the present application effectively improves the stability and accuracy of the risk prediction model, and provides more reliable technical support for laboratory safety management.

[0119] In the embodiment, the S4 specifically comprises:

[0120] S41, by adaptively segmenting the time sequence of the node state characteristic value, a dynamic time window model is constructed, the time lag effect of the state change between nodes is calculated in each time window, the lag effect factor is generated in real time, and the lag effect factor is used as a correction factor of the node state correlation strength to adjust the risk correlation strength between nodes;

[0121] S42, for the risk correlation of each pair of nodes, a multi-level weighted matrix modeling is carried out, a direct risk propagation path between nodes is preliminarily defined, and the path weight in the matrix is adaptively adjusted by a back propagation strategy to form a complex risk propagation path with multiple levels and multiple stages;

[0122] S43, the nonlinear relationship among equipment, environment and personnel in the laboratory is comprehensively considered, the heterogeneous influence factor of upstream and downstream nodes is introduced in the construction process of the risk correlation matrix, the cross-layer influence of each node on adjacent nodes is dynamically calculated, and the risk propagation direction between nodes is adjusted;

[0123] S44, based on the dynamically adjusted risk correlation matrix, the risk propagation path between nodes is geometrically modeled by a spatial embedding technique, and an adaptive path mapping strategy is introduced, so that the weight and direction of the path can be real-time corrected according to the dynamic change of the laboratory environment;

[0124] S45, through multiple rounds of iterative optimization calculation, the propagation ability of each node and the correlation strength between nodes are updated in real time, and the global risk propagation path is dynamically generated driven by the changes of the laboratory environment, equipment and personnel state.

[0125] By dynamically constructing the node risk correlation matrix and combining the time series analysis of the node state, the accurate adjustment of the risk propagation path of each node in the laboratory environment is realized. Compared with the prior art, the present application effectively considers the nonlinear influence of the environment and personnel through a multi-level and cross-level risk correlation adjustment mechanism, ensuring the accurate simulation of the complex interaction relationship between nodes. The method can adaptively optimize the risk propagation path according to the changes of the laboratory environment and the behavior of the personnel, significantly improving the accuracy and flexibility of risk assessment, thereby effectively improving the intelligence and dynamic adjustment capability of laboratory risk control.

[0126] In the embodiment, the S5 specifically comprises:

[0127] S51, according to the real-time updated node risk propagation path, the potential influence of the state change of each node on the adjacent node is calculated, and the preliminary probability estimation of the state transition of the node is generated by weighted accumulation of the state propagation coefficient between adjacent nodes;

[0128] S52, based on the state change characteristics of each node, the propagation path between nodes and the time sequence relationship between node states, the weighted average method is used to smooth the preliminary transition probability of the node, and the instantaneous fluctuation is eliminated;

[0129] S53, using historical node state transition data, the state change of each node is fitted by a regression analysis model, the prediction function of the node state transition is dynamically adjusted, and the update frequency and update strategy of the transition probability are adjusted in real time according to the environmental change;

[0130] S54, according to the updated transition probability, the state risk prediction value of each node is corrected by a Bayesian inference method according to the real-time data of the laboratory, and a risk prediction model based on the current laboratory environment and equipment state is formed;

[0131] S55, according to the prediction value, the state change amplitude of each node is calculated, and the abnormal state risk index of the node is determined according to the comprehensive evaluation value of the state change amplitude and the transition probability of the node;

[0132] S56, dynamically generating the state risk prediction value of each node in the laboratory according to the abnormal state risk index of the node.

[0133] By combining the dynamic updating of the node risk propagation path and the analysis of the node state change, the accuracy and adaptability of laboratory risk control are effectively improved. Compared with the prior art, the present application dynamically adjusts the risk propagation model of the node by calculating the influence of the state change of the node on the adjacent node in real time, and comprehensively analyzing the node state change characteristics and historical data. This method can adaptively adjust the model parameters to optimize the prediction accuracy, thereby providing more accurate risk assessment in the constantly changing laboratory environment. Compared with the traditional method, the present application has stronger real-time response ability and higher prediction accuracy in the laboratory environment with multiple factors changing.

[0134] In the embodiment, the S55 specifically includes:

[0135] S551, based on the real-time prediction value of each node, calculating the change amplitude of the node state between the current time window and the last time window, defined as the state change amplitude of the node:

[0136] ΔX i = |X i (t) - X i (t-1) |;

[0137] Wherein, ΔX i is the state change amplitude of node i between time t and t-1, X i (t) is the state value of node i at time t;

[0138] S552, according to the node state change amplitude, combining the transition probability of the node, defining the node state risk index as the weighted product of the node state change amplitude and the node transition probability:

[0139] RI i = ΔX i · P i ;

[0140] Wherein, RI is the abnormal state risk index of node i, ΔX

[0001] is the state change amplitude of the node, and P is the transition probability of node i to the abnormal state;

[0141] S553, according to the node state risk index, marking the node whose risk index exceeds the predetermined threshold as a high-risk node, and assigning a corresponding risk level to each node, the risk level is divided into low risk, medium risk and high risk, and the nodes are managed according to the risk level.

[0142] The state change amplitude and transition probability of each node are calculated, and the prediction value of the node state is combined to realize accurate evaluation of the abnormal risk of the node. Compared with the prior art, the risk index calculation method based on the node state change amplitude is adopted to dynamically evaluate the risk of the node in real time, and the nodes are managed according to the change characteristics of each node. The method effectively improves the accuracy and response speed of laboratory risk prediction, especially under the complex interaction of multiple factors, which can effectively avoid the error accumulation problem in the traditional method, and ensures that the laboratory risk control is more accurate and flexible.

[0143] In the embodiment, the S6 specifically comprises:

[0144] S61, calculating the deviation value of each node according to the deviation between the state risk prediction value of each node and the actual node state;

[0145] S62, inputting the wolf optimization algorithm for global optimization with the deviation value of the node as the target function, and calculating the optimal weight factor of each node;

[0146] S63, updating the weight parameter in the graph convolution operation according to the optimal weight factor, adjusting the information propagation strength between nodes, and correcting the graph convolution operation with the optimized weight parameter;

[0147] S64, adopting the adaptive mechanism of the wolf optimization algorithm to dynamically adjust the learning rate and exploration range in each optimization process;

[0148] S65, gradually correcting the early warning feature generation mechanism through multiple iterations, and generating a new laboratory risk prediction result based on the new optimization parameter after each update.

[0149] By combining the wolf optimization algorithm with the deviation correction mechanism of the laboratory node state, dynamic optimization of the node state risk prediction is realized. Compared with the prior art, the deviation between the node prediction value and the actual state is calculated in real time, and the adaptive optimization algorithm is used to dynamically adjust the weight factor, which ensures the accuracy and adaptability of the node state risk prediction. Especially in the optimization process, the global optimization is combined with the change and propagation characteristics of the node state, which improves the accuracy and robustness of the laboratory risk prediction model. The method effectively solves the problem that the traditional technology cannot quickly adapt to changes in the dynamic environment, and can realize rapid response and accurate prediction of changes in the laboratory environment.

[0150] In the embodiment, the S64 specifically comprises:

[0151] S641, in each optimization process, the adjustment factor of the current learning rate is calculated according to the error value of the current optimization result and the convergence of the historical result, and the learning rate is adjusted by the adjustment factor;

[0152] S642, in each round of optimization process, the exploration range is dynamically adjusted according to the amplitude and stability of the node state change, if the amplitude of the node state change is greater than the preset threshold, the exploration range is expanded, if the amplitude of the node state change is smaller, the exploration range is reduced;

[0153] S643, different exploration strategies are dynamically selected according to the current optimization stage and error convergence speed;

[0154] S644, the convergence speed of each node in the optimization process is monitored in real time, and the exploration range is adjusted according to the convergence speed. If the convergence speed of the node in the current optimization step is slow, the global search is enhanced and the exploration range is increased, otherwise the exploration range is reduced.

[0155] By introducing the adaptive mechanism of wolf optimization algorithm, the optimization accuracy and efficiency of node state risk prediction are significantly improved. Compared with the prior art, the learning rate and search range in the optimization process are adjusted in real time, different search strategies are dynamically selected to cope with the changes of node state in the laboratory environment. By accurately controlling the search range, the optimization process can be adaptively adjusted, avoiding the calculation redundancy problem in the traditional method when dealing with large range data. In addition, by monitoring the convergence speed in the optimization process, the ratio of exploration and development is effectively balanced, thereby improving the accuracy and efficiency of laboratory risk prediction.

[0156] Embodiment 1:

[0157] In order to verify the feasibility of the application in implementation, the application is applied to the equipment risk control task in a laboratory, and the running state of the equipment in the laboratory, the environmental change and the personnel operation behavior are monitored and predicted in real time. The laboratory includes multiple high-precision equipment, environmental sensors and personnel operation behavior, all these factors may affect the safety of the laboratory. The traditional risk control method mainly depends on a single data monitoring system, which can usually only handle single parameter monitoring, such as temperature, humidity, pressure and other physical quantities of the equipment, and cannot effectively identify the complex interaction between equipment, environment and personnel. The existing method has low prediction accuracy when facing multi-factor interaction, and often needs a long time for data processing and analysis, lacking real-time performance.

[0158] In practical applications, researchers first collect real-time operation data of equipment, environmental monitoring data and personnel operation data through a data collection system in the laboratory. After processing, these data are input into the graph neural network (GNN) model proposed in the application to construct an interaction event graph. Each node represents the state information of equipment, environment or personnel, and the relationship between nodes is represented by the edges in the graph. The graph neural network can update the state information of each node in real time by utilizing the dependency relationship between nodes. For the monitoring of equipment state, the traditional method can only provide single monitoring data such as temperature change, humidity change, etc., but cannot capture the potential risks hidden behind these single data. Therefore, the application can integrate multi-dimensional information of equipment, environment and personnel through the graph neural network, dynamically model on time series data, and provide risk prediction of multi-factor interaction.

[0159] Specifically, the application comprehensively models the state information of equipment, environment and personnel by constructing an interaction graph based on event sequence. The state of each equipment (such as temperature, humidity, running state, etc.) is represented as a node in the graph, the dependency relationship between equipment is represented by the connection between nodes, and the interaction between equipment and environment and personnel is also reflected by the edges in the graph. The graph convolution operation updates the node state and predicts the state by combining the information of adjacent nodes. Through this modeling method, the system can calculate the risk prediction value of each equipment in real time, and further optimize the prediction result by wolf optimization algorithm (WOA).

[0160] In the process of risk prediction, the graph neural network generates a set of warning features for each node according to the trend of node state change, which reflects the complex interaction between equipment, environment and personnel. Through the analysis of the risk propagation path between nodes, the system can identify potential fault points between equipment and generate risk prediction values in time. In order to further improve the accuracy of prediction, the application introduces the wolf optimization algorithm to dynamically adjust the parameters in the graph convolution operation, optimize the network model, and thus improve the prediction accuracy.

[0161] To verify the effectiveness of the application, we compared the application method with the traditional risk prediction method in the experiment. In the experiment, we selected 5 key equipment in the laboratory for risk prediction, and tested them using the method of the application and the traditional threshold-based prediction method respectively. The experimental results show that when using the traditional method to predict equipment failure, the prediction accuracy and lead time are low. Specifically, the lead time of equipment failure prediction of the traditional method is 3 hours, while the lead time of the prediction of the application method is 8 hours, and the error is reduced by about 60%. In addition, the prediction accuracy of the traditional method decreases significantly when the equipment running state changes complexly, while the application method can respond to the dynamic change of equipment state in real time, and maintains high accuracy and stability.

[0162] Table 1: Comparison results of the method of the present application and the traditional method in equipment failure prediction

[0163]

[0164]

[0165] As can be seen from Table 1, the method of the present application has a significant improvement in the advance of equipment failure prediction, especially in prediction accuracy. The failure prediction accuracy of all equipment in the laboratory is significantly better than that of the traditional method. The method of the present application greatly increases the advance of equipment failure prediction and significantly improves the prediction accuracy through real-time updating of the graph structure and deep optimization mechanism.

[0166] Through the optimized prediction model, researchers can identify equipment that may fail in time and take appropriate measures in advance to avoid the impact of equipment failure on the laboratory. In addition, the laboratory risk prediction system provided by the present application can continuously adjust in real time according to changes in laboratory environment, equipment status and personnel behavior, providing more accurate and reliable risk warning.

[0167] In the implementation process of the present application, the optimization process of the system has undergone multiple rounds of iterative optimization, and the number of graph convolution layers is preliminarily determined to be 4 layers, the node embedding dimension is 128, the learning rate is 0.003, and the number of attention heads is 4. Through these optimization configurations, the system performs significantly better in the task of equipment failure prediction, especially in complex environments and multi-factor interactions, it can provide more accurate risk prediction. The adaptive adjustment function of the system enables timely adjustment and optimization of the model when the laboratory environment changes, improving the stability and long-term operation capability of the system.

[0168] In summary, the present application provides a more accurate and real-time laboratory risk prediction method by combining graph neural networks with wolf swarm optimization algorithm, significantly improving the advance and accuracy of equipment failure prediction, especially in complex scenarios with multiple factor interactions and environmental changes, having significant technical advantages.

[0169] The above is only the preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any skilled person in the art can make equivalent substitutions or changes to the technical solutions and inventive concepts of the present application within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A laboratory risk control method based on a graph neural network, characterized in that, The method comprises the following steps: S1, based on the interaction event sequence between laboratory equipment, personnel and environment, a dynamically updated interaction event graph is constructed, and the node connection relationship is formed in the time sequence of event occurrence; S2, the state space of each node in the interaction event graph is defined, the state space of the node is dynamically updated according to the antecedent relationship of event occurrence, and the node state sequence synchronized with time is generated; S3, based on the node state sequence, a graph convolution operation is used to generate a warning feature set facing the trend of node state change in a fixed time window in the future; S4, a risk correlation matrix between nodes facing time evolution is constructed according to the warning feature set, and the risk correlation strength between nodes is dynamically updated to generate a real-time updated node risk propagation path; S5, according to the real-time updated node risk propagation path, the probability of the state of each node changing to an abnormal state is gradually evaluated through the warning feature set, and the state risk prediction value of each node is dynamically generated; S6, the state risk prediction value of each node is input into the wolf swarm optimization algorithm, the minimization of the deviation between the risk prediction value and the real node state is taken as the optimization target, the warning feature generation mechanism is continuously corrected, the parameter combination of the graph convolution operation is dynamically optimized, and the laboratory risk prediction result is obtained.

2. The laboratory risk control method based on a graph neural network according to claim 1, characterized in that, The S1 specifically comprises: S11, collecting the running event data of all equipment in the laboratory, dynamically clustering the periodic characteristics of the event according to the periodic characteristics of repeated event occurrence, and forming the periodic characteristic sequence of the equipment running event; S12, for laboratory environment data, a multi-dimensional change trend synchronous detection mechanism is used to determine the multi-dimensional feature intersection point at which the environment index simultaneously reaches the threshold in dynamic change in real time, and the multi-dimensional feature intersection point is marked as an environment state mutation event; S13, according to the correlation analysis of the type of personnel operation behavior and the real-time influence on the equipment running state, the grading rules of the influence of operation events on the equipment state are constructed according to the equipment state change amplitude and response time caused by operation actions; S14, the dynamic event chain fusion is carried out according to the equipment periodic characteristic sequence, the environment state mutation event mark and the grading rules of the influence of operation events on the equipment state, and the event fusion sequence reflecting the interaction of equipment, environment and personnel and having time sequence is formed; S15, each event in the event fusion sequence is defined as a graph node with an influence factor weight, the quantitative correlation connection relationship between nodes is constructed in turn according to the size of the influence factor weight, the interaction event graph with quantified influence factor weight is formed, and the influence factor weight is calculated and determined according to the historical contribution rate of the event to the risk occurrence and the time sequence closeness between events; S16, using a real-time dynamic sliding window adaptive mechanism, taking the change rate of the influence factor weight of each node event in the window exceeding a preset threshold as a criterion, the quantitative correlation connection relationship between the nodes of the interaction event graph is updated in real time.

3. The laboratory risk control method based on a graph neural network according to claim 1, characterized in that, The S2 specifically comprises: S21, the basic state space and the derived state space of each node in the interaction event graph are defined at the same time, the basic state space includes event type identification, event duration value, event intensity level, and the derived state space is dynamically generated according to the real-time calculated correlation strength value between the current node and the adjacent node; S22, according to the number of previous nodes of the node in the event chain and the node influence factor weight, the state dependence depth value of each node is calculated as a supplementary feature of the node state space; S23, by monitoring the propagation speed of the node state feature to the adjacent node in real time, the state diffusion coefficient of each node is calculated, and the state diffusion coefficient value reflects the immediate influence degree of the node state change on the subsequent node; S24, an independent state evolution trajectory is constructed for each event type in the node basic state space, and the state evolution trend is predicted in real time by using the historical change data of the state space feature, and then the event intensity level and the event duration value of the node state space are dynamically adjusted; S25, based on the real-time autocorrelation analysis of the change of the node state space feature, the state stability index value of each node is calculated, and when the state stability index value is lower than the predetermined stability threshold, the node state space is immediately updated; S26, according to the state dependence depth value, the state diffusion coefficient value and the state stability index value of each node, the node state sequence strictly synchronized with time is generated in sequence with the event accurate time stamp corresponding to the node.

4. The laboratory risk control method based on a graph neural network according to claim 1, characterized in that, The S3 specifically comprises: S31, based on the node state sequence, the node state trend curve is generated by the numerical change of each node state feature in the time window, and the continuous evolution feature of the node state is obtained by using the curve fitting method; S32, the slope, first derivative and second derivative values of each node state trend curve are calculated, which respectively represent the change rate, change acceleration and change trend stability of the node state, and form a trend dynamics feature set of the node state; S33, the cosine similarity between the node state trend curves is used to calculate the similarity value of the node state change, and a node state trend correlation matrix is constructed according to the similarity value, and each element value in the correlation matrix represents the similarity of the state trend between nodes; S34, the aggregation weight of the graph convolution operation is determined by using the node state trend correlation matrix, the trend dynamics feature set of the node with high similarity is weighted, the weighted graph convolution operation is performed on the target node state feature set, and the node state space is updated; S35, the trend sensitivity value of the node state is calculated by the difference value of the trend dynamics feature before and after the node state is updated, the node feature after the graph convolution operation is screened by using the trend sensitivity value, and the node state with the trend sensitivity value exceeding the preset sensitivity threshold is extracted as the key node warning feature; S36, the key node warning features are weighted and aggregated according to the trend sensitivity value to form a warning feature set containing the state change sensitivity information of the node individual and the node group.

5. The laboratory risk control method based on a graph neural network according to claim 4, characterized in that, The S31 specifically comprises: S311, the state feature values of the node state sequence are divided into a plurality of micro trend segments in a fixed time window at a set time interval, each micro trend segment is composed of two adjacent state feature values in succession, and the feature change direction and change amplitude value of each micro trend segment are recorded; S312, defining a micro-trend mode category according to the change direction and amplitude value of the micro-trend segment, and clustering and dividing all micro-trend segments according to the category to form a micro-trend mode set of node state characteristic values; S313, calculating transition probability values between micro-trend modes according to the frequency and time sequence before and after the occurrence of all micro-trend modes of the node state sequence within the time window, and constructing a micro-trend transition probability matrix taking micro-trend modes as nodes and transition probability values as edge weights; S314, defining a node state trend evolution path as a continuous combination of each possible micro-trend mode based on the micro-trend transition probability matrix, and calculating the entropy values of the transition probability values of each path by using an information entropy calculation method, judging the uncertainty of the path according to the information entropy value, and screening out the trend path with the lowest information entropy and the highest stability; S315, screening out the trend evolution path with the lowest information entropy and the highest certainty according to the information entropy calculation result of the trend path, and defining it as the optimal stable trend path of the node state characteristic; S316, taking the characteristic change amplitude value of each micro-trend mode in the optimal stable trend path as a fitting reference point, and generating a continuous evolution characteristic of the node state by smoothing fitting through weighted interpolation.

6. The laboratory risk control method based on a graph neural network according to claim 1, characterized in that, The S4 specifically includes: S41, constructing a dynamic time window model by adaptively segmenting the time sequence of the node state characteristic values, calculating the time lag effect of the state change between nodes in each time window, generating a lag effect factor in real time, and taking the lag effect factor as a correction factor of the node state correlation strength to adjust the risk correlation strength between nodes; S42, for the risk correlation of each pair of nodes, a multi-level weighted matrix modeling is performed to preliminarily define the direct risk propagation path between nodes, and a reverse propagation strategy is used to adaptively adjust the path weight in the matrix to form a multi-level and multi-stage complex risk propagation path; S43, introducing the heterogeneous influence factor of upstream and downstream nodes in the construction process of the risk correlation matrix to dynamically calculate the cross-layer influence of each node on adjacent nodes and adjust the risk propagation direction between nodes; S44, based on the dynamically adjusted risk correlation matrix, the risk propagation path between nodes is geometrically modeled by using a spatial embedding technology, and an adaptive path mapping strategy is introduced to enable the weight and direction of the path to be real-time corrected according to the dynamic changes of the laboratory environment; S45, through multiple rounds of iterative optimization calculation, the propagation ability of each node and the correlation strength between nodes are updated in real time, and the global risk propagation path is dynamically generated driven by the state changes of the environment, equipment and personnel in the laboratory.

7. The laboratory risk control method based on a graph neural network according to claim 1, characterized in that, The S5 specifically includes: S51, according to the real-time updated node risk propagation path, calculating the potential influence of the state change of each node on adjacent nodes, generating a preliminary probability estimate of the node state transition to an abnormal state by weighted accumulation through the state propagation coefficient between adjacent nodes; S52, based on the state change characteristics of each node, the propagation path between nodes and the timing relationship between node states, the preliminary transition probability of the node is smoothed by using weighted average method to eliminate transient fluctuations; S53, using historical node state transition data, fitting the state change of each node through regression analysis model, dynamically adjusting the prediction function of node state transition, and adjusting the update frequency and update strategy of transition probability in real time according to environmental changes; S54, according to the updated transition probability, the state risk prediction value of each node is corrected by using Bayesian inference method according to real-time data in the laboratory, and a risk prediction model based on the current laboratory environment and equipment state is formed; S55, according to the prediction value, the state change amplitude of each node is calculated, and the abnormal state risk index of the node is determined according to the state change amplitude of the node and the comprehensive evaluation value of the transition probability; S56, based on the abnormal state risk index of the node, the state risk prediction value of each node in the laboratory is dynamically generated.

8. The laboratory risk control method based on a graph neural network according to claim 7, characterized in that, The S55 specifically includes: S551, based on the real-time prediction value of each node, the change amplitude of the node state between the current time window and the last time window is calculated; S552, according to the node state change amplitude, combined with the transition probability of the node, the node state risk index is defined as the weighted product of the node state change amplitude and the node transition probability; S553, according to the node state risk index, the nodes whose risk index exceeds the predetermined threshold are marked as high-risk nodes, and each node is assigned a corresponding risk level, the risk level is divided into low risk, medium risk and high risk, and the nodes are managed by level based on the risk level.

9. The laboratory risk control method based on a graph neural network according to claim 1, characterized in that, The S6 specifically includes: S61, according to the deviation between the state risk prediction value of each node and the actual node state, the deviation value of each node is calculated; S62, taking the deviation value of the node as the objective function, inputting the wolf optimization algorithm for global optimization to calculate the optimal weight factor of each node; S63, according to the optimal weight factor, updating the weight parameter in the graph convolution operation, adjusting the information propagation intensity between nodes, and correcting the graph convolution operation with the optimized weight parameter; S64, using the adaptive mechanism of wolf optimization algorithm, dynamically adjusting the learning rate and exploration range in each optimization process; S65, through multiple rounds of iterative optimization, gradually correcting the early warning feature generation mechanism, and generating new laboratory risk prediction results based on new optimization parameters after each update.

10. The laboratory risk control method based on a graph neural network according to claim 9, characterized in that, The S64 specifically includes: S641, in each optimization process, according to the error value of the current optimization result and the convergence of the historical result, the adjustment factor of the current learning rate is calculated, and the learning rate is adjusted by the adjustment factor; S642, in each optimization process, the exploration range is dynamically adjusted according to the amplitude and stability of the node state change, if the amplitude of the node state change is greater than the preset threshold, the exploration range is expanded, and if the amplitude of the node state change is smaller, the exploration range is reduced; S643, according to the current optimization stage and the error convergence speed, different exploration strategies are dynamically selected; S644, monitoring convergence speed of each node in the optimization process in real time, and adjusting the exploration range according to the convergence speed. If the convergence speed of the node in the current optimization step is slow, the global search is enhanced, and the exploration range is increased, otherwise the exploration range is contracted.

Images