Risk early warning method and system suitable for intelligent auditing and storage medium

Abstract

The invention discloses a risk early warning method and system suitable for intelligent auditing and a storage medium, relates to the technical field of big data analysis, and solves the technical problem that highly refined risk detection is difficult to provide for an intelligent auditing system due to the fact that complex auditing data features are not considered in the prior art. The method comprises the following steps: preprocessing an economic benefit audit information data set stored in an audit database; performing secondary processing on the preprocessed economic benefit auditing information data set according to a principal component analysis method; constructing a GRU neural network prediction model; outputting a time sequence corresponding to the predicted principal component index according to the GRU neural network prediction model; constructing a risk early warning model in the GRU neural network prediction model by using a BR-GAN algorithm, and performing risk early warning on prediction results of different indexes output by the GRU neural network prediction model according to the risk early warning model; the auditing efficiency is improved, and the auditing accuracy is enhanced.

Description

Technical Field

[0001] The present invention belongs to the field of big data analysis, relates to principal component analysis technology, and specifically is a risk early warning method applicable to intelligent auditing. Background Art

[0002] Constructing a risk early warning method applicable to intelligent auditing is conducive to enhancing the intelligent auditing ability, making full use of modern information technology to carry out auditing, and improving the auditing quality and efficiency; as an emerging direction in the auditing field, intelligent auditing uses advanced technical means such as big data and artificial intelligence to achieve comprehensive, in-depth, and efficient auditing of audit objects; data mining, as an important part of intelligent auditing, is of great significance for improving auditing efficiency, reducing auditing risks, and enhancing auditing capabilities; at the same time, with the digital transformation of auditing work and the continuous improvement of the auditing supervision system, higher requirements are put forward for the quality, accuracy, and timeliness of audit data.

[0003] The prior art (CN116980162A) discloses a data detection method, device, equipment, medium, and program product for cloud auditing, which relates to fields such as artificial intelligence, cloud auditing, and maps, and application scenarios include but are not limited to the abnormal detection scenario of the log data of cloud auditing. The method includes: obtaining the historical log data of cloud auditing; performing scenario correlation analysis processing on the historical log data to determine the correlation relationship and corresponding correlation rules between the historical log data and the preset scenario types; where the scenario types are at least one of normal business types and abnormal business types; obtaining a trained cloud auditing detection model through the correlation rules, and based on the trained cloud auditing detection model, detecting the real-time log data of cloud auditing to determine the first probability of the correlation relationship between the real-time log data and the scenario type. However, the cloud auditing detection model of the prior art can only detect the occurred auditing abnormalities and risks, and cannot early warn and process the possible risk situations in intelligent auditing;

[0004] At the same time, the prior art only obtains the first probability of the correlation relationship through historical log data, without considering the problem that it is difficult to provide highly refined risk detection for the intelligent auditing system due to the complex characteristics of audit data and the complex correlation relationships between features.

[0005] The present invention provides a risk early warning method applicable to intelligent auditing to solve the above technical problems. Summary of the Invention

[0006] The present invention aims to solve at least one of the technical problems existing in the prior art; for this purpose, the present invention proposes a risk early warning method, system and storage medium applicable to intelligent auditing, which is used to solve the technical problem that the prior art only obtains the first probability of the association relationship through historical log data, and does not consider that due to the complexity of audit data characteristics, it is difficult to provide highly refined risk detection for the intelligent auditing system.

[0007] To achieve the above object, a first aspect of the present invention provides a risk early warning method applicable to intelligent auditing, including:

[0008] Collecting an economic benefit audit information data set through data interface technology;

[0009] Preprocessing the economic benefit audit information data set stored in the audit database; wherein, the preprocessing includes: missing value filling, data organization format conversion and data normalization;

[0010] Performing secondary processing on the preprocessed economic benefit audit information data set according to the principal component analysis method to obtain a secondary processed data set; constructing a GRU neural network prediction model;

[0011] Taking the secondary processed data set as the input of the GRU neural network prediction model, and outputting a time series corresponding to the predicted principal component index according to the GRU neural network prediction model;

[0012] Constructing a risk early warning model using the BR-GAN algorithm in the GRU neural network prediction model, and performing risk early warning on the prediction results of different indicators output by the GRU neural network prediction model according to the risk early warning model.

[0013] Preferably, the performing secondary processing on the preprocessed economic benefit audit information data set according to the principal component analysis method includes:

[0014] S310: Retrieving the preprocessed economic benefit audit information data as a sample x i , marking the number of samples as n, and marking the number of sample features as m, where each sample x i = {x i1 , x i2 , …, x im} is an m-dimensional vector;

[0015] S320: Marking the feature of the m-th column as x m , and calculating the mean value of each feature through the formula ;

[0016] Calculating the difference between each feature and the corresponding feature mean value to obtain the centralized data marked as x km ';

[0017] Through the formula The element σ in the covariance matrix is calculated i，j ; where, i is the first variable and j is the second variable;

[0018] Through the formula The covariance matrix E is calculated; where, the elements on the diagonal are the variances of the features, and the off-diagonal elements are the covariances between the features;

[0019] S330: Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues λ and corresponding eigenvectors a;

[0020] S340: Calculate the principal component contribution rate and the cumulative contribution rate;

[0021] S350: Arrange the principal component contribution rates in descending order. Through the formula F q = a 1q X 1 + a 2q X 2 + … + a pq X p The q-th principal component is calculated; select the first y principal components corresponding to the eigenvalues whose cumulative contribution rate exceeds the threshold; project the high-dimensional original data onto the selected principal components, eliminate the selected principal components, and use the data corresponding to the remaining principal components as the data of the low-dimensional representation; where, y ∈ [1, m], q = 1, 2, 3, …, m, …, p; q represents the number of all features in the dataset; both m and p are positive integers; m ≤ p.

[0022] In the present invention, by selecting several important principal components according to the principal component contribution rate and the cumulative contribution rate, the first several principal components will contain most of the information of the dataset, and the remaining principal components can be ignored to reduce the data dimension.

[0023] Preferably, the calculating of the principal component contribution rate and the cumulative contribution rate includes:

[0024] Retrieve the eigenvalues λ and eigenvectors a corresponding to the covariance matrix;

[0025] Through the formula The principal component contribution rate is calculated;

[0026] Through the formula The cumulative contribution rate is calculated.

[0027] In the present invention, by selecting the principal components corresponding to the eigenvalues whose cumulative contribution rate exceeds 80%, since the first several principal components will contain most of the information of the dataset, and the remaining principal components can be ignored, the data dimension is reduced.

[0028] Preferably, the constructing of the GRU neural network prediction model includes:

[0029] The activation function is calculated through the formula ;

[0030] The candidate state of the GRU neural network structure is calculated through the formula ;

[0031] The hidden state h is obtained through the formula t ;

[0032] where x t is the input value, h t-1 is the last output value of the hidden layer, W h and V h represent weights, and b h is the bias vector;

[0033] The state function of the reset gate R t is calculated through the formula R = α(W r ×x t +V r ×h t-1 +b r ); where W t and V r represent weights, and b r is the bias vector; r is the bias vector;

[0034] The state U of the update gate is calculated through the formula U t = α(W u ×x t +V u ×h t-1 +b u ); where W t and V u represent weights, and b u is the bias vector; u is the bias vector;

[0035] Record the state update process of the update gate and the reset gate into the GRU cell. Construct a GRU layer through several GRU cells. Receive the data processed in the model input data format through the GRU layer, and analyze and predict the results.

[0036] Through the GRU neural network, the present invention can automatically adjust and learn the implicit relationship between time series by calculating the relevant weight coefficients of the hidden layer. Compared with the long short-term memory (LSTM) neural network, the structure is simpler, contains fewer parameters, and is easier to train; the selection, forgetting, and update of feature information are jointly completed through the update gate and the reset gate.

[0037] Preferably, the time series corresponding to the predicted principal component indicators output according to the GRU neural network prediction model includes:

[0038] S410: Retrieve the dataset after secondary processing as the input of the GRU neural network, and divide the processed dataset into a training set, a validation set, and a test set;

[0039] S420: Build a GRU neural network prediction model, use the mean squared error as the loss function of the model; use the Adam optimization algorithm as the model optimizer, and optimize the model multiple times according to the model evaluation results;

[0040] S430: Use the training dataset to train the GRU model, set hyperparameters, and adjust the hyperparameters through the validation set to optimize the model performance;

[0041] S440: Use the optimized GRU neural network model to predict the future time series of the selected principal component indicator data in an iterative manner.

[0042] The present invention performs predictive analysis on the principal component indicators highly correlated with the benefit risk anomalies in the economic benefit audit information dataset according to the GRU neural network, and outputs the future time series of different indicators.

[0043] Preferably, the risk warning model is constructed using the BR-GAN algorithm in the GRU neural network prediction model, including:

[0044] S510: Retrieve the multivariate time series of different indicators output by the GRU neural network prediction model, use a sliding window to split the continuous current and predicted multivariate time series into multiple subsequences of length T, and randomly divide the subsequences into a training dataset and a test dataset;

[0045] S520: Use a bidirectional recursive generative adversarial network to build a BR-GAN anomaly detection model;

[0046] S530: Use the training dataset to train the BR-GAN anomaly detection model, and use the Adam optimizer to optimize the loss function and the objective function;

[0047] S540: Use 5-fold cross-validation for the test set, use 4-fold datasets, select the anomaly threshold by maximizing the F1 score to test the remaining 1-fold data, and generate a risk warning model.

[0048] The present invention uses a standard long short-term recursive neural network structure through the BR-GAN anomaly detection model, which is beneficial to capturing the correlation of time series; evaluating the performance of the machine learning model through the 5-fold cross-validation method is beneficial to quickly selecting the anomaly threshold of the multivariate time series.

[0049] Preferably, the BR-GAN anomaly detection model built using a bidirectional recursive generative adversarial network includes:

[0050] Retrieve the encoder, decoder, and discriminator, and mark the generator network as the first sub-network as G; among them, the first sub-network includes the encoder G E and the decoder G D ; The generator network learns the input data representation and reconstructs the input time series by using the encoder and decoder respectively;

[0051] Mark the encoder network as the second sub-network as E, and compress the time series reconstructed by network G;

[0052] Mark the discriminator network as the third sub-network as D, and classify the original input data x and the generated by the generator network G as real or fake respectively;

[0053] The encoder, decoder, and discriminator all use a bidirectional recursive generative adversarial network. The bidirectional recursive generative adversarial network maps the sample space to the latent variable space through an encoder-decoder-encoder structure, and simultaneously learns the distributions of samples in the original data space and the latent variable space;

[0054] Construct the BR-GAN anomaly detection model through three sub-networks.

[0055] In the present invention, by simultaneously performing the mapping process and the process of learning the distributions of samples in the original data space and the latent variable space, it is beneficial to avoid the problem of excessive time consumption in gradient backpropagation inference; by adopting a standard long short-term recursive neural network structure, it is convenient to capture the correlation of time series;

[0056] Preferably, training the BR-GAN anomaly detection model using the training dataset includes:

[0057] Divide the objective function for training the BR-GAN anomaly detection model into a generator objective function and a discriminator objective function. Among them, the additional encoder E and the encoder G in the generator E and the decoder G D are jointly trained;

[0058] Through the formula Calculate the objective function L of the final generator for joint training G ; where α 1 , α 2 , α 3 are the weights of the loss functions of the three sub-networks, x is the original input, G(x) is the input generated by the discriminator, f θThe function value of () is the value of the output unit of the recurrent neural network in the LSTM-RNN model. represents the loss function;

[0059] The objective function of the discriminator is calculated through the formula ; The BR-GAN anomaly detection model is constructed through the objective function of the final generator and the objective function of the discriminator.

[0060] To achieve the above object, the second aspect of the present invention provides a risk warning system applicable to intelligent auditing, including:

[0061] Data acquisition module: Collect the economic benefit audit information dataset through data interface technology;

[0062] Data processing module: Preprocess the economic benefit audit information dataset stored in the audit database; wherein, the preprocessing includes: missing value filling, data organization format conversion, and data normalization;

[0063] Perform secondary processing on the preprocessed economic benefit audit information dataset according to the principal component analysis method to obtain the secondarily processed dataset;

[0064] Model construction module: Construct a GRU neural network prediction model; Use the secondarily processed dataset as the input of the GRU neural network prediction model, and according to the GRU neural network prediction model, output the time series corresponding to the predicted principal component indicators;

[0065] Use the BR-GAN algorithm to construct a risk warning model in the GRU neural network prediction model, and perform risk warning on the prediction results of different indicators output by the GRU neural network prediction model according to the risk warning model.

[0066] To achieve the above object, the third aspect of the present invention provides a risk warning storage medium applicable to intelligent auditing, on which a computer-readable storage medium is stored, and when the computer-readable storage medium is executed by a processor, it implements a risk warning method applicable to intelligent auditing as described in the first aspect above.

[0067] Compared with the prior art, the beneficial effects of the present invention are:

[0068] The present invention determines a multivariate time series data mining and risk early warning method for economic benefit audit, which not only has a high risk prediction accuracy, strong stability and flexibility, but also can predict the risk situation of economic benefit audit in advance and make corresponding treatments in advance to ensure the safety and stability of intelligent audit. After analyzing and predicting the principal component indicators based on the GRU neural network, the present invention provides a quantitative measure, anomaly detection and risk early warning for the development trend of each principal component indicator. The present invention introduces a bidirectional recursive generative adversarial network algorithm to analyze the audit data set after data preprocessing and principal component analysis dimensionality reduction and the future multivariate time series output by the GRU neural network, screens out abnormal sequences, and completes data mining, anomaly detection and risk early warning for economic benefit audit. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0070] Figure 1 It is a schematic diagram of the specific steps for risk early warning of the present invention;

[0071] Figure 2 It is a schematic diagram of the specific process for data set processing of the present invention;

[0072] Figure 3 It is a schematic diagram of the specific process for time series prediction of the present invention;

[0073] Figure 4 It is a schematic diagram of the module relationship included in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0074] The following will clearly and completely describe the technical solutions of the present invention in combination with the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0075] Please refer to Figure 1 , an embodiment of the first aspect of the present invention provides a risk early warning method applicable to intelligent audit, including:

[0076] Collect the economic benefit audit information data set through the data interface technology;

[0077] Preprocess the economic benefit audit information dataset stored in the audit database; wherein, the preprocessing includes: missing value completion, data organization format conversion, and data normalization;

[0078] Perform secondary processing on the preprocessed economic benefit audit information dataset according to the principal component analysis method to obtain a secondary processed dataset; construct a GRU neural network prediction model;

[0079] Use the secondary processed dataset as the input of the GRU neural network prediction model, and output the time series corresponding to the predicted principal component indicators according to the GRU neural network prediction model;

[0080] Construct a risk warning model using the BR-GAN algorithm in the GRU neural network prediction model, and perform risk warning on the prediction results of different indicators output by the GRU neural network prediction model according to the risk warning model.

[0081] Please refer to Figure 2 for the specific process of dataset processing, and preprocess the economic benefit audit information dataset stored in the audit database; wherein, the preprocessing includes: missing value completion, data organization format conversion, and data normalization;

[0082] S310: Retrieve the preprocessed economic benefit audit information data as sample x i , mark the number of samples as n, and mark the number of sample features as m, where each sample x i ={x i1 , x i2 , …, x im} is an m-dimensional vector;

[0083] S320: Mark the feature in the m-th column as x m , and calculate the mean value of each feature through the formula ;

[0084] Calculate the difference between each feature and the corresponding feature mean value to obtain the centralized data marked as x km ';

[0085] Calculate the element σ in the covariance matrix through the formula i，j ; where i is the first variable and j is the second variable;

[0086] Calculate the covariance matrix E through the formula ; where the elements on the diagonal are the variances of the features, and the off-diagonal elements are the covariances between the features;

[0087] S330: Perform eigenvalue decomposition on the covariance matrix to obtain the eigenvalues λ and the corresponding eigenvectors a;

[0088] S340: Calculate the contribution rate and cumulative contribution rate of the principal components;

[0089] S350: Arrange the contribution rates of the principal components in descending order. Through the formula F i = a 1i X 1 + a 2i X 2 + … + a pi X p Calculate the i-th principal component; select the first, second, …, m-th principal components corresponding to the eigenvalues whose cumulative contribution rate exceeds 80%; project the high-dimensional original data onto the selected principal components, remove the selected principal components, and use the data corresponding to the remaining principal components as the data of the low-dimensional representation; where i = 1, 2, 3, …, m, …, p; i represents the total number of all features in the dataset; both m and p are positive integers; m ≤ p.

[0090] For example, this method preferably uses the PostgreSQL relational database and the Hive non-relational database to store audit data; adopts the median filling method for missing data processing. First, select a certain index data that needs to be processed for missing values, traverse and calculate the median of this index data, and then use the median to fill the missing positions of this index data; when the data processing of all indexes is completed, the processing of the missing values of the original data is completed; after the original data is processed for missing values, then convert the data organization format to meet the requirements of model training; first obtain the dataset after missing value processing, define the file organization format for storing the original data as the form of storing the date and various index data in a single row; select the required audit dataset, traverse each index data value within a period of time row by row, write the date and time corresponding to the index data value, and each index data value into the file, and find the maximum and minimum values of the corresponding index data; traverse this index data again, and perform normalization calculation on each value according to the following formula;

[0091] Through the formula for Perform normalization processing on each index data in the dataset;

[0092] where Z is the current data, Zmax is the maximum value in the current index data, Zmin is the minimum value in the current index data, is the value after data normalization processing;

[0093] According to the principal component analysis method, perform secondary processing on the preprocessed economic benefit audit information dataset to obtain the dataset after secondary processing.

[0094] It should be noted that since the values of each index are distributed within a limited range, this method uses the deviation normalization method for data normalization processing.

[0095] Before collecting the economic benefit audit information data, the present invention processes the missing values, data format conversion, and data normalization of the original economic benefit audit information data, thereby improving the quality of the training data and making the training effect of the subsequent model more accurate; by using principal component analysis, some of the most important features of the high-dimensional data are retained, and the noise and unimportant features are removed, thereby achieving an improvement in the data processing speed.

[0096] Please refer to Figure 3 , for the specific process of time series prediction, through the formula to calculate the activation function;

[0097] Through the formula to calculate the candidate state of the GRU neural network structure;

[0098] Through the formula to obtain the hidden state h t ;

[0099] Wherein, x t is the input value, h t-1 is the last output value of the hidden layer, W h and V h represent weights, b h is the bias vector;

[0100] Through the formula R t = α (W r × x t + V r × h t-1 + b r ) to calculate the state function of the reset gate R t ; wherein, W r and V r represent weights, b r is the bias vector;

[0101] Through the formula U t = α (W u × x t + V u × h t-1 + b u ) to calculate the state U t of the update gate; wherein, W u and V u represent weights, b u is the bias vector;

[0102] Record the state update process of the update gate and the reset gate into the GRU unit, construct a GRU layer through several GRU units, and receive the data processed by the model input data format through the GRU layer, and analyze and predict the results;

[0103] S410: Retrieve the dataset after secondary processing as the input of the GRU neural network, and divide the processed dataset into a training set, a validation set, and a test set;

[0104] S420: Build a GRU neural network prediction model, use the mean squared error as the loss function of the model; use the Adam optimization algorithm as the model optimizer, and optimize the model multiple times according to the model evaluation results;

[0105] S430: Use the training dataset to train the GRU model, set hyperparameters, and adjust the hyperparameters through the validation set to optimize the model performance;

[0106] S440: Use the optimized GRU neural network model to predict the future time series of the selected principal component index data in an iterative manner.

[0107] For example, according to the training requirements of the built GRU neural network prediction model, use the principal component index data of each day from the 1st to the 20th in the audit information dataset as the features of the model input data, and the principal component index data of the 21st day as the label; then use the principal component index data of each day from the 2nd to the 21st in the audit information data as the features of the model input data, and the target index data of the 22nd day as the label, and so on; perform prediction in an iterative manner. After converting the audit data after data preprocessing and dimensionality reduction by principal component analysis according to the supervised learning mode, input it into the GRU model for multivariate time series prediction, obtain the time series of each future principal component index and record it; then add the actual output result to the input features of the next model training, and use it together with other input features as the training sample to predict the future multivariate time series again; and so on, to obtain the data mining and prediction results of the economic benefit audit information in intelligent audit using the GRU neural network.

[0108] The present invention is beneficial to achieving the purpose of supervised learning by continuously learning the features and label data in the sample data through the process of continuously learning the prediction model.

[0109] Please refer to Figure 4 , the second aspect of the present invention provides a risk warning system applicable to intelligent audit, including: a data acquisition module, a data processing module, and a model construction module;

[0110] Data acquisition module: Collect the economic benefit audit information dataset through data interface technology;

[0111] Data processing module: Preprocess the economic benefit audit information dataset stored in the audit database; wherein, the preprocessing includes: missing value filling, data organization format conversion, and data normalization;

[0112] Perform secondary processing on the preprocessed economic benefit audit information data set according to the principal component analysis method to obtain the secondary processed data set;

[0113] Model construction module: Construct a GRU neural network prediction model; Use the secondary processed data set as the input of the GRU neural network prediction model, and according to the GRU neural network prediction model, output the time series corresponding to the predicted principal component indicators;

[0114] Use the BR-GAN algorithm to construct a risk warning model in the GRU neural network prediction model, and perform risk warning on the prediction results of different indicators output by the GRU neural network prediction model according to the risk warning model.

[0115] The third aspect of the present invention provides a risk warning storage medium suitable for intelligent auditing, on which a computer-readable storage medium is stored. When the computer-readable storage medium is executed by a processor, it implements a risk warning method for intelligent auditing as described in the first aspect above.

[0116] Some of the data in the above formula is the numerical value after removing the dimension. The formula is obtained by software simulation of a large amount of collected data to get a formula closest to the actual situation; The preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

[0117] The working principle of the present invention: The present invention collects the economic benefit audit information data set through the data interface technology; Preprocess the economic benefit audit information data set stored in the audit database; Perform secondary processing on the preprocessed economic benefit audit information data set according to the principal component analysis method to obtain the secondary processed data set; Construct a GRU neural network prediction model; Use the secondary processed data set as the input of the GRU neural network prediction model, and according to the GRU neural network prediction model, output the time series corresponding to the predicted principal component indicators; Use the BR-GAN algorithm to construct a risk warning model in the GRU neural network prediction model, and perform risk warning on the prediction results of different indicators output by the GRU neural network prediction model according to the risk warning model.

[0118] The above embodiments are only used to illustrate the technical method of the present invention and not to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical method of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A risk warning method suitable for smart auditing, characterized in that: include: Collect economic benefit audit information data sets through data interface technology; Preprocessing the economic benefit audit information data set stored in the audit database; the preprocessing includes: missing value completion, data organization format conversion and data normalization; According to the principal component analysis method, the pre-processed economic benefit audit information data set is processed again to obtain the secondary processed data set; a GRU neural network prediction model is constructed; The secondary processed data set is used as the input of the GRU neural network prediction model, and the time series corresponding to the main component index is predicted according to the output of the GRU neural network prediction model; The BR-GAN algorithm is used in the GRU neural network prediction model to build a risk warning model, and risk warning is performed on the prediction results of different indicators output by the GRU neural network prediction model according to the risk warning model.

2. According to claim 1, a risk early warning method suitable for smart auditing is characterized in that: The secondary processing of the pre-processed economic benefit audit information data set according to the principal component analysis method includes: S310: Retrieve the pre-processed economic benefit audit information data as sample x i , mark the number of samples as n, and the number of sample features as m, where each sample x i ={x i1 , x i2 , …, x im } is an m-dimensional vector; S320: Mark the feature of the mth column as x m , through the formula Calculate the mean of each feature; Calculate the difference between each feature and the corresponding feature mean, and get the centralized data marked as x km '; By formula Calculate the element σ in the covariance matrix i，j ; Where i is the first variable and j is the second variable; By formula The covariance matrix E is calculated, where the elements on the diagonal are the variances of the features and the off-diagonal elements are the covariances between the features. S330: Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalue λ and corresponding eigenvector a; S340: Calculate the principal component contribution rate and cumulative contribution rate; S350: Arrange the principal component contribution rates in descending order, and use formula F q =a 1q X1+a 2q X2+…+a pq X p Calculate the qth principal component; select the first y principal components corresponding to the eigenvalues ​​whose cumulative contribution rate exceeds the threshold; project the high-dimensional original data to the selected principal components, remove the selected principal components, and use the data corresponding to the remaining principal components as the data represented in the low dimension; where y∈[1,m], q=1,2,3,…,m,…,p; q represents the number of all features in the data set; m and p are both positive integers; m≤p.

3. A risk warning method suitable for smart auditing according to claim 2, characterized in that: The calculation of the principal component contribution rate and the cumulative contribution rate includes: Retrieve the eigenvalue λ and eigenvector a corresponding to the covariance matrix; By formula Calculate the principal component contribution rate; By formula Calculate the cumulative contribution rate.

4. A risk warning method suitable for smart auditing according to claim 1, characterized in that: The construction of the GRU neural network prediction model includes: By formula Calculate the activation function; By formula Calculate the candidate states of the GRU neural network structure; By formula Get the hidden state h t ; Among them, x t is the input value, h t-1 is the last output value of the hidden layer, W h and V h represents the weight, b h is the bias vector; By formula R t =α(W r × t +V r ×h t-1 +b r ) Calculate the reset gate R t The state function of r and V r represents the weight, b r is the bias vector; By formula U t =α(W u × t +V u ×h t-1 +b u ) Calculate the state U of the update gate t ; Among them, W u and V u represents the weight, b u is the bias vector; The state update process of the update gate and the reset gate is recorded in the GRU unit, and a GRU layer is constructed through several GRU units. The data processed by the model input data format is received through the GRU layer, and the data is analyzed and the results are predicted.

5. According to claim 1, a risk early warning method suitable for smart auditing is characterized in that: The time series corresponding to the main component index outputted by the GRU neural network prediction model includes: S410: Retrieving the data set after secondary processing as the input of the GRU neural network, and dividing the processed data set into a training set, a validation set, and a test set; S420: Build a GRU neural network prediction model, use mean square error as the model's loss function; use the Adam optimization algorithm as the model optimizer, and optimize the model multiple times based on the model evaluation results; S430: train the GRU model using the training data set, set hyperparameters, and adjust the hyperparameters through the validation set to optimize model performance; S440: Use the optimized GRU neural network model to iteratively predict the future time series of the corresponding indicator for the selected principal component indicator data.

6. A risk warning method suitable for smart auditing according to claim 1, characterized in that: The GRU neural network prediction model uses the BR-GAN algorithm to build a risk warning model, including: S510: Retrieve the multivariate time series of different indicators output by the GRU neural network prediction model, use a sliding window to split the continuous current and predicted multivariate time series into multiple subsequences of length T, and randomly divide the subsequences into a training data set and a test data set; S520: Use bidirectional recursive generative adversarial network to build BR-GAN anomaly detection model; S530: training the BR-GAN anomaly detection model using the training data set, and using the Adam optimizer to optimize the loss function and the objective function; S540: Use 5-fold cross validation on the test set, use 4-fold data set, select the abnormal threshold by maximizing the F1 score, and use it to test the remaining 1-fold data and generate a risk warning model.

7. A risk warning method suitable for smart auditing according to claim 6, characterized in that: The use of a bidirectional recursive generative adversarial network to build a BR-GAN anomaly detection model includes: Call the encoder, decoder and discriminator, and mark the generator network as the first subnetwork as G; the first subnetwork contains the encoder G E and decoder G D ;The generator network learns the input data representation and reconstructs the input time series by using the encoder and decoder respectively; The encoder network is labeled as the second sub-network E, which compresses the time series reconstructed by the network G; The discriminator network is the third sub-network labeled D, which combines the original input data x and the generator network G Classify as genuine or forged respectively; The encoder, decoder and discriminator all use a bidirectional recursive generative adversarial network, which maps the sample space to the latent variable space through an encoding-decoding-encoding structure, and simultaneously learns the distribution of samples in the original data space and the latent variable space; The BR-GAN anomaly detection model is constructed through three sub-networks.

8. A risk warning method suitable for smart auditing according to claim 6, characterized in that: The method of using the training data set to train the BR-GAN anomaly detection model includes: The objective function of the BR-GAN anomaly detection model training is divided into the generator objective function and the discriminator objective function, where the additional encoder E is combined with the encoder G in the generator E and decoder G D training together; By formula Calculate the objective function L of the final generator trained together G ; Among them, α1, α2, α3 are the weights of the three sub-network loss functions, x is the original input, G(x) is the input generated by the discriminator, and f θ The function value of () is the value of the output unit of the recurrent neural network in the LSTM-RNN model. represents the loss function; By formula The objective function of the discriminator is calculated; the BR-GAN anomaly detection model is constructed through the objective function of the final generator and the objective function of the discriminator.

9. A risk warning system suitable for smart auditing, adapted to the risk warning method suitable for smart auditing described in claims 1-8, characterized in that: include: Data acquisition module, data processing module, model building module; Data collection module: collects economic benefit audit information data sets through data interface technology; Data processing module: preprocessing the economic benefit audit information data set stored in the audit database; the preprocessing includes: missing value completion, data organization format conversion and data normalization; According to the principal component analysis method, the economic benefit audit information data set after preprocessing is processed again to obtain the second processed data set; Model building module: build a GRU neural network prediction model; use the secondary processed data set as the input of the GRU neural network prediction model, and output the prediction of the time series corresponding to the principal component index according to the GRU neural network prediction model; The BR-GAN algorithm is used in the GRU neural network prediction model to build a risk warning model, and risk warning is performed on the prediction results of different indicators output by the GRU neural network prediction model according to the risk warning model.

10. A risk warning storage medium suitable for smart auditing, on which a computer-readable storage medium is stored, characterized in that: When the computer-readable storage medium is executed by a processor, it implements a risk warning method suitable for smart auditing as described in any one of claims 1-8.

Images