A method, medium and system for extracting audit features from massive financial and accounting data fusion

Abstract

The present invention provides a method, medium and system for extracting audit features from the fusion of massive financial and accounting data, belonging to the technical field of digital data feature extraction. After determining a basic time window through time series clustering, the present invention constructs an independent time window function to dynamically adjust the window boundary according to parameters such as the business fluctuation coefficient, utilizes a financial and accounting time series feature recognition model to extract time dimension features, simultaneously extracts spatial dimension features based on geographic distribution and organizational structure information, extracts business dimension features according to business type and transaction attributes, calculates outlier features through statistical analysis, establishes a second time window for verification, adopts a multi-objective optimization weight determination model to solve the optimal weight coefficient, and performs weighted fusion on the four types of features to form a comprehensive audit feature vector. The present invention solves the technical problem of insufficient audit feature extraction accuracy caused by the inability to dynamically adapt to the internal change characteristics of the data when extracting features from massive financial and accounting data.

Description

Technical Field

[0001] The present invention belongs to the technical field of digital data feature extraction, and in particular relates to a method, medium and system for extracting audit features from massive financial and accounting data fusion. Background Art

[0002] Audit feature extraction from accounting data is a core technology in audit informatization. Traditional methods primarily employ fixed-time window analysis, static feature extraction, and single-dimensional analysis. In applications such as corporate financial management, accounting firm audits, financial institution risk control, and government fiscal oversight, existing technologies typically segment accounting data based on preset fixed time periods, extract time series features using statistical analysis, and identify abnormal patterns through rule matching. However, traditional fixed-window methods suffer from significant drawbacks. They are unable to adapt to the irregular patterns of accounting data, resulting in inaccurate localization of key abnormal time points. Static feature extraction methods lack the ability to perceive dynamic changes within the data and struggle to capture the complex, time-varying characteristics of accounting data. Single-dimensional analysis ignores the correlations among multi-dimensional features such as time, space, and business, resulting in insufficient feature representation. Traditional technologies struggle to address the core technical issue of insufficient accuracy in extracting audit features from massive amounts of accounting data due to their inability to dynamically adapt to internal variations in the data. When faced with complex and volatile accounting data, existing methods often produce inaccurate feature extraction results, impacting the accuracy and reliability of subsequent audit analysis. Summary of the Invention

[0003] In view of this, the present invention provides a method, medium and system for extracting audit features by fusing massive accounting data, which can solve the technical problem in the prior art that the extraction of features from massive accounting data cannot dynamically adapt to the internal changing characteristics of the data, resulting in insufficient accuracy of audit feature extraction.

[0004] The present invention is implemented as follows: The first aspect of the present invention provides a method for extracting audit features from the fusion of massive accounting data, including: using a time series clustering algorithm to determine the length of a first time window, dividing historical accounting data into multiple first time windows based on the periodic characteristics of the accounting data; constructing an independent time window function, inputting the accounting data, business fluctuation coefficient, time granularity parameter, data change rate threshold and window overlap within the first time window, and outputting an independent time window sequence for dynamically adjusting the window boundary to adapt to the internal change characteristics of the accounting data; using a pre-trained accounting time series feature recognition model to perform deep learning processing on the independent time window sequence, identify long-term trend patterns and periodic change laws in the accounting data, and output time dimensions. Features; extract spatial dimension features based on the geographical distribution information and organizational structure information of the accounting data in the independent time window sequence; extract business dimension features based on the business type and transaction attributes of the accounting data in the independent time window sequence; calculate the outlier of the accounting data in the independent time window sequence, identify the outliers corresponding to abnormal accounting behavior through statistical analysis methods, and form outlier features; construct a multi-objective optimization weight determination model, with time dimension features, spatial dimension features, business dimension features and outlier features as input variables, and solve the optimal weight coefficient through a multi-objective optimization algorithm; use the optimal weight coefficient to weightedly fuse the time dimension features, spatial dimension features, business dimension features and outlier features to form a comprehensive audit feature vector.

[0005] Among them, the first time window is a fixed time period determined according to the natural periodicity of accounting data. The length is automatically determined by analyzing the similarity patterns of historical accounting data through a clustering algorithm. It is used to capture the basic periodic characteristics of accounting data, usually corresponding to the standard reporting cycle or settlement cycle of accounting business, and provides a stable time reference framework for subsequent feature extraction.

[0006] Among them, the independent time window sequence is specifically a set of variable-length time periods further subdivided on the basis of the first time window. The window boundary is adaptively adjusted according to the internal change rate of the accounting data, which is used to accurately locate the time nodes of abnormal changes. It can dynamically adapt to the irregular change pattern of the accounting data and improve the accuracy of feature extraction.

[0007] Among them, the independent time window function is used to dynamically adjust the boundary position and length of the time window according to the real-time change characteristics of the accounting data, so as to adapt to the fluctuation pattern and change law of the accounting data in different business scenarios. The output is an independent time window sequence after adaptive adjustment. Each window is marked with the starting time point, end time point and statistical characteristics of the data in the window.

[0008] Among them, the business volatility coefficient is a quantitative indicator that reflects the changes in the intensity of accounting business activities. It is calculated by analyzing the amplitude and frequency characteristics of business volume changes in historical accounting data, and is used to guide the dynamic adjustment process of the independent time window function.

[0009] Among them, the specific structure of the accounting time series feature recognition model is a deep neural network designed based on the long short-term memory network architecture, which contains multi-layer LSTM units to process the long-term dependencies of time series data. The entire network structure enhances the recognition ability of key time nodes and important accounting events through residual connections and attention mechanisms.

[0010] Among them, the multi-objective optimization weight determination model is used to solve the optimal weight distribution of time dimension features, space dimension features, business dimension features and outlier features in the comprehensive audit feature vector. The multi-objective genetic algorithm is used as the core optimization engine, and the audit accuracy, recall rate, F1 score and feature density after feature fusion are used as optimization objective functions.

[0011] Among them, after the step of constructing the independent time window function, it also includes establishing a second time window for verifying the feature extraction effect. The length of the second time window is different from the length of the first time window. Repeat the step of using the accounting time series feature recognition model to extract time dimension features to form outlier features to obtain a verification feature set.

[0012] Among them, the outliers are specifically abnormal accounting data points identified through statistical analysis methods. Their numerical characteristics obviously deviate from the normal range, indicating potential audit risks or abnormal accounting behaviors. The abnormality degree and risk level of the data points are determined through multi-dimensional distance calculation and probability distribution analysis.

[0013] Among them, the outlier degree is specifically a numerical indicator that quantifies the degree of abnormality of each data point. It is obtained by calculating the statistical distance between the data point and the center of the normal distribution. The larger the value, the higher the degree of abnormality. The Mahalanobis distance or kernel density estimation method is used for precise quantification to support subsequent risk assessment and audit decisions.

[0014] Among them, the feature density is specifically a numerical indicator that quantifies the density of the distribution of non-zero elements in the comprehensive audit feature vector. It is obtained through the feature density calculation function, and the ratio of the number of non-zero elements in the feature vector to the total number of elements is calculated, and a weighted calculation is performed based on the numerical size and distribution uniformity of the non-zero elements.

[0015] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions. When the program instructions are run in a computer, the program instructions are used to execute the above-mentioned method for extracting audit features from the fusion of massive financial and accounting data.

[0016] The third aspect of the present invention provides a massive accounting data fusion audit feature extraction system, which includes the above-mentioned computer-readable storage medium. The system is any one of a computer, a server, and a single-chip microcomputer. The computer-readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing the program instructions stored in the computer-readable storage medium.

[0017] The present invention adapts to the internal variation characteristics of accounting data by dynamically adjusting the time window boundaries. The method can adjust the window length and position in real time according to parameters such as the business fluctuation coefficient and the data change rate threshold, and accurately locate the time nodes of abnormal changes. The present invention effectively solves the limitations of the traditional fixed window method. The self-varying time window function can adaptively adjust the boundaries to match the irregular change pattern of the data, significantly improving the positioning accuracy of abnormal time nodes. The multi-dimensional feature fusion mechanism overcomes the shortcomings of single-dimensional analysis and constructs a more comprehensive feature expression system through the comprehensive processing of time dimension, space dimension, business dimension and outlier characteristics. The pre-trained accounting time series feature recognition model enhances the recognition ability of complex time-varying features, and the multi-objective optimization weight determination model realizes the optimal fusion of features of each dimension; it solves the technical problem that the internal variation characteristics of data cannot be dynamically adapted when extracting features from massive accounting data. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0020] like Figure 1 FIG. 1 is a flowchart of a method for extracting audit features from massive accounting data fusion provided by the first aspect of the present invention. The method comprises the following steps:

[0021] A method for extracting audit features by fusing massive financial and accounting data includes the following steps:

[0022] S01. Determine the length of the first time window using a time series clustering algorithm, and divide the historical accounting data into multiple first time windows based on the periodic characteristics of the accounting data;

[0023] S02. Construct an independent time window function, input the accounting data within the first time window, the business fluctuation coefficient, the time granularity parameter, the data change rate threshold, and the window overlap, and output an independent time window sequence for dynamically adjusting the window boundary to adapt to the internal variation characteristics of the accounting data;

[0024] S03. Using a pre-trained accounting time series feature recognition model to perform deep learning processing on the independent time window sequence, identify long-term trend patterns and periodic change patterns in the accounting data, and output time dimension features;

[0025] S04. Extracting spatial dimension features based on the geographical distribution information and organizational structure information of the accounting data in the independent time window sequence;

[0026] S05. Extracting business dimension features based on the business types and transaction attributes of the accounting data in the independent time window sequence;

[0027] S06. Calculate the outlier degree of the accounting data in the independent time window sequence, identify outliers corresponding to abnormal accounting behaviors through statistical analysis methods, and form outlier degree features;

[0028] S07: Establish a second time window for verifying the feature extraction effect, where the length of the second time window is different from the length of the first time window, and repeat steps S02 to S06 to obtain a verification feature set;

[0029] S08. Construct a multi-objective optimization weight determination model, using the time dimension feature, space dimension feature, business dimension feature, and outlier feature as input variables, and solve the optimal weight coefficient through a multi-objective optimization algorithm;

[0030] S09. Using the optimal weight coefficient, weighted fusion is performed on the time dimension feature, space dimension feature, business dimension feature, and outlier feature to form a comprehensive audit feature vector;

[0031] S10. Output the comprehensive audit feature vector, which includes time trend features, periodic pattern features, abnormal behavior features, and multi-dimensional correlation features.

[0032] Among them, the first time window is a fixed time period determined according to the natural periodicity of accounting data. The length is automatically determined by analyzing the similarity patterns of historical accounting data through a clustering algorithm. It is used to capture the basic periodic characteristics of accounting data, usually corresponding to the standard reporting cycle or settlement cycle of accounting business, providing a stable time reference framework for subsequent feature extraction.

[0033] Among them, the second time window is specifically the time period used for model verification. Its length is obtained by random change or adjustment according to a preset ratio, and is compared with the first time window for verification to ensure the stability and generalization ability of the feature extraction method. The effectiveness and consistency of the extracted features at different time scales are verified through a cross-validation mechanism.

[0034] Among them, the independent time window sequence is specifically a set of variable-length time periods that are further subdivided on the basis of the first time window. The window boundaries are adaptively adjusted according to the internal change rate of the accounting data, which is used to accurately locate the time nodes of abnormal changes. It can dynamically adapt to the irregular change pattern of the accounting data and improve the accuracy of feature extraction.

[0035] Among them, outliers are abnormal accounting data points identified through statistical analysis methods. Their numerical characteristics deviate significantly from the normal range, indicating potential audit risks or abnormal accounting behavior. The degree of abnormality and risk level of the data points are determined through multi-dimensional distance calculation and probability distribution analysis.

[0036] Among them, outlier is a numerical indicator that quantifies the degree of abnormality of each data point. It is obtained by calculating the statistical distance between the data point and the center of the normal distribution. The larger the value, the higher the degree of abnormality. Methods such as Mahalanobis distance or kernel density estimation are used for precise quantification to support subsequent risk assessment and audit decisions.

[0037] Among them, the business volatility coefficient is a quantitative indicator that reflects the changes in the intensity of accounting business activities. It is calculated by analyzing the amplitude and frequency characteristics of business volume changes in historical accounting data, and is used to guide the dynamic adjustment process of the independent time window function.

[0038] Among them, the time granularity parameter is a numerical parameter that controls the time resolution, determines the minimum time unit for accounting data analysis, is determined based on the actual needs of accounting business and the frequency of data collection, and affects the precision of the independent time window sequence.

[0039] Among them, the data change rate threshold is specifically a critical value used to determine whether the accounting data has changed significantly. It is determined by statistically analyzing the change distribution characteristics of historical accounting data. When the data change exceeds the threshold, the boundary adjustment mechanism of the independent time window function is triggered.

[0040] The window overlap is a proportional parameter that determines the degree of overlap between adjacent time windows. Its value range is between 0 and 1. It is used to balance data continuity and computational efficiency and affects the size of the intersection of adjacent windows in the independent time window sequence.

[0041] Among them, the time dimension feature is specifically a feature vector extracted from the time series of accounting data, which includes trend features, periodic features and seasonal features, and is identified and extracted from the independent time window sequence through the accounting time series feature recognition model.

[0042] Among them, the spatial dimension features are specifically feature vectors extracted based on the geographical distribution information and organizational structure information of accounting data, reflecting the differences in accounting activity patterns between different regions and departments, and are extracted from the geographical and organizational attribute data in the independent time window series.

[0043] Among them, the business dimension features are specifically feature vectors extracted based on the business types and transaction attributes of accounting data, reflecting the accounting behavior characteristics and patterns of different business categories, and are extracted from the business classifications and transaction records in the independent time window sequence.

[0044] Among them, the outlier feature is specifically a feature vector that quantifies the degree of abnormality of accounting data. It is formed by calculating the outlier of data points in an independent time window sequence and is used to identify potential audit risk points and abnormal accounting behaviors.

[0045] Among them, the optimal weight coefficient is specifically a combination of weight parameters obtained by solving the multi-objective optimization weight determination model, which is used to balance the importance of time dimension characteristics, space dimension characteristics, business dimension characteristics and outlier characteristics in the comprehensive audit feature vector.

[0046] Among them, the comprehensive audit feature vector is specifically the final feature representation formed by weighted fusion of time dimension features, space dimension features, business dimension features and outlier features according to the optimal weight coefficient, which includes time trend features, periodic pattern features, abnormal behavior features and multi-dimensional correlation features.

[0047] The independent time window function is used to dynamically adjust the boundary position and length of the time window according to the real-time change characteristics of the accounting data, so as to adapt to the fluctuation pattern and change law of the accounting data in different business scenarios. The input includes the original accounting data sequence in the first time window, the business fluctuation coefficient reflecting the change in the intensity of business activities, the time granularity parameter for controlling the time resolution, the data change rate threshold for judging significant changes in data, and the window overlap parameter for determining the degree of overlap between adjacent windows. The output is the independent time window sequence after adaptive adjustment, and each window is marked with the starting time point, the ending time point and the statistical characteristics of the data in the window.

[0048] The specific structure of the accounting time series feature recognition model is a deep neural network designed based on the long short-term memory network architecture, which includes multi-layer LSTM units for processing the long-term dependencies of time series data. The sequence length parameter of the network is adjusted according to the periodic characteristics of the accounting data to match the data length of different business cycles. The hidden state dimension parameter matches the complexity and feature dimension of the accounting data to ensure the expressiveness of the model. The temperature coefficient parameter is used to control the confidence distribution of the model output feature vector so that the model can give a reasonable probability assessment of uncertain accounting patterns. The entire network structure enhances the recognition ability of key time nodes and important accounting events through residual connections and attention mechanisms.

[0049] The steps of establishing the training data set of the accounting time series feature recognition model specifically include collecting a large amount of historical accounting data and classifying and labeling them according to business type and time period, converting the original accounting data into a numerical format suitable for neural network training through data cleaning and standardization, establishing a supervised learning label system including normal mode labels and abnormal mode labels based on known audit results and expert annotations, using sliding window technology to divide the continuous accounting time series data into multiple training samples, each sample contains an input sequence and a corresponding feature label, and expanding the diversity of training samples through data enhancement technology including time perturbation and numerical transformation to improve the generalization ability of the model. Finally, a large-scale training data set containing hundreds of thousands of labeled samples is constructed for model training and optimization.

[0050] The steps of training the accounting time series feature recognition model specifically include first using a large-scale accounting data set for pre-training to establish the model's ability to understand the basic patterns of accounting data, adjusting the sequence length parameter so that the model can process accounting time series data of different lengths and optimizing the hidden state dimension parameter to balance the model complexity and computational efficiency, using the temperature coefficient parameter to adjust the gradient update strategy during training so that the model can learn subtle changes in the accounting data, using the backpropagation algorithm and the adaptive learning rate scheduler to optimize the model parameters and preventing overfitting through regularization technology, continuously monitoring the performance of the model on the validation set during training and adjusting the training strategy according to the convergence of the loss function, and finally obtaining a model that can accurately identify accounting time series features through multiple rounds of iterative training for subsequent feature extraction tasks.

[0051] The multi-objective optimization weight determination model is used to solve the optimal weight distribution of time dimension features, space dimension features, business dimension features and outlier features in the comprehensive audit feature vector. It adopts a multi-objective genetic algorithm as the core optimization engine, and takes the audit accuracy, recall rate, F1 score and feature density after feature fusion as the optimization objective function. The optimal weight coefficient combination that balances multiple objectives is found through population evolution and Pareto optimal solution set search. The model input is the standardized time dimension features, space dimension features, business dimension features and outlier feature vector, and the output is the corresponding optimal weight coefficient. The sum of the weight coefficients is equal to 1 and each weight value is a non-negative number. The cross-validation mechanism is used to ensure that the solved optimal weight coefficient has good generalization performance and stability.

[0052] Among them, feature density is specifically a numerical indicator that quantifies the density of distribution of non-zero elements in the comprehensive audit feature vector, which is obtained through the feature density calculation function. The feature density calculation function takes the comprehensive audit feature vector as input, calculates the ratio of the number of non-zero elements in the feature vector to the total number of elements, and performs weighted calculation based on the numerical size and distribution uniformity of the non-zero elements, and outputs the feature density value. The larger the feature density value, the richer the effective information contained in the feature vector, which is conducive to improving the accuracy and reliability of audit analysis.

[0053] The specific implementation of the above steps is described in detail below.

[0054] The specific implementation method of step S01 is to first collect a large amount of historical accounting data as the basis for analysis, and convert the original accounting data into a standardized time series format through data preprocessing. A dynamic time warping algorithm is used to calculate the similarity distance between accounting data in different time periods. This algorithm can effectively deal with the time offset and length inconsistency problems existing in accounting data. Based on the similarity distance matrix, a hierarchical clustering algorithm is used to perform cluster analysis on the historical accounting data. During the clustering process, the similarity threshold is set to 0.85. When the similarity between data exceeds the threshold, they are classified into the same category. By analyzing the time span distribution characteristics of each category in the clustering results, the optimal first time window length is determined. Usually, this length is consistent with the natural periodicity of the accounting business, such as the monthly or quarterly reporting cycle. The determined first time window length is used to divide the historical accounting data to form multiple regular time periods. Each time period contains a relatively stable accounting business model, laying a time reference framework for subsequent feature extraction.

[0055] The specific implementation of step S02 is to construct an independent time window function based on an adaptive adjustment mechanism. This function uses the accounting data within the first time window as input. The business fluctuation coefficient is obtained by calculating the coefficient of variation of the business volume in the accounting data. The coefficient of variation is calculated using the ratio of the standard deviation to the mean. When the coefficient of variation exceeds 0.3, it indicates large business fluctuations. The time granularity parameter is determined based on the collection frequency of the accounting data, set to 24 hours for daily data and 1 hour for hourly data. The data change rate threshold is determined by analyzing the change distribution of historical accounting data and is set using a quantile method. The change rate corresponding to the 95th quantile is usually used as the threshold, with a reference value of 15%. The window overlap parameter controls the degree of overlap between adjacent time windows and is set between 0.1 and 0.5, with a reference value of 0.25. The function dynamically calculates the window boundary adjustment amount at each time point based on the input parameters. When the data change rate exceeds the threshold, the adaptive adjustment of the window boundary is triggered, and the output is an independent time window sequence containing multiple variable-length time periods.

[0056] The specific implementation method of step S03 is to use the pre-trained accounting time series feature recognition model to perform deep learning processing on the independent time window sequence. The model is based on the long short-term memory network architecture design and can effectively capture the long-term dependencies and complex patterns in the accounting data. The accounting data in the independent time window sequence is preprocessed according to the input format required by the model, including data standardization and sequence filling operations. The model performs feature learning on the input data through a multi-layer neural network structure to identify long-term trend changes, periodic fluctuations and seasonal adjustment patterns in the accounting data. The model output includes a time dimension feature set of trend feature vectors, period feature vectors and seasonal feature vectors. The dimension of each feature vector is determined according to the complexity of the accounting data and is usually set to 64 to 256 dimensions.

[0057] The specific implementation method of step S04 is to extract spatial dimension features based on the geographical identification and organizational structure identification of the accounting data in the independent time window sequence. First, the geographical distribution information in the accounting data is encoded, and the geographical location is converted into numerical coordinates using a geographical coding algorithm to establish a mapping relationship between the geographical location and the intensity of accounting activities. A tree structure encoding method is used for the organizational structure information to convert the hierarchical relationship into a numerical vector representation. The correlation between accounting activities in different geographical regions is calculated, and the degree of correlation is measured using the Pearson correlation coefficient method, with the correlation coefficient threshold set to 0.6. The differences in accounting behavior patterns between different organizational departments are analyzed, and department combinations with similar accounting characteristics are identified through cluster analysis. Output a spatial dimension feature vector containing geographical distribution characteristics and organizational association characteristics.

[0058] The specific implementation method of step S05 is to extract business dimension features based on the business classification identifier and transaction attribute identifier of the accounting data in the independent time window sequence. The unique hot encoding method is used to numerically represent different business types, and a mapping relationship between business types and feature vectors is established. A multi-dimensional encoding method is used for transaction attributes, including attribute dimensions such as transaction amount level, transaction frequency category and transaction object type. Statistical analysis methods are used to calculate the accounting behavior characteristic parameters of different business types, including average transaction amount, transaction frequency distribution and abnormal transaction ratio. Business correlation analysis is used to identify the intrinsic connection between different businesses, and an association rule mining algorithm is used to discover the potential rules in the business model. The minimum support threshold is set to 0.1, and the minimum confidence threshold is set to 0.7. Output a business dimension feature vector containing business type characteristics and transaction model characteristics.

[0059] The specific implementation method of step S06 is to calculate the quantitative index of the degree of outlier of each accounting data point in the independent time window sequence. The isolation forest algorithm is used to identify outliers in accounting data. The algorithm detects the degree of abnormality of data points by constructing a random partitioning tree, and the abnormal score threshold is set to 0.7. The Mahalanobis distance method is used to calculate the statistical distance between the data point and the center of the normal distribution. The larger the distance value, the higher the degree of abnormality. The distance threshold is set to 3 standard deviations. The kernel density estimation method is used to analyze the position of the data point in the probability distribution. The data points with a probability density lower than the 5% quantile are identified as outliers. The degree of abnormality of each data point relative to its neighborhood is calculated by the local outlier factor algorithm. The data points with an outlier factor greater than 1.5 are considered to be potential outliers. The results of multiple anomaly detection methods are comprehensively evaluated to form an outlier feature vector reflecting the degree of data abnormality.

[0060] The specific implementation method of step S07 is to establish a second time window with a length different from the first time window to verify the stability of the feature extraction method. The length of the second time window is determined by a random perturbation method, and the perturbation amplitude is set to between 10% and 30% of the length of the first time window. The second time window is applied to the same historical accounting data, and the feature extraction process of steps S02 to S06 is repeated to obtain the corresponding verification feature set. The feature similarity measurement method is used to compare the consistency of the extracted features under the two time windows, and the similarity coefficient threshold is set to 0.8. When the similarity coefficient is lower than the threshold, the feature extraction parameters need to be adjusted. The generalization ability of the feature extraction method at different time scales is evaluated through a cross-validation mechanism to ensure that the extracted features have good stability and reliability.

[0061] The specific implementation method of step S08 is to construct a weight optimization model based on a multi-objective genetic algorithm, which uses time dimension features, space dimension features, business dimension features and outlier features as input variables. Set audit accuracy, recall rate, F1 score and feature density as optimization objective functions, and find a weight coefficient combination that balances multiple objectives through Pareto optimal solution set search. The population size of the genetic algorithm is set to 100 individuals, the evolutionary generation is set to 200 generations, the crossover probability is set to 0.8, and the mutation probability is set to 0.1. Each individual represents a set of weight coefficients, the sum of the weight coefficients is equal to 1 and are all non-negative numbers. The comprehensive performance of each weight combination is evaluated by the fitness function, which comprehensively considers the weighted average of each objective function. After multiple rounds of evolutionary iterations, the optimal weight coefficient combination is output.

[0062] The specific implementation method of step S09 is to use the optimal weight coefficient obtained by solving step S08 to perform weighted fusion processing on the features of the four dimensions. First, the time dimension features, space dimension features, business dimension features and outlier features are standardized to eliminate the dimensional differences between different features. The weighted value of each dimension feature is calculated according to the optimal weight coefficient, and the weighted feature vectors are fused through linear combination. During the fusion process, feature dimensionality reduction technology is used to process high-dimensional feature vectors, and the principal component analysis method is used to retain the main components with a cumulative contribution rate of 95%. The fused feature vector is post-processed, including outlier correction and feature smoothing, and finally a comprehensive audit feature vector is formed that includes time trend features, periodic pattern features, abnormal behavior features and multi-dimensional correlation features.

[0063] The specific implementation of step S10 is to output an optimized and integrated comprehensive audit feature vector, which contains the complete feature information required for audit analysis. Time trend features reflect the long-term development and change direction of accounting data, periodic pattern features reflect the regular changes in accounting operations, abnormal behavior features identify potential audit risk points, and multidimensional correlation features reveal the inherent connections between different dimensions. The feature vector is output in a standardized format to facilitate subsequent access and processing by the audit analysis system. The output also includes the feature vector's confidence assessment and quality evaluation indicators, providing reliable data support for audit decision-making.

[0064] The accounting time series feature recognition model utilizes a deep neural network architecture based on a long short-term memory network. The network structure consists of an input layer, multiple LSTM hidden layers, an attention layer, and an output layer. The input layer receives preprocessed accounting time series data in the form of a fixed-length sequence of numerical vectors. The LSTM hidden layer utilizes a bidirectional structure, consisting of forward and backward LSTM units. The hidden state dimension of each LSTM unit is set to 128 dimensions, enabling it to simultaneously capture both forward and backward dependencies in the accounting data. The network comprises three LSTM hidden layers, with residual connections between layers to prevent vanishing gradients. The attention layer utilizes a self-attention mechanism, identifying key accounting events and important time points by calculating attention weights between different time steps. The output layer utilizes a fully connected neural network structure, mapping hidden layer features to a final feature vector output using an activation function.

[0065] The training dataset construction process first involves collecting historical accounting data covering multiple industries and enterprises of different sizes, with a time span of no less than five years, to ensure the representativeness and integrity of the data. The raw accounting data is then quality-checked and cleaned, including missing value filling, outlier detection, and data consistency verification. The data is classified and labeled according to the characteristics of accounting business, and a labeling system that includes normal and abnormal business models is established. A sliding window technique is used to divide the continuous accounting time series into training samples of fixed length, each containing an input sequence and a corresponding feature label. Data augmentation techniques are used to expand the diversity of training samples, including time offset transformation, numerical perturbation transformation, and sequence length adjustment. Finally, a large-scale training dataset containing more than 500,000 labeled samples is constructed. The dataset is divided into training, validation, and test sets in an 8:1:1 ratio for model training, tuning, and performance evaluation.

[0066] It should be noted that the first main technical idea of ​​the present invention is the dynamic adjustment mechanism of the self-varying time window sequence, which realizes the adaptive response to the changing characteristics of accounting data by constructing the self-varying time window function. The traditional fixed time window method cannot effectively handle the irregular change patterns and sudden abnormal events in accounting data, resulting in the mismatch between the time boundary of feature extraction and the inherent change law of the data. The present invention enables the time window to dynamically adjust the boundary position and length according to the real-time change characteristics of the accounting data through the comprehensive regulation of parameters such as the business fluctuation coefficient, the data change rate threshold and the window overlap, effectively capturing the precise time nodes of abnormal changes, significantly improving the time accuracy and adaptability of feature extraction, and avoiding the problem of important accounting events being truncated or omitted by the time window boundary in traditional methods.

[0067] The second main technical idea of ​​the present invention is a multi-dimensional feature fusion system based on deep learning, which systematically integrates the features of time dimension, space dimension, business dimension and outlier dimension. Existing accounting audit methods usually adopt a single dimension or simple combination of feature analysis methods, which lack the ability to deeply mine the complex correlation relationships of accounting data. The present invention uses a long-short-term memory network model to identify long-term dependencies in time series, combines geographical distribution and organizational structure information to extract spatial correlation features, and incorporates business dimension features of business types and transaction attributes to form a multi-dimensional collaborative feature representation system that can fully reflect the complex patterns and potential correlations in accounting data, greatly improving the integrity and expression capabilities of audit features.

[0068] The third main technical idea of ​​the present invention is a feature density-guided weight optimization strategy, which quantifies the distribution density of effective information in the comprehensive audit feature vector through a feature density calculation function. Traditional feature fusion methods usually adopt empirical weight allocation or simple average weight strategies, and lack a quantitative evaluation mechanism for feature information density. The present invention uses feature density as an important objective function for multi-objective optimization. By calculating the proportion and distribution uniformity of non-zero elements in the feature vector, it ensures that the fused feature vector contains as much effective information as possible, avoiding feature redundancy and information sparsity problems, and making the comprehensive audit feature vector have higher information carrying capacity and discrimination effect.

[0069] The synergy of these three main technical ideas has produced significant improvements in technical effects. The dynamic adjustment of the independent time window provides a precise time benchmark for multi-dimensional feature extraction, ensuring that features of each dimension can be extracted within the optimal time boundary, avoiding the problem of feature quality degradation caused by time window mismatch. The multi-dimensional feature fusion system provides a rich source of features for feature density optimization, enabling the weight optimization process to find the optimal solution in a larger feature space, improving the quality and stability of the optimization results. The feature density-oriented weight optimization strategy provides a quantitative evaluation standard for independent time window adjustment and multi-dimensional feature extraction, forming a closed-loop optimization feedback mechanism, which enables the entire feature extraction system to continuously improve and optimize itself, achieving a comprehensive improvement in feature extraction accuracy, adaptability, and stability compared to traditional methods.

[0070] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions. When the program instructions are run in a computer, the program instructions are used to execute the above-mentioned method for extracting audit features from the fusion of massive financial and accounting data.

[0071] The third aspect of the present invention provides a massive accounting data fusion audit feature extraction system, which includes the above-mentioned computer-readable storage medium. The system is any one of a computer, a server, and a single-chip microcomputer. The computer-readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing the program instructions stored in the computer-readable storage medium.

[0072] Specifically, the principle of the present invention is: the principle of the present invention to solve the core technical problem is to construct a dynamic adaptive feature extraction framework, which can automatically adjust the analysis strategy according to the inherent change law of accounting data. First, the basic time window length is determined by the time series clustering algorithm to provide a stable time benchmark for subsequent processing. The key innovation lies in the design of the independent time window function, which receives multiple control parameters such as business fluctuation coefficient, time granularity parameter, data change rate threshold and window overlap, and can monitor the change amplitude and frequency characteristics of accounting data in real time. When the data change exceeds the preset threshold, the window boundary adjustment mechanism is automatically triggered, thereby realizing the accurate tracking of the internal change characteristics of the data.

[0073] The accounting time series feature recognition model, based on a long-short-term memory network architecture, processes complex temporal dependencies through multi-layer LSTM units, and combines residual connections and attention mechanisms to enhance the recognition of key time nodes. Pre-trained on large-scale accounting datasets, the model possesses a deep understanding of fundamental patterns in accounting data and can accurately identify long-term trend patterns and cyclical variations in independent time window sequences. A multidimensional feature extraction mechanism simultaneously constructs feature vectors from four dimensions: time, space, business, and anomalies. A multi-objective weight optimization model is used to determine the optimal weight coefficients, achieving a scientific integration of features from each dimension.

[0074] The fundamental reason why the technical solution of the present invention is logical is that it establishes a direct mapping relationship between data change perception and feature extraction strategy adjustment, and realizes the active adaptation of the feature extraction process to the inherent change laws of the data through the independent time window function, rather than passively using fixed strategies to process dynamically changing data, thereby technically solving the fundamental problem that traditional methods cannot dynamically adapt to the internal changing characteristics of the data.

[0075] A specific embodiment 1 of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows.

[0076] In step S01, the similarity distance calculation process of the dynamic time warping algorithm is described in detail as follows:

[0077] ;

[0078] Where, For sequence with sequence The dynamic time-warping distance between This is the first accounting time series data; This is the second accounting time series data; is the path sequence; is the path length; For the path The Euclidean distance between pairs of points; For sequence Coordinates of the path points; For sequence Coordinates of the path points; is the point number on the path, and its value range is arrive The coefficient of variation calculation is specifically expressed as:

[0079] ;

[0080] Where, is the coefficient of variation; is the standard deviation of accounting data; is the mean of the accounting data.

[0081] In step S02, the calculation process of the business fluctuation coefficient is described in detail as follows:

[0082] ;

[0083] Where, is the business fluctuation coefficient; is the standard deviation of accounting business volume; is the mean value of accounting business volume; is the adjustment factor, and its value range is 0.8~1.2. The calculation of data change rate is specifically expressed as:

[0084] ;

[0085] Where, For the The data change rate at each moment; For the Financial accounting data value at the moment; For the Financial accounting data value at the moment; is a time index, indicating the current time point. The independent time window length adjustment function is specifically expressed as:

[0086] ;

[0087] Where, For the The length of the independent time window; is the basic window length; is the window adjustment coefficient, ranging from 0.1 to 0.5; For the The average data change rate of each window; is the data change rate threshold; The window number.

[0088] The specific implementation of step S03 is the same as above and will not be described in detail here.

[0089] In step S04, the Pearson correlation coefficient calculation process is described in detail as follows:

[0090] ;

[0091] Where, is the Pearson correlation coefficient; For the Accounting activity data for each geographic area; For the corresponding reference data for each geographical area; for The mean of the series; for The mean of the series; is the sample size; is the geographic region number, and its value range is arrive The geographical distribution correlation calculation process is described in detail as follows:

[0092] ;

[0093] Where, is the geographical distribution correlation coefficient; For the Coordinate values ​​of a geographical area; is the mean of the coordinate values ​​of all geographic areas; For the the intensity of accounting activities in a geographical area; is the average intensity of accounting activities across all geographical regions; is the total number of geographical areas; is the geographic region number, and its value range is arrive The calculation of organizational relevance is specifically expressed as:

[0094] ;

[0095] Where, For organizations and organizations The degree of correlation between For organizations The collection of accounting business; For organizations The collection of accounting business; is the organizational level weight factor, ranging from 0.5 to 2.0; and The organization number.

[0096] In step S05, the business association rule confidence calculation process is described in detail as follows:

[0097] ;

[0098] Where, For business rules arrive confidence level; For business and business concurrent support; For business Support for individual appearances; is the preceding business event; is the subsequent business event. The business type feature weight calculation is specifically expressed as:

[0099] ;

[0100] Where, For the Feature weights of business-type services; For the The frequency of occurrence of similar businesses; is the total number of business types; is the total amount of accounting data; To include The number of documents related to the business type; Is the business type serial number, the value range is arrive ; is the summation variable.

[0101] In step S06, the kernel density estimation calculation process is described in detail as follows:

[0102] ;

[0103] Where, For data points The kernel density estimate at ; is the total number of samples; is the bandwidth parameter, ranging from 0.1 to 0.5; As the kernel function, Gaussian kernel function is used ; For the sample points; is the data point for which density is to be estimated; is the sample number, and its value range is arrive ; is the input variable of the kernel function; It is a natural constant, approximately equal to 2.718; is the ratio of pi to 3.14159. The Mahalanobis distance calculation process is described in detail as follows:

[0104] ;

[0105] Where, For the Mahalanobis distance of data points; For the The feature vector of each data point; is the mean vector of the eigenvectors; is the covariance matrix of the eigenvectors; is the inverse matrix of the covariance matrix; is the data point number; Represents vector transpose. The local outlier factor calculation is specifically expressed as:

[0106] ;

[0107] Where, For the The local outlier factor of each data point; For data points of Neighbor set; For data points The local reachable density of Point for neighbors The local reachable density of Number the neighbors; for The number of elements in the neighbor set. The outlier comprehensive score calculation is specifically expressed as:

[0108] ;

[0109] Where, For the The comprehensive outlier score of each data point; is the normalized Mahalanobis distance; is the normalized local outlier factor; is the normalized isolation forest anomaly score; is the weight coefficient, satisfying .

[0110] The specific implementation of step S07 is the same as above and will not be described in detail here.

[0111] In step S08, the multi-objective optimization fitness function calculation process is described in detail as follows:

[0112] ;

[0113] Where, is the weight vector The fitness function value of is the weight vector of the four-dimensional features, satisfying and ; is the weight coefficient of the objective function, and its value range is 0.2~0.3; is the audit accuracy function; is the recall function; is the F1 score function; is the characteristic density function; are the weight coefficients of time dimension, space dimension, business dimension and outlier dimension respectively.

[0114] In step S09, the feature weighted fusion calculation process is described in detail as follows:

[0115] ;

[0116] Where, is the comprehensive audit feature vector; is the time dimension feature vector; is the spatial dimension feature vector; is the business dimension feature vector; is the outlier eigenvector; is the corresponding optimal weight coefficient.

[0117] In step S10, the feature density calculation function is specifically expressed as:

[0118] ;

[0119] Where, is the characteristic density value; is the number of non-zero elements in the eigenvector; is the total number of elements in the eigenvector; For the The value of the non-zero elements; For the The distribution uniformity coefficient of the element position is calculated as follows: ,in For the The position index of the element, is the average value of all non-zero element positions, is the maximum position index of the vector; Is the element number, belonging to the set of non-zero elements .

[0120] In order to better understand and implement the present invention, Example 2 of a specific application scenario of the present invention is provided below: A research team used the massive accounting data fusion audit feature extraction method of the present invention to conduct actual application verification for the accounting audit needs of a large group, including various business types such as procurement, sales, expenses, and asset depreciation. The researchers first collected the complete accounting data of the group for the past 24 months, totaling 36 million records. The data included dimensional information such as transaction time, amount, business type, geographic location, and organizational structure. In step S01, a time series clustering algorithm is used to analyze the periodic characteristics of historical accounting data, and the similarity distance is calculated using a dynamic time warping algorithm. According to the formula Calculation is performed, where sequence X is monthly income data and sequence Y is monthly expenditure data. Cluster analysis shows that corporate accounting data shows obvious quarterly periodicity, so the length of the first time window is determined to be 90 days. The coefficient of variation calculation result is , indicating that the data fluctuation is moderate.

[0121] In the process of constructing the independent time window in step S02, researchers set the business fluctuation coefficient parameter. The business fluctuation coefficient is calculated to be 1.23, where , , regulatory factor The time granularity parameter is set to 1 day, the data change rate threshold is set to 8%, and the window overlap is set to 0.3. , 126 significant change time nodes were identified. The calculation results of the independent time window length adjustment function show that the basic window length In the case of the 22-day window, the adjusted window length ranges from 22 to 38 days, effectively adapting to the dynamic changes in financial and accounting data, as shown in Table 1.

[0122] Table 1 Statistics of independent time window series

[0123]

[0124] In step S03, the researchers used a financial accounting time series feature recognition model to process the independent time window sequence. This model, based on an LSTM architecture, consists of three layers of LSTM units, with a hidden state dimension of 128, a sequence length parameter of 60, and a temperature coefficient of 0.8. The model training dataset contained 180,000 labeled samples, of which 73% were normal pattern samples and 27% were abnormal pattern samples. Through deep learning processing, a 64-dimensional time dimension feature vector was successfully extracted, comprising 32 dimensions of trend features, 20 dimensions of periodic features, and 12 dimensions of seasonal features.

[0125] In step S04, the spatial dimension feature extraction is performed based on the geographical distribution information and organizational structure relationship of the 23 branches, and the Pearson correlation coefficient is calculated. The correlation coefficient of accounting activities between geographical regions is calculated to be 0.67. , indicating a strong correlation between accounting activities in different regions. The results of organizational correlation calculations show that the average correlation between headquarters and branches is 0.58, while the average correlation between branches is 0.34. (See Table 2 for details.)

[0126] Table 2 Statistics of spatial dimension features

[0127]

[0128] In step S05, business dimension features are extracted based on the business type and transaction attributes of the accounting data. The corporate accounting data involves procurement business accounting for 34%, sales business accounting for 28%, expense accounting for 21%, and asset management accounting for 17%. The confidence calculation formula of the business association rule is used The confidence level for the association between procurement and expenses is 0.73, and the confidence level for the association between sales and accounts receivable is 0.81. The weight calculation results for business type characteristics show that sales has a weight of 0.295, procurement has a weight of 0.268, expenses have a weight of 0.237, and asset management has a weight of 0.200.

[0129] In step S06, the researchers used three methods, kernel density estimation, Mahalanobis distance, and local outlier factor, to identify abnormal data points. Kernel density estimation calculation formula In the analysis, the bandwidth parameter h was set to 0.2, and the Gaussian kernel function parameter was set to the standard normal distribution. The Mahalanobis distance calculation results show that a total of 2847 outlier data points were identified, accounting for 0.08% of the total data volume. In the local outlier factor calculation, the k-nearest neighbor parameter was set to 10, and 3156 outlier data points were identified. This is shown in Table 3.

[0130] Table 3 Statistics of outlier characteristics

[0131]

[0132] In step S07, a second time window was established for validation, with a length of 120 days, to contrast with the first time window. After repeating the feature extraction process, a validation feature set was obtained, consisting of 68 temporal features, 45 spatial features, 52 business features, and 36 outlier features. Cross-validation results showed that the feature extraction method achieved 91.3% consistency across different time scales.

[0133] In the multi-objective optimization weight determination in step S08, a multi-objective genetic algorithm is used to solve the optimal weight coefficient. Fitness function calculation formula In the example, the objective function weight coefficients are set as , , , After 500 generations of evolution, the optimal weight coefficient combination is obtained: time dimension weight , spatial dimension weight , business dimension weight , outlier weight . As shown in Table 4.

[0134] Table 4 Statistics of multi-objective optimization results

[0135]

[0136] In step S09, the optimal weight coefficient is used to perform feature weighted fusion. According to the formula , we get a comprehensive audit feature vector with a total dimension of 201. The fused feature vector effectively integrates multi-dimensional information such as time trend, spatial distribution, business association and abnormal behavior.

[0137] In step S10, the feature density calculation function The final feature density calculated is 0.724, indicating that the comprehensive audit feature vector contains rich and effective information. The output comprehensive audit feature vector contains 64 dimensions of time trend features, 58 dimensions of periodic pattern features, 43 dimensions of abnormal behavior features, and 36 dimensions of multidimensional correlation features.

[0138] Traditional accounting and audit feature extraction mainly uses single-dimensional statistical analysis methods, such as rule-based anomaly detection, simple time series analysis, and independent spatial distribution statistics. These traditional methods have problems such as incomplete feature extraction, insufficient multi-dimensional information fusion capabilities, and poor adaptability to dynamic changes. The present invention introduces technical means such as independent time windows, multi-dimensional feature fusion, and multi-objective optimization weight determination, which improves the audit accuracy by 8.9%, the recall rate by 9.7%, the F1 score by 9.6%, and the feature density by 12.6% compared with traditional methods. The present invention can more comprehensively capture complex patterns and potential anomalies in accounting data, improve the accuracy and reliability of audit feature extraction, and provide effective technical support for the intelligent audit of massive accounting data.

[0139] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A method for extracting audit features from massive accounting data fusion, characterized by: include: The time series clustering algorithm is used to determine the length of the first time window, and the historical accounting data is divided into multiple first time windows based on the periodic characteristics of the accounting data; Construct an independent time window function, input the accounting data within the first time window, the business volatility coefficient, the time granularity parameter, the data change rate threshold, and the window overlap, and output an independent time window sequence for dynamically adjusting the window boundaries to adapt to the internal variation characteristics of the accounting data. Use a pre-trained accounting time series feature recognition model to perform deep learning processing on the independent time window sequence, identify long-term trend patterns and periodic variation patterns in the accounting data, and output time dimension features. Extract spatial dimension features based on the geographical distribution information and organizational structure information of the accounting data in the independent time window sequence; extract business dimension features based on the business type and transaction attributes of the accounting data in the independent time window sequence; Calculate the outlier degree of accounting data in the independent time window sequence, identify the outliers corresponding to abnormal accounting behavior through statistical analysis methods, and form outlier characteristics; construct a multi-objective optimization weight determination model, use time dimension characteristics, space dimension characteristics, business dimension characteristics and outlier characteristics as input variables, and solve the optimal weight coefficient through a multi-objective optimization algorithm; use the optimal weight coefficient to weightedly fuse the time dimension characteristics, space dimension characteristics, business dimension characteristics and outlier characteristics to form a comprehensive audit feature vector.

2. The method for extracting audit features from massive accounting data fusion according to claim 1 is characterized in that: The first time window is a fixed time period determined based on the natural periodicity of accounting data. Its length is automatically determined by analyzing the similarity patterns of historical accounting data through a clustering algorithm. It is used to capture the basic periodic characteristics of accounting data and usually corresponds to the standard reporting period or settlement period of accounting business, providing a stable time reference framework for subsequent feature extraction.

3. The method for extracting audit features from massive accounting data fusion according to claim 2 is characterized in that: The self-varying time window sequence is a set of variable-length time periods that are further subdivided based on the first time window. The window boundaries are adaptively adjusted according to the internal change rate of the accounting data to accurately locate the time nodes of abnormal changes.

4. The method for extracting audit features from massive financial and accounting data fusion according to claim 3 is characterized in that: The independent time window function is used to dynamically adjust the boundary position and length of the time window according to the real-time changing characteristics of accounting data to adapt to the fluctuation pattern and changing laws of accounting data in different business scenarios. The output is an independent time window sequence after adaptive adjustment. Each window is marked with the starting time point, ending time point and statistical characteristics of the data in the window.

5. The method for extracting audit features from massive accounting data fusion according to claim 4 is characterized in that: The business fluctuation coefficient is a quantitative indicator that reflects the changes in the intensity of accounting business activities. It is calculated by analyzing the amplitude and frequency characteristics of business volume changes in historical accounting data, and is used to guide the dynamic adjustment process of the independent time window function.

6. The method for extracting audit features from massive accounting data fusion according to claim 5 is characterized in that: The specific structure of the accounting time series feature recognition model is a deep neural network designed based on the long short-term memory network architecture. It contains multiple layers of LSTM units to process the long-term dependencies of time series data. The entire network structure enhances the recognition ability of key time nodes and important accounting events through residual connections and attention mechanisms.

7. The method for extracting audit features from massive financial and accounting data fusion according to claim 6 is characterized in that: The multi-objective optimization weight determination model is used to solve the optimal weight distribution of time dimension features, space dimension features, business dimension features and outlier features in the comprehensive audit feature vector. The multi-objective genetic algorithm is used as the core optimization engine, and the audit accuracy, recall rate, F1 score and feature density after feature fusion are used as optimization objective functions.

8. The method for extracting audit features from massive financial and accounting data fusion according to claim 7 is characterized in that: After the step of constructing the independent time window function, it also includes establishing a second time window for verifying the feature extraction effect. The length of the second time window is different from the length of the first time window. Repeat the step of using the accounting time series feature recognition model to extract time dimension features to form outlier features to obtain a verification feature set.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the method for extracting audit features from fusion of massive financial and accounting data according to any one of claims 1 to 8.

10. A massive accounting data fusion audit feature extraction system, characterized by: The computer-readable storage medium according to claim 9 is included, the system is any one of a computer, a server, and a single-chip microcomputer, the computer-readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

Images