Decision-making method and system based on knowledge graph

Abstract

The invention relates to the technical field of intelligent decision making of production equipment, and discloses a decision making method and system based on a knowledge graph, and the decision making method based on the knowledge graph comprises the following steps: processing input data through a multi-granularity knowledge graph construction system, and obtaining a multi-scale knowledge graph; processing the real-time sensor data and the multi-scale knowledge graph through a time sequence knowledge dual representation learning framework to obtain a dynamic representation model; processing the dynamic representation model and the multi-scale knowledge graph through a time-varying causal propagation network decision system to obtain a decision analysis result; processing the decision analysis result and the multi-scale knowledge graph through a multi-hypothesis reasoning algorithm to obtain a root cause analysis report; a multi-scale knowledge graph is constructed by fusing multi-source heterogeneous data, and intelligent decision with high accuracy and strong interpretation is realized by combining time sequence knowledge dual representation learning and time-varying causal propagation network analysis.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent decision-making for production equipment, and more specifically, it relates to a decision-making method and system based on a knowledge graph. Background Art

[0002] The key production equipment in large petrochemical plants includes catalytic cracking units, distillation columns, heat exchanger networks, etc., all of which play important roles in the production process of petrochemical plants. Sudden failures of key equipment may lead to production line shutdowns, product quality degradation, and even safety accidents, causing significant economic losses and environmental risks. However, current intelligent decision-making for equipment fault prediction and root cause analysis scenarios in petrochemical plants has the following problems:

[0003] Data silo problem: Information such as equipment monitoring data, historical fault records, maintenance experience, and equipment knowledge bases is scattered in different systems and is difficult to effectively integrate, resulting in the inability to comprehensively consider various relevant factors in the decision-making process and insufficient accuracy;

[0004] Complex causal relationships: There are complex temporal dependencies and causal relationships among process parameters, which are difficult to characterize using traditional statistical methods or single physical models;

[0005] Decision reliability problem: Existing decision-making methods often lack interpretability and are difficult to provide confidence assessment, resulting in insufficient decision reliability. Summary of the Invention

[0006] The present invention provides a decision-making method and system based on a knowledge graph to solve the technical problems of data silos, difficulty in characterizing complex causal relationships, and insufficient decision reliability in related technologies.

[0007] The present invention provides a decision-making method based on a knowledge graph, including the following steps:

[0008] Processing input data through a multi-granularity knowledge graph construction system to obtain a multi-scale knowledge graph;

[0009] Processing real-time sensor data and the multi-scale knowledge graph through a temporal knowledge dual representation learning framework to obtain a dynamic representation model;

[0010] Processing the dynamic representation model and the multi-scale knowledge graph through a time-varying causal propagation network decision-making system to obtain a decision analysis result;

[0011] Processing the decision analysis result and the multi-scale knowledge graph through a multi-hypothesis reasoning algorithm to obtain a root cause analysis report.

[0012] Further, the step of processing input data through a multi-granularity knowledge graph construction system to obtain a multi-scale knowledge graph includes:

[0013] The device specification parameters and the process flow chart are processed by a multi-source heterogeneous data fusion algorithm to obtain a device process knowledge network;

[0014] The historical case database is processed by a hierarchical association rule mining algorithm to obtain an association network;

[0015] The time-series data is processed by a parameter causal network learning algorithm to generate a parameter time-series causal diagram;

[0016] The above three types of knowledge networks are processed by a multi-scale knowledge graph integration algorithm to obtain a unified multi-scale knowledge graph.

[0017] Furthermore, the mathematical expression of the hierarchical association rule mining algorithm is:

[0018] ;

[0019] where represents the output association network, represents the hierarchical association rule mining algorithm, represents the historical case database, represents the minimum support threshold, represents the minimum confidence threshold;

[0020] The generation of association rules follows the following constraints:

[0021] ;

[0022] In the formula, represents the association rule from symptom to cause , represents symptom , represents cause , represents the rule confidence calculation function, represents the rule support calculation function, represents the minimum support threshold, represents the minimum confidence threshold.

[0023] Furthermore, the steps of processing the real-time sensor data and the multi-scale knowledge graph through the time-series knowledge dual representation learning framework to obtain a dynamic representation model include:

[0024] The multi-source sensor time-series data is processed by a dynamic time window feature extraction algorithm to obtain a parameter time-series feature vector set;

[0025] The multi-scale knowledge graph is processed by a knowledge-aware graph embedding algorithm to obtain a parameter knowledge semantic vector set;

[0026] Process the temporal feature vector set and the knowledge semantic vector set through the temporal knowledge joint representation model to obtain a dual representation feature space;

[0027] Process the dual representation feature space and the historical state sequence through the state dynamic evolution modeling algorithm to obtain a dynamic representation model.

[0028] Furthermore, the mathematical expression of the knowledge-aware graph embedding algorithm is:

[0029] ;

[0030] where, represents the output parameter knowledge semantic vector set, represents the multi-scale knowledge graph, represents the embedding dimension, represents the learning rate, represents the regularization parameter, represents the number of training epochs, represents the knowledge-aware graph embedding algorithm;

[0031] The optimization objective function of the algorithm is defined as:

[0032] ;

[0033] where, represents the head entity in the knowledge graph, represents the relationship in the knowledge graph, represents the tail entity in the knowledge graph, represents the head entity of negative sampling in the knowledge graph, represents the tail entity of negative sampling in the knowledge graph, represents the set of triples in the knowledge graph, represents the set of reverse triples in the knowledge graph, represents the relaxation factor in the loss function, represents the head entity in the knowledge graph , the relationship in the knowledge graph and the tail entity in the knowledge graph the distance between them, represents the head entity of negative sampling in the knowledge graph , the relationship and the tail entity of negative sampling in the knowledge graph the distance between them, represents the regularization parameter; represents the maximum operation, that is, when the value of the expression is greater than 0, take its value, otherwise take 0; represents the summation symbol; is the optimized objective function of the knowledge-aware graph embedding algorithm. represents the causal relationship loss function, and its specific definition is:

[0034] ;

[0035] where represents the set of causal relationships in the knowledge graph, represents the node and the node the edge weight between them, represents the node the embedding vector of, represents the node the embedding vector of, represents the node and the node the square of the Euclidean distance between them.

[0036] Furthermore, the steps of processing the dynamic representation model and the multi-scale knowledge graph by the time-varying causal propagation network decision system to obtain the decision analysis result include:

[0037] Processing the dynamic representation model through the prediction deviation analysis algorithm to obtain the parameter prediction deviation matrix;

[0038] Processing the parameter prediction deviation matrix and the multi-scale knowledge graph through the graph-guided causal propagation modeling algorithm to obtain the time-varying causal propagation network;

[0039] Processing the time-varying causal propagation network through the multi-scale pattern detection algorithm to obtain the propagation pattern and the matching result;

[0040] Processing the propagation pattern and the matching result through the key factor location and evaluation algorithm to obtain the decision analysis result.

[0041] Furthermore, the graph-guided causal propagation modeling algorithm defines the time-varying causal propagation network as a five-tuple:

[0042] ;

[0043] where represents the set of parameter nodes, represents the set of directed edges, represents the set of edge weights, represents the set of edge delays, represents the set of node states, represents the time-varying causal propagation network;

[0044] For any two parameter nodes, the propagation intensity between them is calculated as follows:

[0045] ;

[0046] Among them, represents the propagation strength from the parameter node to the parameter node . represents the causal relationship strength extracted from the knowledge graph, represents the transfer entropy calculated based on historical data, represents the balance factor.

[0047] Furthermore, the step of processing the decision analysis result and the multi-scale knowledge graph through the multi-hypothesis reasoning algorithm to obtain the root cause analysis report includes:

[0048] Processing the decision analysis result and the multi-scale knowledge graph through the hypothesis space generation algorithm to obtain a set of candidate root cause hypotheses;

[0049] Processing the set of candidate root cause hypotheses, the decision analysis result, and the multi-scale knowledge graph through the evidence chain construction algorithm to obtain a set of hypothesis evidence chains;

[0050] Processing the set of hypothesis evidence chains through the multi-hypothesis scoring and ranking algorithm to obtain a ranked set of root cause hypotheses;

[0051] Processing the ranked set of root cause hypotheses, the decision analysis result, and the multi-scale knowledge graph through the explanatory root cause report generation algorithm to obtain the root cause analysis report.

[0052] Furthermore, the multi-hypothesis scoring and ranking algorithm scores each hypothesis from four dimensions, including evidence support, explanatory power, consistency, and simplicity, and the specific calculations are as follows:

[0053] ;

[0054] ;

[0055] ;

[0056] ;

[0057] Among them, represents the evidence support of hypothesis , represents the explanatory power of hypothesis , represents the consistency of hypothesis , represents the simplicity of hypothesis , represents the evidence chain of hypothesis . Represents the credibility of evidence , Represents the relevance of evidence to the hypothesis , Represents the set of decision analysis results that the hypothesis can explain Represents the set of decision analysis results Represents the number of rules in the knowledge graph that the hypothesis is consistent with Represents the multi-scale knowledge graph Represents the total number of rules in the knowledge graph related to the hypothesis , Represents the cardinality or number of elements of a set Represents the logarithmic function

[0058] A decision-making system based on a knowledge graph for implementing a decision-making method based on a knowledge graph as described above, including:

[0059] A multi-granularity knowledge graph construction system for processing input data to obtain a multi-scale knowledge graph;

[0060] A temporal knowledge dual representation learning framework for processing real-time sensor data and the multi-scale knowledge graph to obtain a dynamic representation model;

[0061] A time-varying causal propagation network decision-making system for processing the dynamic representation model and the multi-scale knowledge graph to obtain decision analysis results;

[0062] A multi-hypothesis reasoning system for processing the decision analysis results and the multi-scale knowledge graph to obtain a root cause analysis report

[0063] The beneficial effects of the present invention are as follows: By fusing multi-source heterogeneous data to construct a multi-scale knowledge graph, combining temporal knowledge dual representation learning and time-varying causal propagation network analysis, intelligent decision-making with high accuracy and strong interpretability is achieved. Among them, the system integrates multi-source information and multiple models, significantly improving decision-making accuracy; it can discover potential problems 12 - 72 hours in advance, providing a sufficient time window for critical decisions and significantly enhancing the timeliness of decision-making responses; each decision result is accompanied by a detailed evidence chain and confidence evaluation, supporting the understanding of the decision-making process and results from multiple perspectives and meeting the transparency requirements of high-reliability scenarios; the system can effectively accumulate domain knowledge and decision-making experience, forming a systematic knowledge graph and reducing the adaptation cost of new scenarios BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 Is a flowchart of a decision-making method based on a knowledge graph in the present invention DETAILED DESCRIPTION OF THE EMBODIMENTS

[0065] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein. Without departing from the scope of protection of the content of this specification, changes can be made to the functions and arrangements of the elements discussed. Each example can omit, substitute, or add various processes or components as needed. Additionally, the features described in some examples can also be combined in other examples.

[0066] In at least one embodiment of the present invention, a decision-making method based on a knowledge graph is disclosed, as Figure 1 shown, including the following steps:

[0067] Step 100, process the input data through a multi-granularity knowledge graph construction system to obtain a multi-scale knowledge graph;

[0068] In one embodiment of the present invention, the input data includes device specification parameters, process flowcharts, historical fault records, maintenance experience, device knowledge bases, and parameter time series datasets;

[0069] Step 101, process the device specification parameters and the process flowchart through a multi-source heterogeneous data fusion algorithm to obtain a device process knowledge network. Specifically, the mathematical expression of the multi-source heterogeneous data fusion algorithm is:

[0070] ;

[0071] where represents the output device process knowledge network, , and respectively represent the device technical manual dataset, the process flowchart dataset, and the piping and instrumentation diagram dataset;

[0072] represents the multi-source heterogeneous data fusion function, which is defined as a combined operation of three sub-functions: entity extraction, relationship mapping, and attribute integration.

[0073] Step 102, process the input historical case base (including historical fault records, maintenance experience, and device knowledge bases) through a hierarchical association rule mining algorithm to obtain a fault-symptom-cause association network;

[0074] Specifically, the hierarchical association rule mining algorithm identifies frequently occurring combinations of problem phenomena based on an improved Apriori algorithm framework, calculates and generates preliminary rules according to support and confidence, and then organizes the generated rules into a multi-level structure through a fault tree analysis (FTA) mechanism. The mathematical expression of this algorithm is:

[0075] ;

[0076] Among them, represents the output association network, represents the hierarchical association rule mining algorithm, represents the historical case database, represents the minimum support threshold, represents the minimum confidence threshold;

[0077] The generation of association rules follows the following constraints:

[0078] ;

[0079] In the formula, represents the association rule from symptom to cause , represents symptom , represents cause , represents the rule confidence calculation function, represents the rule support calculation function, represents the minimum support threshold, represents the minimum confidence threshold.

[0080] Step 103, process the time series data through the parametric causal network learning algorithm to generate a parametric time series causal graph;

[0081] In an embodiment of the present invention, a parametric time series dataset is used as input, and the time series dataset includes two key parameters: a causal strength threshold and a maximum time delay window;

[0082] Specifically, the mathematical expression of the parametric causal network learning algorithm is:

[0083] ;

[0084] Among them, represents the output parametric time series causal graph, represents the parametric causal network learning function, represents the parametric time series dataset, represents the causal strength threshold, represents the maximum time delay window;

[0085] The core calculation of the algorithm is based on the time-delay causal strength function:

[0086] ;

[0087] In the formula, Represents the causal strength function, and respectively represent the parameters and at time values (including parameter values and parameter value confidence levels), and respectively represent the parameters and at time values, represents the time delay, represents the transfer entropy calculation function, defined as:

[0088] ;

[0089] wherein represents the time-delay causal strength of , represents , , the joint probability distribution that co-occurs in the time series, represents the conditional probability distribution of and under the condition of knowing , represents the conditional probability distribution of under the condition of knowing , and respectively represent the parameters and at time values, represents the summation symbol, represents the logarithmic operation.

[0090] Step 104, process the three types of knowledge networks in the above steps 101, 102, and 103 through a multi-scale knowledge graph integration algorithm to obtain a unified multi-scale knowledge graph, forming a multi-level and multi-relational complex network structure that can completely represent the multi-dimensional information of the industrial system, including physical structure, process flow, parameter relationships, and fault modes;

[0091] Specifically, the mathematical expression of the multi-scale knowledge graph integration algorithm is:

[0092] ;

[0093] wherein, represents the integrated multi-scale knowledge graph, represents the output multi-scale knowledge graph integration function, Represents the device process knowledge network output by step 101, Represents the fault - symptom - cause association network output by step 102, Represents the parameter time - series causal graph output by step 103, Represents the knowledge fusion rule set;

[0094] The integrated multi - scale knowledge graph can be formally represented as:

[0095] ;

[0096] Wherein Represents the integrated multi - scale knowledge graph, Represents the entity set, Represents the relationship set, Represents the attribute set, Represents the hierarchical division, Represents the cross - layer connection set.

[0097] Step 200, process the real - time sensor data and the multi - scale knowledge graph of step 100 through the time - series knowledge dual - representation learning framework to obtain the dynamic representation model of the industrial equipment operation state;

[0098] In an embodiment of the present invention, the real - time sensor data includes device operation parameters, process parameters, and environmental parameters;

[0099] Step 201, process the multi - source sensor time - series data through the dynamic time - window feature extraction algorithm to obtain the parameter time - series feature vector set;

[0100] Specifically, the mathematical expression of the dynamic time - window feature extraction algorithm is:

[0101] ;

[0102] Wherein, Represents the output parameter time - series feature vector set, Represents the multi - source sensor data stream set (i.e., real - time sensor data, including device operation parameters, process parameters, and environmental parameters), Represents the initial window size, Represents the window adjustment coefficient, Represents the feature importance threshold, Represents the dynamic time - window feature extraction function;

[0103] The window adaptive adjustment unit calculates the optimal window size based on the variance volatility:

[0104] ;

[0105] Wherein, Represents the parameter The optimal window size of Represents the window adjustment coefficient, Represents the initial window size, Represents the parameter The variance volatility of, and the calculation formula is:

[0106] ;

[0107] Among them, Represents the parameter The variance within the current time window of Represents the current time window size, Represents the parameter The historical variance baseline of

[0108] Step 202, process the multi-scale knowledge graph in step 100 through the knowledge-aware graph embedding algorithm to obtain a set of parameter knowledge semantic vectors;

[0109] Specifically, the knowledge-aware graph embedding algorithm is based on an improved TransE (Translation-based Distance Model) model, and the mathematical expression is:

[0110] ;

[0111] Among them, Represents the output set of parameter knowledge semantic vectors, Represents the multi-scale knowledge graph output in step 100, Represents the embedding dimension, Represents the learning rate, Represents the regularization parameter, Represents the number of training rounds, Represents the knowledge-aware graph embedding function;

[0112] To ensure that the model can effectively capture the complex relationships in the knowledge graph and improve the representation ability of the embedding vectors, the optimization objective function of the algorithm is defined as:

[0113] ;

[0114] Among them, Represents the head entity in the knowledge graph, Represents the relationship in the knowledge graph, Represents the tail entity in the knowledge graph, Represents the head entity of negative sampling in the knowledge graph, Represents the tail entity of negative sampling in the knowledge graph, Represents the set of triples in the knowledge graph, represents the set of inverse triples in the knowledge graph, represents the relaxation factor in the loss function, represents the head entity in the knowledge graph , the relationship in the knowledge graph and the tail entity in the knowledge graph the distance between them, represents the head entity of negative sampling in the knowledge graph , the relationship and the tail entity of negative sampling in the knowledge graph the distance between them, represents the regularization parameter; represents the maximum operation, that is, when the value of the expression is greater than 0, take its value, otherwise take 0; represents the summation symbol; is the optimization objective function of the knowledge-aware graph embedding algorithm, represents the causal relationship loss function, and the specific definition is:

[0115] ;

[0116] Among them, represents the set of causal relationships in the knowledge graph, represents the node and the node the edge weight between them, represents the embedding vector of the node , represents the embedding vector of the node , represents the node and the node the square of the Euclidean distance between them.

[0117] Step 203, process the temporal feature vector set and the knowledge semantic vector set through the temporal knowledge joint representation model to obtain a dual representation feature space;

[0118] Specifically, the mathematical expression of the temporal knowledge joint representation model is:

[0119] ;

[0120] Among them, represents the output dual representation feature space (including temporal features and knowledge semantic features), represents the temporal knowledge joint representation model function, represents the parameter temporal feature vector set output in step 201, represents the parameter knowledge semantic vector set output in step 202, Represents the model weight matrix (including the temporal feature weight matrix and the knowledge semantic weight matrix), represents the model bias vector (including the temporal feature bias vector and the knowledge semantic bias vector);

[0121] The model for each parameter calculates the dual representation vector as follows:

[0122] ;

[0123] In the formula, represents the dual representation vector of parameter , represents the temporal feature vector of parameter , represents the knowledge semantic vector of parameter , represents the adaptive fusion weight, represents the non - linear interaction function, implemented by a multi - layer perceptron.

[0124] The adaptive fusion weight is calculated by the formula:

[0125] ;

[0126] where represents the adaptive fusion weight, represents the temporal feature vector of parameter , represents the knowledge semantic vector of parameter , represents the sigmoid activation function, and represent the weight vector and the bias term respectively, represents the vector concatenation operation, represents the transpose operation.

[0127] Step 204: Process the dual representation feature space and the historical state sequence through the state dynamic evolution modeling algorithm to obtain the dynamic representation model;

[0128] Specifically, the state dynamic evolution modeling algorithm is based on an improved LSTM (Long Short - Term Memory Network) structure, and the mathematical expression of the algorithm is:

[0129] ;

[0130] where represents the output dynamic representation model (i.e., the industrial equipment operation state dynamic representation model), represents the state dynamic evolution modeling function, is the dual-representation feature space output by step 203, is the historical state sequence output by step 204, is the set of model parameters;

[0131] Specifically, the dynamic representation model can predict the state change trend of the device at future time points, providing a basis for subsequent anomaly detection and fault prediction. The calculation process of the dynamic representation model includes four main steps:

[0132] Input layer: Receive the dual-representation feature sequence ,

[0133] ;

[0134] Encoding layer: Use bidirectional LSTM for temporal encoding:

[0135] ;

[0136] ;

[0137] ;

[0138] Attention layer: Calculate the importance weights at different time steps:

[0139] ;

[0140] ;

[0141] ;

[0142] Prediction layer: Generate the current dynamic representation vector:

[0143] ;

[0144] where represents the dual-representation feature sequence, is the time window size, represents the transpose operation, is the current time step, is the dual-representation feature vector at the current time step, represents the dual-representation feature sequence, and respectively represent the hidden states of the forward and backward LSTMs at time step , and respectively represent the forward and backward LSTM cell functions, and respectively represent the forward and backward LSTMs at time step hidden state, represents the time step bidirectional LSTM hidden state of, represents the vector concatenation operation, represents the importance weight of the current time step, represents the importance weight of the current time step, represents the context vector after attention aggregation, represents the exponential function, 、 、 、 and respectively represent the learnable parameters of the attention vector, hidden state transformation matrix, hidden state bias vector, state representation transformation matrix, and state representation bias vector, is the current time step dynamic representation vector of, represents the sigmoid activation function, 、 、 respectively represent the time step attention weight, attention score, bidirectional LSTM hidden state of, represents the hyperbolic tangent activation function.

[0145] Step 300, process the dynamic representation model in Step 200 and the multi-scale knowledge graph in Step 100 through the time-varying causal propagation network decision system to obtain the decision analysis result.

[0146] Step 301, process the dynamic representation model through the prediction deviation analysis algorithm to obtain the parameter prediction deviation matrix;

[0147] Specifically, the mathematical expression of the prediction deviation analysis algorithm is:

[0148] ;

[0149] where, represents the output parameter prediction deviation matrix (including parameter prediction deviation values and prediction deviation confidence levels), represents the prediction deviation analysis function, represents the dynamic representation model output in Step 200, represents historical normal data, represents the prediction time window.

[0150] Step 302, process the parameter prediction deviation matrix and the multi-scale knowledge graph through the graph-guided causal propagation modeling algorithm to obtain the time-varying causal propagation network (including propagation mode and propagation delay);

[0151] The atlas-guided causal propagation modeling algorithm consists of a knowledge constraint extraction unit, a time-varying propagation graph construction unit, and a propagation intensity calculation unit; the knowledge constraint extraction unit extracts the causal relationship constraints between parameters from the knowledge graph; the time-varying propagation graph construction unit establishes the directed connections between parameter nodes; the propagation intensity calculation unit calculates the probability and time delay of causal propagation between different nodes; the mathematical expression of the algorithm is:

[0152] ;

[0153] Among them, represents the output time-varying causal propagation network, represents the atlas-guided causal propagation modeling function, represents the parameter prediction deviation matrix output in step 301, represents the multi-scale knowledge graph output in step 100, represents the propagation intensity threshold;

[0154] This network is defined as a five-tuple:

[0155] ;

[0156] Among them, represents the set of parameter nodes, represents the set of directed edges, represents the set of edge weights, represents the set of edge time delays, represents the set of node states, represents the time-varying causal propagation network.

[0157] The propagation intensity between any two parameter nodes is calculated as follows:

[0158] ;

[0159] Among them represents any two parameter nodes and the propagation intensity between them, represents the causal relationship intensity extracted from the knowledge graph, represents the transfer entropy calculated based on historical data, represents the balance factor;

[0160] The propagation delay is calculated as the time difference between the peaks of the two parameter deviations:

[0161] ;

[0162] Among them represents the propagation delay, is the maximum value operation, Represents a time series related function, Represents a time delay variable, and respectively represent the parameters and at time and deviation values.

[0163] Step 303, process the time-varying causal propagation network through a multi-scale pattern detection algorithm to obtain the propagation pattern and the matching result;

[0164] Specifically, the mathematical expression of the multi-scale pattern detection algorithm is:

[0165] ;

[0166] Among them, Represents the output propagation pattern and matching result, Represents the multi-scale pattern detection function, Represents the time-varying causal propagation network output in step 302, Represents the predefined pattern library, Represents the multi-scale detection parameter set;

[0167] The core of the algorithm is to use a stochastic process to simulate the propagation in the network. For the starting node set , at the th moment, the state of node is calculated as follows:

[0168] ;

[0169] In the formula, Represents the state of node at the th moment, Represents the in-neighbor set of node , Represents the time decay factor, Represents the propagation delay from node to node , Represents the propagation intensity from node to node , Represents the maximum value operation.

[0170] Pattern matching uses a graph matching algorithm to calculate the similarity between the simulated propagation pattern and the predefined pattern :

[0171] ;

[0173] Among them, and respectively represent the edge sets in the simulated propagation mode and the predefined mode, and respectively represent the weights of the edges in the simulated propagation mode and the predefined mode . represents the simulated propagation mode and the predefined mode the similarity between them, and respectively represent the intersection and the union.

[0174] Step 304, process the propagation mode and the matching result through the key factor location evaluation algorithm to obtain the decision analysis result;

[0175] Specifically, the mathematical expression of the algorithm is:

[0176] ;

[0177] Among them, represents the output decision analysis result (including the key source parameter set, the propagation path set, the risk score), represents the key factor location evaluation function, represents the propagation mode and the matching result output in step 303, represents the confidence requirement.

[0178] The output decision analysis result contains the following key information:

[0179] The key source parameter set, representing the source parameters of causal propagation:

[0180] ;

[0181] Among them, represents the key source parameter set, represents the state of node at time , represents the in-degree of node in the causal propagation subgraph, represents the state threshold, represents the detected set of influencing nodes.

[0182] The propagation path set, representing the propagation path in the system:

[0183] ;

[0184] Among them, Represents the set of propagation paths, Represents a directed path from parameter To parameter ; Represents the propagation strength from node To node ; Represents the propagation strength threshold, Represents the state of node At time ; Represents the state threshold, Represents the set of detected affected nodes;

[0185] Risk score, considering the matching pattern, propagation range, and historical data information comprehensively:

[0186] ;

[0187] Among them, Represents the risk score, Represents the similarity of the best - matching pattern, Represents the proportion of affected parameters, Represents the relevance of historical cases, , And Are the pattern - matching weight coefficient, propagation - range weight coefficient, and historical - data weight coefficient respectively.

[0188] The finally generated decision - analysis result is a structured report, containing key source parameters, propagation paths, matching patterns, event - probability estimates, and key information on the expected occurrence time, providing a decision - making basis for users.

[0189] Step 400: Process the decision - analysis result and the multi - scale knowledge graph through the multi - hypothesis reasoning algorithm to obtain a root - cause analysis report; combine the decision - analysis result in step 300 and the multi - scale knowledge graph constructed in step 100, and through multi - dimensional evidence - chain analysis, identify the most likely root cause and provide an explanatory report.

[0190] Step 401: Process the decision - analysis result and the multi - scale knowledge graph through the hypothesis - space generation algorithm to obtain a set of candidate root - cause hypotheses;

[0191] Specifically, the mathematical expression of the hypothesis - space generation algorithm is:

[0192] ;

[0193] Among them, Represents the output set of candidate root - cause hypotheses, Represents the hypothesis - space generation function, Represents the decision analysis result output by step 300, Represents the multi-scale knowledge graph output by step 100, Represents the domain constraint rule set, Represents the maximum number of hypotheses;

[0194] For each key source parameter , Represents the set of key source parameters output by step 300. The algorithm first finds the set of possible cause nodes in the knowledge graph:

[0195] ;

[0196] Among them, Represents the set of possible cause nodes, and represents the path in the knowledge graph from node to node , Represents the type of node , Represents the set of node types of possible causes, Represents the multi-scale knowledge graph output by step 100.

[0197] Then, the algorithm generates candidate hypotheses from all sets of possible cause nodes and considers domain constraints:

[0198] ;

[0199] Among them, Represents the set of all candidate root cause hypotheses, Represents that the hypothesis satisfies the domain constraint rule set , Represents the candidate root cause hypothesis, Represents the set of possible cause nodes, Represents the union of all sets of possible cause nodes, Represents the logical "AND" operation;

[0200] Finally, the algorithm sorts the hypotheses according to the prior score and selects the top as the output:

[0201]

[0202] Among them, Represents the set of output candidate root cause hypotheses, Represents the function of selecting the top hypotheses, Represents the function of sorting according to the prior score, Represents descending order, is the maximum number of hypotheses, is the prior score function;

[0203] Calculate the hypothesis prior score:

[0204] ;

[0205] Among them, represents the association strength between the hypothesis and the key source parameter set ; represents the occurrence frequency of the hypothesis in historical cases; is the balance weight coefficient of the association strength and the historical frequency; is the prior score function.

[0206] Step 402: Process the candidate root cause hypothesis set, decision analysis results, and multi-scale knowledge graph through the evidence chain construction algorithm to obtain the hypothesis evidence chain set;

[0207] Specifically, the mathematical expression of the algorithm is:

[0208] ;

[0209] Among them, represents the output hypothesis evidence chain set (including observation evidence, knowledge evidence, historical case evidence, comprehensive confidence evaluation results); represents the evidence chain construction function; represents the candidate root cause hypothesis set; represents the multi-scale knowledge graph; represents the decision analysis results; represents the historical case database;

[0210] For each candidate hypothesis , the algorithm constructs three types of evidence:

[0211] ;

[0212] ;

[0213] ;

[0214] Among them, , and respectively represent the association relationship determination function, support relationship determination function, and similarity calculation function; represents the evidence; represents the candidate root cause hypothesis; , and respectively represent observational evidence, knowledge evidence, and historical case evidence, represents a multi-scale knowledge graph, represents the decision analysis result, represents the historical case database, represents the logical "AND" operation;

[0215] Then, the algorithm evaluates the reliability and relevance of each piece of evidence:

[0216] ;

[0217] ;

[0218] Among them, and respectively represent the reliability and relevance evaluation functions, 、 and respectively represent the evidence source, type, and confidence level, and respectively represent the logical distance and association strength between the evidence and the hypothesis, and respectively represent the reliability and relevance of the evidence;

[0219] Finally, the algorithm constructs a complete evidence chain:

[0220] ;

[0221] Among them, represents the comprehensive confidence level evaluation result, represents the structured representation of the evidence chain, represents the evidence chain, 、 and respectively represent observational evidence, knowledge evidence, and historical case evidence.

[0222] Step 403, process the set of hypothesis evidence chains through a multi-hypothesis scoring and ranking algorithm to obtain a ranked set of root cause hypotheses;

[0223] The mathematical expression of the multi-hypothesis scoring and ranking algorithm is:

[0224] ;

[0225] Among them, represents the output ranked set of root cause hypotheses, represents the multi-hypothesis scoring and ranking function, represents the set of hypothesis evidence chains output in Step 402, represents the expert experience weight, Represents the multi-dimensional scoring weights;

[0226] The algorithm scores each hypothesis from four dimensions:

[0227] ;

[0228] ;

[0229] ;

[0230] ;

[0231] Among them, represents the degree of evidence support for hypothesis , represents the explanatory power of hypothesis , represents the consistency of hypothesis , represents the simplicity of hypothesis , represents the evidence chain of hypothesis , represents the credibility of evidence , represents the relevance of evidence to hypothesis , represents the set of decision analysis results that hypothesis can explain, represents the set of decision analysis results, represents the number of rules that hypothesis is consistent with in the knowledge graph, represents the multi-scale knowledge graph, represents the total number of rules related to hypothesis in the knowledge graph, represents the cardinality or number of elements of the set, represents the logarithmic function;

[0232] Then, the algorithm calculates the weighted total score:

[0233] ;

[0234] Among them, represents the total score, represents the th scoring dimension, represents the weight of the th scoring dimension, satisfying

[0235] ;

[0236] Finally, the algorithm sorts the hypotheses according to the total score to complete the calculation of the MHSR function:

[0237] ;

[0238] Among them, represents the returned set of root cause hypotheses after sorting, represents the sorting function, represents the set of candidate root cause hypotheses, represents the total score, represents descending order.

[0239] Step 404: Process the sorted set of root cause hypotheses, the decision analysis results, and the multi-scale knowledge graph through the interpretive root cause report generation algorithm to obtain the root cause analysis report;

[0240] The mathematical expression of this algorithm is:

[0241] ;

[0242] Among them, represents the output root cause analysis report (including the diagnosis conclusion, the subset of key evidence supporting the hypothesis , the decision development timeline, the scope of influence, and optimization suggestions), represents the interpretive root cause report generation function, represents the sorted set of root cause hypotheses output in step 403, represents the multi-scale knowledge graph output in step 100, represents the decision analysis result output in step 300, represents the report template configuration; this report contains the following key contents:

[0243] ;

[0244] ;

[0245] ;

[0246] ;

[0247] ;

[0248] Among them, , , , and respectively represent the diagnosis conclusion, the subset of key evidence supporting the hypothesis , the decision development timeline, the scope of influence, and optimization suggestions;

[0249] , , , and respectively represent the functions of extracting diagnostic conclusions from the sorted hypothesis set, extracting key evidence from the sorted hypothesis set, constructing a decision development timeline based on the decision analysis results and diagnostic hypotheses, analyzing the influence scope based on the diagnostic results and knowledge graph, and generating optimization suggestions based on the diagnostic results, knowledge graph and historical cases;

[0250] represents the confidence threshold, ;

[0251] The final root cause analysis report provides an in-depth explanation of the decision-making reasons, a clear description of the decision-making development process, and targeted optimization suggestions, providing comprehensive decision-making diagnostic information for users.

[0252] In an embodiment of the present invention, a decision-making system based on a knowledge graph is used to execute the steps in the above-mentioned decision-making method based on a knowledge graph, and includes:

[0253] A multi-granularity knowledge graph construction system for processing input data to obtain a multi-scale knowledge graph;

[0254] A temporal knowledge dual representation learning framework for processing real-time sensor data and the multi-scale knowledge graph to obtain a dynamic representation model;

[0255] A time-varying causal propagation network decision-making system for processing the dynamic representation model and the multi-scale knowledge graph to obtain decision analysis results;

[0256] A multi-hypothesis reasoning system for processing the decision analysis results and the multi-scale knowledge graph to obtain a root cause analysis report.

[0257] In an embodiment of the present invention, an example of the above-mentioned decision-making method based on a knowledge graph is provided:

[0258] This example has been verified through actual application on the catalytic cracking unit of a large petrochemical enterprise. The catalytic cracking unit is a large industrial unit with a processing capacity of 1.5 million tons / year, equipped with 12 core devices, including reactors, regenerators, main blowers, fractionating towers, etc.;

[0259] A total of 427 sensors are installed in the entire system, including 132 temperature sensors, 98 pressure sensors, 87 flow sensors, 42 liquid level sensors, 36 concentration sensors, and 32 vibration sensors. The data acquisition frequency is divided into three levels according to the parameter importance: high-frequency parameters (once per second), medium-frequency parameters (once per 10 seconds), and low-frequency parameters (once per minute);

[0260] This system has accumulated three years of operation historical data, with a total data volume of 21.6 TB, and has recorded 89 well-documented historical fault cases;

[0261] To support the construction of the intelligent decision-making system, a rich expert knowledge base has also been integrated, including 32 equipment technical manuals, 18 process specification documents, 6 fault diagnosis manuals, as well as 24 process flow diagrams (PFD) and 156 piping and instrumentation diagrams (P&ID).

[0262] The main decision-making objectives of this implementation method are: predicting potential equipment failures 12 - 72 hours in advance; providing a root cause analysis accuracy rate of more than 90% for the occurred anomalies.

[0263] During the application process, a total of 89 equipment failures have occurred in this fluid catalytic cracking unit in the past three years, causing varying degrees of production losses, including:

[0264] 18 main blower failures (impeller wear, bearing problems, etc.);

[0265] 23 reactor / regenerator-related failures (catalyst circulation problems, temperature anomalies, etc.);

[0266] 16 fractionation system failures (tray carbon deposition, heat transfer efficiency decline, etc.);

[0267] 14 compressor system failures (valve leakage, shaft seal problems, etc.);

[0268] 11 heat exchanger system failures (scaling, leakage, etc.);

[0269] 7 other types of failures;

[0270] Among them, each failure has caused an average of 24 hours of shutdown and maintenance time, and the annual economic loss is about 32 million yuan.

[0271] Traditional fault detection methods can often only detect problems when the fault occurs or within a short time after the occurrence, lacking the ability of early warning, and the root cause analysis seriously depends on expert experience, making it difficult to form a standardized and automated analysis process.

[0272] Implementation examples of each step:

[0273] Step 100 is implemented to construct a multi-granularity knowledge graph. In the application of the fluid catalytic cracking unit, the system processes multi-source heterogeneous data and constructs a multi-scale knowledge graph including the equipment layer, process layer, parameter layer, and fault layer. The specific data implementation is as follows:

[0274] The system processes 32 equipment technical manuals and obtains 1,456 equipment entities and 5,842 attribute relationships through named entity recognition and relationship extraction;

[0275] Process 24 process flow diagrams (PFDs). Through image recognition, symbol detection, and connection relationship extraction, 128 process units and 342 flow direction relationships are obtained;

[0276] Analyze 156 piping and instrumentation diagrams (P&IDs) and extract 1,248 equipment connection relationships and 368 control loops;

[0277] Use hierarchical association rule mining to process 89 historical fault cases and generate 423 fault-symptom rules and 216 fault-cause rules;

[0278] Process 3-month sample data (about 1.2 TB) of 427 parameters through the parameter causal network learning algorithm, establish 1,864 parameter causal relationships, and the average confidence level reaches 0.83.

[0279] Algorithm implementation parameters:

[0280] Hierarchical association rule mining parameters: The minimum support is 0.05, and the minimum confidence is 0.7;

[0281] Parameter causal network learning parameters: The causal strength threshold is 0.6, and the maximum time delay window is 30 min;

[0282] The statistical data of the finally constructed multi-scale knowledge graph is shown in the following table:

[0283]

[0284] Step 200 is implemented for dual representation learning of time series knowledge:

[0285] The system constructs a dual representation model that fuses time series features and knowledge semantics for 427 monitoring parameters of the fluid catalytic cracking unit. The specific data implementation is as follows:

[0286] The system processes the real-time sensor data stream of 427 parameters (sampling period 1 s - 60 s) and generates a 427×24-dimensional time series feature vector set through the dynamic time window feature extraction algorithm;

[0287] Process the multi-scale knowledge graph containing 1,125 nodes and 4,093 relationships using the knowledge-aware graph embedding algorithm to obtain a knowledge semantic vector set of 427×32 dimensions;

[0288] Process the 427 groups of dual feature vectors with 56 dimensions per group through the temporal knowledge joint representation to generate a fusion feature space of 427×64 dimensions;

[0289] Finally, process the historical state sequences of 427 parameters for 90 days×24 hours through the state dynamic evolution modeling algorithm to construct a dynamic representation model with a prediction accuracy RMSE < 0.08.

[0290] Algorithm implementation parameters:

[0291] Dynamic time window parameters: The initial window is 60 minutes, the adjustment coefficient is 0.4, and the feature importance threshold is 0.1;

[0292] Knowledge graph embedding parameters: The embedding dimension is 32, the learning rate is 0.01, the regularization parameter is 0.1, and the number of training epochs is 50;

[0293] Joint representation parameters: The initial weight of adaptive fusion is 0.6, and the number of non-linear interaction layers is 3;

[0294] State evolution model parameters: The LSTM hidden layer dimension is 128, the time window length is 12 hours, and the number of attention heads is 8;

[0295] The final model training and evaluation results are shown in the following table:

[0296]

[0297] Implementation of step 300: Time-varying causal propagation network decision system:

[0298] Based on the dynamic representation model constructed in step 200, the system realizes time-varying causal propagation network decision analysis; the specific implementation data is as follows:

[0299] The system conducts a prediction deviation analysis on the predicted values of the dynamic representation model for 427 parameters in the next 72 hours, generating a deviation matrix of 427×72 dimensions (including deviation values and deviation confidence levels);

[0300] Combined with the multi-scale knowledge graph, a time-varying causal propagation network containing 427 nodes and 1,864 edges is constructed through the graph-guided causal propagation modeling algorithm;

[0301] Apply the multi-scale pattern detection algorithm to this network to obtain the propagation pattern matching results and confidence evaluation;

[0302] Finally, through the key factor positioning evaluation algorithm, a decision analysis result report is generated based on the matching scores of 12 typical patterns.

[0303] Algorithm implementation parameters:

[0304] Forecast deviation analysis parameters: the historical normal data window is 90 days, and the forecast time window is 72 hours;

[0305] Graph-guided causal propagation parameters: propagation intensity threshold is 0.65, balance factor is 0.4;

[0306] Multi-scale pattern parameters: The predefined pattern library contains 12 typical fault modes, and the multi-scale detection parameter distribution is [0.2, 0.5, 0.3];

[0307] Key factor evaluation parameters: Confidence requirement is 0.8, and risk score weight distribution is [0.4, 0.3, 0.3];

[0308] Among them, some cases successfully detected by the system in actual applications are shown in the following table:

[0309]

[0310] Step 400 implements: multiple hypothesis reasoning:

[0311] Based on the decision analysis results of step 300, a detailed root cause analysis report is generated through multi-hypothesis root cause reasoning. The specific implementation data is shown in the following table:

[0312]

[0313] Algorithm implementation parameters:

[0314] Hypothesis space generation parameters: The domain constraint rule set contains 28 rules and the maximum number of hypotheses is 20;

[0315] Evidence chain construction parameters: The historical case library contains 89 failure cases, and the evidence reliability threshold is 0.75;

[0316] Multi-hypothesis scoring parameters: expert experience weight distribution is [0.3, 0.3, 0.2, 0.2], and multi-dimensional scoring weight distribution is [0.4, 0.3, 0.2, 0.1];

[0317] Report generation parameters: Report template configuration includes report templates for 5 fault types.

[0318] Among them, a sample snippet of the main fan bearing fault analysis report detected once:

[0319] Root cause diagnosis conclusion:

[0320] Excessive wear exists on the inlet side of the main blower (EQ-K101) bearing (P-K101-02), with a confidence level of 0.91.

[0321] Key evidence:

[0322] 1. The bearing vibration value (V-K101-02) has shown a stepwise increase in the past 48 hours, rising from 2.1 mm / s to 4.8 mm / s, exceeding the warning threshold of 4.5 mm / s;

[0323] 2. Vibration spectrum analysis shows characteristic peaks at 43 Hz and 86 Hz, which match the characteristic frequencies of the bearing inner ring fault;

[0324] 3. There is a significant correlation between the bearing temperature (T-K101-02) and the vibration value (r = 0.87), and the temperature rise lags behind the vibration by about 4 hours;

[0325] 4. The vibration pattern of the similar fault case DC-2021-03 in the historical case library highly matches the current fault (similarity 0.92);

[0326] 5. The equipment maintenance record shows that the bearing has been in operation for 8,760 hours, approaching the recommended replacement cycle (10,000 hours).

[0327] Decision development schedule is as shown in the following table:

[0328]

[0329] Impact analysis:

[0330] If not dealt with in time, it is expected that the blower will experience severe vibration within 48 hours, which may cause damage to the shaft seal, and then lead to shutdown for maintenance. The affected scope includes the catalytic cracking reaction system and the air supply system.

[0331] Suggested measures:

[0332] 1. Arrange for planned shutdown and maintenance of the main blower to replace the inlet side bearing;

[0333] 2. Optimize the lubricating oil replacement cycle, adjusting it from the current quarterly replacement to monthly replacement;

[0334] 3. Increase the online vibration monitoring points to achieve full coverage.

[0335] After this implementation method was applied to the catalytic cracking unit for 6 months, by comparing the changes in key indicators before and after deployment, the technical effect of the method was verified. The specific data is as follows:

[0336] Decision accuracy assessment: The system generated 127 decision analysis results during the actual operation process. The verified accuracy indicators are as shown in the following table:

[0337]

[0338] Decision lead assessment. Through the analysis of the fault events successfully predicted by the system, the decision lead distribution is shown in the following table:

[0339]

[0340] Compared with the traditional method, the increasing trend of the decision lead is shown in the following table:

[0341]

[0342] Operating efficiency assessment. The computing resource usage and operating efficiency indicators of the system are shown in the following table:

[0343]

[0344] Economic benefit analysis. The direct economic benefits created for the enterprise within 6 months after the system deployment are shown in the following table:

[0345]

[0346] Long-term adaptability assessment. The continuous improvement effect after the system has run for 6 months is shown in the following table:

[0347]

[0348] As the running time of the system increases, through knowledge accumulation and model optimization, the performance indicators continue to improve, fully verifying its adaptive learning ability and long-term stability.

[0349] Generally speaking, the application of this embodiment in the fluid catalytic cracking unit verifies that the system can meet the requirements of high reliability, high precision and high efficiency in the industrial production environment, successfully realizes the intelligent decision-making method based on the knowledge graph, and solves the problems of data islands, complex causal relationship modeling and insufficient decision reliability faced by the traditional method.

[0350] The above describes the embodiments of the present invention, but these embodiments are not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative rather than restrictive. Under the inspiration of this embodiment, those of ordinary skill in the art can also make more equivalent embodiments in various forms, all of which fall within the protection scope of this embodiment.

Claims

1. A decision-making method based on a knowledge graph, characterized in that, It includes the following steps: Process the input data through a multi-granularity knowledge graph construction system to obtain a multi-scale knowledge graph; Process the real-time sensor data and the multi-scale knowledge graph through a temporal knowledge dual representation learning framework to obtain a dynamic representation model; Process the dynamic representation model and the multi-scale knowledge graph through a time-varying causal propagation network decision-making system to obtain a decision analysis result; Process the decision analysis result and the multi-scale knowledge graph through a multi-hypothesis reasoning algorithm to obtain a root cause analysis report.

2. The decision-making method based on a knowledge graph according to claim 1, characterized in that, The step of processing the input data through a multi-granularity knowledge graph construction system to obtain a multi-scale knowledge graph includes: Process the device specification parameters and the process flow chart through a multi-source heterogeneous data fusion algorithm to obtain a device process knowledge network; Process the historical case library through a hierarchical association rule mining algorithm to obtain an association network; Process the temporal data through a parameter causal network learning algorithm to generate a parameter temporal causal graph; Process the above three types of knowledge networks through a multi-scale knowledge graph integration algorithm to obtain a unified multi-scale knowledge graph.

3. The decision-making method based on a knowledge graph according to claim 2, characterized in that, The mathematical expression of the hierarchical association rule mining algorithm is: ； Among them, represents the output association network, represents the hierarchical association rule mining algorithm, represents the historical case database, represents the minimum support threshold, represents the minimum confidence threshold; The generation of association rules follows the following constraints: ； In the formula, represents the association rule from symptom to cause . represents symptom , represents cause , represents the rule confidence calculation function, represents the rule support calculation function, represents the minimum support threshold, represents the minimum confidence threshold.

4. The decision-making method based on a knowledge graph according to claim 1, characterized in that, The step of processing the real-time sensor data and the multi-scale knowledge graph through a temporal knowledge dual representation learning framework to obtain a dynamic representation model includes: Process the multi-source sensor temporal data through a dynamic time window feature extraction algorithm to obtain a parameter temporal feature vector set; Process the multi-scale knowledge graph through a knowledge-aware graph embedding algorithm to obtain a parameter knowledge semantic vector set; Process the temporal feature vector set and the knowledge semantic vector set through a temporal knowledge joint representation model to obtain a dual representation feature space; Process the dual representation feature space and the historical state sequence through a state dynamic evolution modeling algorithm to obtain a dynamic representation model.

5. The decision-making method based on a knowledge graph according to claim 4, characterized in that, The mathematical expression of the knowledge-aware graph embedding algorithm is: ； Among them, represents the output parameter knowledge semantic vector set, represents the multi-scale knowledge graph, represents the embedding dimension, represents the learning rate, represents the regularization parameter, represents the number of training epochs, represents the knowledge-aware graph embedding algorithm; The optimization objective function of the algorithm is defined as: ； Among them, represents the head entity in the knowledge graph, represents the relationship in the knowledge graph, represents the tail entity in the knowledge graph, represents the head entity of negative sampling in the knowledge graph, represents the tail entity of negative sampling in the knowledge graph, represents the set of triples in the knowledge graph, represents the set of reverse triples in the knowledge graph, represents the relaxation factor in the loss function, represents the head entity in the knowledge graph , the relationship in the knowledge graph and the tail entity in the knowledge graph the distance between them, represents the head entity of negative sampling in the knowledge graph , the relationship and the tail entity of negative sampling in the knowledge graph the distance between them, represents the regularization parameter; represents the maximum operation, that is, when the expression is greater than 0, take its value, otherwise take 0; represents the summation symbol; is the optimization objective function of the knowledge-aware graph embedding algorithm, represents the causal relationship loss function, and the specific definition is: ； Among them, represents the set of causal relationships in the knowledge graph, represents a node and the edge weight between nodes represents the embedding vector of node ; represents the embedding vector of node ; represents the square of the Euclidean distance between node and node .

6. The decision-making method based on a knowledge graph according to claim 1, characterized in that, The step of processing the dynamic representation model and the multi-scale knowledge graph through a time-varying causal propagation network decision-making system to obtain a decision analysis result includes: Process the dynamic representation model through a prediction deviation analysis algorithm to obtain a parameter prediction deviation matrix; Process the parameter prediction deviation matrix and the multi-scale knowledge graph through a graph-guided causal propagation modeling algorithm to obtain a time-varying causal propagation network; Process the time-varying causal propagation network through a multi-scale pattern detection algorithm to obtain a propagation pattern and a matching result; Process the propagation pattern and the matching result through a key factor location evaluation algorithm to obtain a decision analysis result.

7. The decision-making method based on a knowledge graph according to claim 6, wherein, The graph-guided causal propagation modeling algorithm defines the time-varying causal propagation network as a five-tuple: ； Among them, represents the set of parameter nodes, represents the set of directed edges, represents the set of edge weights, represents the set of edge delays, represents the set of node states, represents a time-varying causal propagation network; The calculation of the propagation intensity between any two parameter nodes is as follows: ； Among them, represents the propagation intensity from the parameter node to the parameter node ; represents the causal relationship intensity extracted from the knowledge graph, represents the transfer entropy calculated based on historical data, represents the balance factor. Note: In English, a semicolon (;) is usually used to separate items in a list when the items are relatively short and the relationship between them is clear. Here, I added a semicolon after the first description for better readability in English while keeping the overall meaning the same as the original Chinese text. If this is not in line with your specific requirements, please let me know.

8. The decision-making method based on a knowledge graph according to claim 1, wherein, The step of processing the decision analysis result and the multi-scale knowledge graph through a multi-hypothesis reasoning algorithm to obtain a root cause analysis report includes: Process the decision analysis result and the multi-scale knowledge graph through a hypothesis space generation algorithm to obtain a candidate root cause hypothesis set; Process the candidate root cause hypothesis set, the decision analysis result and the multi-scale knowledge graph through an evidence chain construction algorithm to obtain a set of hypothesis evidence chains; Process the set of hypothesis evidence chains through a multi - hypothesis scoring and ranking algorithm to obtain a ranked set of root - cause hypotheses; Process the ranked set of root - cause hypotheses, the decision - analysis result, and the multi - scale knowledge graph through an explanatory root - cause report generation algorithm to obtain a root - cause analysis report.

9. The decision-making method based on a knowledge graph according to claim 8, wherein, The multi - hypothesis scoring and ranking algorithm scores each hypothesis from four dimensions, including evidence support, explanatory power, consistency, and simplicity, and the specific calculations are as follows: ； ； ； ； Among them, represents the degree of evidence support for the hypothesis , represents the explanatory power of the hypothesis , represents the consistency of the hypothesis , represents the simplicity of the hypothesis , represents the evidence chain of the hypothesis , represents the credibility of the evidence , represents the relevance between the evidence and the hypothesis , represents the set of decision analysis results that the hypothesis can explain, represents the set of decision analysis results, represents the number of rules in the knowledge graph that the hypothesis is consistent with, represents the multi-scale knowledge graph, represents the total number of rules in the knowledge graph related to the hypothesis , represents the cardinality or number of elements of a set, represents the logarithmic function.

10. A decision-making system based on a knowledge graph for executing the decision-making method based on a knowledge graph according to any one of claims 1-9, wherein, Including: A multi - granularity knowledge graph construction system for processing input data to obtain a multi - scale knowledge graph; A temporal knowledge dual - representation learning framework for processing real - time sensor data and the multi - scale knowledge graph to obtain a dynamic representation model; A time - varying causal propagation network decision - making system for processing the dynamic representation model and the multi - scale knowledge graph to obtain a decision - analysis result; A multi - hypothesis reasoning system for processing the decision - analysis result and the multi - scale knowledge graph to obtain a root - cause analysis report.

Images