A method, system, device and medium for enterprise intelligent diagnosis based on large model

Abstract

The present application provides a method, system, device and medium for intelligent diagnosis of an enterprise based on a large model, which belongs to the field of enterprise diagnosis technology. The method is as follows: enterprise heterogeneous data is collected through a multi-source data interface, and cleaning, feature extraction and cross-modal alignment and fusion are performed to obtain real-time enterprise data; then, based on the industry indicators to which the enterprise belongs, a knowledge graph is constructed through a graph attention network, and the association weights between nodes are updated regularly; the real-time enterprise data is input into a pre-trained multimodal large model for preliminary analysis, and a risk heat map is output; key abnormal indicators are identified from the risk heat map, and subgraphs related to the key abnormal indicators are extracted from the knowledge graph, and then structured prompt words are generated for the subgraphs, and then the subgraphs are input into the multimodal large model together with standardized real-time enterprise data for joint reasoning analysis to generate a visual diagnosis report. The present invention realizes accurate identification and intelligent diagnosis of enterprise risks and improves decision-making efficiency.

Description

Technical Field

[0001] The present application belongs to the field of enterprise diagnosis technology, and specifically relates to a large model-based enterprise intelligent diagnosis method, system, equipment and medium. Background Art

[0002] With the acceleration of digital transformation, the amount of data generated during business operations is rapidly increasing, covering multiple dimensions including finance, production, sales, supply chain, and customer interactions. This data is often multi-source and heterogeneous. How to quickly and accurately identify business operational risks from this complex data structure and provide actionable diagnostic recommendations has become a critical requirement for intelligent decision-making.

[0003] Traditional enterprise diagnostics rely on expert experience, identifying issues through manual statistics and report analysis. This approach is not only inefficient and unable to cope with massive amounts of enterprise data, but also relies on human subjectivity, making it prone to overlooking hidden management issues. Furthermore, traditional diagnostics rely on predefined rules for risk detection, making them unable to adapt to dynamic business changes and struggling to handle anomalies in the coordination of multiple indicators. Machine learning-based enterprise diagnostics employ a single-modal analysis paradigm that models specific types of data separately. This often overlooks cross-modal management, leading to incomplete diagnosis. Another approach involves constructing static knowledge graphs to aid analysis of enterprise entity relationships. This results in delayed node and edge weight updates and a failure to reflect real-time business dynamics.

[0004] In addition, the current multi-source heterogeneous data in enterprise operations lacks a unified standardization method, resulting in feature alignment errors. Traditional knowledge graphs are unable to update node weights in real time, making it difficult to capture sudden risks. Summary of the Invention

[0005] In a first aspect, an embodiment of the present application provides an enterprise intelligent diagnosis method based on a large model, comprising the following steps:

[0006] S1. Collect heterogeneous enterprise data through multi-source data interfaces, cleanse, extract features, and perform cross-modal alignment and fusion on the collected heterogeneous enterprise data to obtain standardized real-time enterprise data.

[0007] S2. Based on the industry indicators of the enterprise and combined with standardized real-time enterprise data, a knowledge graph with enterprise entities, business indicators, and industry standards as nodes is constructed through a graph attention network. The association weights between nodes are regularly updated based on the real-time enterprise data.

[0008] S3. Input standardized real-time enterprise data into a pre-trained multimodal large model for preliminary analysis, outputting a risk heat map containing abnormal indicators and their probabilities.

[0009] S4. Identify key anomaly indicators from the risk heat map and extract subgraphs related to these indicators from the knowledge graph. Then, generate structured prompts from the subgraphs and input them into a multimodal large model along with standardized real-time enterprise data for joint reasoning and analysis.

[0010] S5. Convert the joint reasoning analysis results into executable diagnostic reports and present them to enterprise users through a visual interactive interface.

[0011] Furthermore, the specific steps of step S1 are as follows:

[0012] S11. Obtain enterprise financial data, sales data, inventory data, and human resources data through the enterprise ERP system interface, obtain production and manufacturing data through the enterprise MES system interface, and obtain customer interaction data through the enterprise CRM system interface;

[0013] S12. Use RPA crawler tools to capture operational data from various business processes within the enterprise and interface sensor data through IoT device protocols.

[0014] S13. Clean the collected data, remove duplicate, erroneous, incomplete, and abnormal data records, and supplement missing values;

[0015] S14. Normalize and standardize the cleaned data to unify the dimensions, and then perform feature screening based on the SHAP value.

[0016] S15. Map the feature-screened data to a unified dimensional space and align them based on timestamps to obtain standardized real-time enterprise data.

[0017] Furthermore, the specific steps of step S2 are as follows:

[0018] S21. Extract enterprise entity nodes from standardized real-time enterprise data using a named entity recognition algorithm, load indicator metadata from the industry benchmark library of the enterprise's industry as business indicator nodes, extract clause entity constraints from the enterprise's industry standard documents as industry standard nodes, and complete node initialization;

[0019] S22. Constructing knowledge graph through graph attention network:

[0020] Perform GRU encoding on the standardized enterprise real-time data to generate the feature vector of the corresponding node;

[0021] Set the initial value of the edge weight between nodes according to industry rules;

[0022] S23. When the update conditions are triggered, the node and edge weights of the knowledge graph are dynamically updated.

[0023] Furthermore, the specific steps of step S23 are as follows:

[0024] S231. When a new enterprise entity or indicator anomaly is detected from standardized real-time enterprise data, a node is automatically created for the knowledge graph and the node's embedding vector is initialized:

[0025]

[0026] in, is a physical feature, Embedding for similar nodes;

[0027] S232. For existing nodes in the knowledge graph that have no associated data in the standardized enterprise real-time data for N consecutive days, they are marked as dormant nodes and stored offline;

[0028] S233. Update the knowledge graph according to the set update cycle and use the temporal graph convolutional network to adjust the edge weights between nodes:

[0029]

[0030] in, is the embedding of node i at time t, is the amount of data change between nodes.

[0031] Furthermore, the specific steps of step S3 are as follows:

[0032] S31. Divide standardized enterprise real-time data into structured data, text data, and time series data;

[0033] S32. Mapping the categorical features in the structured data into dense vectors of a set dimension through an embedding algorithm, and extracting local pattern features from the numerical features in the structured data using a 1D convolutional neural network;

[0034] S33. The text data is divided into table-like text and plain text. The table-like text is processed using the LayoutXLM model to obtain the table structure and content semantics. The plain text is extracted using the RoBERTa-wwm model, and the token vector is extracted for the Chinese text.

[0035] S34. Use the Informer model to capture long-term dependencies for sensor time series data, and use the N-HiTS model to decompose trend and periodic features for business time series data.

[0036] S35. The features extracted from structured data, text data, and time series data are used to calculate the correlation weights between each modality through the cross-modal attention mechanism algorithm:

[0037]

[0038] in, is the query vector of modality i, is the key vector of mode j, is the dimension scaling factor;

[0039] S36. Use the gated fusion algorithm to generate joint features for the feature vectors of structured data, text data, and time series data:

[0040]

[0041]

[0042] in, represents the gate value, σ(*) is the activation function, is the trainable parameter matrix, is the structured data feature vector, is the text data feature vector, Time series data feature vector, is the fused cross-modal feature vector;

[0043] S37. Calculate the abnormal probability of each business indicator through the multi-layer perceptron model:

[0044]

[0045]

[0046] in, is the abnormal probability of the i-th indicator, , is the weight matrix of the first layer of the multilayer perceptron model, is the weight matrix of the second layer of the multilayer perceptron model, is the bias vector of the first layer of the multilayer perceptron model, is the bias value of the second layer of the multilayer perceptron model, d is the dimension of the input feature, m is the number of hidden units in the first layer, and 1 is the dimension of the output;

[0047] Determine the abnormal business indicator based on the abnormal probability of each business indicator, and generate an abnormal probability vector by combining the abnormal business indicator and the corresponding abnormal probability;

[0048] S38. Use a deconvolutional network to map the anomaly probability vector to the dimensions of the original standardized enterprise real-time data to generate a pixel-level risk heat map.

[0049] S39. Display the pixel-level risk heat map in layers.

[0050] Furthermore, the specific steps of step S4 are as follows:

[0051] S41. The specific steps for identifying key abnormal indicators from the risk heat map are as follows:

[0052] Dynamic probability thresholds are pre-set to determine business indicators that meet any of the following conditions as key abnormal indicators:

[0053] Business indicators whose abnormal probability exceeds the dynamic probability threshold;

[0054] Affect business indicators of more than N business departments;

[0055] The abnormal probability increases for M consecutive detection cycles exceeding the preset amplitude threshold;

[0056] S42. Specific steps for extracting subgraphs related to key anomaly indicators from the knowledge graph:

[0057] Starting from the key anomaly indicator node, perform an N-level deep search in the knowledge graph:

[0058] The first layer consists of nodes directly connected to the key anomaly indicator nodes, and the next layer consists of nodes connected to the previous layer nodes;

[0059] The search priority of nodes in each layer is calculated by edge weight and SHAP value:

[0060]

[0061] in, is the weight between node i and node j, SHAP is the feature screening result, which indicates the contribution of the feature to the model output. and is an adjustable weight coefficient used to balance the influence of edge weight and SHAP value;

[0062] S43. The specific steps of generating structured prompt words from the subgraph are as follows:

[0063] Pre-build a prompt word template and fill the relevant nodes and relationships in the searched subgraph into the prompt word template, thereby converting the subgraph into a structured prompt word as a natural language description;

[0064] S44. The specific steps for inputting structured prompt words and standardized real-time enterprise data into a multimodal large model for joint reasoning analysis are as follows:

[0065] Construct a multimodal large model of hierarchical hybrid coding, including parallel text coding layer, table coding layer, time series coding layer and data fusion layer;

[0066] The structured prompt words are concatenated with the structured data in the standardized real-time data and then input into the text encoding layer of the multimodal large model to generate a structured query vector.

[0067] The structured prompt words are concatenated with the standardized text data in real-time data and then input into the table encoding layer of the multimodal large model to generate a text query vector.

[0068] Input the time series prompt words and the time series data in the standardized real-time data into the time series encoding layer of the multimodal large model to generate a time series query vector;

[0069] The structured query vector, text query vector, and temporal query vector are input into the data fusion layer, and the joint reasoning analysis results are generated through cross-modal attention fusion and gated feature fusion.

[0070] Furthermore, the specific steps of step S5 are as follows:

[0071] S51. Use the T5 model to compress the text of the joint reasoning analysis results, generate an executive summary containing key anomaly indicators, root cause analysis, and risk prediction, and highlight key data in the executive summary;

[0072] S52. Output a hierarchical risk heat map and set up a drill-down interface to connect to standardized enterprise real-time data and knowledge graphs;

[0073] S53. In response to the user's click operation on the risk heat map, the corresponding standardized enterprise real-time data and related subgraphs in the knowledge graph are displayed in a linked manner, and the similarity ranking results of the related case nodes are displayed.

[0074] In a second aspect, an embodiment of the present application further provides an enterprise intelligent diagnosis system based on a large model, comprising:

[0075] The data collection and processing module is used to collect heterogeneous enterprise data through multi-source data interfaces, and clean, extract features, and align and fuse the collected heterogeneous enterprise data across modalities to obtain standardized real-time enterprise data;

[0076] The knowledge graph construction and update module is used to build a knowledge graph with enterprise entities, business indicators and industry standards as nodes based on the industry indicators of the enterprise and standardized real-time enterprise data through the graph attention network, and regularly update the association weights between nodes according to the real-time enterprise data;

[0077] The risk heat map generation module is used to input standardized real-time enterprise data into a pre-trained multimodal large model for preliminary analysis and output a risk heat map containing abnormal indicators and abnormal indicator probabilities;

[0078] The reasoning analysis module is used to identify key abnormal indicators from the risk heat map and extract subgraphs related to the key abnormal indicators from the knowledge graph. The subgraphs are then converted into structured prompt words and input into a multimodal large model together with standardized real-time enterprise data for joint reasoning analysis.

[0079] The diagnosis display module is used to convert the joint reasoning analysis results into executable diagnosis reports and display them to enterprise users through a visual interactive interface.

[0080] In a third aspect, an embodiment of the present application further provides an electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the program, the steps of the enterprise intelligent diagnosis method based on a large model as described in the first aspect are implemented.

[0081] In a fourth aspect, an embodiment of the present application further provides a storage medium on which a computer program is stored, and when the computer program is executed by a processor, the steps of the enterprise intelligent diagnosis method based on a large model as described in the first aspect are implemented.

[0082] It can be seen from the above technical solutions that this application has the following advantages:

[0083] The enterprise intelligent diagnosis method, system, equipment and medium based on large models provided in this application realize the accurate identification, intelligent analysis and real-time monitoring of enterprise operation risks through multi-source data collection and processing, dynamic knowledge graph construction, multimodal deep analysis and visual display; it not only improves the accuracy, efficiency and operability of diagnosis, but also reduces labor costs and the possibility of risk omissions, provides support for enterprise decision-making, and enhances the level of digital transformation of enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] In order to more clearly illustrate the technical solution of the present application, the following is a brief introduction to the drawings required for the description. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0085] Figure 1 Schematic diagram of the process of the enterprise intelligent diagnosis method based on large model of the present invention.

[0086] Figure 2 Schematic diagram of the enterprise intelligent diagnosis system based on large model of the present invention. DETAILED DESCRIPTION

[0087] The specific steps of the enterprise intelligent diagnosis method based on the large model will be described in detail below, and various embodiments of the present disclosure will be described more comprehensively. The present disclosure may have various embodiments, and adjustments and changes may be made therein. However, it should be understood that there is no intention to limit the various embodiments of the present disclosure to the specific embodiments disclosed herein, but rather that the present disclosure should be understood to cover all adjustments, equivalents and / or alternatives that fall within the spirit and scope of the various embodiments of the present disclosure.

[0088] For example, as digital transformation accelerates, businesses are accumulating massive amounts of heterogeneous, multi-source data during their operations, covering finance, production, sales, supply chain, customer interactions, and more. Faced with this complex data, businesses urgently need a method that can quickly and accurately identify operational risks and provide practical diagnostic recommendations to meet the needs of intelligent decision-making.

[0089] Traditional diagnostic methods have obvious limitations. On the one hand, manual statistics and report analysis methods that rely on expert experience are not only inefficient and difficult to process large-scale data, but also prone to missing potential problems due to human subjective judgment. On the other hand, risk detection methods based on fixed rules lack flexibility and have difficulty adapting to rapid changes in the business, especially when multiple indicators are abnormal in collaboration, they often seem powerless. Existing machine learning methods are usually limited to the analysis of a single data type and tend to ignore the association between cross-modal data, resulting in one-sided diagnostic results. At the same time, static knowledge graph methods cannot reflect business dynamics in real time due to the untimely update of node and edge weights, making it difficult to capture sudden risks.

[0090] Furthermore, enterprises lack a unified, standardized approach when processing multi-source, heterogeneous data, leading to errors in feature alignment. Traditional knowledge graph technology cannot update node weights in real time, making it difficult to adapt to the rapid changes in the business environment, thus affecting the timeliness and accuracy of risk identification.

[0091] To address the above issues, this embodiment provides an enterprise intelligent diagnosis method based on a large model. Through multi-source data fusion, dynamic knowledge graphs and multimodal analysis, it can achieve accurate identification and intelligent diagnosis of enterprise risks and improve decision-making efficiency.

[0092] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0093] See also Figure 1FIG2 is a flowchart of a method for intelligent enterprise diagnosis based on a large model in a specific embodiment, the method comprising the following steps:

[0094] S1. Collect heterogeneous enterprise data through multi-source data interfaces, cleanse, extract features, and perform cross-modal alignment and fusion on the collected heterogeneous enterprise data to obtain standardized real-time enterprise data.

[0095] It should be noted that by collecting heterogeneous enterprise data through multi-source data interfaces and performing cleaning, feature extraction, and cross-modal alignment and fusion, data scattered across various enterprise systems can be integrated into a unified standard format, improving data quality, reducing data dimensions, and providing a data foundation for subsequent analysis.

[0096] S2. Based on the industry indicators of the enterprise and combined with standardized real-time enterprise data, a knowledge graph with enterprise entities, business indicators, and industry standards as nodes is constructed through a graph attention network. The association weights between nodes are regularly updated based on the real-time enterprise data.

[0097] It should be noted that based on the industry indicators and standardized real-time data of the enterprise, a knowledge graph with enterprise entities, business indicators and industry standards as nodes is constructed through the graph attention network, which can intuitively display the complex relationships and business logic within the enterprise; the association weights between nodes are regularly updated according to the enterprise's real-time data to ensure that the knowledge graph can reflect changes in the enterprise's operating status in real time, providing the latest structured information for subsequent risk analysis and diagnosis; using enterprise entities, business indicators and industry standards as graph nodes facilitates in-depth analysis of various relationships and influencing factors in enterprise operations, providing semantic information for identifying anomalies and risks;

[0098] S3. Input standardized real-time enterprise data into a pre-trained multimodal large model for preliminary analysis, outputting a risk heat map containing abnormal indicators and their probabilities.

[0099] It should be noted that inputting standardized real-time data into a pre-trained multimodal large model for preliminary analysis can fully utilize the model's ability to process multiple data types and identify potential patterns and abnormal information in the data. The output includes a risk heat map containing abnormal indicators and abnormal probabilities, which presents the risk distribution in the enterprise operation in an intuitive and visual way, helping to quickly locate problem areas. Through comprehensive analysis of multimodal data, potential abnormal indicators can be preliminarily identified, providing direction for in-depth diagnosis and analysis.

[0100] S4. Identify key anomaly indicators from the risk heat map and extract subgraphs related to these indicators from the knowledge graph. Then, generate structured prompts from the subgraphs and input them into a multimodal large model along with standardized real-time enterprise data for joint reasoning and analysis.

[0101] It should be noted that by identifying key abnormal indicators from the risk heat map, it is possible to identify the problem points with the greatest impact on enterprise operations, thereby improving the pertinence and efficiency of diagnosis; extracting subgraphs related to key abnormal indicators from the knowledge graph and generating structured prompt words facilitates in-depth analysis of the background and related factors of abnormal indicators, providing rich contextual information for subsequent reasoning and analysis; inputting structured prompt words and standardized real-time data into a multimodal large model for joint reasoning and analysis, combining data-driven and knowledge-guided approaches, can more comprehensively and accurately analyze the causes and impacts of abnormal indicators and provide in-depth diagnostic results;

[0102] S5. Convert the joint reasoning analysis results into an executable diagnostic report and present it to enterprise users through a visual interactive interface;

[0103] It should be noted that converting the joint reasoning analysis results into executable diagnostic reports provides enterprises with clear and specific improvement suggestions and action plans; displaying the diagnostic reports through a visual interactive interface makes complex data and analysis results intuitive, making it easier for enterprise users to quickly obtain key information and improve user experience.

[0104] This embodiment collects and standardizes heterogeneous enterprise data through a multi-source data interface to improve data quality, builds a knowledge graph based on the industry indicators of the enterprise, and dynamically updates the association weights between nodes to reflect the enterprise's operating status in real time; performs preliminary analysis through a pre-trained multimodal large model to generate a risk heat map to intuitively display the enterprise's abnormal indicators and probabilities; conducts joint reasoning analysis by integrating structured prompt words generated by knowledge graph subgraphs with standardized real-time data to improve the accuracy and depth of diagnosis; converts the results of the joint reasoning analysis into an executable diagnostic report and displays it through a visual interactive interface to facilitate enterprise users to understand and use the diagnostic results.

[0105] Furthermore, as a refinement and extension of the specific implementation of the above embodiment, in order to fully illustrate the specific implementation process of this embodiment, another enterprise intelligent diagnosis method based on a large model is provided, which includes the following steps:

[0106] S1. Collect heterogeneous enterprise data through multi-source data interfaces, cleanse, extract features, and perform cross-modal alignment and fusion on the collected heterogeneous enterprise data to obtain standardized real-time enterprise data. The specific steps of step S1 are as follows:

[0107] S11. Obtain enterprise financial data, sales data, inventory data, and human resources data through the enterprise ERP system interface, obtain production and manufacturing data through the enterprise MES system interface, and obtain customer interaction data through the enterprise CRM system interface;

[0108] For example, corporate financial data includes core financial indicators such as the balance sheet, income statement, and cash flow statement that can demonstrate the company's financial health. Corporate sales data includes product line sales, sales volume, sales region distribution, and customer purchase data, which help companies grasp market dynamics and sales trends. Corporate inventory data includes the quantity, turnover rate, and inventory cost of raw materials, work-in-progress, and finished goods inventory, which help companies optimize inventory management. Human resources data includes employee attendance, performance evaluation, staff turnover, and human resource allocation data for each department.

[0109] S12. Use RPA crawler tools to capture operational data from various business processes within the enterprise and interface sensor data through IoT device protocols.

[0110] Exemplarily, the sensor data includes temperature data, humidity data, and pressure data;

[0111] S13. Clean the collected data, remove duplicate, erroneous, incomplete, and abnormal data records, and supplement missing values;

[0112] S14. Normalize and standardize the cleaned data to unify the dimensions, and then perform feature screening based on the SHAP value.

[0113] Specifically, the normalization process uses the Min-Max normalization method to scale the data to the [0,1] interval. The specific formula is as follows:

[0114]

[0115] The standardization process uses the Z-score standardization method to convert the data into a distribution with a mean of 0 and a standard layer of 1. The specific formula is as follows:

[0116]

[0117] Where μ is the mean of the data and σ is the standard deviation of the data;

[0118] The feature screening based on SHAP value is as follows:

[0119] Each piece of collected data is treated as a feature, and the SHAP value is used to analyze the contribution of each feature to the output of the pre-trained model. A simple machine learning model (such as linear regression or decision tree) is trained to calculate the SHAP value of each feature. The features are sorted according to the absolute value of the SHAP value, and the top K features that contribute most to the model output are selected.

[0120] The specific formula is as follows:

[0121]

[0122] in, is the SHAP value of feature i, is the set of all features, is a subset of features, is a feature subset The model predicted value of

[0123] S15. Map the feature-screened data to a unified dimensional space and align them based on timestamps to obtain standardized real-time enterprise data.

[0124] Specifically, data from different sources (such as financial data, sales data, inventory data, etc.) are mapped into a unified multidimensional space. For example, the "sales," "cost," and "profit" indicators in financial data, the "sales volume" and "customer purchasing behavior" indicators in sales data, and the "inventory turnover rate" and "inventory cost" indicators in inventory data are uniformly mapped into a space with "time," "business type," and "indicator value" as the dimensions.

[0125] S2. Based on the enterprise's industry indicators and standardized real-time enterprise data, a knowledge graph with enterprise entities, business indicators, and industry standards as nodes is constructed through a graph attention network. The association weights between nodes are regularly updated based on the enterprise's real-time data. The specific steps of step S2 are as follows:

[0126] S21. Extract enterprise entity nodes from standardized real-time enterprise data using a named entity recognition algorithm, load indicator metadata from the industry benchmark library of the enterprise's industry as business indicator nodes, extract clause entity constraints from the enterprise's industry standard documents as industry standard nodes, and complete node initialization;

[0127] For example, the enterprise entity node can extract the department, production line, and equipment number obtained from the original ERP system or MES system interface from the standardized enterprise real-time data and generate a unique URI identifier;

[0128] S22. Constructing knowledge graph through graph attention network:

[0129] Perform GRU encoding on the standardized enterprise real-time data to generate the feature vector of the corresponding node;

[0130] Set the initial value of the edge weight between nodes according to industry rules;

[0131] For example, the weight of the association between the Finance Department and the Purchasing Department, both of which are enterprise entity nodes, is set to 0.8;

[0132] S23. When the update condition is triggered, dynamically update the node and edge weights of the knowledge graph;

[0133] The specific steps of step S23 are as follows:

[0134] S231. When a new enterprise entity or indicator anomaly is detected from standardized real-time enterprise data, a node is automatically created for the knowledge graph and the node's embedding vector is initialized:

[0135]

[0136] in, is a physical feature, Embedding for similar nodes;

[0137] S232. For existing nodes in the knowledge graph that have no associated data in the standardized enterprise real-time data for N consecutive days, they are marked as dormant nodes and stored offline;

[0138] S233. Update the knowledge graph according to the set update cycle and use the temporal graph convolutional network to adjust the edge weights between nodes:

[0139]

[0140] in, is the embedding of node i at time t, is the amount of data change between nodes;

[0141] S3. Input the standardized real-time enterprise data into the pre-trained multimodal large model for preliminary analysis, and output a risk heat map containing abnormal indicators and abnormal indicator probabilities. The specific steps of step S3 are as follows:

[0142] S31. Divide standardized enterprise real-time data into structured data, text data, and time series data;

[0143] S32. Mapping the categorical features in the structured data into dense vectors of a set dimension through an embedding algorithm, and extracting local pattern features from the numerical features in the structured data using a 1D convolutional neural network;

[0144] Specifically, a 1D convolutional neural network is a neural network architecture that processes sequence data (such as time series, signals, etc.). It uses a one-dimensional convolution kernel to slide on the sequence and calculates the dot product between the convolution kernel and the sequence fragment to extract local features.

[0145] For example, taking enterprise sales data (sales recorded monthly) as an example, the specific data format is as follows:

[0146] [0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0]

[0147] These data have been processed by Min-Max normalization and range between [0,1];

[0148] The convolution operation of the 1D convolutional neural network slides the convolution kernel on the sequence and calculates the dot product between the convolution kernel and the sequence fragment. For example, a convolution kernel of size 3 and a step size of 1 are used. The weights of the convolution kernel can be randomly initialized, for example: [0.2, 0.5, 0.3]

[0149] The calculation process of the convolution operation is as follows:

[0150] First convolution: 0.1×0.2+0.2×0.5+0.3×0.3=0.23

[0151] Second convolution: 0.2×0.2+0.3×0.5+0.4×0.3=0.31

[0152] And so on, until the entire sequence is covered;

[0153] Through convolution operations, 1D CNN can extract local patterns in sequences; for example, in sales data, convolution operations can identify local features such as upward trends, downward trends, or cyclical changes in sales;

[0154] S33. The text data is divided into table-like text and plain text. The table-like text is processed using the LayoutXLM model to obtain the table structure and content semantics. The plain text is extracted using the RoBERTa-wwm model, and the token vector is extracted for the Chinese text.

[0155] It should be noted that the LayoutXLM model is a pre-trained multimodal model for processing table-like text, which can simultaneously capture the structural information and content semantics of the table;

[0156] Take the table text containing sales data as shown in Table 1 as an example:

[0157] Table 1

[0158]

[0159] Use the LayoutXLM model to capture the 4×3 row and column relationships of the table and the semantics of the product name, sales volume, and sales amount fields;

[0160] It should be noted that RoBERTa-wwm is a pre-trained Chinese language model used to process plain text data and can extract semantic features of text;

[0161] S34. Use the Informer model to capture long-term dependencies for sensor time series data, and use the N-HiTS model to decompose trend and periodic features for business time series data.

[0162] S35. The features extracted from structured data, text data, and time series data are used to calculate the correlation weights between each modality through the cross-modal attention mechanism algorithm:

[0163]

[0164] in, is the query vector of modality i, is the key vector of mode j, is the dimension scaling factor;

[0165] S36. Use the gated fusion algorithm to generate joint features for the feature vectors of structured data, text data, and time series data:

[0166]

[0167]

[0168] in, Represents the gate value, which is used to determine the mixing ratio of structured data and text data. σ(*) is the activation function. is the trainable parameter matrix, is the structured data feature vector, which comes from the processing output of 1D CNN logarithmic features. is the text data feature vector, the [CLS] tag embedding from the RoBERTa model, Time series data feature vector, the final hidden state mean from the Informer model, It is the fused cross-modal feature vector, which is used for subsequent anomaly detection tasks;

[0169] S37. Calculate the abnormal probability of each business indicator through the multi-layer perceptron model:

[0170]

[0171]

[0172] in, is the abnormal probability of the i-th indicator, , , is the weight matrix of the first layer of the multilayer perceptron model, is the weight matrix of the second layer of the multilayer perceptron model, is the bias vector of the first layer of the multilayer perceptron model, is the bias value of the second layer of the multilayer perceptron model, d is the dimension of the input feature, m is the number of hidden units in the first layer, and 1 is the dimension of the output;

[0173] Determine the abnormal business indicator based on the abnormal probability of each business indicator, and generate an abnormal probability vector by combining the abnormal business indicator and the corresponding abnormal probability;

[0174] S38. Use a deconvolutional network to map the anomaly probability vector to the dimensions of the original standardized enterprise real-time data to generate a pixel-level risk heat map.

[0175] Specifically, a deconvolution operation is performed on the anomaly probability vector:

[0176]

[0177] in, is the abnormal probability vector, K is the number of business indicators, It is a matrix constructed from original standardized enterprise real-time data;

[0178]

[0179] in, It is an upsampling operation that interpolates P from the K dimension to the original data dimension m×n. AttentionMask(X) is a binary mask generated based on the validity of the original data. Represents element-wise multiplication;

[0180] The upsampling operation is to interpolate P from K dimensions to m×n dimensions using bilinear interpolation:

[0181]

[0182] in, is the abnormal probability of the kth business indicator, u and v are interpolation coordinates, indicating the position of the kth element in P in the m×n matrix, is the upsampled matrix The (i,j)th element in ;

[0183] The binary mask is based on Generate a binary mask:

[0184]

[0185] in, is the indicator function, when 1 if yes, 0 otherwise. is the original data matrix The (i,j)th element in ;

[0186] The upsampled matrix Z and attention mask AttentionMask( X ) Element-wise multiplication suppresses the heat value of the invalid area:

[0187]

[0188] Finally, the wind heat map matrix is ​​output ;

[0189] S39. Displaying the pixel-level risk heat map in layers;

[0190] For example, the first-level risk heat map shows an overview of department-level anomalies, and the second-level risk heat map shows the degree of deviation of specific indicators;

[0191] S4. Identify key abnormal indicators from the risk heat map, extract subgraphs related to the key abnormal indicators from the knowledge graph, generate structured prompt words from the subgraphs, and then input them into the multimodal large model together with standardized real-time enterprise data for joint reasoning and analysis. The specific steps of step S4 are as follows:

[0192] S41. The specific steps for identifying key abnormal indicators from the risk heat map are as follows:

[0193] Dynamic probability thresholds are pre-set to determine business indicators that meet any of the following conditions as key abnormal indicators:

[0194] Business indicators whose abnormal probability exceeds the dynamic probability threshold;

[0195] Affect business indicators of more than N business departments;

[0196] The abnormal probability increases for M consecutive detection cycles exceeding the preset amplitude threshold;

[0197] For example, the dynamic probability threshold can be set to , Represents the average value of all current abnormal probabilities, Represents the standard deviation of all current abnormal probabilities;

[0198] Business indicators that affect more than three business objectives can be set as key abnormality indicators. Business indicators with an abnormality probability increase of more than 15% for two consecutive detection cycles can be set as abnormal business indicators.

[0199] S42. Specific steps for extracting subgraphs related to key anomaly indicators from the knowledge graph:

[0200] Starting from the key anomaly indicator node, perform an N-level deep search in the knowledge graph:

[0201] The first layer consists of nodes directly connected to the key anomaly indicator nodes, and the next layer consists of nodes connected to the previous layer nodes;

[0202] The search priority of nodes in each layer is calculated by edge weight and SHAP value:

[0203]

[0204] in, is the weight between node i and node j, SHAP is the feature screening result, which indicates the contribution of the feature to the model output. and is an adjustable weight coefficient used to balance the influence of edge weight and SHAP value, e.g. Take 0.6, Take 0.4;

[0205] S43. The specific steps of generating structured prompt words from the subgraph are as follows:

[0206] Pre-build a prompt word template and fill the relevant nodes and relationships in the searched subgraph into the prompt word template, thereby converting the subgraph into a structured prompt word as a natural language description;

[0207] S44. The specific steps for inputting structured prompt words and standardized real-time enterprise data into a multimodal large model for joint reasoning analysis are as follows:

[0208] Construct a multimodal large model of hierarchical hybrid coding, including parallel text coding layer, table coding layer, time series coding layer and data fusion layer;

[0209] The structured prompt words are concatenated with the structured data in the standardized real-time data and then input into the text encoding layer of the multimodal large model to generate a structured query vector.

[0210] The structured prompt words are concatenated with the standardized text data in real-time data and then input into the table encoding layer of the multimodal large model to generate a text query vector.

[0211] Input the time series prompt words and the time series data in the standardized real-time data into the time series encoding layer of the multimodal large model to generate a time series query vector;

[0212] Specifically, in each encoding layer, the query vector Q is generated by the prompt word and the data:

[0213]

[0214] in, The hidden state of the prompt word, is the data feature;

[0215] The structured query vector, text query vector, and temporal query vector are input into the data fusion layer, and the joint reasoning analysis results are generated through cross-modal attention fusion and gated feature fusion.

[0216] Specifically, the inter-modal attention weight is calculated by the following formula:

[0217]

[0218] in, is the query vector of modality i (such as text query vector), is the key vector of mode j (such as the time series key vector), is the vector dimension scaling factor (e.g. d k =64);

[0219] Use gated dynamic weighting of each modal feature:

[0220]

[0221] in, is the characteristic line of mode i, is the global average pooling vector of the prompt word;

[0222] Extract key entities from the prompt words to perform abnormal root cause analysis:

[0223]

[0224] in, is the weight matrix used to calculate the probability distribution of key entities, is the bias vector used to adjust the probability distribution;

[0225] Use a temporal convolutional network to predict risk within time T:

[0226]

[0227] Among them, TCN is a temporal convolutional network, which is used to capture long-term dependencies in time series data. It is a time series feature, which represents the feature representation of time series data;

[0228] Retrieve case nodes in the knowledge graph whose similarity with the current fused cross-modal feature vector is greater than a threshold, filter out the top-N case nodes, and generate output prompt words;

[0229] Output the output prompt words, root cause analysis results, and risk transmission prediction results as structured joint reasoning analysis results;

[0230] S5. Convert the joint reasoning analysis results into an executable diagnostic report and present it to the enterprise user through a visual interactive interface. The specific steps of step S5 are as follows:

[0231] S51. Use the T5 model to compress the text of the joint reasoning analysis results, generate an executive summary containing key anomaly indicators, root cause analysis, and risk prediction, and highlight key data in the executive summary;

[0232] For example, key data such as abnormal probability exceeding dynamic thresholds or indicators affecting multiple departments;

[0233] S52. Output a hierarchical risk heat map and set up a drill-down interface to connect to standardized enterprise real-time data and knowledge graphs;

[0234] S53. In response to the user's click operation on the risk heat map, the corresponding standardized enterprise real-time data and related subgraphs in the knowledge graph are displayed in a linked manner, and the similarity ranking results of the related case nodes are displayed.

[0235] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0236] like Figure 2 As shown, the following is an embodiment of the enterprise intelligent diagnosis system based on a big model provided by the embodiment of the present disclosure. This system and the enterprise intelligent diagnosis method based on a big model in the above-mentioned embodiments belong to the same inventive concept. For details not fully described in the embodiment of the enterprise intelligent diagnosis system based on a big model, please refer to the embodiment of the above-mentioned enterprise intelligent diagnosis method based on a big model.

[0237] The system includes:

[0238] The data collection and processing module is used to collect heterogeneous enterprise data through multi-source data interfaces, and clean, extract features, and align and fuse the collected heterogeneous enterprise data across modalities to obtain standardized real-time enterprise data;

[0239] The knowledge graph construction and update module is used to build a knowledge graph with enterprise entities, business indicators and industry standards as nodes based on the industry indicators of the enterprise and standardized real-time enterprise data through the graph attention network, and regularly update the association weights between nodes according to the real-time enterprise data;

[0240] The risk heat map generation module is used to input standardized real-time enterprise data into a pre-trained multimodal large model for preliminary analysis and output a risk heat map containing abnormal indicators and abnormal indicator probabilities;

[0241] The reasoning analysis module is used to identify key abnormal indicators from the risk heat map and extract subgraphs related to the key abnormal indicators from the knowledge graph. The subgraphs are then converted into structured prompt words and input into a multimodal large model together with standardized real-time enterprise data for joint reasoning analysis.

[0242] The diagnosis display module is used to convert the joint reasoning analysis results into executable diagnosis reports and display them to enterprise users through a visual interactive interface.

[0243] This embodiment realizes accurate identification and intelligent diagnosis of enterprise risks and improves decision-making efficiency through the interactive collaboration of the data collection and processing module, the knowledge graph construction and update module, the risk heat map generation module, the reasoning analysis module and the diagnosis display module.

[0244] The enterprise intelligent diagnosis method based on a large model provided in the embodiment of the present application can be applied to electronic devices. Those skilled in the art will understand that the electronic device structure involved in the embodiment of the present invention does not constitute a limitation on the electronic device, and the electronic device may include more or fewer components than shown in the figure, or combine certain components, or arrange components differently. In the embodiment of the present invention, the electronic device includes but is not limited to a laptop computer, a desktop computer, a workbench, a personal digital assistant, a server, a blade server, a mainframe computer, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the embodiments of the present application described and / or required herein.

[0245] The electronic device may include a processor, an external memory interface, an internal memory, a universal serial bus (USB) interface, a charging management module, a power management module, a battery, a wireless communication module, an audio module, a speaker, a microphone, a sensor module, a button, a camera, a display, and a SIM card interface, etc.

[0246] It is understood that the structures illustrated in the embodiments of the present application do not constitute specific limitations on the electronic device. In other embodiments of the present application, the electronic device may include more or fewer components than shown, or combine or separate certain components, or arrange the components differently. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0247] A processor may include one or more processing units, such as a central processing unit (CPU), an application processor (AP), a modem processor, a graphics processing unit (GPU), an image signal processor (ISP), a controller, a memory, a video codec, a digital signal processor (DSP), a baseband processor, and / or a neural-network processing unit (NPU). Different processing units may be independent devices or integrated into one or more processors.

[0248] The processor can be the nerve center and command center of the electronic device. The controller can generate operation control signals based on the instruction opcode and timing signal to complete the control of instruction fetching and execution.

[0249] The processor may also include a memory for storing instructions and data. In some embodiments, the memory in the processor is a cache memory. This memory can store instructions or data that the processor has just used or is reusing. If the processor needs to use the instruction or data again, it can directly call it from the memory. This avoids repeated accesses, reduces processor latency, and thus improves system efficiency.

[0250] The electronic device implements the enterprise intelligent diagnosis method based on a large model of the present application, which collects heterogeneous enterprise data through a multi-source data interface, cleans, extracts features, and aligns and fuses the collected heterogeneous enterprise data across modalities to obtain standardized enterprise real-time data; constructs a knowledge graph with enterprise entities, business indicators, and industry standards as nodes through a graph attention network based on the enterprise's industry indicators and standardized enterprise real-time data, and regularly updates the association weights between nodes based on the enterprise real-time data; inputs the standardized enterprise real-time data into a pre-trained multimodal large model for preliminary analysis, and outputs a risk heat map containing abnormal indicators and abnormal indicator probabilities; identifies key abnormal indicators from the risk heat map, extracts subgraphs related to the key abnormal indicators from the knowledge graph, generates structured prompt words from the subgraphs, and then inputs them into the multimodal large model together with the standardized enterprise real-time data for joint reasoning and analysis; converts the joint reasoning and analysis results into an executable diagnostic report and presents it to enterprise users through a visual interactive interface. The technical solution achieves the beneficial effects of accurately identifying and intelligently diagnosing enterprise risks and improving decision-making efficiency through multi-source data fusion, dynamic knowledge graph, and multimodal analysis.

[0251] The storage medium provided in this application stores a program product that can implement a large-model-based enterprise intelligent diagnosis method.

[0252] The enterprise intelligent diagnosis method based on the big model includes: collecting enterprise heterogeneous data through multi-source data interfaces, and cleaning, feature extraction, and cross-modal alignment and fusion of the collected enterprise heterogeneous data to obtain standardized enterprise real-time data; based on the industry indicators of the enterprise and combined with standardized enterprise real-time data, a knowledge graph with enterprise entities, business indicators and industry standards as nodes is constructed through a graph attention network, and the association weights between nodes are updated regularly according to the enterprise real-time data; the standardized enterprise real-time data is input into the pre-trained multimodal big model for preliminary analysis, and a risk heat map containing abnormal indicators and abnormal indicator probabilities is output; key abnormal indicators are identified from the risk heat map, and subgraphs related to the key abnormal indicators are extracted from the knowledge graph, and then structured prompt words are generated from the subgraphs and input into the multimodal big model together with the standardized enterprise real-time data for joint reasoning analysis; the results of the joint reasoning analysis are converted into executable diagnostic reports and displayed to enterprise users through a visual interactive interface.

[0253] In some possible implementations, the big model-based enterprise intelligent diagnosis method disclosed herein can be implemented in the form of a program product, which includes program code. When the program product is run on a terminal device, the program code is used to enable the terminal device to execute the steps described in the above "Exemplary Method" section of this specification according to various exemplary implementations of the present disclosure.

[0254] The storage medium of the present disclosure can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can be, for example, but not limited to, a system, device or component of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination thereof. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0255] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for enterprise intelligent diagnosis based on a large model, characterized by: The steps include: S1. Collect heterogeneous enterprise data through multi-source data interfaces, cleanse, extract features, and perform cross-modal alignment and fusion on the collected heterogeneous enterprise data to obtain standardized real-time enterprise data. S2. Based on the industry indicators of the enterprise and combined with standardized real-time enterprise data, a knowledge graph with enterprise entities, business indicators, and industry standards as nodes is constructed through a graph attention network. The association weights between nodes are regularly updated based on the real-time enterprise data. S3. Input standardized real-time enterprise data into a pre-trained multimodal large model for preliminary analysis, outputting a risk heat map containing abnormal indicators and their probabilities. The specific steps of step S3 are as follows: S31. Divide standardized enterprise real-time data into structured data, text data, and time series data; S32. Mapping the categorical features in the structured data into dense vectors of a set dimension through an embedding algorithm, and extracting local pattern features from the numerical features in the structured data using a 1D convolutional neural network; S33. The text data is divided into table-like text and plain text. The table-like text is processed using the LayoutXLM model to obtain the table structure and content semantics. The plain text is extracted using the RoBERTa-wwm model, and the token vector is extracted for the Chinese text. S34. Use the Informer model to capture long-term dependencies for sensor time series data, and use the N-HiTS model to decompose trend and periodic features for business time series data. S35. The features extracted from structured data, text data, and time series data are used to calculate the correlation weights between each modality through the cross-modal attention mechanism algorithm: in, is the query vector of modality i, is the key vector of mode j, is the dimension scaling factor; S36. Use the gated fusion algorithm to generate joint features for the feature vectors of structured data, text data, and time series data: in, represents the gate value, σ(*) is the activation function, is the trainable parameter matrix, is the structured data feature vector, is the text data feature vector, Time series data feature vector, is the fused cross-modal feature vector; S37. Calculate the abnormal probability of each business indicator through the multi-layer perceptron model: in, is the abnormal probability of the i-th indicator, , is the weight matrix of the first layer of the multilayer perceptron model, is the weight matrix of the second layer of the multilayer perceptron model, is the bias vector of the first layer of the multilayer perceptron model, is the bias value of the second layer of the multilayer perceptron model, d is the dimension of the input feature, m is the number of hidden units in the first layer, and 1 is the dimension of the output; Determine the abnormal business indicator based on the abnormal probability of each business indicator, and generate an abnormal probability vector by combining the abnormal business indicator and the corresponding abnormal probability; S38. Use a deconvolutional network to map the anomaly probability vector to the dimensions of the original standardized enterprise real-time data to generate a pixel-level risk heat map. S39. Displaying the pixel-level risk heat map in layers; S4. Identify key anomaly indicators from the risk heat map and extract subgraphs related to these indicators from the knowledge graph. Then, generate structured prompts from the subgraphs and input them into a multimodal large model along with standardized real-time enterprise data for joint reasoning and analysis. The specific steps of step S4 are as follows: S41. The specific steps for identifying key abnormal indicators from the risk heat map are as follows: Dynamic probability thresholds are pre-set to determine business indicators that meet any of the following conditions as key abnormal indicators: Business indicators whose abnormal probability exceeds the dynamic probability threshold; Affect business indicators of more than N business departments; The abnormal probability increases for M consecutive detection cycles exceeding the preset amplitude threshold; S42. Specific steps for extracting subgraphs related to key anomaly indicators from the knowledge graph: Starting from the key anomaly indicator node, perform an N-level deep search in the knowledge graph: The first layer consists of nodes directly connected to the key anomaly indicator nodes, and the next layer consists of nodes connected to the previous layer nodes; The search priority of nodes in each layer is calculated by edge weight and SHAP value: in, is the weight between node i and node j, SHAP is the feature screening result, which indicates the contribution of the feature to the model output. and It is an adjustable weight coefficient used to balance the influence of edge weight and SHAP value; S43. The specific steps of generating structured prompt words from the subgraph are as follows: Pre-build a prompt word template and fill the relevant nodes and relationships in the searched subgraph into the prompt word template, thereby converting the subgraph into a structured prompt word as a natural language description; S44. The specific steps for inputting structured prompt words and standardized real-time enterprise data into a multimodal large model for joint reasoning analysis are as follows: Construct a multimodal large model of hierarchical hybrid coding, including parallel text coding layer, table coding layer, time series coding layer and data fusion layer; The structured prompt words are concatenated with the structured data in the standardized real-time data and then input into the text encoding layer of the multimodal large model to generate a structured query vector. The structured prompt words are concatenated with the standardized text data in real-time data and then input into the table encoding layer of the multimodal large model to generate a text query vector. Input the time series prompt words and the time series data in the standardized real-time data into the time series encoding layer of the multimodal large model to generate a time series query vector; The structured query vector, text query vector, and temporal query vector are input into the data fusion layer, and the joint reasoning analysis results are generated through cross-modal attention fusion and gated feature fusion. S5. Convert the joint reasoning analysis results into executable diagnostic reports and present them to enterprise users through a visual interactive interface.

2. The enterprise intelligent diagnosis method based on a large model according to claim 1 is characterized in that: The specific steps of step S1 are as follows: S11. Obtain enterprise financial data, sales data, inventory data, and human resources data through the enterprise ERP system interface, obtain production and manufacturing data through the enterprise MES system interface, and obtain customer interaction data through the enterprise CRM system interface; S12. Use RPA crawler tools to capture operational data from various business processes within the enterprise and interface sensor data through IoT device protocols. S13. Clean the collected data, remove duplicate, erroneous, incomplete, and abnormal data records, and supplement missing values; S14. Normalize and standardize the cleaned data to unify the dimensions, and then perform feature screening based on the SHAP value. S15. Map the feature-screened data to a unified dimensional space and align them based on timestamps to obtain standardized real-time enterprise data.

3. The enterprise intelligent diagnosis method based on a large model according to claim 2 is characterized in that: The specific steps of step S2 are as follows: S21. Extract enterprise entity nodes from standardized real-time enterprise data using a named entity recognition algorithm, load indicator metadata from the industry benchmark library of the enterprise's industry as business indicator nodes, extract clause entity constraints from the enterprise's industry standard documents as industry standard nodes, and complete node initialization; S22. Constructing knowledge graph through graph attention network: Perform GRU encoding on the standardized enterprise real-time data to generate the feature vector of the corresponding node; Set the initial value of the edge weight between nodes according to industry rules; S23. When the update conditions are triggered, the node and edge weights of the knowledge graph are dynamically updated.

4. The enterprise intelligent diagnosis method based on a large model according to claim 3 is characterized in that: The specific steps of step S23 are as follows: S231. When a new enterprise entity or indicator anomaly is detected from standardized real-time enterprise data, a node is automatically created for the knowledge graph and the node's embedding vector is initialized: in, is a physical feature, Embedding for similar nodes; S232. For existing nodes in the knowledge graph that have no associated data in the standardized enterprise real-time data for N consecutive days, they are marked as dormant nodes and stored offline; S233. Update the knowledge graph according to the set update cycle and use the temporal graph convolutional network to adjust the edge weights between nodes: in, is the embedding of node i at time t, is the amount of data change between nodes.

5. The enterprise intelligent diagnosis method based on large model according to claim 3 is characterized in that: The specific steps of step S5 are as follows: S51. Use the T5 model to compress the text of the joint reasoning analysis results, generate an executive summary containing key anomaly indicators, root cause analysis, and risk prediction, and highlight key data in the executive summary; S52. Output a hierarchical risk heat map and set up a drill-down interface to connect to standardized enterprise real-time data and knowledge graphs; S53. In response to the user's click operation on the risk heat map, the corresponding standardized enterprise real-time data and related subgraphs in the knowledge graph are displayed in a linked manner, and the similarity ranking results of the related case nodes are displayed.

6. A large-model-based enterprise intelligent diagnosis system, applying the large-model-based enterprise intelligent diagnosis method according to any one of claims 1 to 5, characterized in that: include: The data collection and processing module is used to collect heterogeneous enterprise data through multi-source data interfaces, and clean, extract features, and align and fuse the collected heterogeneous enterprise data across modalities to obtain standardized real-time enterprise data; The knowledge graph construction and update module is used to build a knowledge graph with enterprise entities, business indicators and industry standards as nodes based on the industry indicators of the enterprise and standardized real-time enterprise data through the graph attention network, and regularly update the association weights between nodes according to the real-time enterprise data; The risk heat map generation module is used to input standardized real-time enterprise data into a pre-trained multimodal large model for preliminary analysis and output a risk heat map containing abnormal indicators and abnormal indicator probabilities; The reasoning analysis module is used to identify key abnormal indicators from the risk heat map and extract subgraphs related to the key abnormal indicators from the knowledge graph. The subgraphs are then converted into structured prompt words and input into a multimodal large model together with standardized real-time enterprise data for joint reasoning analysis. The diagnosis display module is used to convert the joint reasoning analysis results into executable diagnosis reports and display them to enterprise users through a visual interactive interface.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the enterprise intelligent diagnosis method based on a large model as described in any one of claims 1 to 5 are implemented.

8. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the enterprise intelligent diagnosis method based on a large model as claimed in any one of claims 1 to 5 are implemented.

Images