Enterprise financial data anomaly detection method and system based on model distillation

Abstract

The application provides a model distillation-based enterprise financial data anomaly detection method and system, relates to the technical field of data processing, and comprises the following steps: obtaining a plurality of window financial data sets of an enterprise, triggering and sorting according to a financial anomaly multidimensional factor, and establishing a plurality of window financial anomaly trigger vectors; performing multi-source training, establishing a financial anomaly detection architecture, including a debt repayment anomaly detection model set, a profit anomaly detection model set and a fund chain anomaly detection model set; performing distillation parameter decision, obtaining a distillation decision multidimensional space; introducing a distillation loss multistage optimization mechanism for iterative optimization to generate a financial anomaly detection channel; and obtaining an enterprise financial anomaly detection report. The application solves the technical problem of low accuracy of enterprise financial data anomaly detection caused by the weak cross-layer knowledge transfer mechanism of the model in the prior art, which leads to the inability of the financial data anomaly detection process to effectively share information and transfer knowledge between different models.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to an enterprise financial data anomaly detection method and system based on model distillation. BACKGROUND

[0002] In the current field of enterprise financial data anomaly detection, financial data analysis and anomaly detection methods mainly rely on traditional statistical analysis models, machine learning models and deep learning algorithms. With the continuous growth and increasing complexity of enterprise financial data, existing financial data anomaly detection methods are facing more and more challenges. In many financial anomaly detection systems, when using deep learning algorithms, the model is often complex and consumes a lot of computing resources, and the training time is long. Especially when the data volume is large, the effect may not be proportional to the increase. In addition, the cross-layer knowledge transfer mechanism of the model in the prior art is relatively weak, which leads to the inability to effectively share information and transfer knowledge between different models in the financial data anomaly detection process, resulting in repeated learning, waste of computing resources, and insufficient model performance, which further leads to low accuracy of enterprise financial data anomaly detection. SUMMARY

[0003] The present application provides an enterprise financial data anomaly detection method and system based on model distillation, aiming to solve the technical problem that the cross-layer knowledge transfer mechanism of the model in the prior art is relatively weak, which leads to the inability to effectively share information and transfer knowledge between different models in the financial data anomaly detection process, and further leads to low accuracy of enterprise financial data anomaly detection.

[0004] The first aspect of the present application provides an enterprise financial data anomaly detection method based on model distillation, the method comprising: obtaining a plurality of window financial data sets of an enterprise, and triggering and combing the plurality of window financial data sets according to a financial anomaly multi-dimensional factor to establish a plurality of window financial anomaly trigger vectors; performing multi-source training according to the financial anomaly multi-dimensional factor to establish a financial anomaly detection architecture, the financial anomaly detection architecture comprising a debt repayment anomaly detection model set, a profit anomaly detection model set and a fund chain anomaly detection model set; performing distillation parameter decision on the financial anomaly detection architecture according to a distillation learning factor to obtain a distillation decision multi-dimensional space; introducing a distillation loss multi-level optimization mechanism to iteratively optimize the distillation decision multi-dimensional space to generate a financial anomaly detection channel, the financial anomaly detection channel comprising a debt repayment anomaly detection channel, a profit anomaly detection channel and a fund chain anomaly detection channel; inputting the plurality of window financial anomaly trigger vectors into the financial anomaly detection channel to obtain an enterprise financial anomaly detection report.

[0005] In a second aspect, the application discloses a model distillation-based enterprise financial data anomaly detection system, which is used for the model distillation-based enterprise financial data anomaly detection method, and comprises a trigger carding module, a multi-source training module, a distillation parameter decision module, an iterative optimization module, and a detection report generation module.

[0006] The one or more technical solutions provided in the application have at least the following beneficial effects:

[0007] By obtaining multiple window financial data sets of an enterprise, the financial data of the enterprise at different time scales can be analyzed in detail, and according to the financial abnormal multi-dimensional factors, the multiple window financial abnormal trigger vectors can be triggered and combed, so that potential financial abnormal signals can be identified from multiple dimensions. Each window financial abnormal trigger vector represents a risk indication of different financial dimensions, thereby providing clear feature information for subsequent abnormal detection, facilitating more accurate identification and classification of abnormal conditions. By training multiple sources according to the financial abnormal multi-dimensional factors, the financial abnormal detection architecture established includes a debt repayment abnormal detection model set, a profit abnormal detection model set and a fund chain abnormal detection model set. This multi-model architecture can comprehensively analyze the financial health of the enterprise from multiple angles, not only can detect different types of financial abnormalities, but also can optimize on different types of financial data, thereby improving detection accuracy and comprehensiveness. By distilling the parameters of the financial abnormal detection architecture according to the distillation learning factor, the knowledge in the teacher model can be transferred to the student model in a more efficient and simplified manner, thereby improving detection accuracy while reducing the consumption of computing resources. The distillation decision multi-dimensional space provides a variety of optimized parameter configurations, which can effectively guide the selection and configuration of the model, so that the distillation process can achieve efficient knowledge transfer in multiple dimensions. By introducing a multi-level optimization mechanism for distillation loss to iteratively optimize the distillation decision multi-dimensional space, the optimal distillation parameter configuration can be found. This mechanism improves the learning effect of the model, so that multiple financial abnormal detection channels achieve a balance between different levels of loss, thereby achieving more accurate financial abnormal detection. By inputting the multiple window financial abnormal trigger vectors into the financial abnormal detection channel, an enterprise financial abnormal detection report is obtained. This process combines data from different windows and the results of multi-channel detection, so that financial abnormal detection can cover multiple dimensions and time scales, enhancing the accuracy and comprehensiveness of the report. Enterprises can understand potential risks in their financial situation through the report and make timely risk management and decision-making accordingly.

[0008] The above description is only a summary of the technical solutions of the present application. In order to more clearly understand the technical means of the present application, the following specific embodiments of the present application can be implemented according to the content of the specification, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following specific embodiments of the present application are described in detail. BRIEF DESCRIPTION OF DRAWINGS

[0009] Figure 1 The enterprise financial data abnormal detection method based on model distillation provided by the embodiments of the present application is shown in the flowchart.

[0010] Figure 2 The enterprise financial data abnormal detection system structure based on model distillation provided by the embodiments of the present application is shown in the schematic diagram.

[0011] Explanation of reference signs: trigger combing module 10, multi-source training module 20, distillation parameter decision module 30, iterative optimization module 40, and detection report generation module 50. DETAILED DESCRIPTION

[0012] The embodiments of the present application provide a model distillation-based enterprise financial data anomaly detection method and system, which solves the technical problem of weak cross-layer knowledge transfer mechanism of the model in the prior art, resulting in that the financial data anomaly detection process cannot effectively share information and transfer knowledge between different models, and further resulting in low efficiency of enterprise financial data anomaly detection.

[0013] After introducing the basic principles of the present application, various non-limiting embodiments of the present application will be specifically introduced in combination with the drawings of the specification. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0014] In embodiment one, as shown in the figure, the embodiments of the present application provide a model distillation-based enterprise financial data anomaly detection method, which comprises: Figure 1

[0015] Obtain a plurality of window financial data sets of an enterprise, and trigger combing of the plurality of window financial data sets according to financial anomaly multi-dimensional factors to establish a plurality of window financial anomaly trigger vectors.

[0016] A plurality of financial data streams of the enterprise are collected through multi-scale windows, the role of the multi-scale window is to obtain different perspectives on the financial health status from different time scales, and the collected plurality of financial data streams are subjected to data cleaning, the goal of which is to remove noise, missing values and any inconsistent or incorrect values in the data, the plurality of financial data streams after cleaning are converted into a plurality of window financial data sets, and the financial data in each window is related to a specific time period.

[0017] The financial anomaly multi-dimensional factor is a set of multi-dimensional characteristics used to describe potential anomalies of enterprise finance, including debt repayment anomalies, profit anomalies and fund chain anomalies, and the trigger combing of the plurality of window financial data sets, that is, according to the financial anomaly multi-dimensional factor, possible financial anomaly patterns are identified and marked, for example, according to historical data and the financial anomaly multi-dimensional factor, a risk point of debt repayment anomaly is identified, and each window financial data set generates a window financial anomaly trigger vector containing detection results of all related financial anomaly factors in the window.

[0018] Multi-source training is performed according to the financial anomaly multi-dimensional factor to establish a financial anomaly detection architecture, which comprises a debt repayment anomaly detection model set, a profit anomaly detection model set and a fund chain anomaly detection model set.

[0019] ​The method of multi-source training is used to train financial data collected from multiple data sources, including the financial statements of enterprises, historical financial data, industry data, market indicators, etc. Through these multi-source data, a financial anomaly detection architecture is established. The goal of multi-source training is to enable the model to capture factors that affect the financial health of enterprises from different angles, such as macroeconomics and industry characteristics. The financial anomaly detection architecture includes a debt repayment anomaly detection model set, a profit anomaly detection model set, and a fund chain anomaly detection model set. These model sets can be built based on different algorithms, such as decision trees, random forests, neural networks, etc. Each model set has multiple models, such as the debt repayment anomaly model set, which includes multiple debt repayment anomaly detection models based on different algorithms. Each debt repayment anomaly detection model is trained on different types of financial data to capture different abnormal characteristics.

[0020] The debt repayment anomaly detection model set includes multiple debt repayment anomaly detection models for analyzing the debt repayment ability of an enterprise, such as by analyzing factors such as debt ratio, cash flow, and interest coverage ratio to detect whether the debt repayment ability is abnormal. The profit anomaly detection model set includes multiple profit anomaly detection models for analyzing the profitability of an enterprise, using indicators such as gross profit margin, net profit, and sales revenue to determine whether the profitability of the enterprise is normal. The fund chain anomaly detection model set includes multiple fund chain anomaly detection models for monitoring the cash flow of an enterprise, analyzing cash flow, accounts receivable, inventory turnover rate, etc., to determine whether there is potential risk in the fund chain of the enterprise.

[0021] According to the distillation learning factor, the distillation parameter decision is made for the financial anomaly detection architecture, and a distillation decision multi-dimensional space is obtained.

[0022] The distillation learning factor is a key factor in the distillation process that adjusts and optimizes model parameters. Distillation is a model compression technique that distills the knowledge of multiple complex models into a smaller, more efficient model to improve the inference speed and accuracy of the model. In this step, the distillation learning factor includes the number of teacher models, student model attributes, and distillation hyperparameters. By adjusting the distillation learning factor for distillation parameter decision, it is determined how to adjust the parameter mapping relationship between the teacher model and the student model during distillation, so that the final student model can effectively learn and retain important information from the teacher model. The distillation decision multi-dimensional space is the parameter space obtained during the distillation process, which includes multi-dimensional decision information and describes the decision space of different models and parameter settings during the distillation process.

[0023] An iterative optimization mechanism for distillation loss is introduced to optimize the distillation decision multi-dimensional space, generating a financial anomaly detection channel. The financial anomaly detection channel includes a debt repayment anomaly detection channel, a profit anomaly detection channel, and a fund chain anomaly detection channel.

[0024] The distillation loss is an index for measuring the gap between the student model and the teacher model during the learning process. In knowledge distillation, the cross-entropy loss or mean square error loss is used to calculate the difference between the output of the student model and the output of the teacher model. The distillation loss multi-level optimization mechanism refers to gradually optimizing the model parameters through multiple iterations during the distillation process. In each iteration, the model parameters are adjusted according to the feedback information of the distillation loss, so that the performance of the student model gradually improves. Specifically, multi-factor distillation loss is introduced at different learning stages to balance and optimize from multiple angles during the optimization process. These factors include model accuracy, computational efficiency, overfitting prevention, and other factors. Through multiple iterations of optimizing the distillation loss, the model adjusts its weights in each iteration to gradually approach the optimal financial anomaly detection capability. In each iteration, the model continuously adjusts the parameters in its decision space according to the distillation loss and feedback until the predetermined optimization goal is reached.

[0025] After distillation optimization, a financial anomaly detection channel is generated, which is responsible for detecting financial anomalies in enterprises. The debt repayment anomaly detection channel focuses on checking the solvency of enterprises, such as debt default risk and cash flow shortage; the profit anomaly detection channel focuses on the profitability of enterprises, such as low profit level and profit volatility; and the fund chain anomaly detection channel focuses on whether there are anomalies in the fund flow of enterprises, such as fund turnover problems and liquidity crisis.

[0026] The plurality of window financial anomaly trigger vectors are input into the financial anomaly detection channel to obtain an enterprise financial anomaly detection report.

[0027] The plurality of window financial anomaly trigger vectors are input into the financial anomaly detection channel to obtain an enterprise financial anomaly detection report.

[0028] Further, the plurality of window financial data sets are triggered and sorted according to the multi-dimensional factors of financial anomalies to establish a plurality of window financial anomaly trigger vectors, including:

[0029] According to the financial anomaly multi-dimensional factor, trigger mining is performed to establish a debt repayment anomaly trigger detection tree, a profit anomaly trigger detection tree, and a fund chain anomaly trigger detection tree; according to the plurality of window financial data sets, an nth window financial data set is extracted, and feature recognition is performed on the nth window financial data set to obtain a plurality of financial data features, n being a positive integer; according to the debt repayment anomaly trigger detection tree, trigger analysis is performed on the plurality of financial data features to obtain an nth debt repayment anomaly trigger vector; according to the profit anomaly trigger detection tree, trigger analysis is performed on the plurality of financial data features to obtain an nth profit anomaly trigger vector; according to the fund chain anomaly trigger detection tree, trigger analysis is performed on the plurality of financial data features to obtain an nth fund chain anomaly trigger vector, and the nth debt repayment anomaly trigger vector and the nth profit anomaly trigger vector are combined to construct an nth window financial anomaly trigger vector.

[0030] Trigger mining is to use financial anomaly multi-dimensional factors to mine abnormal trigger points according to abnormal patterns in financial data. Through the mining of these abnormal trigger points, a recognizable model can be formed to detect future abnormalities. The financial anomaly multi-dimensional factors include debt repayment anomaly, profit anomaly, and fund chain anomaly. These factors serve as key features to help identify and mine financial anomalies.

[0031] According to the financial anomaly multi-dimensional factor, different types of anomalies are processed separately, and a trigger detection tree is established for each type of anomaly. The debt repayment anomaly trigger detection tree is used to detect the debt repayment ability of an enterprise, such as analyzing financial data such as debt ratio, cash flow, interest payment, etc., to determine whether a debt crisis exists. The profit anomaly trigger detection tree is used to analyze the profitability of an enterprise, such as gross profit margin, net profit, sales revenue, etc., to determine whether the enterprise is facing a decline or fluctuation in profitability risk. The fund chain anomaly trigger detection tree is used to check the fund flow of an enterprise, such as cash flow, accounts receivable, inventory turnover, etc., to determine whether there is a fund chain break or liquidity crisis.

[0032] The trigger detection tree is composed of nodes and branches. Each node represents a condition judgment, and each branch represents a different decision path. Through the traversal of the trigger detection tree, it can ultimately determine whether a certain financial data has an abnormal risk.

[0033] The n-th window financial data set is extracted from the plurality of window financial data sets, n is a positive integer, representing the number of the extracted window, and the n-th window financial data set can represent the financial data of any specific time period. Feature recognition is performed on the extracted n-th window financial data set, i.e., financial data features such as debt ratio, net profit, accounts receivable turnover rate, etc. are extracted from the data set. These features are important inputs for model training and anomaly detection. The features can also be transformed or combined to improve the performance of the model, such as standardizing certain features or combining multiple features to form new composite features such as financial leverage ratio. After feature recognition, a plurality of financial data features are obtained, which reflect the main financial situation of the n-th window financial data set.

[0034] The debt repayment anomaly trigger detection tree is used to detect the debt repayment ability of the enterprise, such as analyzing the debt ratio, cash flow, interest payment, etc. to judge whether there is a debt repayment crisis. The debt repayment anomaly trigger detection tree is used to trigger and analyze a plurality of financial data features, and the tree structure is used to determine whether these features meet the conditions of debt repayment anomaly. The trigger analysis process is performed by sequentially traversing the nodes of the tree, and determining whether each feature meets the trigger condition according to the rules, such as whether the debt ratio is greater than a certain threshold, to determine whether there is a debt repayment anomaly for the corresponding feature. If the condition is met, the node outputs a debt repayment anomaly trigger signal, indicating that the feature triggers a debt repayment anomaly. By traversing the entire debt repayment anomaly trigger detection tree, a set of debt repayment anomaly trigger signals is obtained, which form the n-th debt repayment anomaly trigger vector. The n-th debt repayment anomaly trigger vector contains all the trigger signals related to debt repayment ability, indicating the risk assessment of the financial data in the window in terms of debt repayment anomaly detection.

[0035] The profit anomaly trigger detection tree is used to analyze the profit situation of the enterprise, such as gross profit margin, net profit, sales revenue, etc. to judge whether the enterprise faces the risk of declining or fluctuating profitability. According to the profit anomaly trigger detection tree, a plurality of financial data features are triggered and analyzed, and the tree structure is used to determine whether these features meet the trigger conditions of profit anomaly. If the condition is met, a profit anomaly trigger signal is output. After triggering and analyzing a plurality of financial data features by the profit anomaly trigger detection tree, the n-th profit anomaly trigger vector is obtained, which contains trigger signals related to profit anomaly, reflecting whether the enterprise has a profit ability anomaly such as loss or declining profit in the n-th window.

[0036] The fund chain abnormal trigger detection tree is used to check the fund flow of the enterprise, such as cash flow, accounts receivable, inventory turnover, etc., to determine whether there is a fund chain break or liquidity crisis. According to the fund chain abnormal trigger detection tree, multiple financial data features are triggered and combed, and whether these features meet the trigger conditions of fund chain abnormality is judged through the tree structure. If the conditions are met, a fund chain abnormal trigger signal is output. After triggering and combing multiple financial data features through the fund chain abnormal trigger detection tree, the nth fund chain abnormal trigger vector is obtained, indicating that a fund chain problem is detected in the window.

[0037] The nth window financial abnormal trigger vector is the result of combining the nth fund chain abnormal trigger vector, the nth debt repayment abnormal trigger vector, and the nth profit abnormal trigger vector. It contains the trigger signals of the three kinds of financial abnormalities and is used to comprehensively reflect the financial health status of the enterprise in the nth window.

[0038] Further, the triggering and mining according to the financial abnormal multi-dimensional factor includes:

[0039] According to the financial data feature sample set, the financial abnormal multi-dimensional factor is triggered and searched to obtain each debt repayment abnormal trigger sample cluster corresponding to each financial data feature sample. The in-cluster confidence evaluation set is obtained by evaluating the in-cluster confidence of each debt repayment abnormal trigger sample cluster. According to the in-cluster confidence evaluation set, the in-cluster confidence of each debt repayment abnormal trigger sample cluster is fused respectively to obtain the debt repayment abnormal trigger confidence set. The financial data feature sample set is taken as a plurality of trigger detection input nodes, and the debt repayment abnormal trigger confidence set is taken as a plurality of trigger detection output nodes to construct the debt repayment abnormal trigger detection tree.

[0040] The financial data feature sample set includes a plurality of financial data feature samples, containing various financial data of the enterprise, such as debt ratio, cash flow, interest coverage ratio, etc. Triggering and searching refers to analyzing the financial data feature sample set according to the financial abnormal multi-dimensional factor, i.e., debt repayment abnormality, profit abnormality, and fund chain abnormality, to identify which sample data meets the characteristics of debt repayment abnormality. Specifically, triggering and searching scans and compares each financial data feature sample according to the set conditions, such as whether the debt ratio is too high or whether the cash flow is tight. If the sample meets the conditions, it is determined as a debt repayment abnormal trigger sample. Further, through a similarity clustering algorithm such as K-means clustering, samples with similar financial abnormal characteristics are grouped into the same group, thereby forming a plurality of sample clusters. Through triggering and searching, each debt repayment abnormal trigger sample cluster is obtained, each of which contains a group of financial data samples with similar trigger characteristics. These samples represent an abnormality in debt repayment ability.

[0041] The cluster confidence evaluation is to evaluate the samples in each debt repayment anomaly trigger sample cluster, and to determine whether the samples in the cluster are reliable. The confidence is an index for measuring the consistency of the samples in the cluster, indicating whether the samples are close enough and representative. The confidence is obtained by calculating the similarity measure in the cluster, for example, by calculating the Euclidean distance between the samples in the cluster, the cosine similarity, etc. If the difference between the samples is small, the confidence is high; if the difference is large, the confidence is low. For each debt repayment anomaly trigger sample cluster, the corresponding cluster confidence is calculated. By this method, a representative confidence value can be assigned to each cluster, which will reflect the quality and consistency of the samples in the cluster. Finally, a cluster confidence evaluation set is obtained, which contains the confidence evaluations of all debt repayment anomaly trigger sample clusters. These evaluation sets can help to determine which sample clusters are more reliable in the model and are worthy of further analysis.

[0042] The cluster confidence fusion refers to the fusion of the confidence information of each debt repayment anomaly trigger sample cluster to obtain a more accurate debt repayment anomaly trigger confidence set. Specifically, the cluster confidence fusion is based on the confidence evaluation of different sample clusters, and the multiple trigger signals in the cluster are weighted and merged to improve the accuracy and reliability of the final detection result. The debt repayment anomaly trigger confidence set is the final result after the confidence information of all debt repayment anomaly trigger sample clusters is fused, including the confidence value of each cluster, which is used to represent the contribution of each cluster to the debt repayment anomaly.

[0043] The financial data feature sample set is used as a plurality of trigger detection input nodes, each trigger detection input node corresponding to a specific financial data feature such as liability ratio, cash flow, etc. The debt repayment anomaly trigger confidence set is used as a plurality of trigger detection output nodes, each trigger detection output node reflecting the trigger confidence of the corresponding financial feature sample cluster, indicating whether the cluster has a debt repayment anomaly. By combining the plurality of trigger detection input nodes and the plurality of trigger detection output nodes, a debt repayment anomaly trigger detection tree is constructed by training a decision tree algorithm. The leaf node of the tree represents the final abnormal judgment result, and the root node represents the initial judgment standard. The decision tree is a series of judgments and decisions on the input nodes, and finally outputs the result of whether to trigger an anomaly. The decision tree is based on the trigger signals of each feature and performs segmentation and judgment on each branch of the tree.

[0044] The constructed debt repayment anomaly trigger detection tree is used to identify whether there is a debt repayment ability anomaly in the enterprise financial data. Through the structure of the tree, it can quickly determine whether the different financial data features meet the trigger conditions of the debt repayment anomaly, and output the final judgment result.

[0045] Further, according to the debt repayment anomaly trigger detection tree, the plurality of financial data features are triggered and sorted to obtain an nth debt repayment anomaly trigger vector, which includes:

[0046] The plurality of financial data features are input into the debt repayment anomaly trigger detection tree to obtain a plurality of debt repayment anomaly trigger coefficients; it is judged whether the plurality of debt repayment anomaly trigger coefficients are greater than or equal to a debt repayment anomaly trigger threshold to obtain a plurality of trigger judgment results; the plurality of financial data features are selected according to the plurality of trigger judgment results to obtain a debt repayment anomaly trigger feature space; and the debt repayment anomaly trigger feature space is subjected to feature fusion to output the nth debt repayment anomaly trigger vector.

[0047] The plurality of financial data features are input into the debt repayment anomaly trigger detection tree, and each financial data feature represents a different financial indicator of an enterprise. These financial data features are input data and are sequentially input into the nodes of the decision tree for judgment. The debt repayment anomaly trigger coefficient is a value for measuring the degree of anomaly corresponding to the financial data feature in the detection tree. Each financial data feature obtains a corresponding debt repayment anomaly trigger coefficient through the judgment rule of the decision tree. The debt repayment anomaly trigger coefficient can be regarded as the strength or confidence of triggering an anomaly, which indicates whether the financial data feature triggers a debt repayment anomaly and to what extent. For example, if the debt ratio is too high, the decision tree assigns a larger debt repayment anomaly trigger coefficient to the feature, indicating that there is a larger problem with the debt repayment ability. The plurality of debt repayment anomaly trigger coefficients contains the trigger coefficients of all input financial data features, reflecting the possibility of debt repayment anomaly of the enterprise.

[0048] The debt repayment anomaly trigger threshold is a preset standard for judging whether a certain financial data feature triggers a debt repayment anomaly. This threshold can be obtained based on historical data, industry standards, or through experimental optimization. The debt repayment anomaly trigger coefficient of each financial data feature is compared with the debt repayment anomaly trigger threshold. If the debt repayment anomaly trigger coefficient is greater than or equal to the debt repayment anomaly trigger threshold, it is considered that the financial data feature triggers a debt repayment anomaly. Through the judgment of all the plurality of debt repayment anomaly trigger coefficients, a plurality of trigger judgment results are output.

[0049] According to the plurality of trigger judgment results, the financial data features with debt repayment anomaly trigger coefficients greater than or equal to the debt repayment anomaly trigger threshold are selected from the set of financial data features. These financial data features are key factors that trigger a debt repayment anomaly. The debt repayment anomaly trigger feature space is a subset that contains those financial data features that trigger a debt repayment anomaly, representing all features that exhibit debt repayment anomaly risk within the window.

[0050] According to the debt repayment anomaly trigger feature space, feature fusion is performed, which is to combine multiple related features into a new feature, aiming to improve the expression ability and accuracy of the model. Feature fusion can be achieved in various ways, such as weighted average, principal component analysis, or feature splicing in neural networks. The specific method used depends on the selected model and task requirements. The nth debt repayment anomaly trigger vector is a new vector generated by the fused financial features. This vector is a multi-dimensional vector, with each dimension representing a different financial anomaly feature, such as debt ratio, cash flow, etc. Through the fusion result, the abnormal situation of the enterprise in debt repayment can be more accurately reflected.

[0051] Furthermore, according to the distillation learning factor, the financial anomaly detection architecture is subjected to distillation parameter decision, obtaining a distillation decision multi-dimensional space, including:

[0052] The distillation learning factor includes the number of teacher models, student model attributes, and distillation hyperparameters. According to the distillation learning factor, the debt repayment anomaly detection model set is subjected to distillation parameter decision, establishing a distillation decision first space. According to the distillation learning factor, the profit anomaly detection model set is subjected to distillation parameter decision, establishing a distillation decision second space. According to the distillation learning factor, the fund chain anomaly detection model set is subjected to distillation parameter decision, establishing a distillation decision third space. Combining the distillation decision first space and the distillation decision second space, the distillation decision multi-dimensional space is generated.

[0053] The distillation learning factor is a key parameter used in the knowledge distillation process, which is used to optimize the knowledge transfer between teacher models and student models. Knowledge distillation is a technique that transfers the knowledge of a complex model (teacher model) to a smaller model (student model), aiming to maintain high performance of the student model while reducing computational overhead and storage requirements.

[0054] The number of teacher models refers to the number of teacher models used in the distillation process. Each teacher model is a larger or more complex model that has been thoroughly trained and can provide accurate predictions and detailed knowledge. The output of these teacher models is used to guide the training of student models. Using multiple teacher models for distillation can increase the diversity and accuracy of the distillation process, allowing the student model to learn the advantages of different teacher models through multi-model integration.

[0055] The student model attribute refers to the structure or configuration of the student model. Student models are usually smaller and simpler than teacher models, but their performance still needs to be acceptable. Student model attributes include the number of layers, the number of neurons in each layer, activation functions, etc. When designing a student model, a trade-off between computational efficiency and model accuracy needs to be made to ensure that the student model can effectively perform anomaly detection while maintaining a small size.

[0056] The distillation hyperparameter is a parameter adjusted in the distillation process to control the knowledge transfer between the teacher model and the student model, and the distillation hyperparameter includes a temperature parameter and a loss function weight.

[0057] The distillation parameter decision is made for the debt repayment anomaly detection model set using the distillation learning factor. In this process, by adjusting the distillation hyperparameter, the number of suitable teacher models and the properties of student models are selected to achieve knowledge transfer between the teacher model and the student model. The distillation decision first space is a decision space established by adjusting the distillation parameter. This space represents the performance and results under different distillation configurations, which can help select the optimal distillation configuration to obtain the most effective student model for the debt repayment anomaly detection model.

[0058] Similarly, the distillation parameter decision is made for the profit anomaly detection model set using the distillation learning factor. The distillation decision second space is the result after adjusting the distillation learning factor, reflecting the distillation effect of the profit anomaly detection model set under different configurations. This space represents the optimal configuration for profit anomaly detection and helps select the best parameter setting.

[0059] Similarly, the distillation parameter decision is made for the fund chain anomaly detection model set using the distillation learning factor. The distillation decision third space is an optimized decision space provided for the fund chain anomaly detection model set by adjusting the distillation parameter. This space reflects the effect of the distillation process in fund chain anomaly detection to select the best configuration.

[0060] The distillation decision first space, the distillation decision second space, and the distillation decision third space are combined to generate a comprehensive distillation decision multi-dimensional space. The distillation decision multi-dimensional space contains distillation decision information for all financial anomaly detection tasks (including debt repayment, profit, and fund chain). In this distillation decision multi-dimensional space, the optimal parameter configuration can be found by combining the optimization results of each model to achieve comprehensive anomaly detection.

[0061] Further, the distillation loss multi-level optimization mechanism includes:

[0062] The debt repayment anomaly detection model set is distilled according to the distillation decision first space to obtain a debt repayment anomaly distillation first space, which includes a plurality of debt repayment anomaly detectors. The debt repayment anomaly distillation first space is optimized and analyzed according to the distillation loss multi-factor to obtain a debt repayment anomaly distillation second space. The global distillation loss function is established by weight distribution according to the distillation loss multi-factor. The debt repayment anomaly distillation second space is globally distilled and loss-minimized to generate the debt repayment anomaly detection channel according to the global distillation loss function.

[0063] The distillation decision first space is used as the basis for optimizing the decision, and the debt repayment anomaly detection model set is subjected to distillation learning to further optimize the model. The goal of distillation learning is to transfer the knowledge in the teacher model (a complex and accurate model) to the student model (a relatively simple and efficient model). The debt repayment anomaly detection model set includes multiple models, each responsible for detecting debt repayment anomalies based on different features. Through distillation learning, the student model can capture the knowledge of the teacher model to maintain high efficiency and accuracy. The debt repayment anomaly distillation first space is the space obtained after distillation learning, which includes multiple debt repayment anomaly detectors that have been subjected to distillation processing. Each debt repayment anomaly detector is an optimized model after distillation learning, used to detect debt-related financial anomalies. The outputs of multiple debt repayment anomaly detectors can provide diverse perspectives for debt repayment anomaly detection based on different feature combinations and detection methods.

[0064] The distillation loss multi-factor is a plurality of factors used to measure the difference between the teacher model and the student model during the distillation process, including: a prediction layer distillation loss for measuring the difference between the student model and the teacher model in output prediction; a hidden layer distillation loss for measuring the difference between the student model and the teacher model in hidden layer representation; an attention layer distillation loss for measuring the difference between the student model and the teacher model in attention mechanism layer; and an embedding layer distillation loss for measuring the difference between the student model and the teacher model in embedding layer.

[0065] According to the above distillation loss multi-factor, the debt repayment anomaly distillation first space is subjected to optimization analysis, that is, by optimizing the values of the distillation loss multi-factor, the parameters of the multiple debt repayment anomaly detectors are adjusted, so that the student model can better learn from the teacher model and reduce errors in the distillation process. Optimization analysis uses optimization algorithms such as gradient descent, genetic algorithm, etc., to make the distillation learning effect more significant and obtain more accurate debt repayment anomaly detection models. After optimization analysis, the debt repayment anomaly distillation second space is obtained, which includes further optimized models. Each model in this space has been optimized for debt repayment anomaly detection for multiple rounds, with high accuracy and low error rate.

[0066] For the prediction layer distillation loss, the hidden layer distillation loss, the attention layer distillation loss and the embedding layer distillation loss in the distillation loss multi-factor, different distillation loss factors have different effects on the model performance, therefore, appropriate weights are assigned to each distillation loss factor to ensure that the model can be balanced and optimized in various aspects, for example, a certain distillation loss factor (such as the prediction layer loss) contributes more to the model performance, and therefore can be assigned a higher weight; while other distillation loss factors (such as the attention layer loss) have less effect on the model, and can be assigned a lower weight. According to the assigned weights, all distillation loss factors are integrated in a linear manner, that is, each distillation loss factor is multiplied by its corresponding weight, and then all the losses are added to obtain the overall global distillation loss function, which is used to measure the overall gap between the student model and the teacher model, and provides an optimization target for the training process. Preferably, the global distillation loss function is: global distillation loss coefficient = prediction layer distillation loss coefficient * prediction layer distillation loss weight + hidden layer distillation loss coefficient * hidden layer distillation loss weight + attention layer distillation loss coefficient * attention layer distillation loss weight + embedding layer distillation loss coefficient * embedding layer distillation loss weight.

[0067] Using the global distillation loss function, the global distillation loss minimization optimization is performed on the debt repayment anomaly distillation second space, which means that the model parameters are continuously adjusted by the optimization algorithm to reduce the value of the global distillation loss function until the optimal model configuration is found. The goal of this process is to minimize the difference between the student model and the teacher model while improving the accuracy and efficiency of the model. In the process of minimizing the loss, the model adjusts its internal parameters according to the weights of each distillation loss factor. Through global distillation loss minimization optimization, the optimized debt repayment anomaly detection channel is finally obtained. The debt repayment anomaly detection channel is a model that has been distilled and optimized, and can determine whether there is a debt repayment anomaly based on the input financial data.

[0068] Further, the debt repayment anomaly distillation first space is optimized and analyzed according to the distillation loss multi-factor to obtain the debt repayment anomaly distillation second space, which includes:

[0069] According to the distillation loss multi-factor, a distillation loss multi-constraint is established, the distillation loss multi-factor including the prediction layer distillation loss, the hidden layer distillation loss, the attention layer distillation loss and the embedding layer distillation loss; based on the distillation loss multi-factor, a distillation loss analysis space is established by analyzing the debt repayment anomaly distillation first space; based on the distillation loss analysis space, the debt repayment anomaly distillation first space is analyzed and optimized according to the distillation loss multi-constraint to generate the debt repayment anomaly distillation second space.

[0070] The distillation loss multi-factor is a plurality of factors used to measure the difference between the teacher model and the student model in the distillation process. The prediction layer distillation loss is used to measure the difference between the student model and the teacher model in the prediction output, which is usually calculated by cross-entropy loss. The hidden layer distillation loss is used to measure the difference between the student model and the teacher model in the hidden layer representation, which is calculated by comparing the output of each layer of the student model and the teacher model. The attention layer distillation loss is used to compare the difference between the teacher model and the student model in the attention mechanism layer, which has a significant impact on performance, especially in natural language processing and visual tasks. The embedding layer distillation loss is used to measure the difference between the student model and the teacher model in the embedding layer. Word embedding is a general term for language models and representation learning techniques in NLP, which refers to embedding a high-dimensional space containing all word quantities into a low-dimensional continuous vector space. Each word or phrase is mapped to a vector in the real number domain. If you want to better represent words, you need to spend more memory space. Compressing the embedding matrix without sacrificing model performance is crucial for practical applications. Therefore, the mean square error is used to represent the distance between the word embeddings of the teacher model and the student model, and the student model performance is improved by minimizing it.

[0071] The distillation loss multi-constraint is a restriction imposed by the distillation loss multi-factor between different layers in the distillation process. For example, it specifies that the loss of a certain layer does not exceed a certain threshold, or requires that the loss difference between certain layers remain within a reasonable range. This ensures that the distillation losses of different layers are balanced, such as maintaining the balance of each layer's loss, ensuring that both the prediction layer loss and the hidden layer loss are fully optimized during training, without allowing a certain layer's loss to dominate the training process; limiting the loss difference to avoid excessive increase or decrease in a certain layer's loss, which can cause the distillation process to become unbalanced.

[0072] Distillation loss analysis is a loss analysis of each model in the debt anomaly distillation first space. The purpose is to determine the gap between each debt anomaly detector and the teacher model and the student model by calculating the loss of each debt anomaly detector at different levels. In this process, the debt anomaly distillation first space is analyzed according to the distillation loss multi-factor, and the distillation loss coefficients of each debt anomaly detector at different layers are calculated. Each distillation loss analysis result includes the prediction layer distillation loss coefficient, the hidden layer distillation loss coefficient, the attention layer distillation loss coefficient, and the embedding layer distillation loss coefficient of each debt anomaly detector. These coefficients describe the loss degree of each layer.

[0073] The distillation loss analysis space is a space composed of all these distillation loss analysis results. By analyzing the distillation loss of each debt anomaly detector at different layers, it can be determined which parts of the distillation process are less effective and which layers need more optimization.

[0074] The optimization analysis is a process of optimizing according to the distillation loss multi-constraint on the basis of the distillation loss analysis space. Through an optimization algorithm, the model parameters are gradually adjusted in the distillation loss analysis space to find the optimal parameter configuration that meets the distillation loss multi-constraint. The goal is to make the distillation loss of all levels within a suitable range, while minimizing the difference between the teacher model and the student model, and ensuring that the student model can effectively learn from the teacher model.

[0075] After the optimization analysis, the debt repayment abnormal distillation second space is obtained, which contains the optimized models that have met the distillation loss multi-constraint and achieved a balance between different levels of loss.

[0076] Further, a plurality of window financial data sets of the enterprise are obtained, including:

[0077] According to the multi-scale window, the financial data of the enterprise is collected to obtain a plurality of financial data streams; and the plurality of financial data streams are data cleaned to generate the plurality of window financial data sets.

[0078] The multi-scale window is used to collect financial data at different time scales, including short-term (such as monthly), medium-term (such as quarterly), and long-term (such as annual) data. Different window sizes help to analyze the financial situation of the enterprise from different time perspectives. The financial data stream is the original financial data collected by the multi-scale window. Each financial data stream represents the financial performance of the enterprise within a certain time window, including indicators such as sales, net profit, liabilities, and cash flow. Through the financial data collection of the multi-scale window, the financial data stream of the enterprise can cover different trends from short-term to long-term, thereby providing a comprehensive data basis for financial anomaly detection.

[0079] The plurality of financial data streams are data cleaned to ensure the quality of the data, including excluding erroneous, abnormal, or inconsistent financial data, such as correcting or deleting input errors; for financial data with missing values, the missing values can be filled by mean filling, median filling, or other methods, or the records containing missing values can be deleted; identifying and processing outliers that affect analysis, such as a certain quarter's net profit being abnormally high or low, which may be caused by data collection errors or extreme events. The cleaned data is organized into a plurality of window financial data sets, each containing processed financial data streams.

[0080] Further, the financial anomaly multi-dimensional factors include debt repayment abnormality, profit abnormality, and fund chain abnormality.

[0081] The financial anomaly multi-dimensional factors include debt repayment anomaly, profit anomaly and fund chain anomaly. The debt repayment anomaly reflects the ability of the enterprise to repay debts, and is usually measured by indicators such as debt ratio, cash flow and interest coverage ratio. The profit anomaly reflects the profitability of the enterprise, and is usually measured by indicators such as net profit, gross profit margin and operating income. The fund chain anomaly reflects the status of the fund flow of the enterprise, and is usually measured by indicators such as cash flow, accounts receivable and inventory turnover.

[0082] In summary, the enterprise financial data anomaly detection method based on model distillation provided by the embodiments has the following technical effects:

[0083] By obtaining multiple window financial data sets of the enterprise, the financial data of the enterprise at different time scales can be analyzed in detail. According to the financial anomaly multi-dimensional factors, the multiple window financial anomaly trigger vectors are established by triggering and combing the multiple window financial data sets. The potential financial anomaly signals can be identified from multiple dimensions. Each window financial anomaly trigger vector represents the risk indication of different financial dimensions, thereby providing clear feature information for subsequent anomaly detection, facilitating more accurate identification and classification of abnormal situations. By training multiple sources according to the financial anomaly multi-dimensional factors, the financial anomaly detection architecture includes a debt repayment anomaly detection model set, a profit anomaly detection model set and a fund chain anomaly detection model set. This multi-model architecture can comprehensively analyze the financial health of the enterprise from multiple angles. It can not only detect different types of financial anomalies, but also optimize different types of financial data, thereby improving detection accuracy and comprehensiveness. By distilling the parameters of the financial anomaly detection architecture according to the distillation learning factors, the knowledge in the teacher model can be transferred to the student model in a more efficient and simplified manner, thereby improving detection accuracy while reducing the consumption of computing resources. The distillation decision multi-dimensional space provides multiple optimization parameter configurations, which can effectively guide the selection and configuration of the model, so that the distillation process can achieve efficient knowledge transfer in multiple dimensions. By introducing the distillation loss multi-level optimization mechanism to iteratively optimize the distillation decision multi-dimensional space, the optimal distillation parameter configuration can be found. This mechanism improves the learning effect of the model, so that multiple financial anomaly detection channels reach a balance between different levels of loss, thereby achieving more accurate financial anomaly detection. By inputting the multiple window financial anomaly trigger vectors into the financial anomaly detection channel, the enterprise financial anomaly detection report is obtained. This process combines the data of different windows and the results of multi-channel detection, so that the financial anomaly detection can cover multiple dimensions and time scales, enhancing the accuracy and comprehensiveness of the report. Enterprises can understand the potential risks in their financial situation through the report and make timely risk management and decision-making accordingly.

[0084] Embodiment two, based on the same inventive concept as the model distillation-based enterprise financial data anomaly detection method in the preceding embodiments, as shown in Figure 2 The embodiment of the present application provides a model distillation-based enterprise financial data anomaly detection system, which comprises:

[0085] A trigger carding module 10 is configured to obtain a plurality of window financial data sets of an enterprise, and trigger carding is performed on the plurality of window financial data sets according to financial anomaly multidimensional factors, thereby establishing a plurality of window financial anomaly trigger vectors.

[0086] A multi-source training module 20 is configured to perform multi-source training according to the financial anomaly multidimensional factors, thereby establishing a financial anomaly detection architecture, wherein the financial anomaly detection architecture comprises a debt repayment anomaly detection model set, a profit anomaly detection model set and a fund chain anomaly detection model set.

[0087] A distillation parameter decision module 30 is configured to perform distillation parameter decision on the financial anomaly detection architecture according to a distillation learning factor, thereby obtaining a distillation decision multidimensional space.

[0088] An iterative optimization module 40 is configured to introduce a distillation loss multi-level optimization mechanism to perform iterative optimization on the distillation decision multidimensional space, thereby generating a financial anomaly detection channel, wherein the financial anomaly detection channel comprises a debt repayment anomaly detection channel, a profit anomaly detection channel and a fund chain anomaly detection channel.

[0089] A detection report generation module 50 is configured to input the plurality of window financial anomaly trigger vectors into the financial anomaly detection channel, thereby obtaining an enterprise financial anomaly detection report.

[0090] Further, the trigger carding module 10 is configured to perform the following operation steps:

[0091] Trigger mining is performed according to the financial anomaly multidimensional factors, thereby establishing a debt repayment anomaly trigger detection tree, a profit anomaly trigger detection tree and a fund chain anomaly trigger detection tree; an nth window financial data set is extracted from the plurality of window financial data sets, and a plurality of financial data features are obtained by performing feature recognition on the nth window financial data set, wherein n is a positive integer; an nth debt repayment anomaly trigger vector is obtained by performing trigger carding on the plurality of financial data features according to the debt repayment anomaly trigger detection tree; an nth profit anomaly trigger vector is obtained by performing trigger carding on the plurality of financial data features according to the profit anomaly trigger detection tree; and an nth fund chain anomaly trigger vector is obtained by performing trigger carding on the plurality of financial data features according to the fund chain anomaly trigger detection tree, and the nth debt repayment anomaly trigger vector and the nth profit anomaly trigger vector are combined to construct an nth window financial anomaly trigger vector.

[0092] Further, the trigger carding module 10 is configured to perform the following operation steps:

[0093] According to the financial data feature sample set, the financial anomaly multidimensional factor is triggered to search, and each debt settlement anomaly trigger sample cluster corresponding to each financial data feature sample is obtained; the cluster confidence evaluation set is obtained by evaluating the cluster confidence of each debt settlement anomaly trigger sample cluster; the cluster confidence fusion is performed on each debt settlement anomaly trigger sample cluster according to the cluster confidence evaluation set, and a debt settlement anomaly trigger confidence set is obtained; and the debt settlement anomaly trigger detection tree is constructed by taking the financial data feature sample set as a plurality of trigger detection input nodes and taking the debt settlement anomaly trigger confidence set as a plurality of trigger detection output nodes.

[0094] Further, the trigger analysis module 10 is used to perform the following operation steps:

[0095] The plurality of financial data features are input into the debt settlement anomaly trigger detection tree, and a plurality of debt settlement anomaly trigger coefficients are obtained; it is judged whether the plurality of debt settlement anomaly trigger coefficients are greater than or equal to a debt settlement anomaly trigger threshold, and a plurality of trigger judgment results are obtained; the plurality of financial data features are selected according to the plurality of trigger judgment results, and a debt settlement anomaly trigger feature space is obtained; and the debt settlement anomaly trigger feature space is used for feature fusion, and the nth debt settlement anomaly trigger vector is output.

[0096] Further, the distillation parameter decision module 30 is used to perform the following operation steps:

[0097] The distillation learning factor includes the number of teacher models, the attribute of the student model, and the distillation hyperparameter; the distillation parameter decision is performed on the debt settlement anomaly detection model set according to the distillation learning factor, a distillation decision first space is established; the distillation parameter decision is performed on the profit anomaly detection model set according to the distillation learning factor, a distillation decision second space is established; the distillation parameter decision is performed on the fund chain anomaly detection model set according to the distillation learning factor, a distillation decision third space is established, and the distillation decision multi-dimensional space is generated by combining the distillation decision first space and the distillation decision second space.

[0098] Further, the iterative optimization module 40 is used to perform the following operation steps:

[0099] The distillation learning is performed on the debt settlement anomaly detection model set according to the distillation decision first space, a debt settlement anomaly distillation first space is obtained, and the debt settlement anomaly distillation first space includes a plurality of debt settlement anomaly detectors; the distillation learning is performed on the debt settlement anomaly distillation first space according to the distillation loss multi-factor, a debt settlement anomaly distillation second space is obtained; the weight distribution is performed according to the distillation loss multi-factor, and a global distillation loss function is established; and the global distillation loss minimization optimization is performed on the debt settlement anomaly distillation second space according to the global distillation loss function, and the debt settlement anomaly detection channel is generated.

[0100] Further, the iteration optimization module 40 is configured to perform the following operation steps:

[0101] According to the constraint configuration of the distillation loss multi-factor, the distillation loss multi-constraint is established, and the distillation loss multi-factor includes the prediction layer distillation loss, the hidden layer distillation loss, the attention layer distillation loss and the embedding layer distillation loss; based on the distillation loss multi-factor, the debt abnormal distillation first space is analyzed for distillation loss to establish a distillation loss analysis space; based on the distillation loss analysis space, the debt abnormal distillation first space is optimized and analyzed according to the distillation loss multi-constraint to generate the debt abnormal distillation second space.

[0102] Further, the trigger analysis module 10 is configured to perform the following operation steps:

[0103] According to the multi-scale window, the financial data of the enterprise is collected to obtain a plurality of financial data streams; the plurality of financial data streams are data cleaned to generate a plurality of window financial data sets.

[0104] Further, the financial abnormal multi-dimensional factor includes debt abnormality, profit abnormality and fund chain abnormality.

[0105] Through the foregoing detailed description of the enterprise financial data anomaly detection method based on model distillation, those skilled in the art can clearly understand the enterprise financial data anomaly detection system based on model distillation in the embodiment. Since it corresponds to the method disclosed in the embodiment, it is described relatively simply, and the relevant part is referred to the method part description.

[0106] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting anomalies in enterprise financial data based on model distillation, characterized in that, The method comprises: obtaining a plurality of window financial data sets of an enterprise, and performing triggered grooming on the plurality of window financial data sets according to financial anomaly multidimensional factors to establish a plurality of window financial anomaly triggered vectors; performing multi-source training according to the financial anomaly multidimensional factors to establish a financial anomaly detection architecture, the financial anomaly detection architecture comprising a debt repayment anomaly detection model set, a profit anomaly detection model set, and a fund chain anomaly detection model set; performing distillation parameter decision on the financial anomaly detection architecture according to a distillation learning factor to obtain a distillation decision multidimensional space; introducing a distillation loss multi-level optimization mechanism to iteratively optimize the distillation decision multidimensional space to generate a financial anomaly detection channel, the financial anomaly detection channel comprising a debt repayment anomaly detection channel, a profit anomaly detection channel, and a fund chain anomaly detection channel; inputting the plurality of window financial anomaly triggered vectors into the financial anomaly detection channel to obtain an enterprise financial anomaly detection report; wherein the distillation loss multi-level optimization mechanism comprises: performing distillation learning on the debt repayment anomaly detection model set according to a distillation decision first space to obtain a debt repayment anomaly distillation first space, the debt repayment anomaly distillation first space comprising a plurality of debt repayment anomaly detectors; performing optimization analysis on the debt repayment anomaly distillation first space according to a distillation loss multi-element factor to obtain a debt repayment anomaly distillation second space; performing weight allocation according to the distillation loss multi-element factor to establish a global distillation loss function; performing global distillation loss minimization optimization on the debt repayment anomaly distillation second space according to the global distillation loss function to generate the debt repayment anomaly detection channel; wherein performing optimization analysis on the debt repayment anomaly distillation first space according to a distillation loss multi-element factor to obtain a debt repayment anomaly distillation second space comprises: performing constraint configuration according to the distillation loss multi-element factor to establish a distillation loss multi-element constraint, the distillation loss multi-element factor comprising a prediction layer distillation loss, a hidden layer distillation loss, an attention layer distillation loss, and an embedding layer distillation loss; performing distillation loss analysis on the debt repayment anomaly distillation first space based on the distillation loss multi-element factor to establish a distillation loss analysis space; based on the distillation loss analysis space, performing optimization analysis on the debt repayment anomaly distillation first space according to the distillation loss multi-element constraint to generate the debt repayment anomaly distillation second space. 2.The model-based distillation based enterprise financial data anomaly detection method of claim 1, wherein, performing triggered grooming on the plurality of window financial data sets according to the financial anomaly multidimensional factors to establish a plurality of window financial anomaly triggered vectors comprises: performing triggered mining according to the financial anomaly multidimensional factors to establish a debt repayment anomaly triggered detection tree, a profit anomaly triggered detection tree, and a fund chain anomaly triggered detection tree; extracting an nth window financial data set from the plurality of window financial data sets, and performing feature recognition on the nth window financial data set to obtain a plurality of financial data features, n being a positive integer; performing triggered grooming on the plurality of financial data features according to the debt repayment anomaly triggered detection tree to obtain an nth debt repayment anomaly triggered vector; performing triggered grooming on the plurality of financial data features according to the profit anomaly triggered detection tree to obtain an nth profit anomaly triggered vector; According to the fund chain abnormal trigger detection tree, the plurality of financial data features are triggered and combed to obtain an n th fund chain abnormal trigger vector. In combination with the n th debt repayment abnormal trigger vector and the n th profit abnormal trigger vector, a n th window financial abnormal trigger vector is constructed. 3.The model-based distillation method for enterprise financial data anomaly detection of claim 2, wherein, According to the financial abnormality multi-dimensional factor, triggered mining is performed, including: According to the financial data feature sample set, the financial abnormality multi-dimensional factor is triggered and searched to obtain each debt repayment abnormal trigger sample cluster corresponding to each financial data feature sample; The cluster confidence evaluation set is obtained by evaluating the cluster confidence of each debt repayment abnormal trigger sample cluster; According to the cluster confidence evaluation set, the cluster confidence of each debt repayment abnormal trigger sample cluster is fused respectively to obtain a debt repayment abnormal trigger confidence set; The financial data feature sample set is taken as a plurality of trigger detection input nodes, and the debt repayment abnormal trigger confidence set is taken as a plurality of trigger detection output nodes to construct the debt repayment abnormal trigger detection tree. 4.The model-based distillation method for enterprise financial data anomaly detection of claim 2, wherein, According to the fund chain abnormal trigger detection tree, the plurality of financial data features are triggered and combed to obtain an n th fund chain abnormal trigger vector. In combination with the n th debt repayment abnormal trigger vector and the n th profit abnormal trigger vector, a n th window financial abnormal trigger vector is constructed. The plurality of financial data features are input into the debt repayment abnormal trigger detection tree to obtain a plurality of debt repayment abnormal trigger coefficients; It is judged whether the plurality of debt repayment abnormal trigger coefficients are greater than or equal to a debt repayment abnormal trigger threshold to obtain a plurality of trigger judgment results; According to the plurality of trigger judgment results, the plurality of financial data features are selected to obtain a debt repayment abnormal trigger feature space; According to the debt repayment abnormal trigger feature space, feature fusion is performed to output the n th debt repayment abnormal trigger vector. 5.The model-based distillation method for enterprise financial data anomaly detection of claim 1, wherein, According to the distillation learning factor, distillation parameter decision is performed on the financial abnormality detection architecture to obtain a distillation decision multi-dimensional space, including: The distillation learning factor includes a teacher model number, a student model attribute, and a distillation hyperparameter; According to the distillation learning factor, distillation parameter decision is performed on the debt repayment abnormality detection model set to establish a distillation decision first space; According to the distillation learning factor, distillation parameter decision is performed on the profit abnormality detection model set to establish a distillation decision second space; According to the distillation learning factor, distillation parameter decision is performed on the fund chain abnormality detection model set to establish a distillation decision third space. In combination with the distillation decision first space and the distillation decision second space, the distillation decision multi-dimensional space is generated. 6.The model-based distillation method for enterprise financial data anomaly detection of claim 1, wherein, A plurality of window financial data sets of an enterprise are obtained, including: According to a multi-scale window, financial data of the enterprise is collected to obtain a plurality of financial data streams; The plurality of financial data streams are cleaned to generate the plurality of window financial data sets. 7.The model-based distillation method for enterprise financial data anomaly detection of claim 1, wherein, The financial abnormality multi-dimensional factor includes debt repayment abnormality, profit abnormality, and fund chain abnormality.

8. A system for detecting anomalies in enterprise financial data based on model distillation, the system comprising: A system for implementing the model distillation-based enterprise financial data abnormality detection method of any one of claims 1-7, the system comprising: A trigger combing module is configured to obtain a plurality of window financial data sets of an enterprise, and to perform trigger combing on the plurality of window financial data sets according to a financial abnormality multi-dimensional factor to establish a plurality of window financial abnormal trigger vectors; The multi-source training module is configured to perform multi-source training according to the financial anomaly multi-dimensional factor, and establish a financial anomaly detection architecture, wherein the financial anomaly detection architecture comprises a debt repayment anomaly detection model set, a profit anomaly detection model set and a fund chain anomaly detection model set; The distillation parameter decision module is configured to perform distillation parameter decision on the financial anomaly detection architecture according to a distillation learning factor, and obtain a distillation decision multi-dimensional space; The iterative optimization module is configured to introduce a distillation loss multi-level optimization mechanism to perform iterative optimization on the distillation decision multi-dimensional space, and generate a financial anomaly detection channel, wherein the financial anomaly detection channel comprises a debt repayment anomaly detection channel, a profit anomaly detection channel and a fund chain anomaly detection channel; The detection report generation module is configured to input the plurality of window financial anomaly trigger vectors into the financial anomaly detection channel, and obtain an enterprise financial anomaly detection report.

Images