Tobacco industry financial data verification method based on multi-source data fusion and knowledge graph

Abstract

The invention discloses a tobacco industry financial data verification method based on multi-source data fusion and a knowledge graph, and the method comprises the steps: collecting multi-source heterogeneous data of a tobacco industry chain, carrying out the preprocessing of the multi-source heterogeneous data of the tobacco industry chain, so as to obtain the preprocessed multi-source heterogeneous data of the tobacco industry chain, and carrying out the verification of the multi-source heterogeneous data of the tobacco industry chain; performing semantic space mapping and standardized integration processing on the preprocessed tobacco industry chain multi-source heterogeneous data to obtain a standardized data set, performing distributed storage on the standardized data set, and performing containerization deployment on the stored data to obtain a plurality of data resource pools, the method comprises the following steps: acquiring an entity set from a tobacco industry core business data table based on a data resource pool, acquiring a relationship set from a structured business log, constructing a graph basic framework according to entities and relationships, and carrying out vectorization coding on the graph basic framework. The technical problems that an existing manual checking method is low in efficiency, high in error rate and difficult to apply to multi-source data with format differences can be solved.

Description

Technical Field

[0001] This invention belongs to the fields of knowledge graph, intelligent operation and data management technology of tobacco industry, and more specifically, relates to a method for verifying financial data of tobacco industry based on multi-source data fusion and knowledge graph. Background Technology

[0002] As a crucial pillar of the agricultural economy, the tobacco industry encompasses the entire supply chain, including planting, processing, storage, logistics, and market transactions. The accuracy of its financial data (such as purchase amounts, purchase volumes, processing costs, and allocation expenses) directly determines the industry's cost control, financial security, and the scientific nature of decision-making. Financial data verification, as a core aspect of tobacco industry operation and management, requires achieving logical consistency verification between "business data" and "financial data," accurate anomaly tracing, and early risk warning. It is a key support for driving the industry's transformation from "manual dependence" to "intelligent management."

[0003] Existing methods for verifying financial data in the tobacco industry can be mainly divided into three categories: The first category is manual verification, which involves financial personnel manually comparing business documents (such as purchase orders and processing orders) with financial ledgers to verify basic logic such as "purchase amount = purchase quantity × unit price"; the second category is fixed rule automation, which uses Excel formulas or simple software programs to preset fixed rules such as numerical ranges (such as purchase unit price thresholds) and basic logic formulas to achieve data format and simple logic verification; the third category is single-stage data association, which relies on local data from a single business system (such as the purchase system or warehousing system) to perform one-way comparison with financial data, lacking cross-stage data collaborative verification.

[0004] However, the aforementioned existing methods for verifying financial data in the tobacco industry all have some significant drawbacks: First, manual verification is inefficient and has a high error rate, requiring a large amount of manpower to verify massive amounts of data one by one. It cannot cover the data correlation verification of the entire chain from planting to processing, warehousing, and logistics, and it is difficult to apply to multi-source data with different formats. Second, the fixed-rule automated method can only handle preset scenarios and cannot adapt to dynamic changes in the industry chain (such as fluctuations in processing losses and regional adjustments in purchase prices), nor can it eliminate semantic barriers in multimodal data. Third, the single-link data correlation method does not integrate business data from the entire industry chain and cannot uncover the deep correlation between financial data and business data (such as the correlation between abnormal purchase volume and yield of planting plots), resulting in anomaly tracing remaining only at the financial level and failing to locate the root cause at the business end. Fourth, the three existing methods lack systematic knowledge representation and reasoning capabilities, cannot capture potential dependencies at nodes in the industry chain (such as the impact of logistics delays on processing costs), and can only "verify" financial anomalies "after the fact," failing to achieve real-time monitoring and risk warning. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method for verifying financial data in the tobacco industry based on multi-source data fusion and knowledge graphs. Its purpose is to solve the technical problems of existing manual verification methods, which are inefficient and have high error rates, requiring significant manpower to verify massive amounts of data one by one, failing to cover the entire chain of data association verification from planting to processing, warehousing, and logistics, and being difficult to apply to multi-source data with format differences; existing fixed-rule automated methods, which can only handle preset scenarios and cannot adapt to dynamic changes in the industry chain (such as fluctuations in processing losses and regional adjustments in purchase prices), and cannot eliminate semantic barriers in multimodal data; existing single-link data association methods, which do not integrate business data across the entire industry chain and cannot uncover deep relationships between financial and business data, resulting in anomaly tracing remaining only at the financial level and failing to pinpoint the root cause at the business end; and the lack of systematic knowledge representation and reasoning capabilities in the three existing methods, which cannot capture potential dependencies at industry chain nodes, thus failing to achieve real-time monitoring and risk warning.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for verifying financial data in the tobacco industry based on multi-source data fusion and knowledge graph is provided, comprising the following steps: (1) Collect multi-source heterogeneous data of the tobacco industry chain and preprocess the multi-source heterogeneous data of the tobacco industry chain to obtain preprocessed multi-source heterogeneous data of the tobacco industry chain. (2) Perform semantic space mapping and standardization integration on the preprocessed multi-source heterogeneous data of the tobacco industry chain obtained in step (1) to obtain a standardized dataset; (3) Distribute the standardized dataset obtained in step (2) and containerize the stored data to obtain multiple data resource pools; (4) Based on the data resource pool obtained in step (3), obtain the entity set from the core business data table of the tobacco industry, obtain the relationship set from the structured business log, construct the graph basic framework according to the entity and relationship, and perform vectorized encoding on the graph basic framework to obtain the structured knowledge graph. (5) The structured knowledge graph obtained in step (4) is processed using a relational reasoning method based on graph neural networks (GNN) to obtain the fusion features between every two nodes in the structured knowledge graph after processing: (6) Use the knowledge graph incremental update algorithm to process the fusion features between every two static nodes in the structured knowledge graph obtained in step (5) to obtain the updated knowledge graph; (7) Obtain the core financial data of the tobacco industry and verify the core financial data of the tobacco industry based on the updated knowledge graph obtained in step (6) to obtain the verified financial data.

[0007] Preferably, step (1) includes the following sub-steps: (1-1) Obtain multi-source heterogeneous data of the tobacco industry chain, including a set of tobacco planting data composed of tobacco planting data collected by multiple sensors at different times. and a collection of tobacco processing data, including multiple tobacco processing data. ;in Let represent the tobacco planting data collected by the i-th sensor at the t-th timestamp, where i ∈ [1, total number of sensors N], t ∈ [1, total number of timestamps T]. This represents the processing data for the j-th tobacco leaf. This represents the field name of the j-th tobacco leaf processing data. The field value represents the j-th tobacco processing data, and j∈[1, the total number of tobacco processing data M]; (1-2) Impute missing values ​​in the multi-source heterogeneous data of the tobacco industry chain obtained in step (1-1) to obtain imputed multi-source heterogeneous data of the tobacco industry chain: (1-3) Determine whether there is any abnormal data in the multi-source heterogeneous data of the tobacco industry chain obtained in step (1-2) after filling. If there is, proceed to step (1-4); otherwise, proceed to step (1-5): (1-4) Mark and remove abnormal data in the multi-source heterogeneous data of the tobacco industry chain to obtain multi-source heterogeneous data of the tobacco industry chain after removing abnormal data, and then proceed to step (1-5): (1-5) Perform optimization processing such as format unification and redundancy removal on the multi-source heterogeneous data of the tobacco industry chain obtained in step (1-3) or step (1-4) to obtain pre-processed multi-source heterogeneous data of the tobacco industry chain.

[0008] Preferably, the tobacco planting data collected by the i-th sensor at the t-th time stamp is... The missing value imputation process in step (1-2) uses the following formula: in This represents the tobacco planting data collected by the i-th sensor at the t-th timestamp after the data filling process. This indicates the timestamp preceding timestamp t. Indicates the next timestamp after timestamp t; For the j-th tobacco leaf processing data Specifically, the missing value imputation process in step (1-2) uses the following formula: in This represents the parameter value that makes the subsequent function reach its maximum value; Indicates the first Fields in tobacco leaf processing data The set of possible values ​​for this field. The set of all possible values. This indicates counting, or statistical analysis of values. In the set The number of times it appears in; For the numerical data in the multi-source heterogeneous data of the tobacco industry chain after imputation processing, the anomaly detection in steps (1-3) is performed using... The principles, related calculations, and judgment logic are as follows: First, calculate the numerical data. mean and standard deviation : Then, determine if the numerical data exists. If so, the numerical data is considered abnormal; otherwise, the numerical data is considered normal. For data strongly correlated with tobacco industry business in the multi-source heterogeneous data of the tobacco industry chain after filling in the gaps, the anomaly judgment in steps (1-3) is based on preset business rules, and the judgment logic is as follows: Data strongly correlated with the tobacco leaf industry satisfy If the data exceeds a reasonable range, then the data strongly correlated with the tobacco industry is considered abnormal; if If the data is strongly correlated with the tobacco industry business, it is considered normal. This represents the lowest historical yield per unit area for tobacco leaves in this production region. This represents the highest single-yield value in the history of tobacco production in this region. Steps (1-4) are as follows: First, add an anomaly identifier field to the abnormal data and record the anomaly type and cause. Then, through data filtering, remove the abnormal data with the anomaly identifier field from the multi-source heterogeneous data of the tobacco industry chain, retain all normal data, and finally obtain the multi-source heterogeneous data of the tobacco industry chain after removing the abnormal data.

[0009] Preferably, step (2) includes the following sub-steps: (2-1) Perform semantic space mapping on the preprocessed multi-source heterogeneous data of the tobacco industry chain obtained in step (1) to obtain cross-modal semantic vectors: (2-2) The cross-modal semantic vectors obtained in step (2-1) are standardized and integrated using a cross-modal semantic alignment algorithm and a unified feature space mapping technique to map them to a unified semantic space and generate a JSON format dataset. As a standardized dataset, This serves as a unique identifier for the entities in this standardized dataset. This represents the k-th JSON-formatted data in the standardized dataset, where k represents the total number of JSON-formatted data in the standardized dataset.

[0010] Preferably, in the text data of the preprocessed multi-source heterogeneous data of the tobacco industry chain obtained in step (1), step (2-1) uses a fine-tuned BERT model to achieve semantic space mapping, with the formula as follows: in For text semantic vectors, For text data, This represents the 768-dimensional real space to which the text semantic vector belongs; For the image data in the preprocessed multi-source heterogeneous data of the tobacco industry chain obtained in step (1), step (2-1) uses the ResNet50 model to achieve semantic space mapping, and the formula is: in For image feature vectors, For image data, This represents the 2048-dimensional real space to which the image feature vector belongs.

[0011] Preferably, step (3) includes the following sub-steps: (3-1) Distribute the standardized dataset obtained in step (2) to obtain the stored data: Specifically, for unstructured data in standardized datasets, the HDFS distributed storage system is used for distributed storage, where the number of replicas stored is... Satisfy the following formula: in It is a rounding function. The probability of losing stored unstructured data is given, and its value ranges from 0.001 to 0.01, preferably 0.005. For structured data in standardized datasets, ClickHouse database is used for distributed storage and partitioned tables are designed. (3-2) Containerize and dynamically manage the stored data obtained in step (3-1) to obtain multiple data resource pools; Specifically, this step involves first, containerizing the stored data obtained in step (3-1), that is, converting the access logic and processing logic of the data into executable code, and then packaging the executable code and the runtime dependent environment together into a standardized container image, which serves as a containerized data service unit that can be elastically scheduled. Subsequently, Prometheus was used to obtain the resource utilization rate of the containerized data service unit. The used resources are the resources actually occupied by the standardized container image, and the total allocated resources are the preset maximum available resources. Subsequently, the resource utilization rate of the obtained containerized data service unit is determined. Whether u is greater than 0.8 or less than 0.3, if u>0.8, the scaling process is automatically triggered, that is, the number of Pods is increased by 30%~50% of the current number of Pods, so as to obtain a containerized data service cluster with elastic resource scaling; if u<0.3, the scaling process is automatically triggered, that is, the number of Pods is reduced by 20%~30% of the current number of Pods, so as to obtain a containerized data service cluster with elastic resource scaling. Finally, the obtained containerized data service cluster is processed using a cluster high availability scheduling algorithm and data access link adaptation technology to obtain multiple stable and available data resource pools.

[0012] Preferably, in step (4), the process of obtaining entity sets and relationship sets from the core business data table and structured business logs of the tobacco industry based on the data resource pool obtained in step (3), and constructing a graph basic framework based on these entities and relationships, specifically involves: First, multiple core business elements of the tobacco industry are extracted from the data resource pool, and an entity set is constructed based on these core business elements. and relation set ,in Represents the first in the entity set One entity, Indicates the size of the entity set. Represents the first in the set of relations A relationship, Indicates the size of the relation set; Then, the basic framework of the graph is constructed based on the entities in the entity set and the relationships in the relation set. : in Represents a set of entities The tail entity in the middle; Step (4) involves vectorizing the basic framework of the knowledge graph to obtain a structured knowledge graph. The specific process is as follows: First, the basic framework of the graph. All head entities Tail entity Mapped to the complex vector space respectively, i.e. , and the relationship It is denoted as the rotation factor in this complex vector space, i.e. ,in The preset vector dimension (which can be set according to the amount of business data in the tobacco industry and the complexity of entity relationships, such as...) or ), Represents the field of complex numbers; Then, vector matching constraints are added to the obtained complex vector space to obtain a vector set with business association constraints: Subsequently, the loss function is obtained based on the resulting vector set with business-related constraints. : in, This represents the boundary hyperparameter, which takes a value of 1 and is used to differentiate between positive and negative samples. Represents the vector distance function. The negative sample tail entity is formed by replacing the basic framework of the graph constructed in step (4). Middle Triple Group Positive sample tail entity Generate, and ensure that the generated negative sample tail entities are generated. With head entity No real business relationship; Subsequently, the stochastic gradient descent (SGD) method was used to adjust the obtained loss function. Iterative processing is performed, that is, the head entity vector is continuously corrected through iteration. Tail entity vector and relational rotation factor The value until the loss function The convergence continues until a preset threshold is reached, thus obtaining a structured knowledge graph.

[0013] Preferably, step (5) specifically includes: First, the structured knowledge graph obtained in step (4) is statically reasoned using the graph attention network GAT to obtain the association results between every two static nodes. Specifically, this step involves obtaining the association result between the i-th static node and the j-th static node in the structured knowledge graph. The following formula is used, where i and j are both ∈ [1, and the total number of nodes in the structured knowledge graph is 1]. ]: ; And there are: ; in The value range is from 0 to L-1. The preset total number of layers for the GAT network. This indicates that the i-th static node is in the GAT network at the _th ... The embedding vector of the layer reflects the association features between the i-th static node and the j-th static node. This represents the activation function. This represents the set of neighboring nodes of the i-th static node in the structured knowledge graph; Represents an exponential function; Indicates a linear unit with leakage correction; Represents the attention vector; Indicates the first in the GAT network The weight matrix of the layer has dimensions of ,in For the first The input dimension of the layer node embedding, = or , For the first Output dimension of layer node embedding; weight matrix The elements in the array are learnable real-valued parameters, and the initial values ​​are randomly generated using a Xavier normal distribution. Indicates the first in the GAT network The original embedding vector of the i-th static node in layer 1. Indicates the first in the GAT network The original embedding vector of the j-th static node in the layer; This represents a vector concatenation operation used to merge the embedded features of two nodes; Indicates the first in the GAT network The original embedding vector of the k-th static node in the layer, where k∈[1, 1], the i-th static node in the GAT network is the first static node in the layer. [Number of all neighboring nodes in the layer]; Then, a temporal graph neural network is used to analyze the association results between every two static nodes in the obtained structured knowledge graph. Perform dynamic reasoning to obtain The dynamic embedding vector of the i-th static node at time i : ; in, Indicates a gated loop unit. express The dynamic embedding vector of the i-th static node at time i; Subsequently, the association results between every two static nodes in the obtained structured knowledge graph were analyzed. and The dynamic embedding vector of the i-th static node at time i Feature alignment is performed to obtain the pre-fusion feature set corresponding to each pair of static nodes in the structured knowledge graph; Subsequently, a feature fusion algorithm is used to integrate the pre-fusion feature sets between every two static nodes in the obtained structured knowledge graph, so as to obtain the preliminary fusion features between every two static nodes in the structured knowledge graph. Finally, the preliminary fusion features between every two static nodes in the obtained structured knowledge graph are validated and corrected to obtain the processed fusion features between every two static nodes in the structured knowledge graph.

[0014] Preferably, step (7) includes the following sub-steps: (7-1) Obtain core financial data for the tobacco industry, including actual acquisition amounts. Actual data on acquisition volume Actual data on purchase price per unit And determine whether the preset business rules are met. If the equation holds true, it means that the core financial data of the tobacco industry conforms to the business logic, and the process ends; otherwise, it is determined that the core financial data of the tobacco industry is abnormal, and then proceed to step (7-2). (7-2) Based on the updated knowledge graph obtained in step (6), perform correlation analysis and root cause localization on the core financial data of the tobacco industry obtained in step (6) to obtain the localization results of abnormal financial data: (7-3) Based on the location results of the abnormal financial data obtained in step (7-2), the abnormal financial data is verified to obtain the verified financial data.

[0015] Preferably, step (7-2) specifically involves: first, obtaining the business entity association chain corresponding to the core financial data of the tobacco industry through the updated knowledge graph, and extracting all business data on the business entity association chain; then, based on the fusion features between every two nodes in the updated knowledge graph obtained in step (6), obtaining the matching degree between the core financial data of the tobacco industry and each business data, and obtaining the location result of abnormal financial data based on the matching degree.

[0016] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: (1) Since the present invention adopts steps (1) to (3), it acquires multi-source heterogeneous data of the tobacco industry chain through automated acquisition interface and sensor network, and combines missing value filling, outlier removal, semantic alignment integration and distributed storage technology to achieve high-quality processing of multi-source data and construction of a unified resource pool. It does not require a large amount of manual intervention in data sorting and format conversion. Therefore, it can solve the technical problems of low efficiency, high error rate and inability to cover the entire chain of data association verification in the existing manual verification method. (2) Since the present invention adopts steps (4) to (5), it constructs a knowledge graph of the tobacco industry through rule extraction and NER / RE technology, and relies on GAT static reasoning and TemporalGNN dynamic reasoning to realize the systematic modeling of business knowledge and node association feature mining. It can adapt to the dynamic changes of the industrial chain (such as fluctuations in processing loss and adjustment of purchase unit price), and eliminate the semantic barriers of multimodal data. Therefore, it can solve the technical problems of existing fixed rule automation methods that can only handle preset scenarios, cannot adapt to dynamic business, and have prominent semantic barriers. (3) Since the present invention adopts steps (6) to (7-2), it uses the knowledge graph incremental update algorithm to obtain real-time knowledge graph, extracts the whole chain business data through the business entity association chain (such as "abnormal purchase amount → tobacco farmer → planting plot → processing batch"), and calculates the matching degree of financial and business data by combining node fusion features, so as to accurately locate the root cause of the abnormality (such as the statistical deviation of purchase volume, and the failure to calculate processing loss). Therefore, it can solve the technical problem that the existing single link data association method does not integrate the whole industry chain business data, cannot explore the deep association between financial data and business data, and causes the abnormal source tracing to only stay at the financial level and cannot locate the root cause of the business. (4) Since the present invention adopts steps (5) to (7-3), it mines the potential dependencies of the industrial chain nodes (such as the impact of logistics delays on processing costs) through GNN relational reasoning, relies on knowledge graph incremental update technology to ensure data real-time performance, and combines financial data dynamic verification and anomaly warning to promote the transformation of financial verification from "post-event verification" to "real-time monitoring". Therefore, it can solve the technical problem that the existing three methods are unable to capture the potential dependencies of the industrial chain nodes due to the lack of systematic knowledge representation and reasoning ability, which leads to the inability to achieve real-time monitoring and risk warning. (5) This invention uses semantic association technology of business and financial data, relying on the relationship between "business entity-financial indicator" in the knowledge graph, to realize automatic classification, verification and structured output of financial data, open up the data channels of acquisition, allocation, budgeting, accounting and settlement, reduce the repetitive work of financial personnel, and further improve the accuracy and efficiency of business and financial data processing; (6) This invention adopts containerized deployment and dynamic resource management technology. It monitors resource utilization through Prometheus and automatically triggers expansion or contraction processes to avoid resource idleness or overload, ensuring the stable operation of multi-source data processing and knowledge graph reasoning, and adapting to the data processing needs of the tobacco industry at different times (such as the scenario of a surge in data volume during the peak acquisition season). (7) This invention records the updates and changes of entities, relationships and fusion features through a knowledge graph version management mechanism, quantifies the differences between different versions of the graph, ensures the consistency and traceability of the knowledge graph in the incremental update process, and provides reliable data support for the historical retrospective of financial data verification results and the iteration of business rules. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the intelligent operation risk factor identification process for the tobacco leaf industry based on multi-source data fusion and knowledge graph. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0019] To address the shortcomings or improvement needs in the existing tobacco industry operations, such as inefficient integration of multi-source data, fragmented business knowledge, and insufficient business-finance integration and collaboration, this invention provides a method for verifying financial data in the tobacco industry based on multi-source data fusion and knowledge graphs. This method not only systematically integrates data resources and business knowledge across the entire tobacco industry chain through multimodal data semantic alignment and knowledge graph construction, breaking down data silos and knowledge barriers, but also leverages GNN relational reasoning and customized applications for financial scenarios to achieve in-depth data value mining and intelligent decision support. This drives the transformation of tobacco industry operations from manual reliance and post-event verification to automated processing and real-time monitoring.

[0020] like Figure 1 As shown, this invention provides a method for verifying financial data in the tobacco industry based on multi-source data fusion and knowledge graphs, including the following steps: (1) Collect multi-source heterogeneous data of the tobacco industry chain and preprocess the multi-source heterogeneous data of the tobacco industry chain to obtain preprocessed multi-source heterogeneous data of the tobacco industry chain. This step includes the following sub-steps: (1-1) Obtain multi-source heterogeneous data of the tobacco industry chain, including a set of tobacco planting data composed of tobacco planting data collected by multiple sensors at different times. and a collection of tobacco processing data, including multiple tobacco processing data. ;in Let represent the tobacco planting data collected by the i-th sensor at the t-th timestamp, where i ∈ [1, total number of sensors N], t ∈ [1, total number of timestamps T]. This represents the processing data for the j-th tobacco leaf. This represents the field name of the j-th tobacco leaf processing data. The field value represents the j-th tobacco processing data, and j∈[1, the total number of tobacco processing data M]; (1-2) Impute missing values ​​in the multi-source heterogeneous data of the tobacco industry chain obtained in step (1-1) to obtain imputed multi-source heterogeneous data of the tobacco industry chain: Specifically, this refers to the tobacco planting data collected by the i-th sensor at the t-th time stamp. The missing value imputation process in this step uses the following formula: in This represents the tobacco planting data collected by the i-th sensor at the t-th timestamp after the data filling process. This indicates the timestamp preceding timestamp t. Indicates the next timestamp after timestamp t; For the j-th tobacco leaf processing data In this step, the missing value imputation is performed using the following formula: in It represents a "parameter", that is, "the parameter value that makes the subsequent function achieve its maximum value". Indicates the first Fields in tobacco leaf processing data The set of possible values ​​for this field. The set of all possible values. This indicates counting, or statistical analysis of values. In the set The number of times it appears in the text.

[0021] (1-3) Determine whether there is any abnormal data in the multi-source heterogeneous data of the tobacco industry chain obtained in step (1-2) after filling. If there is, proceed to step (1-4); otherwise, proceed to step (1-5): Specifically, for numerical data (such as temperature and humidity data of tobacco growth environment collected by sensors, material consumption data in processing stages, etc.) in the multi-source heterogeneous data of the tobacco industry chain after filling in the gaps, the anomaly detection in this step is performed by using... The principles, related calculations, and judgment logic are as follows: First, calculate the numerical data. mean and standard deviation : Then, determine if the numerical data exists. If so, the numerical data is considered abnormal; otherwise, the numerical data is considered normal. For data strongly correlated with tobacco industry operations (such as tobacco yield per unit area data, purchase weight data, etc.) in the multi-source heterogeneous data of the tobacco industry chain after filling in the data, the anomaly judgment in this step is based on preset business rules, and the judgment logic is as follows: Data strongly correlated with the tobacco leaf industry satisfy If the data exceeds a reasonable range, then the data strongly correlated with the tobacco industry is considered abnormal. If the data is strongly correlated with the tobacco industry business, it is considered normal. This represents the lowest historical yield per unit area for tobacco leaves in this production region. This represents the highest single-yield value in the history of tobacco production in this region. The advantage of steps (1-3) is that by combining data types and business scenarios to formulate targeted anomaly judgment rules, different types of abnormal data can be accurately identified, ensuring data quality.

[0022] (1-4) Mark and remove abnormal data in the multi-source heterogeneous data of the tobacco industry chain to obtain multi-source heterogeneous data of the tobacco industry chain after removing abnormal data, and then proceed to step (1-5): Specifically, this step involves first adding an anomaly flag field (e.g., marking "abnormal_flag=1") to the abnormal data, and recording the anomaly type (e.g., "3σ principle determines anomaly" and "business rule determines anomaly") and the cause of the anomaly (e.g., "temperature and humidity data exceed 3 times the standard deviation" and "tobacco leaf yield per unit area is higher than the historical highest threshold"). Then, through data filtering, the abnormal data with anomaly flag fields in the multi-source heterogeneous data of the tobacco industry chain are removed, and all normal data are retained, finally obtaining the multi-source heterogeneous data of the tobacco industry chain after removing the abnormal data.

[0023] The advantage of steps (1-4) is that it not only removes abnormal data, but also marks and records the reasons for the abnormal data in detail, providing a basis for subsequent data quality analysis and problem tracing.

[0024] (1-5) Perform optimization processing such as format unification and redundancy removal on the multi-source heterogeneous data of the tobacco industry chain obtained in step (1-3) or step (1-4) to obtain pre-processed multi-source heterogeneous data of the tobacco industry chain.

[0025] The advantage of steps (1-5) is that it eliminates format differences between multiple data sources by unifying the format, reduces data redundancy by removing redundancy, and improves data processing efficiency and the accuracy of subsequent analysis.

[0026] The advantage of the above sub-steps (1-1) to (1-5) is that they form a complete multi-source heterogeneous data preprocessing process, which progresses step by step from data acquisition, missing value imputation, anomaly judgment and removal to format unification and redundancy removal, comprehensively improving data quality and laying a solid data foundation for subsequent semantic alignment, knowledge graph construction and other work.

[0027] (2) Perform semantic space mapping and standardization integration on the preprocessed multi-source heterogeneous data of the tobacco industry chain obtained in step (1) (the purpose is to eliminate the semantic barriers of multimodal data) to obtain a standardized dataset; This step includes the following sub-steps: (2-1) Perform semantic space mapping on the preprocessed multi-source heterogeneous data of the tobacco industry chain obtained in step (1) to obtain cross-modal semantic vectors: Specifically, for the text data in the preprocessed multi-source heterogeneous data of the tobacco industry chain obtained in step (1), this step uses a fine-tuned BERT model to achieve semantic space mapping, as shown in the formula: in For text semantic vectors, For text data, This represents the 768-dimensional real space to which the text semantic vector belongs; For the image data in the preprocessed multi-source heterogeneous data of the tobacco industry chain obtained in step (1), the ResNet50 model is used to realize semantic space mapping, and the formula is: in For image feature vectors, For image data, This represents the 2048-dimensional real space to which the image feature vector belongs; The advantage of step (2-1) is that it selects an appropriate advanced model for semantic space mapping for two different modalities of data, namely text and images, and can accurately extract the core semantic features of each modality of data.

[0028] (2-2) The cross-modal semantic vectors obtained in step (2-1) are standardized and integrated using a cross-modal semantic alignment algorithm and a unified feature space mapping technique to map them to a unified semantic space and generate a JSON format dataset. As a standardized dataset, This serves as a unique identifier for the entities in this standardized dataset. This represents the k-th JSON-formatted data in the standardized dataset, where k represents the total number of JSON-formatted data in the standardized dataset.

[0029] The advantage of step (2-2) is that it effectively eliminates the semantic barriers between different modal data through cross-modal semantic alignment and unified feature space mapping, realizes the standardized integration of multimodal data, and facilitates subsequent distributed storage and knowledge graph construction.

[0030] The advantage of the above sub-steps (2-1) to (2-2) is that they construct a complete link from semantic extraction of multimodal data to standardized integration, which not only ensures the accurate extraction of semantic features of each modality of data, but also realizes the effective fusion of multimodal data, providing a unified and standardized data form for subsequent data applications.

[0031] (3) Distribute the standardized dataset obtained in step (2) and containerize the stored data to obtain multiple data resource pools; This step includes the following sub-steps: (3-1) Distribute the standardized dataset obtained in step (2) to obtain the stored data: Specifically, for unstructured data (such as tobacco leaf images and operation logs) in the standardized dataset, the Hadoop Distributed File System (HDFS) is used for distributed storage, where the number of replicas is... Satisfy the following formula: in It is a rounding function. The probability of losing stored unstructured data is given, and its value ranges from 0.001 to 0.01, preferably 0.005. For structured data (such as tobacco leaf purchase volume, tobacco leaf processing parameters, etc.) in the standardized dataset, ClickHouse database is used for distributed storage and partitioned tables are designed.

[0032] The advantage of step (3-1) is that it selects a suitable distributed storage solution based on the structured characteristics of the data. For unstructured data, HDFS is used to ensure storage reliability and scalability, while for structured data, ClickHouse is used to ensure query efficiency. At the same time, the number of replicas is set reasonably through the replica number formula to reduce the risk of data loss.

[0033] (3-2) Containerize and dynamically manage the stored data obtained in step (3-1) to obtain multiple data resource pools; Specifically, this step involves first, containerizing the stored data (unstructured data stored in HDFS and structured data stored in ClickHouse) obtained in step (3-1). This involves converting the data access logic (such as HDFS file read / write interface and ClickHouse partition table query script) and processing logic (such as data format conversion program and basic integrity verification algorithm) into executable code. Then, the executable code and the runtime dependent environment are packaged together into a standardized container image, which serves as a containerized data service unit that can be elastically scheduled. Subsequently, Prometheus was used to obtain the resource utilization rate of the containerized data service unit. The used resources are the resources actually occupied by the standardized container image, and the total allocated resources are the preset maximum available resources. Subsequently, the resource utilization rate of the obtained containerized data service unit is determined. If u > 0.8, an automatic scaling-up process is triggered, increasing the number of Pods by 30% to 50% of the current number (rounded up, and the total number of Pods after a single scaling-up does not exceed a preset maximum threshold, such as twice the initial number of Pods), thus obtaining a containerized data service cluster with elastic resource scaling. If u < 0.3, an automatic scaling-down process is triggered, decreasing the number of Pods by 20% to 30% of the current number (rounded down, and the total number of Pods after a single scaling-down is not less than a preset minimum threshold, such as 50% of the initial number of Pods), thus obtaining a containerized data service cluster with elastic resource scaling (which can avoid resource idleness or overload). Finally, the obtained containerized data service cluster is processed using a cluster high availability scheduling algorithm and data access link adaptation technology to obtain multiple stable and available data resource pools.

[0034] The advantage of step (3-2) is that it achieves standardization and portability of data services through containerized deployment, realizes elastic scaling of resources through dynamic resource management to avoid resource idleness or overload, and, combined with high availability scheduling algorithms, ensures the stability and availability of the data resource pool.

[0035] The advantages of the above sub-steps (3-1) to (3-2) are that they build an efficient, reliable and scalable data storage and deployment system, adopt differentiated storage solutions for different types of data, and improve resource utilization and system stability through containerization and dynamic resource management, providing stable and efficient data support for subsequent knowledge graph construction and financial data verification.

[0036] (4) Based on the data resource pool obtained in step (3), obtain the entity set from the core business data table of the tobacco industry, obtain the relationship set from the structured business log, construct the graph basic framework according to the entity and relationship, and perform vectorized encoding on the graph basic framework to obtain the structured knowledge graph. In this step, based on the data resource pool obtained in step (3), the entity set and relation set are obtained from the core business data table and structured business log of the tobacco industry, and the graph basic framework is constructed based on these entities and relations. Specifically, this process involves extracting multiple core business elements of the tobacco industry from the data resource pool and constructing an entity set based on these core business elements of the tobacco industry. (e.g., "tobacco farmers", "planting plots", "processing equipment") and sets of relationships (e.g., "tobacco farmer - planting - planting plot", "processing equipment - processing - tobacco raw material"), among which Represents the first in the entity set One entity, Indicates the size of the entity set. Represents the first in the set of relations A relationship, Indicates the size of the relation set; Then, the basic framework of the graph is constructed based on the entities in the entity set and the relationships in the relation set. : in Represents a set of entities The tail entity in the middle.

[0037] The process of vectorizing the basic framework of the knowledge graph to obtain a structured knowledge graph in this step is as follows: First, the basic framework of the graph. All head entities Tail entity Mapped to the complex vector space respectively, i.e. , and the relationship It is denoted as the rotation factor in this complex vector space, i.e. ,in The preset vector dimension (which can be set according to the amount of business data in the tobacco industry and the complexity of entity relationships, such as...) or ), Represents the field of complex numbers; Then, vector matching constraints are added to the obtained complex vector space (including the mapped head entity vector, tail entity vector, and relation rotation factor) to obtain a vector set with business association constraints: Specifically, it employs knowledge graph embedding vector matching constraints. (in (Hadamard product), through this vector matching constraint, the head entity vector is rotated by the relational rotation factor. After rotation, it is matched with the tail entity vector to ensure that the triple accurately represents the "head entity-relationship-tail entity" business logic of the tobacco industry; Subsequently, the loss function is obtained based on the resulting vector set with business-related constraints. : in, This represents the boundary hyperparameter, which takes a value of 1 and is used to differentiate between positive and negative samples. This represents a vector distance function (such as Euclidean distance, used to calculate vector similarity). The negative sample tail entity is formed by replacing the basic framework of the graph constructed in step (4). Middle Triple Group Positive sample tail entities (i.e., positive sample triples) Generate, and ensure that the generated negative sample tail entities are generated. With head entity There is no real business relationship (for example, if the positive sample triple is "tobacco farmer-planting-planting plot", then the negative sample tail entity can be replaced with "processing equipment", and "processing equipment" and "tobacco farmer" have no real business relationship corresponding to "planting"). Subsequently, the obtained loss function was processed using the stochastic gradient descent (SGD) method. Perform iterative processing (i.e., continuously refine the head entity vector through iteration) Tail entity vector and relational rotation factor (the value), until the loss function Converging to a preset threshold (i.e.) This process continues until a structured knowledge graph is obtained.

[0038] The advantage of this step (4) is that it transforms the business data in the data resource pool into a structured knowledge graph, realizes the systematic integration and accurate representation of business knowledge, breaks the situation of scattered business knowledge, and provides a solid knowledge foundation for cross-link business correlation analysis and intelligent reasoning.

[0039] (5) The structured knowledge graph obtained in step (4) is processed using a relational reasoning method based on graph neural networks (GNN) to obtain the fusion features between every two nodes in the structured knowledge graph after processing: This step is specifically as follows: First, a graph attention network (GAT, a typical branch of GNN) is used to perform static reasoning on the structured knowledge graph obtained in step (4) to obtain the association results between every two static nodes. Specifically, this step involves obtaining the i-th and j-th static nodes in the structured knowledge graph (where i and j are both ∈ [1, 2, 3, 4, 5, 6, 7, 8, 9, 1, 1, 1, 2, 1 ... Association results of ]) The following formula is used: ; And there are: ; in The value range is 0 to L-1 ( The total number of layers preset for the GAT network can be set according to the number of nodes and the complexity of relationships in the tobacco industry knowledge graph. When the number of nodes is less than 1000 and the number of relationship types is less than 50, the maximum number of layers can be set accordingly. Take values ​​of 2 to 4; when the number of nodes is ≥1000 and the number of relation types is ≥50. Take 4~6). This indicates that the i-th static node is in the GAT network at the _th ... The embedding vector of the layer reflects the association features between the i-th static node and the j-th static node. This represents the activation function (such as Sigmoid or ReLU). This represents the set of neighboring nodes (including the j-th static node) of the i-th static node in the structured knowledge graph. This represents an exponential function used to normalize attention scores; This indicates a linear unit with leakage correction, used to introduce nonlinear characteristics; This represents the attention vector, used to calculate the attention score between nodes; Indicates the first in the GAT network The weight matrix of the layer has dimensions of ,in For the first The input dimension of the layer node embedding, = or , For the first The output dimension of the layer node embedding can be set according to the feature extraction requirements, and is usually related to... Equal to or twice that, preferred To maintain consistency in vector dimensions; weight matrix The elements in the array are learnable real-valued parameters, and the initial values ​​are obtained through a Xavier normal distribution (…). The data is randomly generated and subsequently, during the training of the GAT network, the loss function in step (4) is minimized. The GAT network is iteratively updated using the stochastic gradient descent (SGD) method until it converges (i.e., the GAT network can accurately capture the optimal value of the associated features of the tobacco industry entity nodes - that is, the size and sign of the element values, corresponding to the weight allocation of different dimensions of the node embedded in the feature transformation process, positive elements strengthen the transmission of the corresponding dimension features, and negative elements suppress the influence of redundant or interfering dimension features). Indicates the first in the GAT network The original embedding vector of the i-th static node in layer 1. Indicates the first in the GAT network The original embedding vector of the j-th static node in the layer; This represents a vector concatenation operation used to merge the embedded features of two nodes; Indicates the first in the GAT network The original embedding vector of the k-th static node in the layer, where k∈[1, 1], the i-th static node in the GAT network is the first static node in the layer. The number of all neighboring nodes in the layer.

[0040] Then, a temporal graph neural network (Temporal GNN, a typical branch of GNN) is used to analyze the association results between every two static nodes in the obtained structured knowledge graph. Perform dynamic reasoning to obtain The dynamic embedding vector of the i-th static node at time i ; Specifically, this step introduces the time dimension. (Covering the entire time series of data collection across the tobacco industry chain, such as planting cycle, processing period, and trading cycle), the following formula is used for calculation. The dynamic embedding vector of the i-th static node at time i: ; in, It reflects the real-time evolution of the relationship with the j-th node. This represents a gated loop unit, used to capture the dynamic transmission of node states over time; express The dynamic embedding vector (historical evolution state) of the i-th static node at time i.

[0041] Subsequently, the association results between every two static nodes in the obtained structured knowledge graph were analyzed. and The dynamic embedding vector of the i-th static node at time i Feature alignment is performed, which unifies the data format and vector dimension of the two types of results to ensure that static association features (such as fixed dependencies between nodes) and dynamic evolution features (such as the changing trend of node relationships over time) can be compatible and fused, so as to obtain the set of features before fusion for each pair of static nodes in the structured knowledge graph (with unified format and adapted dimensions). Subsequently, a feature fusion algorithm is used to integrate the pre-fusion feature sets between every two static nodes in the obtained structured knowledge graph (i.e., through weighted summation or attention fusion mechanism, the weights of key related features and significant evolutionary trends are highlighted) to obtain the preliminary fusion features between every two static nodes in the structured knowledge graph. Finally, the preliminary fusion features between every two static nodes in the obtained structured knowledge graph are validated and corrected (i.e., redundant information and logically conflicting data generated during the fusion process are removed) to obtain the processed fusion features between every two static nodes in the structured knowledge graph.

[0042] The advantage of this step (5) is that by combining static and dynamic reasoning, it fully captures the correlation features and dynamic evolution patterns between knowledge graph nodes. Through feature alignment, fusion and verification correction, high-quality node fusion features are obtained, providing accurate feature support for subsequent incremental updates of the knowledge graph and verification of financial data.

[0043] (6) Use the incremental knowledge graph update algorithm to process the fusion features between every two static nodes in the structured knowledge graph obtained in step (5) to obtain the updated knowledge graph. The advantage of this step (6) is that, through the incremental update algorithm, new node fusion features can be absorbed in real time, ensuring the real-time nature and consistency of the knowledge graph, so that it can continuously adapt to the business changes in the tobacco industry.

[0044] (7) Obtain the core financial data of the tobacco industry and verify the core financial data of the tobacco industry based on the updated knowledge graph obtained in step (6) to obtain the verified financial data.

[0045] This step includes the following sub-steps: (7-1) Obtain core financial data for the tobacco industry, including actual acquisition amounts. Actual data on acquisition volume Actual data on purchase price per unit And determine whether the preset business rules are met. If the equation holds true, it means that the core financial data of the tobacco industry conforms to the business logic, and the process ends; otherwise, it is determined that the core financial data of the tobacco industry is abnormal, and then proceed to step (7-2). The advantage of step (7-1) is that it allows for a quick and preliminary assessment of whether financial data is abnormal through simple and intuitive business rules, providing a screening basis for subsequent precise positioning and improving the efficiency of financial verification.

[0046] (7-2) Based on the updated knowledge graph obtained in step (6), perform correlation analysis and root cause localization on the core financial data of the tobacco industry obtained in step (6) to obtain the localization results of abnormal financial data: Specifically, this step involves first obtaining the business entity association chain corresponding to the core financial data of the tobacco industry (such as "abnormal purchase amount → corresponding tobacco farmer → planting plot → processing batch → storage record") through the updated knowledge graph, and extracting all business data on the business entity association chain (such as actual purchase weight, grade evaluation results, processing loss rate, logistics and transportation records, etc.); then, based on the fusion characteristics between each two nodes in the updated knowledge graph obtained in step (6), obtaining the matching degree between the core financial data of the tobacco industry and each business data, and obtaining the location result of abnormal financial data based on the matching degree (such as purchase volume statistical deviation, unit price entry error, processing loss not calculated, etc.).

[0047] Specifically, obtaining anomaly localization results based on matching degree involves: 1. First, setting a acceptable matching degree threshold. (Set based on historical data patterns in the tobacco industry); 2. Examine the matching degree between financial data and each business data point (such as purchase weighing and unit price records); if it is lower than... The business data is the "problematic related data", and the corresponding links (such as acquisition statistics and unit price entry) are the root causes of the anomalies; 3. From these problematic related data, select the ones that have the greatest impact on financial anomalies (contributing more than half), and clarify the specific problems (such as "the acquisition volume is too small" or "the unit price is wrong"); 4. Finally, summarize the anomaly location results into "where the problem is and what the problem is", which will facilitate subsequent data modification.

[0048] The advantage of step (7-2) is that, relying on the updated knowledge graph, it can quickly trace the business entity relationship chain corresponding to the financial data, calculate the matching degree through node fusion features, accurately locate the root cause of the anomaly, and provide a clear direction for the verification and correction of the financial data.

[0049] (7-3) Based on the location results of the abnormal financial data obtained in step (7-2), the abnormal financial data is verified to obtain the verified financial data.

[0050] The advantage of the above sub-steps (7-1) to (7-3) is that they construct a complete financial data verification process of "preliminary judgment - root cause location - precise verification". Combined with the correlation analysis capabilities of business rules and knowledge graphs, they realize the rapid identification, precise location and effective correction of financial data anomalies, thereby improving the efficiency and accuracy of financial data verification.

[0051] The advantages of the above steps (1) to (7) are that a complete technical solution has been constructed, from multi-source data processing, knowledge graph construction and updating to financial data verification. This has achieved efficient integration of multi-source heterogeneous data, systematic modeling of business knowledge, in-depth mining of node association features, and accurate verification of financial data. It has effectively solved problems such as difficulty in integrating multi-source data, fragmentation of business knowledge, difficulty in identifying risks, and disconnection between financial data and business data in the tobacco industry. It has promoted the transformation of tobacco industry operation and management from "post-event verification" to "real-time monitoring" and improved the scientific nature and timeliness of decision-making.

[0052] By formally defining data processing, knowledge representation, reasoning and learning through the above formula, this invention realizes the full-process quantification and technological innovation from multi-source data to intelligent operation, providing rigorous and efficient technical support for the digital transformation of the tobacco industry.

[0053] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for verifying financial data in the tobacco leaf industry based on multi-source data fusion and knowledge graph, characterized in that, Includes the following steps: (1) Collect multi-source heterogeneous data of the tobacco industry chain and preprocess the multi-source heterogeneous data of the tobacco industry chain to obtain preprocessed multi-source heterogeneous data of the tobacco industry chain. (2) Perform semantic space mapping and standardization integration on the preprocessed multi-source heterogeneous data of the tobacco industry chain obtained in step (1) to obtain a standardized dataset; (3) Distribute the standardized dataset obtained in step (2) and containerize the stored data to obtain multiple data resource pools; (4) Based on the data resource pool obtained in step (3), obtain the entity set from the core business data table of the tobacco industry, obtain the relationship set from the structured business log, construct the graph basic framework according to the entity and relationship, and perform vectorized encoding on the graph basic framework to obtain the structured knowledge graph. (5) The structured knowledge graph obtained in step (4) is processed using a relational reasoning method based on graph neural networks (GNN) to obtain the fusion features between every two nodes in the structured knowledge graph after processing: (6) Use the knowledge graph incremental update algorithm to process the fusion features between every two static nodes in the structured knowledge graph obtained in step (5) to obtain the updated knowledge graph; (7) Obtain the core financial data of the tobacco industry and verify the core financial data of the tobacco industry based on the updated knowledge graph obtained in step (6) to obtain the verified financial data.

2. The method for verifying financial data in the tobacco industry based on multi-source data fusion and knowledge graph as described in claim 1, characterized in that, Step (1) includes the following sub-steps: (1-1) Obtain multi-source heterogeneous data of the tobacco industry chain, including a set of tobacco planting data composed of tobacco planting data collected by multiple sensors at different times. and a collection of tobacco processing data, including multiple tobacco processing data. ;in Let represent the tobacco planting data collected by the i-th sensor at the t-th timestamp, where i ∈ [1, total number of sensors N], t ∈ [1, total number of timestamps T]. This represents the processing data for the j-th tobacco leaf. This represents the field name of the j-th tobacco leaf processing data. The field value represents the j-th tobacco processing data, and j∈[1, the total number of tobacco processing data M]; (1-2) Impute missing values ​​in the multi-source heterogeneous data of the tobacco industry chain obtained in step (1-1) to obtain imputed multi-source heterogeneous data of the tobacco industry chain: (1-3) Determine whether there is any abnormal data in the multi-source heterogeneous data of the tobacco industry chain obtained in step (1-2) after filling. If there is, proceed to step (1-4); otherwise, proceed to step (1-5): (1-4) Mark and remove abnormal data in the multi-source heterogeneous data of the tobacco industry chain to obtain multi-source heterogeneous data of the tobacco industry chain after removing abnormal data, and then proceed to step (1-5): (1-5) Perform optimization processing such as format unification and redundancy removal on the multi-source heterogeneous data of the tobacco industry chain obtained in step (1-3) or step (1-4) to obtain pre-processed multi-source heterogeneous data of the tobacco industry chain.

3. The method for verifying financial data in the tobacco industry based on multi-source data fusion and knowledge graph as described in claim 1 or 2, characterized in that, For tobacco planting data collected by the i-th sensor at the t-th time stamp The missing value imputation process in step (1-2) uses the following formula: ； in This represents the tobacco planting data collected by the i-th sensor at the t-th timestamp after the data filling process. This indicates the timestamp preceding timestamp t. Indicates the next timestamp after timestamp t; For the j-th tobacco leaf processing data Specifically, the missing value imputation process in step (1-2) uses the following formula: ； in This represents the parameter value that makes the subsequent function reach its maximum value; Indicates the first Fields in tobacco leaf processing data The set of possible values ​​for this field. The set of all possible values. This indicates counting, or statistical analysis of values. In the set The number of times it appears in; For the numerical data in the multi-source heterogeneous data of the tobacco industry chain after imputation processing, the anomaly detection in steps (1-3) is performed using... The principles, related calculations, and judgment logic are as follows: First, calculate the numerical data. mean and standard deviation : Then, determine if the numerical data exists. If so, the numerical data is considered abnormal; otherwise, the numerical data is considered normal. For data strongly correlated with tobacco industry business in the multi-source heterogeneous data of the tobacco industry chain after filling in the gaps, the anomaly judgment in steps (1-3) is based on preset business rules, and the judgment logic is as follows: Data strongly correlated with the tobacco leaf industry satisfy If the data exceeds a reasonable range, then the data that is strongly correlated with the tobacco industry is deemed abnormal. like If the data is strongly correlated with the tobacco industry business, it is considered normal. This represents the lowest historical yield value for tobacco leaves in this production area. This represents the highest single-yield value in the history of tobacco production in this region. Steps (1-4) are as follows: First, add an anomaly identifier field to the abnormal data and record the anomaly type and cause. Then, through data filtering, remove the abnormal data with the anomaly identifier field from the multi-source heterogeneous data of the tobacco industry chain, retain all normal data, and finally obtain the multi-source heterogeneous data of the tobacco industry chain after removing the abnormal data.

4. The method for verifying financial data in the tobacco industry based on multi-source data fusion and knowledge graphs according to any one of claims 1 to 3, characterized in that, Step (2) includes the following sub-steps: (2-1) Perform semantic space mapping on the preprocessed multi-source heterogeneous data of the tobacco industry chain obtained in step (1) to obtain cross-modal semantic vectors: (2-2) The cross-modal semantic vectors obtained in step (2-1) are standardized and integrated using a cross-modal semantic alignment algorithm and a unified feature space mapping technique to map them to a unified semantic space and generate a JSON format dataset. As a standardized dataset, This serves as a unique identifier for the entities in this standardized dataset. This represents the k-th JSON-formatted data in the standardized dataset, where k represents the total number of JSON-formatted data in the standardized dataset.

5. The method for verifying financial data in the tobacco industry based on multi-source data fusion and knowledge graph as described in claim 4, characterized in that, The text data obtained in step (1) from the preprocessed multi-source heterogeneous data of the tobacco industry chain, and step (2-1) using the fine-tuned BERT model to achieve semantic space mapping, the formula is: ； in For text semantic vectors, For text data, This represents the 768-dimensional real space to which the text semantic vector belongs; For the image data in the preprocessed multi-source heterogeneous data of the tobacco industry chain obtained in step (1), step (2-1) uses the ResNet50 model to achieve semantic space mapping, and the formula is: ； in For image feature vectors, For image data, This represents the 2048-dimensional real space to which the image feature vector belongs.

6. The method for verifying financial data in the tobacco industry based on multi-source data fusion and knowledge graph as described in claim 5, characterized in that, Step (3) includes the following sub-steps: (3-1) Distribute the standardized dataset obtained in step (2) to obtain the stored data: Specifically, for unstructured data in standardized datasets, the HDFS distributed storage system is used for distributed storage, where the number of replicas stored is... Satisfy the following formula: ； in It is a rounding function. The probability of losing stored unstructured data is given, and its value ranges from 0.001 to 0.01, preferably 0.

005. For structured data in standardized datasets, ClickHouse database is used for distributed storage and partitioned tables are designed. (3-2) Containerize and dynamically manage the stored data obtained in step (3-1) to obtain multiple data resource pools; Specifically, this step involves first, containerizing the stored data obtained in step (3-1), that is, converting the access logic and processing logic of the data into executable code, and then packaging the executable code and the runtime dependent environment together into a standardized container image, which serves as a containerized data service unit that can be elastically scheduled. Subsequently, Prometheus was used to obtain the resource utilization rate of the containerized data service unit. The used resources are the resources actually occupied by the standardized container image, and the total allocated resources are the preset maximum available resources. Subsequently, the resource utilization rate of the obtained containerized data service unit is determined. Whether u is greater than 0.8 or less than 0.3, if u>0.8, the scaling process is automatically triggered, that is, the number of Pods is increased by 30%~50% of the current number of Pods, so as to obtain a containerized data service cluster with elastic resource scaling; if u<0.3, the scaling process is automatically triggered, that is, the number of Pods is reduced by 20%~30% of the current number of Pods, so as to obtain a containerized data service cluster with elastic resource scaling. Finally, the obtained containerized data service cluster is processed using a cluster high availability scheduling algorithm and data access link adaptation technology to obtain multiple stable and available data resource pools.

7. The method for verifying financial data in the tobacco industry based on multi-source data fusion and knowledge graph as described in claim 6, characterized in that, In step (4), based on the data resource pool obtained in step (3), the process of obtaining entity sets and relationship sets from the core business data table and structured business logs of the tobacco industry, and constructing a graph framework based on these entities and relationships, specifically involves the following steps: First, multiple core business elements of the tobacco industry are extracted from the data resource pool, and an entity set is constructed based on these core business elements. and relation set ,in Represents the first in the entity set One entity, Indicates the size of the entity set. Represents the first in the set of relations A relationship, Indicates the size of the relation set; Then, the basic framework of the graph is constructed based on the entities in the entity set and the relationships in the relation set. : ； in Represents a set of entities The tail entity in the middle; Step (4) involves vectorizing the basic framework of the knowledge graph to obtain a structured knowledge graph. The specific process is as follows: First, the basic framework of the graph. All head entities Tail entity Mapped to the complex vector space respectively, i.e. , and the relationship It is denoted as the rotation factor in this complex vector space, i.e. ,in The preset vector dimension (which can be set according to the amount of business data in the tobacco industry and the complexity of entity relationships, such as...) or ), Represents the field of complex numbers; Then, vector matching constraints are added to the obtained complex vector space to obtain a vector set with business association constraints: Subsequently, the loss function is obtained based on the resulting vector set with business-related constraints. : ； in, This represents the boundary hyperparameter, which takes a value of 1 and is used to differentiate between positive and negative samples. Represents the vector distance function. The negative sample tail entity is formed by replacing the basic framework of the graph constructed in step (4). Middle Triple Group Positive sample tail entity Generate, and ensure that the generated negative sample tail entities are generated. With head entity No real business relationship; Subsequently, the stochastic gradient descent (SGD) method was used to adjust the obtained loss function. Iterative processing is performed, that is, the head entity vector is continuously corrected through iteration. Tail entity vector and relational rotation factor The value until the loss function The convergence continues until a preset threshold is reached, thus obtaining a structured knowledge graph.

8. The method for verifying financial data in the tobacco industry based on multi-source data fusion and knowledge graph as described in claim 7, characterized in that, Step (5) is as follows: First, the structured knowledge graph obtained in step (4) is statically reasoned using the graph attention network GAT to obtain the association results between every two static nodes. Specifically, this step involves obtaining the association result between the i-th static node and the j-th static node in the structured knowledge graph. The following formula is used, where i and j are both ∈ [1, and the total number of nodes in the structured knowledge graph is 1]. ]: ； And there are: ； in The value range is from 0 to L-1. The preset total number of layers for the GAT network. This indicates that the i-th static node is in the GAT network at the _th ... The embedding vector of the layer reflects the association features between the i-th static node and the j-th static node. This represents the activation function. This represents the set of neighboring nodes of the i-th static node in the structured knowledge graph; Represents an exponential function; Indicates a linear unit with leakage correction; Represents the attention vector; Indicates the first in the GAT network The weight matrix of the layer has dimensions of ,in For the first The input dimension of the layer node embedding, = or , For the first Output dimension of layer node embedding; weight matrix The elements in the array are learnable real-valued parameters, and the initial values ​​are randomly generated using a Xavier normal distribution. Indicates the first in the GAT network The original embedding vector of the i-th static node in layer 1. Indicating the first in the GAT network The original embedding vector of the j-th static node in the layer; This represents a vector concatenation operation used to merge the embedded features of two nodes; Indicating the first in the GAT network The original embedding vector of the k-th static node in the layer, where k∈[1, 1], the i-th static node in the GAT network is the first static node in the layer. [Number of all neighboring nodes in the layer]; Then, a temporal graph neural network is used to analyze the association results between every two static nodes in the obtained structured knowledge graph. Perform dynamic reasoning to obtain The dynamic embedding vector of the i-th static node at time i : ； in, Indicates a gated loop unit. express The dynamic embedding vector of the i-th static node at time i; Subsequently, the association results between every two static nodes in the obtained structured knowledge graph were analyzed. and The dynamic embedding vector of the i-th static node at time i Feature alignment is performed to obtain the pre-fusion feature set corresponding to each pair of static nodes in the structured knowledge graph; Subsequently, a feature fusion algorithm is used to integrate the pre-fusion feature sets between every two static nodes in the obtained structured knowledge graph, so as to obtain the preliminary fusion features between every two static nodes in the structured knowledge graph. Finally, the preliminary fusion features between every two static nodes in the obtained structured knowledge graph are validated and corrected to obtain the processed fusion features between every two static nodes in the structured knowledge graph.

9. The method for verifying financial data in the tobacco industry based on multi-source data fusion and knowledge graph as described in claim 8, characterized in that, Step (7) includes the following sub-steps: (7-1) Obtain core financial data for the tobacco leaf industry, including actual acquisition amounts. Actual data on acquisition volume Actual data on purchase price per unit And determine whether the preset business rules are met. If the equation holds true, it means that the core financial data of the tobacco industry conforms to the business logic, and the process ends; otherwise, it is determined that the core financial data of the tobacco industry is abnormal, and then proceed to step (7-2). (7-2) Based on the updated knowledge graph obtained in step (6), perform correlation analysis and root cause localization on the core financial data of the tobacco industry obtained in step (6) to obtain the localization results of abnormal financial data: (7-3) Based on the location results of the abnormal financial data obtained in step (7-2), the abnormal financial data is verified to obtain the verified financial data.

10. The method for verifying financial data in the tobacco industry based on multi-source data fusion and knowledge graph as described in claim 9, characterized in that, Step (7-2) is as follows: First, obtain the business entity association chain corresponding to the core financial data of the tobacco industry through the updated knowledge graph, and extract all business data on the business entity association chain; then, according to the fusion feature between each two nodes in the updated knowledge graph obtained in step (6), obtain the matching degree between the core financial data of the tobacco industry and each business data, and obtain the location result of abnormal financial data according to the matching degree.

Images