Intelligent amount and fee auditing method based on artificial intelligence large model

Abstract

The invention discloses an intelligent amount and fee auditing method based on an artificial intelligence large model, and the method comprises the steps: S1, carrying out the automatic collection and arrangement of needed data through an intelligent corpus collection technology and a natural language processing technology, and forming a structured and unstructured multi-modal data set; s2, after the multi-modal data set is preprocessed, key features are extracted, and fusion processing is conducted on the multi-modal data set with the key features extracted; s3, through a graph neural network, based on the multi-modal data set, establishing a relation model for the power consumption behavior and the charging relation between the users, and using the relation model to automatically detect and identify the amount and fee abnormity problem; and S4, establishing an auditing rule model, and training the graph neural network output vector and the auditing rule model based on a reinforcement learning method to improve the accuracy of the amount and cost anomaly detection. According to the invention, the problems of insufficient scene adaptability, low data utilization rate and lagging auditing rule and technology update of the existing amount and cost auditing mode are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of electricity charges, and in particular relates to an intelligent audit method for electricity charges based on an artificial intelligence large model. Background Art

[0002] Big model technology is an artificial intelligence model based on deep learning, which can realize the automated processing and intelligent decision-making of complex tasks through the training of massive data. In the audit of electricity charges, big model technology can integrate multi-source information such as user billing files, algorithm rules and historical abnormal data, automatically identify abnormal electricity consumption, optimize audit rules, improve the accuracy and efficiency of audits, solve the abnormal quantity and charge problems in complex business scenarios, and promote the intelligentization and automation of the audit process.

[0003] However, with the increase in the number of power grid users, the complexity of user electricity consumption behavior and the digital transformation of the power system, the traditional quantity and fee audit model is facing many challenges: (1) Insufficient scenario adaptability: When dealing with quantity and fee audit, the existing large models mainly target some standardized business scenarios. However, in the actual environment where electricity consumption behavior is complex and changeable, the model's generalization ability is insufficient and it is difficult to cover all application scenarios. (2) Insufficient data utilization: The existing large models rely more on structured electricity consumption data and fail to make full use of multimodal data, such as user behavior, historical abnormal cases and expert knowledge, which results in the model's inability to fully analyze complex business scenarios, affecting the accuracy and coverage of the audit results. (3) Insufficient audit rules and technology updates: Most of the existing fee audit rules rely on manual settings, and the model lacks an automated rule update and optimization mechanism, which cannot adapt to the ever-changing business scenarios and new electricity consumption behaviors.

[0004] The Chinese patent with the publication number CN114154719A discloses a method for predicting the basic electricity charges for large industrial electricity consumption based on a neural network, which includes the following steps: (1) firstly statistically preprocessing the historical electricity load data of the enterprise; (2) then normalizing the load data and converting the load data into an input vector in the interval [0,1]; (3) then constructing an error back propagation algorithm neural network model based on the input vector and the required output vector; (4) after completing the training and passing the test, the method is put into use, the maximum value of the output result is selected and normalized and restored; (5) the predicted load value is converted into the basic electricity charge prediction value of the demand. The feature extraction of this invention is relatively simple, mainly relying on the statistical preprocessing and normalization of the historical electricity load data, and lacks comprehensive consideration of multiple factors; and the invention uses a neural network model of the error back propagation algorithm, whose structure and parameters are relatively fixed, and it is difficult to quickly adapt to the changing electricity consumption behavior and external environment, and cannot flexibly adapt to complex electricity consumption behavior relationships. Summary of the invention

[0005] The present invention provides an intelligent quantity and fee audit method based on an artificial intelligence big model, aiming to solve the problems of insufficient scenario adaptability of existing quantity and fee audit models, low data utilization, and lagging audit rules and technology updates.

[0006] In order to solve the above technical problems, the present invention provides a method for intelligent auditing of quantity and fee based on an artificial intelligence big model, comprising the following steps:

[0007] S1: Use intelligent corpus collection technology and natural language processing technology to automatically collect and organize user billing files, user behavior logs, algorithm rules, exception review rules, historical exception cases and expert knowledge to form structured and unstructured multimodal data sets.

[0008] S2: After preprocessing the multimodal dataset, the key features are extracted using feature extraction technology, and the multimodal dataset with the extracted key features is fused.

[0009] S3: Through graph neural networks and multimodal data sets, a relational model is established for the electricity consumption behavior and billing relationship between users to automatically detect and identify abnormal quantity and fee issues.

[0010] S4: Establish an audit rule model, train the graph neural network output vector and the audit rule model based on the reinforcement learning method, and adjust the reinforcement learning parameters through hyperparameter optimization technology and back propagation algorithm to improve the accuracy of volume and fee anomaly detection.

[0011] Preferably, the step S2 is specifically:

[0012] The multimodal datasets were preprocessed through data cleaning including outlier detection, missing value filling and normalization.

[0013] The key features are extracted from the preprocessed multimodal dataset using any of the principal component analysis, linear discriminant analysis and feature selection algorithms.

[0014] A hybrid model based on convolutional neural networks and long short-term memory networks is used to fuse multimodal datasets with extracted key features.

[0015] Preferably, the hybrid model structure of the convolutional neural network and the long short-term memory network is specifically:

[0016] f(x)=α*f struct (x struct )+β*f unstruct (x unstruct )+γ*f time (x time )

[0017] Among them, f structRepresents the feature extraction function of structured data, which is mainly processed by a fully connected network; f unstruct The feature extraction function representing unstructured data is extracted by the convolutional neural network; f time represents the feature extraction function of time series data, which is extracted by the long short-term memory network; α, β, and γ are dynamic weight coefficients, which represent the importance of various types of data. These weights are adaptively optimized through the back propagation algorithm during the training process; x struct is structured data; x unstruct is unstructured data; x time For time series data.

[0018] Preferably, the step S3 is specifically:

[0019] The user network is represented as an undirected graph G = (V, E), where V is the user node and E is the relationship between users. i The feature vector of ∈V is the feature vector in the multimodal dataset Information propagation and aggregation are performed through the graph neural network layers to update the representation of the nodes.

[0020] For each user node v i , whose feature vector at the k+1 layer is It is expressed as:

[0021]

[0022] In the formula, Aggregate is an aggregation function, which means average, maximum or weighted sum; is the feature vector of the u-th user node in the k-th layer; N(v i ) is the node v i The set of neighbor nodes of (k) is the weight matrix of the kth layer, and σ is the activation function.

[0023] Through the propagation of multi-layer graph neural networks, the representation of user nodes is updated, and the information of its neighbor nodes is gradually integrated to capture the association of user electricity consumption behaviors.

[0024] Preferably, for each user node, an anomaly detection function A(v) is constructed to evaluate the abnormality of the user node, indicating the degree of abnormality of the node in the network. Combined with the updated representation of the user node, the anomaly metric is defined as:

[0025]

[0026] Among them, h mean It is the average representation of all user nodes. The larger the anomaly metric, the greater the difference between the power consumption behavior of the user node and its neighbor nodes, and there may be abnormal behavior.

[0027] Preferably, in step S4, the audit rule model is trained by a reinforcement learning method as follows:

[0028] The reinforcement learning method is based on the Q-learning model. By interacting with the environment, the strategy π(s) is continuously optimized. Its goal is to maximize the cumulative reward R = ∑γ t r t , where γ is the discount factor, r t is the immediate reward at step t.

[0029] In the Q-learning model training, the audit rule model is set as action a, the graph neural network output vector is set as state s, and the state-action value function Q(s,a) is defined. Its update rule is:

[0030] Q(s,a)←Q(s,a)+α[r+γmaxQ(s′,a′)-Q(s,a)]

[0031] Where r is the immediate reward; α is the learning rate; γ is the discount factor; Q(s′,a′) is the state-action value function, which specifically refers to the expected value of the cumulative reward that can be obtained in the future after taking action a in state s.

[0032] At each time step, the Q-learning model calculates the current state s t Select action a t , through environmental feedback reward r t , and optimize the audit rule model according to the Q value update formula.

[0033] Preferably, the method further comprises:

[0034] S5: Based on rule reasoning technology and expert system, a dynamic optimization mechanism is constructed to automatically optimize the generation and update of initial audit rules, and combined with reinforcement learning algorithms, the audit rules can be adaptively adjusted to cope with changes in new business scenarios.

[0035] On the other hand, the present invention provides an intelligent quantity and fee audit system based on an artificial intelligence big model, including a data acquisition module, a data extraction and fusion module, a relationship modeling and anomaly detection module, and a model training and quantity and fee audit module.

[0036] The data collection module is used to automatically collect and organize user billing files, user behavior logs, algorithm rules, exception review rules, historical exception cases and expert knowledge using intelligent corpus collection technology and natural language processing technology to form structured and unstructured multimodal data sets.

[0037] The data extraction and fusion module is used to extract key features using feature extraction technology after preprocessing the multimodal data set, and to fuse the multimodal data set with extracted key features.

[0038] The relationship modeling and anomaly detection module is used to establish a relationship model for the electricity consumption behavior and billing relationship between users based on multimodal data sets through graph neural networks, so as to automatically detect and identify abnormal quantity and fee problems.

[0039] The quantity and fee audit module is used to establish an audit rule model, train the graph neural network output vector and the audit rule model based on the reinforcement learning method, and adjust the reinforcement learning parameters through hyperparameter optimization technology and back propagation algorithm to improve the accuracy of quantity and fee anomaly detection.

[0040] On the other hand, the present invention also provides 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 computer program, the method for intelligent quantity and fee audit based on the artificial intelligence big model as described in any embodiment of the present invention is implemented.

[0041] On the other hand, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the intelligent quantity and fee audit method based on the artificial intelligence big model as described in any embodiment of the present invention.

[0042] Compared with the prior art, the present invention has the following technical effects:

[0043] The present invention is based on multimodal data fusion, deep learning model construction and reinforcement learning optimization, which greatly improves the accuracy and efficiency of quantity and fee audit. It uses intelligent corpus collection technology to comprehensively integrate user billing data, build a hybrid deep learning model to process complex data, optimize audit rules to adapt to new business scenarios, and achieve real-time and accurate anomaly detection. This method not only improves the audit capabilities of the power industry, but also supports the continuous optimization and self-adjustment of the model, significantly improving the intelligence level and business responsiveness of quantity and fee audit. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is an overall flow chart of the intelligent audit method for quantity and fee based on the artificial intelligence big model described in the present invention. DETAILED DESCRIPTION

[0045] In order to make the objectives, technical solutions and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in combination with specific embodiments of the present application and with reference to the accompanying drawings.

[0046] Embodiment 1

[0047] This embodiment provides a method for intelligent auditing of volume and fee based on an artificial intelligence big model, see Figure 1 As shown, the following steps are included:

[0048] S1: Use intelligent corpus collection technology and natural language processing technology to automatically collect and organize user billing files, user behavior logs, algorithm rules, abnormal audit rules, historical abnormal cases and expert knowledge to form structured and unstructured multimodal data sets. In addition, distributed storage technology (such as Hadoop and NoSQL) can be used to build an efficient data storage structure to ensure parallel processing and integration of multimodal data. The multimodal data sets collected above (including user billing files, user behavior logs, etc.) can be used as the basis for extracting key features in step S2, which is helpful for subsequent fusion processing; in step S3, a more accurate user behavior relationship model is established through the fused multimodal data set; in step S4, an audit rule model is established through abnormal audit rules, and the performance of the audit rule model in different policy scenarios is optimized and adjusted; in the S5 audit rule optimization step, the initial audit rules can be automatically generated based on the existing algorithm rules, historical abnormal cases and expert knowledge, and feedback optimization is continuously achieved.

[0049] The quantity and fee audit involves multiple types of data, including structured electricity billing data, unstructured user behavior data, and relational electricity usage connections between users, and includes multiple modalities such as text, audio, and video. In order to effectively process these data, this embodiment provides a multi-modal quantity and fee audit data adaptive fusion processing algorithm, which is designed to process the multi-modal data involved in the quantity and fee audit process, such as structured electricity data, unstructured user behavior data, and historical abnormal case data. Dynamically and adaptively fuse multiple modal data to build a unified data representation space.

[0050] S2: After preprocessing the multimodal dataset, the key features are extracted using feature extraction technology, and the multimodal dataset with the extracted key features is fused.

[0051] Preferably, the step S2 is specifically:

[0052] The multimodal datasets were preprocessed through data cleaning including outlier detection, missing value filling and normalization (based on KNN and interpolation methods).

[0053] The key features are extracted from the preprocessed multimodal dataset using any of principal component analysis, linear discriminant analysis, and feature selection algorithms (such as L1 regularization).

[0054] The hybrid model based on convolutional neural network and long short-term memory network is used to fuse the multimodal data sets with key features. The structured data (such as user billing archive data) and unstructured data (such as text rules) are fused to achieve unified expression of different data sources.

[0055] As a preferred implementation of this embodiment, the hybrid model structure of the convolutional neural network and the long short-term memory network is specifically:

[0056] f(x)=α*f struct (x struct )+β*f unstruct (x unstruct )+γ*f time (x time )

[0057] Among them, f struct Represents the feature extraction function of structured data, which is mainly processed by a fully connected network; f unstruct The feature extraction function representing unstructured data is extracted by the convolutional neural network; f time represents the feature extraction function of time series data, which is extracted by the long short-term memory network; α, β, and γ are dynamic weight coefficients, which represent the importance of various types of data. These weights are adaptively optimized through the back propagation algorithm during the training process; x struct is structured data; x unstruct is unstructured data; x time For time series data.

[0058] Based on the fusion processing of multimodal data sets based on the hybrid model of convolutional neural network and long short-term memory network, the effect of time series data analysis is improved based on the Transformer model. The application of the Transformer model aims to further improve the analysis ability of time series data, such as the dynamic change law in power and loss, so as to make up for the problems of information loss and low efficiency when the long short-term memory network faces data with too long time. The input data of the Transformer model is the time series feature data output by the long short-term memory network; after the Transformer model is processed by the self-attention mechanism, the output processed time series feature data is used as one of the input data of the graph neural network.

[0059] S3: A relational model is established between electricity consumption behaviors and billing relationships among users through graph neural networks, which is used to automatically detect and identify abnormal quantity and billing issues.

[0060] As a preferred implementation of this embodiment, the relationship model established by the graph neural network is used for electricity consumption association, billing anomaly propagation and multi-level dependencies between users.

[0061] The user network can be represented as an undirected graph G = (V, E), where V is the user node and E is the relationship between users (e.g., region, behavior similarity). i The feature vector of is the feature vector in the multimodal dataset Information propagation and aggregation are performed through the graph neural network layers to update the representation of the nodes.

[0062] Node update of graph neural network: For each power user node v i , whose feature vector at the k+1 layer is It is expressed as:

[0063]

[0064] In the formula, Aggregate is an aggregation function, which means average, maximum or weighted sum; is the characteristic vector of the u-th power user node in the k-th layer; N(v i ) is the node v i The set of neighbor nodes of (k) is the weight matrix of the kth layer, and σ is the activation function.

[0065] Through the propagation of multi-layer graph neural networks, the representation of user nodes is continuously updated, gradually integrating the information of its neighbor nodes and capturing the complex associations of user electricity usage behaviors.

[0066] As a preferred implementation of this embodiment, for each user node, an anomaly detection function A(v) is constructed to evaluate the abnormality of the user node, indicating the degree of abnormality of the node in the network. Combined with the updated representation of the node, the anomaly metric is defined as:

[0067]

[0068] Among them, h mean It is the average representation of all user nodes. The larger the anomaly metric, the greater the difference between the power consumption behavior of the user node and its neighbor nodes, and there may be abnormal behavior.

[0069] The relational model established by the graph neural network has as input the time series feature data processed by the Transformer model and the structured data and time series data output by the hybrid model. Its output is the embedding vector of the behavioral relationship and billing relationship between users, which is specifically manifested in the form of similarity measurement between users, abnormal user behavior detection, and clustering relationship of power consumption patterns.

[0070] S4: Establish an audit rule model, train the graph neural network output vector and the audit rule model based on the reinforcement learning method, and adjust the reinforcement learning parameters through hyperparameter optimization technology and back propagation algorithm to improve the accuracy of volume and fee anomaly detection.

[0071] Specifically, in order to enable the audit rules to automatically adjust as the business scenarios change, an audit rule model is established. The audit rule model includes a series of rules and standards for evaluating user electricity consumption behavior, such as the fluctuation range of electricity consumption (whether it is within the normal consumption range), electricity consumption habits in time periods (whether it conforms to the user's historical electricity consumption pattern), detection of abnormal events (such as sudden high power consumption, inconsistent billing, etc.), and billing standards (whether it conforms to the price and settlement policy formulated by the company). Through the reinforcement learning method, it interacts with all the data of the actual multimodal data set (for example: structured data such as the fluctuation range of electricity consumption, unstructured data such as real-time business data, and time series data such as historical business data), and dynamically adjusts these audit rules to make them more adaptive. The reinforcement learning method continuously optimizes the audit rule model through feedback loops, so that the model can adaptively adjust the audit standards.

[0072] The reinforcement learning method is based on the Q-learning model. By interacting with the environment, the strategy π(s) is continuously optimized. Its goal is to maximize the cumulative reward R = ∑γ t r t , where γ is the discount factor, r t is the immediate reward at step t.

[0073] In the Q-learning model training, the audit rule is set as action a, the audit scenario is set as state s, and the state-action value function Q(s,a) is defined. Its update rule is:

[0074] Q(s,a)←Q(s,a)+α[r+γmaxQ(s′,a′)-Q(s,a)]

[0075] Where r is the immediate reward; α is the learning rate; γ is the discount factor; Q(s′,a′) is the state-action value function, which specifically refers to the expected value of the cumulative reward that can be obtained in the future after taking action a in state s.

[0076] At each time step, the Q-learning model calculates the current state s t Select action a t , through environmental feedback reward r t , and optimize the audit rule model according to the Q value update formula. In the long-term operation, the model can learn the optimal audit rule adjustment strategy to adapt to the ever-changing electricity consumption behavior and business scenarios.

[0077] As a preferred implementation of this embodiment, the method may further include:

[0078] S5: Based on rule reasoning technology and expert system, a dynamic optimization mechanism is constructed to automatically optimize the generation and update of initial audit rules, and combined with reinforcement learning algorithms, the audit rules can be adaptively adjusted to cope with changes in new business scenarios.

[0079] Furthermore, the step S5 is specifically as follows:

[0080] Based on expert systems and rule reasoning technology, a dynamic audit rule optimization engine is built, and fuzzy logic and Bayesian networks are used to automatically optimize the generation and update of audit rules to achieve intelligent reasoning of user electricity consumption behavior in different business scenarios.

[0081] Based on historical abnormal cases and the audit rule optimization engine, the abnormal detection rules are automatically updated to improve the detection accuracy of transformer loss and power anomalies; at the same time, the existing algorithm rules will be incorporated into the optimization framework, and the audit rules will be optimized through reinforcement learning and the audit rule optimization engine. The abnormal detection of transformer loss is mainly carried out in the dynamic optimization stage of the relationship model and audit rules. When establishing the relationship model, the correlation between the user's electricity consumption behavior and billing is analyzed through the graph neural network. In the abnormal processing optimization stage, the rules and accuracy of transformer loss detection are further optimized in combination with historical abnormal cases and the rule optimization engine. The above is a supplement and refinement of the preliminary recognition results of the graph neural network stage.

[0082] As a preferred implementation of this embodiment, the deployment and application of the method may be specifically as follows:

[0083] Containerized deployment: Use container technologies such as Docker to encapsulate large models and deploy them to the power company's fee and quantity audit system to achieve cross-platform operation.

[0084] API interface integration: Seamlessly integrate large models with existing business systems through the RESTful API interface to ensure real-time feedback of audit results.

[0085] Microservice architecture: The microservice architecture is designed to support load balancing and distributed computing, ensuring the stability and real-time performance of the audit system under high concurrency conditions.

[0086] Continuous monitoring and feedback: Combine the log monitoring system and A / B testing to track the actual performance of the large model and continuously optimize the review process through the feedback mechanism.

[0087] Embodiment 2

[0088] Correspondingly, this embodiment provides an intelligent quantity and fee audit system based on an artificial intelligence big model, including a data acquisition module, a data extraction and fusion module, a relationship modeling and anomaly detection module, and a model training and quantity and fee audit module.

[0089] The data collection module is used to automatically collect and organize user billing files, user behavior logs, algorithm rules, exception review rules, historical exception cases and expert knowledge using intelligent corpus collection technology and natural language processing technology to form structured and unstructured multimodal data sets. This module is used to implement the function of step S1 in Example 1 and will not be repeated here.

[0090] The data extraction and fusion module is used to extract key features using feature extraction technology after preprocessing the multimodal data set, and to fuse the multimodal data set with extracted key features. This module is used to implement the function of step S2 in Example 1 and will not be repeated here.

[0091] The relationship modeling and anomaly detection module is used to establish a relationship model for the electricity consumption behavior and billing relationship between users based on a multimodal data set through a graph neural network, and is used to automatically detect and identify abnormal quantity and fee problems. This module is used to implement the function of step S3 in Example 1 and will not be repeated here.

[0092] The quantity and fee audit module is used to establish an audit rule model, train the graph neural network output vector and the audit rule model based on the reinforcement learning method, and adjust the reinforcement learning parameters through hyperparameter optimization technology and back propagation algorithm to improve the accuracy of quantity and fee anomaly detection. This module is used to implement the function of step S4 in Example 1 and will not be repeated here.

[0093] Embodiment 3

[0094] This embodiment provides an electronic device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for intelligent auditing of quantity and charges based on an artificial intelligence big model as described in any embodiment of the present invention is implemented.

[0095] Embodiment 4

[0096] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method for intelligent auditing of quantity and fees based on an artificial intelligence big model as described in any embodiment of the present invention is implemented.

[0097] In the embodiments of the present application, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent the existence of A alone, the existence of A and B at the same time, and the existence of B alone. Among them, A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b and c can be represented by: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, c can be single or multiple.

[0098] Those of ordinary skill in the art will appreciate that the various units and algorithm steps described in the embodiments disclosed herein can be implemented in a combination of electronic hardware, computer software, and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0099] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0100] In several embodiments provided in the present application, if any function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory; hereinafter referred to as: ROM), random access memory (Random Access Memory; hereinafter referred to as: RAM), disk or optical disk, and other media that can store program codes.

[0101] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An intelligent audit method for quantity and fee based on an artificial intelligence big model, characterized in that: The following steps are involved: S1: Use intelligent corpus collection technology and natural language processing technology to automatically collect and organize user billing files, user behavior logs, algorithm rules, exception review rules, historical exception cases and expert knowledge to form structured and unstructured multimodal data sets; S2: After preprocessing the multimodal data set, the key features are extracted using feature extraction technology, and the multimodal data set with the extracted key features is fused; S3: Through graph neural networks and multimodal data sets, a relationship model is established between users’ electricity consumption behaviors and billing relationships to automatically detect and identify abnormal electricity consumption and billing issues. S4: Establish an audit rule model, train the graph neural network output vector and the audit rule model based on the reinforcement learning method, and adjust the reinforcement learning parameters through hyperparameter optimization technology and back propagation algorithm to improve the accuracy of volume and fee anomaly detection.

2. The method for intelligent auditing of quantity and fee based on artificial intelligence big model according to claim 1 is characterized in that: The step S2 is specifically as follows: Preprocessing of multimodal datasets through data cleaning including outlier detection, missing value filling, and normalization; Extract key features from the preprocessed multimodal data set using any of principal component analysis, linear discriminant analysis, and feature selection algorithms; A hybrid model based on convolutional neural networks and long short-term memory networks is used to fuse multimodal datasets with extracted key features.

3. The method for intelligent auditing of quantity and fee based on artificial intelligence big model according to claim 2 is characterized in that: The hybrid model structure of the convolutional neural network and the long short-term memory network is specifically as follows: f(x)=α*f struct (x struct )+β*f unstruct (x unstruct )+γ*f time (x time ) Among them, f dtruct Represents the feature extraction function of structured data, which is mainly processed by a fully connected network; f unstruct The feature extraction function representing unstructured data is extracted by the convolutional neural network; f time represents the feature extraction function of time series data, which is extracted by the long short-term memory network; α, β, and γ are dynamic weight coefficients, which represent the importance of various types of data. These weights are adaptively optimized through the back propagation algorithm during the training process; x struct is structured data; x unstruct is unstructured data; x time For time series data.

4. The method for intelligent auditing of quantity and fee based on artificial intelligence big model according to claim 1 is characterized in that: The step S3 is specifically as follows: The user network is represented as an undirected graph G = (V, E), where V is the user node and E is the relationship between users. i The feature vector of ∈V is the feature vector in the multimodal dataset Information propagation and aggregation are performed through the graph neural network layer to update the representation of the nodes; For each user node v i , whose feature vector at the k+1 layer is It is expressed as: In the formula, Aggregate is an aggregation function, which means average, maximum or weighted sum; is the feature vector of the u-th user node in the k-th layer; N(v i ) is the node v i The set of neighbor nodes; W (k) is the weight matrix of the kth layer, σ is the activation function; Through the propagation of multi-layer graph neural networks, the representation of user nodes is updated, and the information of its neighbor nodes is gradually integrated to capture the association of user electricity consumption behaviors.

5. The method for intelligent auditing of quantity and fee based on artificial intelligence big model according to claim 4 is characterized in that: For each user node, an anomaly detection function A(v) is constructed to evaluate the abnormality of the user node, indicating the degree of abnormality of the node in the network. Combined with the updated representation of the user node, the anomaly metric is defined as: Among them, h mean It is the average representation of all user nodes. The larger the anomaly metric, the greater the difference between the power consumption behavior of the user node and its neighbor nodes, and there may be abnormal behavior.

6. The method for intelligent auditing of quantity and fee based on artificial intelligence big model according to claim 1 is characterized in that: In step S4, the audit rule model is trained by the reinforcement learning method as follows: The reinforcement learning method is based on the Q-learning model. By interacting with the environment, the strategy π(s) is continuously optimized. Its goal is to maximize the cumulative reward R = ∑γ t r t , where γ is the discount factor, r t is the immediate reward at step t; In the Q-learning model training, the audit rule model is set as action a, the graph neural network output vector is set as state s, and the state-action value function Q(s,a) is defined. Its update rule is: Q(s,a)←Q(s,a)+α[r+γmaxQ(s′,a′)-Q(s,a)] Where r is the immediate reward; α is the learning rate; γ is the discount factor; Q(s′,a′) is the state-action value function, which specifically refers to the expected value of the cumulative reward that can be obtained in the future after taking action a in state s; At each time step, the Q-learning model calculates the current state s t Select action a t , through environmental feedback reward r t , and optimize the audit rule model according to the Q value update formula.

7. The method for intelligent auditing of quantity and cost based on artificial intelligence big model according to claim 1 is characterized in that: The method further comprises: S5: Based on rule reasoning technology and expert system, a dynamic optimization mechanism is constructed to automatically optimize the generation and update of initial audit rules, and combined with reinforcement learning algorithms, the audit rules can be adaptively adjusted to cope with changes in new business scenarios.

8. An intelligent fee audit system based on artificial intelligence big model, characterized in that: The system implements intelligent audit of quantity and fee based on the intelligent audit method of quantity and fee based on the artificial intelligence big model as described in any one of claims 1 to 7, including a data acquisition module, a data extraction and fusion module, a relationship modeling and anomaly detection module, and a model training and quantity and fee audit module; The data collection module is used to automatically collect and organize user billing files, user behavior logs, algorithm rules, exception review rules, historical exception cases and expert knowledge using intelligent corpus collection technology and natural language processing technology to form structured and unstructured multimodal data sets; The data extraction and fusion module is used to extract key features using feature extraction technology after preprocessing the multimodal data set, and to fuse the multimodal data set with the extracted key features; The relationship modeling and anomaly detection module is used to establish a relationship model between users' electricity consumption behaviors and billing relationships based on multimodal data sets through graph neural networks, and is used to automatically detect and identify abnormal quantity and fee issues; The quantity and fee audit module is used to establish an audit rule model, train the graph neural network output vector and the audit rule model based on the reinforcement learning method, and adjust the reinforcement learning parameters through hyperparameter optimization technology and back propagation algorithm to improve the accuracy of quantity and fee anomaly detection.

9. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the intelligent audit method for quantity and fee based on an artificial intelligence big model as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the intelligent audit method for quantity and fee based on an artificial intelligence big model as described in any one of claims 1 to 7.

Images