Business process standard dynamic optimization method and system based on large language model

Abstract

The invention relates to a business process standard dynamic optimization method and system based on a large language model, and belongs to the technical field of computer application, and the method comprises the following steps: constructing a data collection strategy template; performing anomaly detection on the key field based on a preset three-level anomaly detection mechanism to obtain standard structured data; constructing a process execution accurate quantitative evaluation model, and constructing a process topological graph by using a process mining technology; calculating an abnormal probability of each node in the process topological graph by using a multi-task graph neural network, and obtaining a high-attention area in the process topological graph based on the abnormal probability; analyzing the multi-source heterogeneous data associated with the high-attention area through a large language model, positioning root causes, calculating root cause influence weights, and obtaining a root cause list; establishing an intelligent bidirectional feedback mechanism to generate optimization suggestions of the standard process; according to the method, the problems of serious data redundancy, insufficient standard dynamics and lack of feedback mechanisms in the traditional flow standard optimization technology in the power industry are effectively solved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of computer application, and particularly relates to a business process standard dynamic optimization method and system based on a large language model. BACKGROUND

[0002] Efficient management and optimization of business processes are crucial for the development of enterprises. Reasonable business processes can improve operational efficiency, reduce costs, and enhance customer satisfaction. With the rapid development of information technology, in-depth analysis and improvement of business processes using advanced technical means have become the focus of enterprises. The emergence of large language models has brought new opportunities for business process management. Its powerful language understanding and processing capabilities make it possible to extract key information from business process-related texts and conduct effective analysis. However, there are still some limitations in the current technical methods in the field of business process management, which need to be further improved and innovated.

[0003] For example, the Chinese invention patent application with publication number CN119988605A discloses a business process information extraction method based on a large language model, which includes the following steps: collecting business process text data; calling a large language model to preprocess the collected business process text data; learning the context examples of the large language model, and extracting the business process information in the preprocessed business process text data through the learned large language model; and generating a visual business process diagram based on the business process information. The above patent makes positive exploration in extracting business process information using a large language model, mainly focusing on extracting information from business process text data and generating a visual business process diagram, but it lacks consideration of abnormal situations that may occur in the actual operation of the business process, and cannot timely discover potential problems and bottlenecks in the process. In addition, in actual business scenarios, the business environment is constantly changing, and the business process needs to be adjusted and improved in real time according to actual conditions. The above patent only stays in the information extraction and visualization stage, and lacks deep mining and analysis of the extracted information, which cannot accurately find the root cause of the business process deviation, and thus it is difficult to provide targeted optimization suggestions, which is not conducive to the continuous improvement of the efficiency and quality of the business process of the enterprise.

[0004] Therefore, there is an urgent need for a business process optimization method to improve the efficiency and quality of enterprise business processes. SUMMARY

[0005] In order to solve the problems existing in the prior art, the present application proposes a business process standard dynamic optimization method and system based on a large language model, which can obtain feedback-driven standard updates from process execution on one hand, and apply the updated standards to process execution in a timely manner to test the effect on the other hand, thereby constructing a closed-loop mechanism of bidirectional feedback and self-adaptive optimization from standards to processes, realizing real-time updating of standards and continuous optimization of processes.

[0006] The technical scheme of the present application is as follows: In one aspect, the present application proposes a business process standard dynamic optimization method based on a large language model, comprising the following steps: A data collection strategy template is constructed, which is used to extract business process documents and standard specification texts to obtain key fields; Based on a preset three-level anomaly detection mechanism, the key fields are subjected to anomaly detection to obtain standard structured data; A process execution precision quantization evaluation model is constructed, which is based on standard structured data and a standard process diagram, and uses process mining technology to construct a process topology graph; the abnormal probability of each node in the process topology graph is calculated using a multi-task graph neural network, and a high attention area in the process topology graph is obtained based on the abnormal probability; a large language model is used to analyze the multi-source heterogeneous data associated with the high attention area, locate the root cause and calculate the root cause influence weight, and obtain a root cause list; An intelligent two-way feedback mechanism is established, and optimization suggestions for the standard process are generated based on the root cause list; the standard process is dynamically adjusted according to the optimization suggestions and the optimization effect is monitored in real time, and the data collection strategy template is dynamically updated based on the optimization effect.

[0007] As a preferred embodiment, the three-level anomaly detection mechanism will use business rules to quickly filter the input key fields as a first-level detection; the isolated forest algorithm is used for unsupervised anomaly detection of the key fields after quick screening as a second-level detection; the key fields after unsupervised anomaly detection are subjected to semantic consistency verification as a third-level detection; wherein the isolated forest algorithm for unsupervised anomaly detection of the key fields after quick screening is as follows: Collect key field data within a preset time period and construct a feature vector; An isolated forest model is constructed, which adopts a random subspace method, i.e., a subset of the feature vector is randomly selected for partitioning and constructing an isolated tree iTree; after the iTree structure is constructed, the key parameters of the isolated forest model are set; The isolated forest model is trained using a pre-collected training sample set corresponding to the key fields to obtain a trained isolated forest model; each key field to be detected is input into the trained isolated forest model, the isolated forest model calculates the average path length required for the key field to be isolated, and outputs the abnormal score corresponding to the key field; the abnormal score is compared with a preset abnormal threshold, and the second-level abnormal data with an abnormal score higher than the preset abnormal threshold is removed.

[0008] As a preferred embodiment, the specific steps for constructing the process topology graph are as follows: Based on the standard structured data, an actual business process model is constructed by using a process mining technique; for any adjacent event pair in the actual business process model, the frequency of the event pair is counted, and the average duration is calculated; Based on the frequency and the average duration, a process topology graph of the standard structured data is constructed, the nodes of the process topology graph are events in the business process, the edges represent the sequence relationship between the events, and the weight of the edge is determined by the frequency and the average duration.

[0009] As a preferred embodiment, based on the anomaly probability, a high-attention area in the process topology graph is obtained, specifically: A quantitative indicator is calculated for each node in the process topology graph, based on the text information of the node, a semantic feature of the node is extracted by using a pre-set large language model, the quantitative indicator and the semantic feature of the node are fused to obtain a fused feature of the node; an edge feature corresponding to a directed edge in the process topology graph is extracted; The node fused feature, the edge feature and the topological connection structure of the process topology graph are combined to construct a computable graph; The computable graph is input into a multi-task graph neural network, the multi-task graph neural network includes an encoder and a prediction head, wherein the encoder is a graph attention structure stacked by several layers, used for updating and aggregating the node fused feature to obtain a node embedding; the prediction head is connected to the top of the encoder and includes a node head, an edge head and a graph head, the node head receives the node embedding output by the encoder as input and outputs the anomaly probability of the node; the edge head calculates the violation probability of each edge based on the node embedding output by the encoder; the graph head comprehensively classifies the graph compliance based on all node embeddings output by the encoder; The node anomaly probability and the edge violation probability are combined and normalized into a heat value, a double screening strategy is applied to obtain a high-heat node set, and a maximum connected subgraph is intercepted to form a high-attention area.

[0010] As a preferred embodiment, the quantitative indicator includes a node deviation rate and a path compliance difference, wherein: Based on the process topology graph, a critical path is analyzed, the high-frequency path and the longest time-consuming path are focused on, the actual average time consumption and the edge set existing in the actual path are calculated, and the bottleneck link is analyzed in terms of processing time and waiting time; A standard process is loaded, the standard process time consumption and the allowed edge set in the standard process are calculated, and a standard process graph is output; The actual average duration of the node in the process topology graph is compared with the standard average duration of the corresponding node in the standard process graph, and the difference is taken as the node execution deviation rate. Each edge in the process topology graph is traversed to determine whether the current edge exists in the preset compliance edge set of the standard process graph. The number of compliant edges is divided by the total number of edges in the actual topology graph to obtain a compliance difference index.

[0011] As a preferred embodiment, the multi-source heterogeneous data associated with the high attention area is analyzed by a large language model to locate the root cause and calculate the root cause influence weight, and a root cause list is obtained, specifically: Unstructured data associated with the high attention area is obtained, and the unstructured data is analyzed by a large language model to obtain a process deviation corresponding root cause; An influence index of each root cause on the current area process deviation is calculated. The root cause with an influence index exceeding a preset influence threshold is marked as a key bottleneck. The influence index is used as the influence weight of the root cause, and a root cause list containing the key bottleneck root cause and its influence weight is constructed.

[0012] As a preferred embodiment, the calculation of the influence index is specifically: The influence index is obtained by collecting the influence index corresponding to each root cause, including collecting the occurrence frequency, average timeout amplitude, topology heat, and deviation correlation, standardizing and normalizing the influence index, and linearly weighting and summing the standardized and normalized influence index.

[0013] On the other hand, the present application also proposes a business process standard dynamic optimization system based on a large language model, comprising the following modules: Data acquisition module: construct a data acquisition strategy template, which is used to extract business process documents and standard specification texts to obtain key fields; Abnormality detection module: based on a preset three-level abnormality detection mechanism, the key fields are detected for abnormality to obtain standard structured data; Process quantification modeling and abnormality positioning module: construct a process execution precision quantification evaluation model based on standard structured data and a standard process graph, and use process mining technology to construct a process topology graph. Use a multi-task graph neural network to calculate the abnormal probability of each node in the process topology graph, and obtain a high attention area in the process topology graph based on the abnormal probability. The multi-source heterogeneous data associated with the high attention area is analyzed by a large language model to locate the root cause and calculate the root cause influence weight, and a root cause list is obtained; Intelligent two-way feedback module: an intelligent two-way feedback mechanism is established, and optimization suggestions of a standard process are generated based on a root cause list; the standard process is dynamically adjusted according to the optimization suggestions, and the optimization effect is monitored in real time, and the data collection strategy template is dynamically updated based on the optimization effect.

[0014] In another aspect, the present application also provides an electronic device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the method according to the present application when executing the program.

[0015] In another aspect, the present application also provides a computer readable storage medium having a computer program stored thereon, wherein the program is executable on a processor to implement the method according to the present application.

[0016] The present application has the following beneficial effects: (1) The present application provides a business process standard dynamic optimization method based on a large language model, which extracts key fields from business process documents and standard specification texts by using the constructed data collection strategy template, ensures the accurate extraction of key information, and provides basic data for subsequent analysis; then, the key fields are processed by means of a preset three-level anomaly detection mechanism, which includes rapid filtering by using business rules, unsupervised anomaly detection by using isolated forest algorithm, and semantic consistency verification, the three-level anomaly detection mechanism can comprehensively and deeply investigate the anomalies in the key fields, effectively remove the noise and abnormal values in the data, so as to obtain high-quality standard structured data, so that the subsequent process analysis and evaluation based on these data are more accurate and reliable, and the credibility and effectiveness of the present application are improved; (2) The present application provides a business process standard dynamic optimization method based on a large language model, which constructs a process execution quantitative evaluation model, based on standard structured data and standard flowchart, constructs a process topology graph by using process mining technology, the process topology graph provides a clear visual model for subsequent analysis, and intuitively displays the structure and node relationship of the business process; the present application also uses a multi-task graph neural network to accurately calculate the anomaly probability of each node in the process topology graph, so as to quickly locate the high attention area in the process, and improve the efficiency of problem discovery; finally, the present application analyzes the multi-source heterogeneous data associated with the high attention area by means of a large language model, locates the root cause and calculates the root cause influence weight, and obtains a root cause list, which provides a targeted direction and basis for process optimization; (3) The application provides a business process standard dynamic optimization method based on a large language model, which establishes an intelligent two-way feedback mechanism, generates optimization suggestions for the standard process based on a root cause list, then dynamically adjusts the standard process according to the optimization suggestions, and monitors the optimization effect in real time, finally, dynamically updates the data collection strategy template according to the optimization effect, realizes the closed-loop management from problem discovery to process optimization, makes the standard process continuously adjust and improve according to the actual situation, improves the adaptability and efficiency of the process. And the application monitors the optimization effect in real time and dynamically updates the data collection strategy template, ensures that the whole method can continuously adapt to the changes of the business environment, continuously improves the accuracy and pertinence of data collection, and further improves the effect and quality of process optimization. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The technical process schematic diagram provided for the embodiments of the application is shown in the figure. Figure 2 The automatic extraction and processing process schematic diagram of data in the process standard provided for the embodiments of the application is shown in the figure. Figure 3 The actual processing path schematic diagram of each link in the three-line technical support service process scenario provided for the embodiments of the application is shown in the figure. Figure 4 The two-way dynamic feedback optimization process schematic diagram of standard knowledge and business process execution provided for the embodiments of the application is shown in the figure. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0019] It should be understood that the step numbers used herein are only for the convenience of description, and are not limited to the execution sequence of the steps.

[0020] It should be understood that the terms used in the specification of the application are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in the specification and the appended claims of the application, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include the plural forms.

[0021] The terms "include" and "contain" indicate the presence of the described features, whole, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, whole, steps, operations, elements, components and / or sets thereof.

[0022] The term "and / or" means any combination of one or more of the associated listed terms and all possible combinations, and includes these combinations.

[0023] Embodiment one: Referring to Figure 1 The business process standard dynamic optimization method based on a large language model specifically includes the following steps: S100, a data collection strategy template is constructed, which is used to extract business process documents and standard specification texts to obtain key fields; Specifically, the business process documents and standard specification texts are automatically analyzed by a multi-modal document parsing technology in this embodiment. The business process documents record in detail each link and step of the business, and the standard specification texts specify the various guidelines to be followed in the business execution process. In this embodiment, the business process documents and standard specification texts constitute a standard knowledge base of the business process. Based on the parsed business process documents and standard specification texts, a data collection strategy template is automatically generated using semantic analysis of a large language model. The data collection strategy template dynamically determines the key fields required in each specific process link based on the business logic and data requirements obtained by semantic analysis, ensuring that the collected key fields are closely related to the business and avoiding the collection of irrelevant or redundant fields.

[0024] In this embodiment, the data collection strategy template has two specific forms. For ease of understanding, the data collection strategy template will be further illustrated as follows: S101, in one form, the data collection strategy template adopts a json The machine-readable configuration file of the format records field names, data types, value ranges, collection weights, trigger events, and other meta information. Taking the "three-line technical support / emergency request" scenario as an example, the key fields include service type, response time, ticket closing method, and problem description. The service type corresponds to the data type "enum", the source is "ticket.meta", and it is used to clearly indicate the type of service involved in this technical support or emergency request. The response time data type is "float", the source is "log.timer", and the unit is "min". This field can reflect the time spent from receiving the request to starting the response. The ticket closing method data type is "bool", the source is "ticket.flag", and it can represent how the ticket is closed. The problem description data type is "string", the source is "ticket.text", and it records the details of the problem encountered by the customer. At the same time, the machine-readable configuration file also contains validation rules for the above fields, such as the maximum value of "response time" is 60 minutes, and the maximum length of "problem description" is 512 characters. The machine-readable configuration file is automatically distributed to the collection end through the pipeline, ensuring that the collection end can collect key fields according to uniform rules.

[0025] In another form of expression, the data collection strategy template is a runtime dynamic form rendered by the front end according to the configuration file in S101. To reduce the input cost of front-line personnel, the runtime dynamic form only displays key fields with a collection weight not less than a preset collection weight threshold. The system only renders the displayed key fields at runtime and supports incremental loading according to the scenario. For operation and maintenance personnel, the interface usually has no more than 5-8 inputs, and all key fields are accompanied by validation and default values. In the "three-line technical support service process", the large language model automatically extracts a series of key fields as collection objects by analyzing standard process documents and actual operation and maintenance cases. In addition to the previously mentioned service type, response time, ticket closing method, and problem description, there are also response personnel, response duration, and customer satisfaction feedback. The response personnel is used to clearly indicate the staff handling the problem. The response duration is used to understand the duration of the entire response process. The customer satisfaction feedback is used to collect the customer's evaluation of the service. The runtime dynamic form only displays key fields with weights meeting the requirements.

[0026] Through this mechanism, while ensuring the integrity of key fields, the risk of human omission and redundant storage is effectively avoided. Although the collection strategy template stores complete fields and logic in the background, the system only renders key fields with weights higher than the set collection weight threshold at runtime and supports incremental loading according to the scenario. For operation and maintenance personnel, the interface usually has no more than 5-8 inputs, and all key fields are accompanied by validation and default values, ensuring that the operation is simple and easy to operate.

[0027] In the three-line technical support service process optimization implementation, the large language model automatically analyzes the service process standard text and the actual operation and maintenance case, automatically extracts the key fields such as "service type", "response personnel", "problem description", "response duration", "work order closing mode", "customer satisfaction feedback" and the like as collection objects; at the same time, the system dynamically assigns weights according to the importance of the scene; in this embodiment, the emergency field collection weight is greater than 0.8, and the conventional field collection weight is between 0.5 and 0.8, thereby significantly reducing redundant data collection and reducing system burden.

[0028] S200, in order to improve the availability of key fields and data analysis effectiveness, the embodiment also presets a three-level anomaly detection mechanism to clean up the data of the key fields, specifically by the three-level anomaly detection mechanism to detect the key fields, eliminate abnormal data, and obtain standard structured data, wherein the three-level anomaly detection mechanism is specifically: S201, rapid screening based on business rules: The step S201 is a first-level detection, which carries out rapid screening work on the key fields according to the pre-set explicit business rules. In actual business scenarios, the setting of business rules is based on industry experience, business logic and specific business needs. Common business rules include duration anomaly judgment, numerical range limitation, data relationship constraint and the like.

[0029] For example, in the technical support process, the response time of a certain service link is set as a rule, and the service response time should be within a reasonable interval, if it exceeds the specified duration range, it can be preliminarily determined as abnormal. For another example, in the order processing business, there is a certain reasonable range for the order amount, if the order amount of a certain order is far higher or lower than the range, it may be abnormal. In addition, the logical relationship between data can also be used as a business rule, such as in inventory management, the outbound quantity should not be greater than the current inventory quantity, otherwise it is an abnormal situation.

[0030] Through this rapid screening based on business rules, data that obviously do not conform to the business logic can be quickly identified.

[0031] S202, using the Isolation Forest algorithm to perform unsupervised anomaly detection on the key fields after rapid screening: The step S201 is a second-level detection, and the Isolation Forest algorithm is an unsupervised learning algorithm specially used for anomaly detection, which is based on the idea of decision tree, and isolates abnormal data by randomly dividing the data space, specifically: S2011, collect the key field related data in a preset time period, and construct a feature vector. In this embodiment, the time period can reflect the recent characteristics and change trend of the data, and is more in line with the current business actual situation. When constructing the feature vector, the key fields and other important information related thereto need to be considered comprehensively, and the information is quantified and combined to form a vector that can comprehensively describe the data characteristics. For example, when analyzing user behavior data, in addition to the key fields such as operation time and operation frequency of the user, the related information such as the geographic location and device type of the user is also considered to jointly construct the feature vector.

[0032] S2012, a isolation forest model is constructed, and the isolation forest model is composed of a plurality of isolation trees iTree, and each tree is a binary tree structure. When constructing each iTree, a random subspace method is used in this embodiment, that is, a subset of the feature vector is randomly selected for splitting and constructing the iTree; after the iTree structure is constructed, the key parameters of the isolation forest model are set, including the number of trees, the maximum sample size and the maximum tree depth (or height limit). The number of isolation trees determines the overall stability and accuracy of the model, the maximum sample size affects the data size of each tree training, and the maximum tree depth controls the growth degree of the tree to avoid overfitting.

[0033] S2013, the isolation forest model is trained by using the pre-collected training sample set corresponding to the key field, and a trained isolation forest model is obtained.

[0034] When performing secondary detection, each key field to be detected is input into the trained isolation forest model, and the isolation forest model calculates the average path length required for the key field to be isolated. In the principle of the isolation forest algorithm, abnormal data is usually distributed far from normal data, and is more likely to be isolated at a shallow level of the tree. Therefore, the shorter the path of the data being isolated, the greater the difference between it and the normal data, and the greater the possibility of abnormality. Based on this principle, the isolation forest model outputs an abnormal score corresponding to the key field, and the shorter the path, the higher the abnormal score.

[0035] The abnormal score is compared with a preset abnormal threshold, and the secondary abnormal data with an abnormal score higher than the preset abnormal threshold is removed.

[0036] Preferably, in order to automatically adapt the preset abnormal threshold to the data distribution difference in each business scenario and avoid its invalidity due to a fixed value, the embodiment further automatically determines the normal range and fluctuation threshold of each key field based on the statistical distribution of historical key field data, and determines the dynamic abnormal threshold based on the normal range and fluctuation threshold. When the abnormal score exceeds the dynamic abnormal threshold, it is determined that the data is abnormal and is rejected. For example, for the key field of "remote access response time", the isolation forest model obtains its normal range of 2-10 minutes through statistical analysis of historical key field data. If the monitoring value exceeds 15 minutes, the key field is automatically marked as a secondary abnormality and processed.

[0037] S203, performing semantic consistency verification on the key fields after unsupervised anomaly detection to obtain standard structured data: Step S203 is a third-level detection, which is implemented by using a large language model (LLM). Specifically: The large language model checks the matching degree of the key fields and the business semantics to ensure that the data is reasonable and consistent at the semantic level. For example, when analyzing medical data, the diagnosis result should have a reasonable logical relationship with the symptom description and examination data. If there is a semantic inconsistency, it may indicate a third-level abnormality of the data.

[0038] Through semantic consistency verification, it can not only further confirm whether the data is truly abnormal, but also classify and process the third-level abnormal data. Different processing methods are adopted for different types of third-level abnormal data, such as correcting third-level abnormal data caused by data entry errors, or rejecting or separately marking third-level abnormal data that actually has problems. Finally, standard structured data is obtained, which meets the requirements in terms of quality and semantics, thereby improving the effectiveness and accuracy of subsequent data analysis.

[0039] Preferably, the embodiment further includes calling a large language model to infer the context of the work order log corresponding to the third-level abnormal key field, and analyzing the specific reasons for the third-level abnormality. For example, when the response time of a certain service link is abnormally high, the system will call the large language model to analyze the related work order log. The work order log contains detailed business operation records, user feedback, processing process, and other information. The large language model performs semantic analysis and understanding on these text information to determine the specific reasons for the delay, such as "network failure" or "inadequate scheduling".

[0040] It should be noted that for each key field (such as response time, processing capacity, success rate, etc.), historical data is continuously collected, and its statistical distribution characteristics (mean, variance, skewness, kurtosis, quantile, etc.) are analyzed to dynamically determine the "normal range" and "fluctuation threshold" of the field in the current context (business line, time period, resource status, etc.).

[0041] When the data points in the data stream are marked as abnormal in any level detection, and the related core field value exceeds the threshold dynamically calculated according to the current scene, the system will perform the related operation: the system automatically triggers an early warning and notifies the relevant personnel, performs an automatic cleaning action, automatically corrects, and rejects.

[0042] Referring to Figure 2 , the input business process document and the standard specification text are analyzed by a large language model, and the key fields of the input business process document and the standard specification text are filtered. The key fields obtained by filtering are input into the three-level anomaly detection, and after sequentially passing through rule detection, isolated forest detection and semantic consistency detection, structured data is output.

[0043] S300, construct a process execution precision quantitative evaluation model, the process execution precision quantitative evaluation model is based on standard structured data and standard process graph, using process mining technology to construct a process topology graph; using a multi-task graph neural network to calculate the anomaly probability of each node in the process topology graph, obtaining a high attention area in the process topology graph based on the anomaly probability; through a large language model, analyzing the multi-source heterogeneous data associated with the high attention area, locating the root cause and calculating the root cause influence weight, obtaining a root cause list; specifically: S301, based on the standard structured data, using process mining technology to construct an actual business process model; for any adjacent event pair in the actual business process model, the frequency of the event pair is counted, and the average duration is calculated, which is expressed in the formula as: ; In the formula, represents the average duration of event to event ; represents the frequency of event jumping to event ; represents the th event, represents the th event, and the two events are adjacent events; represents the sum of the related quantities of the th event to the th event, represents the duration of event to event ; Based on the above frequency and average duration, a process topology graph of the standard structured data is constructed, the nodes of the process topology graph are events in the business process, and the edges represent the sequence relationship between events. The weight of the edge is determined by the frequency and the average duration.

[0044] S302, analyze the process topology graph to find high-frequency paths and the longest time-consuming path, and determine the critical path. At the same time, analyze the processing time and waiting time of the possible bottleneck link in detail; through the above analysis, obtain the process topology graph marked with the critical path and the bottleneck link; S303, according to the process topology graph marked with the critical path and the bottleneck link, calculate the weight of the node (such as the importance and activity of the node) and the weight of the edge (such as the importance of the jump), generate the process topology graph, which clearly shows the frequency, time length and other detailed information of the specific steps and execution paths in the actual business process; S304, load the standard process graph, automatically compare the process topology graph with the standard process graph, and calculate the quantitative indicators, including node deviation rate and path compliance difference.

[0045] Preferably, in the present embodiment, the node execution deviation is the difference between the actual node execution time and the standard time, which can be expressed in the formula as follows: ; In the formula, is the actual average time consumption of node (event ), is the standard time consumption of node (event ); is the node execution deviation of node (event ); The path compliance difference is to traverse each edge in the process topology graph, judge whether the edge exists in the preset compliance edge set of the standard process graph, divide the number of compliance edges by the total number of edges in the actual topology graph, and express it in the formula as follows: ; In the formula, is the path compliance difference, E is the edge set existing in the actual path, is the edge set allowed in the standard process, indicates the number of actual path edges that meet the standard edge rules, |E| indicates the total number of edges in the actual path; the path compliance difference index equal to 1 indicates that all actual edges exist in the standard process; less than 1 indicates that there is an "out-of-bound" jump.

[0046] S305, output absolute duration deviation and path compliance difference index, after the construction of the directed process topology graph is realized, the directed process topology graph needs to be automatically and intelligently positioned (such as bottleneck nodes, illegal paths, and overall risk areas) to identify hidden problems in complex business processes that are difficult to find through manual inspection or simple statistics.

[0047] That is, after the automatic comparison of the process topology graph and the standard process graph is completed, the process topology graph is converted into a computable graph rich in multi-dimensional information, and a multi-task graph neural network is used for training and reasoning, and finally a high attention area of node-level abnormal probability distribution is output. Specifically: S3051, feature extraction and fusion, construct a computable graph In this embodiment, a relevant quantitative indicator is calculated for each activity node in the process topology graph. In this embodiment, the relevant quantitative indicator includes: the number of occurrences, i.e. the number of occurrences of the node in the entire process, the average time, i.e. the average time required for the node to execute once.

[0048] The text attributes (such as: name, role, type) of the node are processed, and the pre-set large language model is used to convert these text information into dense semantic vectors.

[0049] The statistical feature vector and the semantic feature vector of the same node are combined to form a new fused node feature vector; the quantitative attributes and the semantic attributes are combined, so that both the objective running data (i.e. quantitative attributes) and the semantic role and meaning in the business are retained, providing rich information for the multi-task graph neural network.

[0050] The directed edges (representing the jump / flow relationship between nodes) in the process topology graph are calculated; such as: jump frequency, i.e. the number of times of jumping from the source node to the target node (reflecting the path heat / importance), average step time, i.e. the average time required to complete a jump (measuring the jump efficiency), cross-departmental label, indicating whether the source node and the target node belong to different departments (reflecting the collaboration complexity of the process link); the calculated original edge features are normalized.

[0051] All node features and edge features generated by the above steps, as well as the topological connection structure of the process topology graph, are combined to construct a structured graph data structure, which completely contains: topological structure, node features, edge features, i.e. a computable graph, as the direct input data source for subsequent multi-task graph neural network learning and reasoning.

[0052] In this embodiment, the process of feature extraction and fusion, and construction of a computable graph is as follows: The work order system draws the actual processing path of each link of "telephone support", "remote access" and "on-site support" based on the real-time recorded service response log, see Figure 3 .

[0053] By automatic comparison and analysis with the standard "three-line technical support service process" topology, the quantitative deviation value of response speed, node jump, rework of each link is calculated, and the specific optimization link is determined (for example, it is found that the "remote access link" consumes 20% more time on average than the standard, which is preliminarily locked as a bottleneck to be optimized). According to the node statistical characteristics, node semantic characteristics and edge characteristics, feature fusion and computable graph construction are performed; the above features and graph-level indicators jointly constitute the multi-task graph neural network supervision signal. Then multi-task graph neural network training and indicator output are performed, and 4-layer GATv2 is used for message passing; the node reading layer outputs the time deviation value and risk score of each link, and the graph reading layer outputs the efficiency, quality, cost and risk four-dimensional scores.

[0054] The optimal model of the verification set is selected through hyperparameter search and using Early Stopping. The node heat map is output, and the visual service is pushed for the next sub-step to directly read.

[0055] S3052、Multi-task graph neural network model design and training In this embodiment, a multi-task graph neural network is used, which includes an encoder and a prediction head, specifically: The encoder uses a graph attention structure with edge information. This structure will stack several layers. When processing the computable graph data, each layer of graph attention structure will consider the edge information in the graph, update and aggregate the node features, and finally obtain the node embedding after multi-layer stacking. These node embeddings contain the feature information of the nodes in the graph and their context information in the graph structure. In this embodiment, three different prediction heads are connected at the top of the encoder, which are node head, edge head and graph head, wherein: The node head is mainly used to predict the abnormal probability of the node, which receives the node embedding output by the encoder as input, and obtains the abnormal probability value of each node through a series of calculations; the edge head is responsible for predicting the violation probability of the edge, which is also based on the node embedding output by the encoder, and obtains the violation probability of each edge through corresponding calculation; the graph head is used to perform regression or classification tasks, and is used to determine the overall compliance classification of the graph, and by comprehensively considering all node embedding information output by the encoder, the compliance of the entire graph is evaluated.

[0056] The core principle of the multi-task graph neural network described in this embodiment is to capture complex relationships in graph data using a graph attention mechanism. The graph attention mechanism can adaptively assign weights to different nodes and edges, allowing the model to focus on information that is more important for the prediction task. By stacking multiple layers of graph attention structures with edge information, the model can learn high-order interactions and hidden patterns between nodes and edges in the graph. The three prediction heads analyze and predict the graph data from different perspectives. The node head focuses on the abnormality of individual nodes, the edge head focuses on the violation of edges, and the graph head focuses on the compliance of the entire graph. This multi-task design allows the model to learn different levels of information simultaneously during training, improving the model's generalization ability and prediction accuracy.

[0057] In this embodiment, the workflow of the multi-task graph neural network is as follows: During the training phase, the encoder processes the input computable graph data and obtains node embeddings through multi-layer stacking of graph attention structures with edge information. The three prediction heads process the node embeddings output by the encoder to calculate the node anomaly probability, edge violation probability, and overall graph compliance rating. Each prediction head calculates its own loss. In this embodiment, the node loss L_node measures the difference between the model's predicted node anomaly probability and the true node anomaly state. The edge loss L_edge measures the difference between the predicted edge violation probability and the true edge violation state. The graph loss L_graph measures the difference between the predicted overall compliance rating and risk score and the true value. These losses are combined into a total loss by weighting coefficients, expressed as: L_total = α*L_node + β*L_edge + γ*L_graph where α, β, and γ are weighting coefficients, L_total is the total loss, L_node is the node loss, L_edge is the edge loss, and L_graph is the graph loss. The total loss is used to drive the backpropagation algorithm to update the model's parameters, so that the model's prediction gradually approaches the true value until training converges.

[0058] After the training converges, the latest process topology is converted into a computable graph of the same format, and a forward inference is performed to obtain the anomaly probability of each node; the anomaly probability of the node is normalized to a heat value between zero and one, and the closer the value is to 1, the more the model considers that the node has an anomaly, a bottleneck or a risk point, and these values are visualized on the graph to form a heat map; in order to extract the area that needs to be focused on, a "double screening" strategy is adopted, that is, a threshold is set and the top several percent of heat values are taken, and the union of the two is taken to obtain a high-heat node set; the high-heat nodes are intercepted in the process topology, all direct connection edges between the high-heat nodes are found, and the nodes and their direct connections are retained; the connectivity of the nodes and their direct connections is analyzed, and the largest connected subgraph composed of them is extracted, which is the "high-attention area".

[0059] The purpose of extracting the high-attention connected subgraph is that in the specific implementation process, a single abnormal node may be important, but multiple abnormal and connected nodes may indicate a complex problem area or a violation path; extracting the connected subgraph containing multiple high-heat nodes and their connections can more clearly present the core area where the problem occurs and its context association, which is very valuable for locating the root cause (in this embodiment, a core node or a path leading to the problem). The subgraph provides local focus visualization, which facilitates manual analysis and decision-making.

[0060] S3053、In this embodiment, the key root cause is located and the index library is updated by combining a large language model with a multi-dimensional quantitative index system, specifically: Automatic analysis of root causes and quantification of weights based on a large language model: sample classification is performed on historical key logs and work order texts related to the high-attention area; after classification, a root cause label library that can describe the deviation source (such as "data check delay", "approval authority conflict", "resource scheduling shortage", "dependent system failure", etc.) is obtained; Obtain the current and high-attention area associated with the standard specification text, work order text and other unstructured data. Utilize the semantic analysis and knowledge reasoning capabilities of the large language model to automatically identify and extract key information segments and patterns related to the above root cause label library from the above unstructured data; automatically mine the main reason list leading to the current area process deviation (such as node delay, error); calculate the influence index of each root cause on the specific area process deviation; usually use the word frequency, co-occurrence relationship, causal relationship indicator and other features learned by the large language model to estimate; finally output a list containing multiple root causes and their preliminary influence degree (weight).

[0061] In this embodiment, the specific calculation of the influence index is as follows: Four core indicators are calculated for each candidate root cause: collection occurrence frequency, which in this embodiment refers to the frequency of the root cause being identified by the large language model in the associated text data, reflecting the frequency of root cause occurrence; average timeout amplitude, which in this embodiment refers to the average degree of timeout or additional delay caused by the root cause each time it occurs, reflecting the length of time delay caused by the root cause; topology heat, which in this embodiment refers to the heat value of the node (or its adjacent node) where the root cause occurs on the node heat map, reflecting the association strength of the root cause with the most abnormal area; and deviation correlation, which in this embodiment refers to the statistical correlation between the occurrence of the root cause and the deviation score of the node / area, reflecting the objective association between the root cause and the current deviation degree; for each root cause, the values of its four core indicators are respectively scaled to the [0, 1] interval.

[0062] Impact index calculation: linearly weighted sum of the above four core indicators, expressed in formula as: ; In the formula, is the impact index of the i-th root cause, is the standardized value of the j-th root cause on the i-th core indicator, is the weight coefficient of the four core indicators, and the cumulative sum of the weight coefficients is 1, i.e. is the i-th standardized indicator value of root cause j; the impact index obtained by the above calculation is a weighted comprehensive value, and the higher the value, the more outstanding the overall performance of the root cause j in the four dimensions of collection occurrence frequency, average timeout amplitude, topology heat, and deviation correlation, and the more significant the comprehensive impact.

[0063] Softmax operation is performed on the impact index set of all k root causes to obtain the final relative weight of each root cause j, expressed in formula as: ; In the formula, is the impact index of the i-th root cause, is the natural exponential form of the impact index of the i-th root cause, is the sum of all root cause index values, is the weight of the i-th root cause, where m is the total number of actual root causes; 0 <1, ∑ = 1; Softmax normalizes the impact index into a probability distribution (i.e. all weights and are 1); The highest root cause has the largest relative impact and is considered the most critical bottleneck; Softmax can amplify the weight of the highest I value and suppress extremely low values, making the ranking result more significant.

[0064] The root cause with the highest weight (or multiple root causes with weights significantly higher than the average) is automatically marked as a key bottleneck; the key bottleneck root cause label obtained by analysis and its calculated weight are written back to the index library and possible related information such as time, region, etc.), and a root cause list is obtained, facilitating automatic monitoring of index changes by the system and triggering a secondary optimization process.

[0065] For ease of understanding, an example is given as follows: In the three-line technical support service process optimization implementation, the model automatically analyzes historical work order texts to summarize the main reasons for process deviation and their weight influences; for example, after analyzing more than 1,000 work orders in the past six months, the model finds that more than 70% of the "on-site support" link delays are closely related to "manufacturer support response not timely", and its influence weight in the overall process delay is more than 60%, providing a basis for subsequent optimization.

[0066] The node heat map is read, and it is found that the average deviation score of the "remote access" node in the past 7 days is greater than the threshold value; the deviation score of the "on-site support" node in the past 30 days exceeds the threshold value multiple times and is on the rise. The system automatically adds the above two nodes and their upstream paths to the high-attention queue, and subsequent analysis will focus on these areas.

[0067] The model automatically selects 1,263 work order texts, customer service chat records, and on-site handover logs associated with the high-attention nodes in the past 180 days to summarize the main reasons for process deviation and their weight influences, and finds that more than 70% of the "on-site support" link delays are closely related to "manufacturer support response not timely", and its influence weight in the overall process delay is more than 60%. The root cause list and its influence weight are written back to the network structure table for detection and improvement.

[0068] S400, an intelligent two-way feedback mechanism is established, optimization suggestions for the standard process are generated based on the root cause list; the standard process is dynamically adjusted according to the optimization suggestions and the optimization effect is monitored in real time, and the data collection strategy template is dynamically updated based on the optimization effect.

[0069] As shown in Figure 4 , to realize dynamic interaction and mutual optimization between the standard knowledge base and the actual operation process, forming a closed-loop feedback system, the embodiment establishes an A / B test framework and a dynamic adjustment algorithm, specifically: S401. This embodiment utilizes reinforcement learning technology to enable the precise application of standard knowledge in business process execution: Based on root cause reports, it automatically identifies key nodes in the standard process execution. If the root cause report shows that a certain step often leads to deviations in business execution due to unclear standards, the system will provide more detailed and focused information on the standards for that step when making personalized recommendations to frontline personnel in different positions. For example, for call center operators, the system will emphasize the dispatch timeliness standard. In the scenario of optimizing the third-line technical support service process, when a call center operator receives a "non-emergency request," the system uses a large model to identify the process standard and generates and clearly displays the specific operation steps and precautions in real time on the work order processing interface, such as "Please be sure to dispatch the work order to the third-line engineer within 5 minutes." For engineers, the system will highlight the key points of the technical specifications on their interface, while reminding them of the details of standard step execution, potential risks, and precautions, and automatically verifying field compliance when filling out the work order. Personalized recommendations ensure that standard knowledge is accurately conveyed to business personnel, laying the foundation for subsequent two-way feedback.

[0070] S402. During business process execution, when using a large language model to identify potential problems in the standards, newly collected problems are categorized and analyzed in depth, referencing root cause reports. If the root cause report has previously pointed out the root cause of a certain type of problem, then when facing similar problems, the optimization direction can be quickly determined based on the report's conclusions. Based on the identification results, the system generates targeted standard optimization suggestions and automatically generates standard revision documents, including a revision comparison table, an impact assessment report, and an expected benefit analysis. In the optimization of the three-tier technical support service process, after collecting execution data and problem feedback, the large language model, combined with root cause reports, identifies potential problems in the standards, generates optimization suggestions, and automatically forms a modification document. After expert confirmation, this document is automatically published to update the standard knowledge base and pushed to the execution stage in real time to ensure rapid response to actual business changes. Subsequently, a canary release mechanism is implemented to verify the effectiveness of the new standard in stages, enabling the standard knowledge to be continuously optimized according to the actual business situation.

[0071] S403, To ensure the continuous effectiveness of the bidirectional dynamic feedback of standard knowledge and business process execution, the embodiment also establishes an indicator monitoring dashboard to display real-time information such as new standard adoption rate, key indicator trend, abnormal fluctuation warning, etc. When the monitoring finds abnormal conditions in the implementation of new standards, a comprehensive judgment is made based on the root cause report. If the root cause report shows that a certain type of abnormality may be caused by multiple factors, the analysis logic in the root cause report is used to gradually investigate and quickly determine whether the problem is with the standard itself, personnel operation, or external environment. The system monitors the changes in process data in real time, automatically tracks and evaluates the implementation effect of new standards in real time, uses hypothesis testing methods (p<0.05) to verify the significance of optimization, and automatically triggers adjustment or rollback mechanisms according to the verification results. For example, in the optimization of the three-line technical support service process, after the "remote access link standard" is optimized, the system automatically collects the optimized work order processing time data in real time, uses a large language model to analyze the improvement effect, and if it finds that the remote access processing efficiency has improved by 30%, it means that the optimization is effective. If it finds that the new standard has a lack of adaptability problem, the large model will automatically generate a new round of optimization suggestions and feedback, thereby dynamically adjusting the standard to ensure that the business process is continuously in the optimal state.

[0072] Through the above series of closely connected operations, the embodiment constructs a complete closed-loop feedback system. The root cause report runs through the bidirectional dynamic feedback process of standard knowledge and business process execution, providing important support for decision-making and optimization at each link, enabling standard knowledge and business process execution to promote each other and continuously optimize, achieving a virtuous cycle of bidirectional dynamic feedback.

[0073] Embodiment Two: The embodiment also proposes a business process standard dynamic optimization system based on a large language model, which specifically includes the following modules: Data collection module: build a data collection strategy template for extracting business process documents and standard specification texts to obtain key fields; Abnormality detection module: based on a preset three-level abnormality detection mechanism, perform abnormality detection on the key fields to obtain standard structured data; Process quantification modeling and abnormality positioning module: build a process execution precise quantification evaluation model based on standard structured data and standard process diagrams, use process mining technology to build a process topology graph, use a multi-task graph neural network to calculate the abnormal probability of each node in the process topology graph, obtain high-attention areas in the process topology graph based on the abnormal probability, and analyze multi-source heterogeneous data associated with the high-attention areas through a large language model to locate the root cause and calculate the root cause influence weight to obtain a root cause list; Intelligent bidirectional feedback module: an intelligent bidirectional feedback mechanism is established, and optimization suggestions of a standard process are generated based on a root cause list; the standard process is dynamically adjusted according to the optimization suggestions, and the optimization effect is monitored in real time, and the data collection strategy template is dynamically updated based on the optimization effect.

[0074] Embodiment three: The embodiment provides an electronic device, including a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the method according to any one of the embodiments of the present application when executing the program.

[0075] Embodiment four: The embodiment provides a computer readable storage medium, and the computer readable storage medium stores a computer program, and the computer program is executable on a processor to implement the method according to any one of the embodiments of the present application.

[0076] In the embodiments of the present application, "at least one" refers to one or more, and "multiple" refers to two or more. The "and / or" describes the association relationship of the associated objects, which means that there can be three kinds of relationships, for example, A and / or B, which can represent the cases of A alone, A and B together, and B alone. Wherein A and B can be singular or plural. The character " / " generally represents that the associated objects before and after it are in an "or" relationship. "At least one of the following" and the like means any combination of these items, including any combination of single or multiple items. For example, at least one of a, b and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, wherein a, b, and c can be single or multiple.

[0077] Those of ordinary skill in the art can realize that the units and algorithm steps described in the embodiments disclosed herein can be realized in electronic hardware, computer software and a combination of electronic hardware and computer software. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0078] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system, device and unit can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0079] In several embodiments provided by the present application, any function, if realized in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in part, or parts of the technical solutions, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (Read-Only Memory, ROM for short), a random access memory (Random Access Memory, RAM for short), a magnetic disk or an optical disk, and various media that can store program codes.

[0080] The above description is only some embodiments of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation, or direct or indirect application in other related technical fields, based on the content of the specification and drawings of the present application, are also included in the patent protection scope of the present application.

Claims

1. A large language model-based business process standard dynamic optimization method, characterized in that, The method comprises the following steps: A data collection strategy template is constructed, which is used to extract business process documents and standard specification texts to obtain key fields; Based on a preset three-level anomaly detection mechanism, the key fields are subjected to anomaly detection to obtain standard structured data; A process execution precision quantization evaluation model is constructed, which is based on standard structured data and a standard process diagram, and uses process mining technology to construct a process topology graph; An abnormal probability of each node in the process topology graph is calculated using a multi-task graph neural network, and a high attention area in the process topology graph is obtained based on the abnormal probability; a multi-source heterogeneous data associated with the high attention area is analyzed by a large language model to locate a root cause and calculate a root cause influence weight, and a root cause list is obtained; An intelligent bidirectional feedback mechanism is established, and optimization suggestions for the standard process are generated based on the root cause list; The standard process is dynamically adjusted according to the optimization suggestions, and the optimization effect is monitored in real time, and the data collection strategy template is dynamically updated based on the optimization effect.

2. The method of claim 1, wherein the method is characterized by, The three-level anomaly detection mechanism uses business rules to quickly filter the input key fields as the first-level detection; an isolated forest algorithm is used for unsupervised anomaly detection of the key fields after quick screening as the second-level detection; The key fields after unsupervised anomaly detection are subjected to semantic consistency verification as the third-level detection; wherein the isolated forest algorithm is used for unsupervised anomaly detection of the key fields after quick screening as the second-level detection specifically comprises: Collecting key field data in a preset time period and constructing a feature vector; An isolated forest model is constructed, and a random subspace method is used, that is, a subset of the feature vector is randomly selected for splitting and constructing an isolated tree iTree; after the iTree structure is constructed, the key parameters of the isolated forest model are set; The isolated forest model is trained using a pre-collected training sample set corresponding to the key fields, and a trained isolated forest model is obtained; each key field to be detected is input into the trained isolated forest model, the isolated forest model calculates the average path length required for the key field to be isolated, outputs an abnormal score corresponding to the key field, and compares the abnormal score with a preset abnormal threshold to eliminate second-level abnormal data with an abnormal score higher than the preset abnormal threshold.

3. The method of claim 1, wherein the method further comprises: The specific steps of constructing the process topology graph are: Based on the standard structured data, a process mining technology is used to construct an actual business process model; for any adjacent event pair in the actual business process model, the frequency of the event pair is counted, and the average duration is calculated; Based on the frequency and average duration, a process topology graph of the standard structured data is constructed, the nodes of the process topology graph are events in the business process, and the edges represent the sequence relationship between the events, and the weight of the edge is determined by the frequency and the average duration.

4. The method of claim 1, wherein the method is characterized by, The abnormal probability of each node in the process topology graph is calculated using a multi-task graph neural network, and a high attention area in the process topology graph is obtained based on the abnormal probability, specifically: The quantitative index of each node in the process topology graph is calculated, the semantic features of the node are extracted based on the text information of the node by using a preset large language model, the quantitative index and the semantic features of the node are fused, and the fused features of the node are obtained; Edge features corresponding to the directed edges in the process topology graph are extracted; The node fused features, edge features and topological connection structure of the process topology graph are combined to construct a computable graph; The computable graph is input into a multi-task graph neural network, the multi-task graph neural network includes an encoder and a prediction head, the encoder is a graph attention structure stacked with several layers, is used for updating and aggregating the node fused features to obtain node embedding, the prediction head is connected to the top of the encoder and includes a node head, an edge head and a graph head, the node head receives the node embedding output by the encoder as input and outputs the abnormal probability of the node, the edge head calculates the violation probability of each edge based on the node embedding output by the encoder, and the graph head integrates all the node embedding output by the encoder to perform graph compliance grading; The node abnormal probability and the edge violation probability are combined and normalized into heat values, a double screening strategy is applied to obtain a high-heat node set, and a maximum connected subgraph of the high-heat node set is intercepted to form a high-attention area.

5. The method of claim 4, wherein the method further comprises: The quantitative index includes a node deviation rate and a path compliance difference, wherein: Based on the process topology graph, the key path is analyzed, the high-frequency path and the longest time-consuming path are focused on, the actual average time consumption and the edge set existing in the actual path are calculated, and the bottleneck link is analyzed in terms of processing time and waiting time; A standard process is loaded, the standard process time consumption and the allowed edge set in the standard process are calculated, and a standard process graph is output; The process topology graph and the standard process graph are compared, the difference between the actual average duration of the node in the process topology graph and the standard average duration of the corresponding node in the standard process graph is taken as the node execution deviation rate, and each edge in the process topology graph is traversed to determine whether the current edge exists in the preset compliance edge set of the standard process graph, the compliance edge number is divided by the total edge number of the actual topology graph to obtain a compliance difference index.

6. The method of claim 4, wherein the method further comprises: The multi-source heterogeneous data associated with the high-attention area is analyzed by using a large language model, the root cause is located, and the root cause influence weight is calculated to obtain a root cause list, specifically: Unstructured data associated with the high-attention area is obtained, the unstructured data is analyzed by using a large language model, and a process deviation corresponding root cause is obtained; The influence index of each root cause on the current area process deviation is calculated, the root cause with an influence index exceeding a preset influence threshold is marked as a key bottleneck, the influence index is taken as the influence weight of the root cause, and a root cause list including the key bottleneck root cause and the influence weight is constructed.

7. The method of claim 6, wherein the method further comprises: The calculation of the influence index is specifically as follows: The influence index is calculated by collecting the influence index corresponding to each root cause, including collecting the occurrence frequency, the average timeout amplitude, the topology heat and the deviation correlation, standard normalizing the influence index, and linearly weighting and summing the standard normalized influence index to obtain the influence index.

8. A large language model-based business process standard dynamic optimization system, characterized in that, The following modules are included: The data collection module: construct a data collection strategy template, which is used to extract business process documents and standard specification texts, and obtain key fields; The anomaly detection module: based on the preset three-level anomaly detection mechanism, the key fields are subjected to anomaly detection, and standard structured data is obtained; The process quantization modeling and anomaly positioning module: construct a process execution precise quantization evaluation model, which is based on standard structured data and standard process diagrams, and uses process mining technology to construct a process topology graph; Use multi-task graph neural network to calculate the anomaly probability of each node in the process topology graph, obtain the high attention area in the process topology graph based on the anomaly probability; analyze the multi-source heterogeneous data associated with the high attention area through the large language model, locate the root cause and calculate the root cause influence weight, and obtain the root cause list; The intelligent bidirectional feedback module: establish an intelligent bidirectional feedback mechanism, and generate optimization suggestions for the standard process based on the root cause list; According to the optimization suggestion, the standard process is dynamically adjusted and the optimization effect is monitored in real time, and the data collection strategy template is dynamically updated based on the optimization effect.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to realize the method of any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to realize the method of any one of claims 1-7.

Images