A financial data verification method based on RPA

Abstract

The application discloses a kind of financial data verification methods based on RPA, it includes extracting financial data from the process of financial system;Data verification result is obtained by logical verification to financial data, the logical verification includes: judging whether there is data fraud risk, if yes, the data is marked as fraud identification;Judge whether there is single field extreme value, if yes, the numerical value of the single field is marked as extreme value identification;Judge whether there is multi-field combination anomaly, if yes, the numerical value of the multi-field is marked as abnormal identification;Finally, the data verification result is aggregated and generates verification report and submits subsequent process.By the method, the technical problem that "garbage in, garbage out" of data source can be solved when using RPA to process financial data lacks identification function.

Description

Technical Field

[0001] This invention relates to the field of data application and verification technology, specifically to a financial data verification method based on RPA, which is mainly applied to the robotic process automation of financial work. Background Technology

[0002] With the development of computer technology and social progress, Robotic Process Automation (RPA) has rapidly developed and been used across various industries. It is an application software technology that simulates human operation on a computer through software programs, automatically executing repetitive and rule-based tasks according to preset rules. Its greatest feature is the redefinition of human value; RPA handles repetitive labor, allowing employees to focus on higher-level tasks requiring judgment and creativity. This value is particularly evident in positions that require processing large amounts of data and performing repetitive tasks.

[0003] Tasks in financial systems, such as reconciliation and verification of statements, cross-system data synchronization, data retrieval from multiple systems, and invoice processing, are characterized by large data volumes and high repetitiveness. Therefore, RPA has been widely used in the financial field. However, despite its advantages of processing speeds far exceeding human capabilities, its tireless nature, and its ability to liberate finance personnel from repetitive tasks, its drawbacks are also obvious. For example, it can lead to a "garbage in, garbage out" data source: while RPA can accurately replicate manual operations, if the input data itself is incorrect (such as incorrect expense reimbursement amounts submitted by upstream business departments), RPA will "faithfully" process the erroneous data and carry the error to downstream systems. This automates errors, accelerates error propagation, and blurs the lines of responsibility. Summary of the Invention

[0004] The main objective of this invention is to provide a financial data verification method based on RPA, which aims to solve the technical problem mentioned in the background art of RPA lacking identification function in the application of financial systems, resulting in "garbage in, garbage out" of data sources.

[0005] To achieve the above objectives, this invention provides a financial data verification method based on RPA, comprising: Financial data is extracted from the financial system's workflow; the financial data is logically validated to obtain data validation results. The logical validation includes: determining whether there is a risk of data fraud, and if so, marking the data as a fraud identifier; determining whether there are extreme values ​​in a single field, and if so, marking the value of the single field as an extreme value identifier; determining whether there are abnormal combinations of multiple fields, and if so, marking the values ​​of the multiple fields as an anomaly identifier; aggregating the data validation results and generating a validation report; and submitting the financial data and the validation report to subsequent processes for execution.

[0006] Furthermore, the method for determining whether there is a risk of data fraud is as follows: Construct a dataset {x1, x2, ..., x} from the financial data of the same type. n}; Extract the first non-zero digit of each data point in the dataset; The number of times the non-zero digits appear is counted, and the ratio of the number of times the non-zero digits appear to n is calculated. When the difference between the ratio and the preset expected value is not less than the preset threshold, there is a risk of data fraud.

[0007] Furthermore, when the ratio is less than a preset threshold, the statistic M is calculated: ; When the statistic M is greater than a preset threshold, there is a risk of data fraud. Among them, O d E represents the number of times the number d is counted. d Let d be the expected number of times. .

[0008] Furthermore, the method for determining whether a single field extreme value exists is as follows: Construct a dataset {x1, x2, ..., x} from the financial data of the same type. n}; Calculate the maximum absolute deviation of the dataset. ; If G > G c The critical value, then the x i For single-field extreme values; The G c The method for calculating the critical value is as follows: ; Where n is the number of data points, t α / (2n),n 2 is the critical value of the t-distribution. The significance level is denoted by s, and the standard deviation of the dataset is s. is the mean of the dataset.

[0009] Furthermore, the method for determining whether there is an anomaly in the combination of multiple fields is as follows: Extract several features from the financial data to form a feature vector group; The feature vector set is input into the trained self-learning model to calculate the anomaly score; When the abnormal score is greater than a preset threshold, the financial data has multiple field combination anomalies.

[0010] Furthermore, the self-learning model has T data sets, and the training method for the T data sets is as follows: Obtain historical financial data from the financial system; S1: Randomly select m data points from the historical financial data to form a data group; S2: Randomly select one feature from the features of the financial data; S3: Randomly select a segmentation value point for one of the features; the value of the segmentation value point is between the minimum and maximum values ​​in the data of the current feature; S4: Based on the values ​​of the segmentation points, divide the current data group into two sub-data groups, left and right; S5: Repeat steps S2 to S4 until the set conditions are met, then the training of the data set is complete; Repeat steps S1 to S5 until T data sets are obtained; The conditions are as follows: The number of sub-data groups reaches the maximum limit value ceil(log2(m)); where ceil represents rounding up; Or the current subgroup contains only one piece of data; Or all feature values ​​of all data in the current subgroup are exactly the same.

[0011] Furthermore, the method for calculating the abnormal score is as follows: Abnormal scores , Where h(x) is the number of times the value x is split from the beginning of a certain data group in the self-learning model until the data is isolated; E(h(x)) represents the average number of splits of data point x in T data groups; and c(p) represents the expected value of the average number of splits given sample p.

[0012] Furthermore, after generating the verification report and before submitting it for subsequent process execution, it is determined whether the financial data is text data. If so, a text comparison method is used to verify the financial data.

[0013] Furthermore, the text comparison method is characterized by: The text data is compared with preset text data. When the similarity between the text data and the preset text data is greater than a threshold, the text data of the financial data is adjusted to the preset text data and the financial data is marked.

[0014] Furthermore, the method for calculating the text data similarity is as follows: Generate a character binary grammar set by combining the text data with the preset text data; Construct a union vector space from the character binary grammar sets of both; Create frequency vectors for each of the two strings respectively; Calculate the cosine similarity between the frequency vectors of the two vectors.

[0015] In the technical solution of this invention, by loading this technical solution into a financial system with RPA, the RPA, while processing financial data according to certain rules, can identify common problems encountered in daily work, such as false expense reports and data entry errors, by judging whether there is a risk of data fraud, whether there are extreme values ​​in a single field, and whether there are abnormal combinations of multiple fields. This gives the RPA the initial ability to identify and verify data logic anomalies. It is understandable that by using the technical solution of this invention, the RPA will no longer be a "rigid" machine, but rather equivalent to equipping its "hands and feet" with a "brain." Using the judgment ability of this "brain," the presence of garbage data will be identified, changing from "garbage in, garbage out" to "garbage in, report out." This not only purifies the garbage data in the financial system but also improves the work efficiency of financial personnel, enabling modern enterprises to move from automation to intelligence in their financial work. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of a financial data verification method based on RPA according to an embodiment of the present invention; Figure 2 This is a chi-square value table at different significance levels according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a training method for a self-learning model according to an embodiment of the present invention; Figure 4 This is a schematic diagram of a text data similarity calculation method according to an embodiment of the present invention; The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Financial systems and operations are an indispensable part of modern corporate governance, characterized by large data volumes, high repetition, and heavy workloads. The emergence of Robotic Process Automation (RPA) has significantly improved efficiency. However, due to RPA's unique characteristic of executing processes solely according to pre-defined rules without addressing the validity of the data itself, if the data is flawed, RPA will simply pass it on unchanged to subsequent processes. This increases the difficulty of tracing problems and introduces new challenges to financial work.

[0020] Common data integrity issues in daily financial work include: misplaced decimal points, confused units, reversed numbers, and incorrect plus or minus signs. In addition, there are subjective human issues, such as fabricating or tampering with data, such as falsifying reimbursement amounts or transaction amounts. There are even more subtle issues, such as individual amounts that seem fine, but when combined, statistical analysis may reveal contradictions in the amounts, transaction times, supplier types, and frequencies.

[0021] The aforementioned issues such as decimal point misalignment, unit confusion, number reversal, and incorrect plus / minus signs can be categorized as single-field extreme values ​​when they occur.

[0022] If data is tampered with, it may be revealed in the statistics of a large amount of historical data.

[0023] Even more subtle data anomalies may be found in combinations of multiple fields, such as the aforementioned combinations of amount, transaction time, supplier type, and frequency of occurrence.

[0024] Therefore, the main objective of this invention is to propose a financial data verification method based on RPA, which aims to solve the aforementioned technical problem of "garbage in, garbage out" in financial systems where RPA only executes logic but does not recognize the rationality of data.

[0025] like Figure 1 As shown, one embodiment of the technical solution of the present invention is as follows: Financial data is extracted from the financial system's workflow; the financial data is logically validated to obtain data validation results. The logical validation includes: determining whether there is a risk of data fraud, and if so, marking the data as a fraud identifier; determining whether there are extreme values ​​in a single field, and if so, marking the value of the single field as an extreme value identifier; determining whether there are abnormal combinations of multiple fields, and if so, marking the values ​​of the multiple fields as an anomaly identifier; aggregating the data validation results and generating a validation report; and submitting the financial data and the validation report to subsequent processes for execution.

[0026] The method for determining whether there is a risk of data fraud is as follows: Construct a dataset {x1, x2, ..., x} from the financial data of the same type. n For example, RPA needs to process the settlement of a batch of 500 invoices. Assume that the amounts of these 500 invoices are {125.40, 489.00, 85.10, 2340.55, 18.90, 4999.00, 312.80, ...] (a total of 500 amounts, the rest are omitted, this is just to provide a data example)}, ignore or remove zero values ​​and negative values, only focus on the absolute value of the amount.

[0027] Extract the first non-zero digit of each data point in the dataset. The resulting list of first digits is: {1, 4, 8, 2, 1, 4, 3, ...} (a total of 500 digits).

[0028] Count the occurrences of the non-zero digits. Count the occurrences of each digit (1-9) in the cleaned list of first digits and calculate their frequency. Calculate the ratio of the occurrences of the non-zero digits to n (where n is 500): Benford's Law is a mathematical law concerning the distribution of the first digit in naturally generated sets of numbers. In a large, unmanipulated sample of numbers, the probability of starting with the digit "1" is the highest (approximately 30.1%), while the probability of starting with "9" is the lowest (approximately 4.6%). Real, naturally occurring financial data (such as transaction amounts and invoice amounts) generally follows Benford's Law. However, artificially fabricated data deviates significantly from this law's expected distribution because the fabricator tends to use numbers evenly (e.g., making numbers start with 3, 4, 5, or 6 to appear more "random"), or to circumvent certain approval thresholds (e.g., a large number of forged invoices below 5000 yuan, leading to an abnormally high frequency of the first digit "4"). Based on this law, the expected frequencies of the first digits 1 through 9 can be calculated: .

[0029] The "actual frequency" of these 500 numbers is compared with the "expected probability P". d "Compare the data to identify significant deviations. If the difference between the ratio and the preset expected value is not less than a preset threshold, there is a risk of data fraud. For example, in these 500 numbers, as shown in the table below:" In this example, the actual frequency of the first digit "4" (18.0%) is significantly higher than the preset probability of 9.7%, a difference of +8.3%. Based on this significant difference, it can be inferred that there may be an authorization threshold of 5,000 yuan in the financial approval process (invoices exceeding 5,000 yuan require higher-level approval). To circumvent this approval process, someone may have forged a large number of invoices with amounts between 4,000 and 4,999 yuan. Without verification, the RPA system would directly process these potentially fraudulent invoices.

[0030] This preset threshold can be set according to the actual situation. Based on practical work experience, it is recommended to set it between 5% and 10%.

[0031] Furthermore, to automate the verification process, the "bias" can be quantified for further verification. A chi-square test is typically used to determine whether the difference between the actual and expected distributions is statistically significant. When the ratio is less than a preset threshold, the statistic M is calculated. ; Among them, O d E represents the number of times the number d is counted. d Let d be the expected number of times. .

[0032] Let's take these 500 numbers as an example again: (n=500) The degrees of freedom (df) are determined because the constraint condition is that the total number of iterations (n) is fixed, so the degrees of freedom = number of categories - 1. In this embodiment, the degrees of freedom = 9 - 1 = 8.

[0033] Choose a significance level (α), which is usually 0.05 or 0.01. Based on the degrees of freedom df and the significance level α, find the critical value from the chi-square distribution table.

[0034] In this embodiment, for example, when df=8 and α=0.05, such as Figure 2 As shown, the corresponding chi-square value is 15.51.

[0035] When the statistic M is greater than a preset threshold, there is a risk of data fraud. As shown in this embodiment, the statistic M is 44.36, which is greater than 15.51. Therefore, it can be concluded that there is a risk of data fraud in this set of financial data with n=500. The preset threshold here is a chi-square distribution table.

[0036] In practice, in addition to paying attention to outliers, it is also necessary to focus on extreme values. Extreme values ​​are those that deviate significantly from normal business patterns and may have a major impact on settlement results. Compared to general outliers, extreme values ​​are of greater concern because their absolute magnitude exceeds a reasonable range, they have a destructive impact on statistical distribution, and they may cause settlement imbalances or audit risks.

[0037] Furthermore, the method for determining whether a single field has extreme values ​​is as follows: Construct a dataset {x1, x2, ..., x} from the financial data of the same type. n}; Calculate the maximum absolute deviation of the dataset. ; If G > G c The critical value, then the x i For single-field extreme values; The G c The method for calculating the critical value is as follows: ; Where n is the number of data points, t α / (2n),n 2 is the critical value of the t-distribution. The significance level is denoted by s, and the standard deviation of the dataset is s. is the mean of the dataset.

[0038] Taking accounts payable as an example, suppose there are 10 accounts payable (n=10), in ten thousand yuan. {15.2, 18.5, 16.8, 17.3, 16.1, 19.2, 17.8, 16.5, 18.1, 152.0}, then Number of data points n=10 mean = (15.2 + 18.5 + 16.8 + 17.3 + 16.1 + 19.2 + 17.8 + 16.5 + 18.1 + 152.0) / 10 = 31.75 Standard deviation s = =42.87; Calculate the maximum absolute deviation G = = 2.805; G c The critical value calculation method is set according to the t-distribution in a preset table, as shown in the appendix. Figure 2 As shown, at the significance level When G is 0.05, corresponding to n=10, c The critical value is 2.290. Therefore... Since 2.808 > 2.290, 152.0 is an extreme value. Therefore, it can be determined that the value of 152.0 is an outlier in the financial system, requiring a reminder or warning.

[0039] In actual financial work, some complex situations exist. For example, fraudsters often ensure that a single field (such as amount) looks "normal," but the combination of multiple fields will expose its irrationality. For instance, a junior employee (low-level position) might reimburse a huge "business entertainment expense," or a logistics department might apply to purchase a large quantity of "core server components." Individual fields may appear fine, but the combination is highly unusual. Therefore, to enable RPA to detect such problems in a large amount of financial data, a multi-dimensional combination judgment model is introduced. This further determines whether there are anomalies in the combination of multiple fields. In this embodiment of the invention, a method for determining whether there are anomalies in the combination of multiple fields is provided as follows: Extract several features from the financial data to form a feature vector group; The feature vector set is input into the trained self-learning model to calculate the anomaly score; When the anomaly score exceeds a preset threshold, the financial data is considered to have a multi-field combination anomaly. Assuming the preset threshold is 0.7, if the calculated anomaly score for this set of feature vectors is less than 0.7, it is considered normal; if the calculated anomaly score for this set of feature vectors is greater than 0.7, it is considered abnormal.

[0040] Furthermore, the self-learning model has T data sets, and the training method for the T data sets is as follows: Obtain historical financial data from the financial system; S1: Randomly select m data points from the historical financial data to form a data group; S2: Randomly select one feature from the features of the financial data; S3: Randomly select a segmentation value point for one of the features; the value of the segmentation value point is between the minimum and maximum values ​​in the data of the current feature; S4: Based on the values ​​of the segmentation points, divide the current data group into two sub-data groups, left and right; S5: Repeat steps S2 to S4 until the set conditions are met, then the training of the data set is complete; Repeat steps S1 to S5 until T data sets are obtained; The conditions are as follows: The number of sub-data groups reaches the maximum limit value ceil(log2(m)); where ceil represents rounding up; Or the current subgroup contains only one piece of data; Or all feature values ​​of all data in the current subgroup are exactly the same.

[0041] Furthermore, the method for calculating the abnormal score is as follows: Abnormal scores , Where h(x) is the number of times the value x is split from the beginning of a certain data group in the self-learning model until the data is isolated; E(h(x)) represents the average number of splits of data point x in T data groups; and c(p) represents the expected value of the average number of splits given sample p.

[0042] The example data provided in this embodiment of the invention assumes the following: historical normal settlement data from the past 1-2 years is exported from the ERP system. Key field combinations are selected: reimbursement amount, reimbursement date, and job level. T=100 data groups are selected. Assume there are 10 historical data entries and 1 pending invoice reimbursement document. For example, the pending reimbursement document data might have the following characteristics: reimbursement amount = 9800, reimbursement date (converted to "days until the end of the month") = 1, job level = 3.

[0043] At this point, the feature vector set is {reimbursement amount, reimbursement date, job level}. This set of values ​​is input into a pre-trained self-learning model. Assuming the model has 100 data sets, the anomaly score of this feature vector set in these 100 data sets is calculated. Then, the average of the anomaly scores obtained from the 100 data sets is calculated as the final result. In this embodiment, X... new =Place {9800, 1, 3} into each data group and iterate through it.

[0044] In the first data group: the root node splitting condition (assuming the selected job level < 2.5, this is a random selection; the job level is randomly selected from the reimbursement amount, reimbursement date, and job level; 2.5 is a random selection from all job level values ​​that appear in the historical data), because 3 > 2.5, proceed to the right sub-data group. The right sub-data group might directly isolate it to a data node using the condition that the amount > 10000. Assume the path length h1 (X new =2.

[0045] In the second data group: it might first be split using the number of days until the end of the month < 0.5 (the chosen splitting condition and value are randomly selected). Since 1 > 0.5, it moves to the right subgroup, and is then separated after another split. Assume the path length h2 (X... new =3.

[0046] This process continues until the path length values ​​for 100 data sets have been calculated.

[0047] Then calculate the average path length across the 100 data sets. Assume E(h(X)) new =3.5.

[0048] The data volume p=11 (10 historical records + 1 new record. In actual training, only historical data may be used. Here, the value is taken to be small for ease of calculation and understanding).

[0049] Calculate the normalization factor: C(11)=2H(10) - ≈2(In(10)+0.5772)-1.818≈2(2.3026+0.5772)-1.818≈3.9416.

[0050] Calculate the abnormal score s(X) new ,11)= ≈ 0.536.

[0051] Therefore, a score of 0.536, slightly higher than 0.5, indicates a certain probability of an anomaly, but it is not extreme. If most normal scores are between 0.3 and 0.45, then 0.536 warrants attention. If the threshold is set at 0.55, this point passes; if the threshold is set at 0.5, this point is marked as "requires review".

[0052] To describe the training process of a self-learning model, such as Figure 3 As shown, the self-learning model training method provided in this embodiment is as follows: Obtain historical financial data from the financial system; S1: Randomly select m data points from the historical financial data to form a data group; S2: Randomly select one feature from the features of the financial data; S3: Randomly select a segmentation value point for one of the features; the value of the segmentation value point is between the minimum and maximum values ​​in the data of the current feature; S4: Based on the values ​​of the segmentation points, divide the current data group into two sub-data groups, left and right; S5: Repeat steps S2 to S4 until the set conditions are met, then the training of the data set is complete; Repeat steps S1 to S5 until T data sets are obtained; The conditions are as follows: The number of sub-data groups reaches the maximum limit value ceil(log2(m)); (ceil means rounding up) or the current sub-data group contains only one data or all feature values ​​of all data in the current sub-data group are exactly the same.

[0053] Suppose we have a historical dataset X containing 50,000 records. From this dataset X, we randomly select ψ samples (ψ is a preset parameter, typically 128 or 256). This sampling process is performed independently for each dataset. We then randomly select a feature q', for example, "reimbursement amount" among reimbursement amount, reimbursement date, and job level.

[0054] Then, randomly select a dividing point, such as 1010 yuan, which is between the minimum and maximum values ​​of all "reimbursement amounts" data in the current node. Then, based on the condition (the value of feature q', such as 1010 yuan), divide the data of the current node into two sub-data groups, left and right.

[0055] Next, randomly select the next feature q'', such as "reimbursement date" (or possibly "reimbursement amount," this process is random). Randomly select a dividing point, such as 5 days, which is between the minimum and maximum values ​​of all "reimbursement date" data in the current node. Then, based on the condition (the value of feature q'', such as 5 days), divide the data of the current node into two sub-data groups, left and right.

[0056] Repeat the two steps above until the number of sub-data sets reaches the maximum limit ceil(log2(ψ)) (ceil means rounding up), or the current sub-data set contains only one data point, or all feature values ​​of all data points in the current sub-data set are exactly the same. At this point, the training of the first data set is complete.

[0057] Then, from the historical dataset X containing 50,000 records, randomly select ψ samples and repeat the above steps until the second dataset is trained. Repeat this process until 100 datasets are trained, thus completing the training process.

[0058] For training self-learning models, training methods can be performed on an annual, quarterly, monthly, or user-defined time basis. Generally, the self-learning model can be updated and trained based on historical data at regular intervals.

[0059] In daily work, in addition to verifying numerical financial data, it is also necessary to verify text-type financial data. For example, "Beijing Science and Technology Co., Ltd." is wrongly entered as "Beijing Science and Technology Company". Therefore, further, after generating the verification report and before submitting it to the subsequent process for execution, it is determined whether the financial data is text data. If so, the text comparison method is also used to verify the financial data.

[0060] The text comparison method is as follows: Compare the text data with the preset text data. When the similarity between the text data and the preset text data is greater than the threshold, adjust the text data of the financial data to the preset text data and mark the financial data.

[0061] As Figure 4 shown, the method for calculating the similarity of the text data is as follows: Generate a character bigram set from the text data and the preset text data; Construct a union vector space from the character bigram sets of the two; Create frequency vectors for each string of the two respectively; Calculate the cosine similarity of the two frequency vectors.

[0062] This method converts two long text strings into two vectors and calculates the cosine value of the two vectors. The closer the cosine value is to 1, the more consistent the vector directions are and the more similar the strings are. Taking "Beijing Science and Technology Co., Ltd." and "Beijing Science and Technology Company" as examples, the text comparison method of the present invention is described below.

[0063] Generate a character bigram set from "Beijing Science and Technology Co., Ltd." and "Beijing Science and Technology Company", that is, take two adjacent characters in sequence as a unit.

[0064] The bigram set of "Beijing Science and Technology Co., Ltd." is: "Beijing", "Jingke", "Science and Technology", "Technology and Ownership", "Limited", "Limited Company", "Company", "Company" (there are various ways to handle the last single character, such as discarding or special processing. For simplicity, the single character "Company" is discarded in this example). The final valid set is: {Beijing, Jingke, Science and Technology, Technology and Ownership, Limited, Limited Company, Company}.

[0065] The bigram set of "Beijing Science and Technology Company" is: "Beijing", "Jingke", "Science and Technology", "Technology and Company", "Company". The final valid set: {Beijing, Jingke, Science and Technology, Technology and Company, Company}.

[0066] Construct the union vector space of the character bigram sets of "Beijing Technology Co., Ltd." and "Beijing Technology Company". Merge all the unique bigrams in the two sets to form the dimensions of the entire vector space. Union = {"Beijing", "Jingke", "Technology", "Youxian", "Limited", "Xiangong", "Company", "Jigong"}.

[0067] Create a frequency vector for each string. Each position in the vector corresponds to a bigram in the union, and the value is the number of times the grammar appears in the string.

[0068] The number of times "Beijing Technology Co., Ltd." appears in the union is: {1, 1, 1, 1, 1, 1, 1, 0}, denoted as vector A; the number of times "Beijing Technology Company" appears in the union is: {1, 1, 1, 0, 0, 0, 1, 1}, denoted as vector B.

[0069] Cosine similarity cosθ = Vector A × Vector B = (1 * 1) + (1 * 1) + (1 * 1) + (1 * 0) + (1 * 0) + (1 * 0) + (1 * 1) + (0 * 1) = 1 + 1 + 1 + 0 + 0 + 0 + 1 + 0 = 4.

[0070] The norm of vector A |A| = sqrt(1^2 + 1^2 + 1^2 + 1^2 + 1^2 + 1^2 + 1^2 + 0^2) = sqrt(7) ≈ 2.6458.

[0071] The norm of vector B |B| = sqrt(1^2 + 1^2 + 1^2 + 0^2 + 0^2 + 0^2 + 1^2 + 1^2) = sqrt(5) ≈ 2.2361.

[0072] Therefore, cosine similarity cosθ = 4 / (2.6458 * 2.2361) ≈ 4 / 5.916 ≈ 0.676.

[0073] It can be seen that the similarity score is 0.676 (ranging from 0 to 1). This indicates that the two supplier names have a relatively high similarity and are likely to refer to the same entity, but there are character differences. In actual work, this situation is generally handed over to manual review, and RPA cannot directly change it. According to work experience, the recommended threshold setting is: > 0.85: Automatically match and fill in the standard name.

[0074] 0.65 - 0.85: Push for manual review and prompt "Possible match: A and B, similarity 0.676".

[0075] <0.65: Treated as different company names, no action is taken.

[0076] By implementing the technical solution of this invention, RPA, in the process of processing financial data according to certain rules, can identify common problems encountered in daily work, such as false expense reports and data entry errors, by judging whether there is a risk of data fraud, whether there are extreme values ​​in a single field, and whether there are anomalies in the combination of multiple fields. This gives RPA the initial ability to identify and verify data logic anomalies. It is understandable that, using the technical solution of this invention, RPA will no longer be a "rigid" machine, but rather equivalent to equipping its "hands and feet" with a "brain." Using the judgment ability of this "brain," it will identify the presence of garbage data, changing from "garbage in, garbage out" to "garbage in, report out." This not only purifies the garbage data in the financial system but also improves the work efficiency of financial personnel, enabling modern enterprises to move from automation to intelligence in their financial work.

[0077] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural transformations made using the contents of the specification and drawings of the present invention under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the scope of patent protection of the present invention.

Claims

1. A financial data verification method based on RPA, characterized in that, include: Extract financial data from the financial system's processes; The financial data is logically validated to obtain the data validation result. The logical validation includes: Determine if there is a risk of data fraud; if so, mark the data as fraudulent. Determine if there is an extreme value in a single field; if so, mark the value of the single field as an extreme value identifier. Determine if there is an anomaly in the combination of multiple fields; if so, mark the values ​​of the multiple fields as an anomaly identifier. The data verification results are aggregated and a verification report is generated. The financial data and the verification report will be submitted for subsequent processing.

2. The RPA-based financial data verification method as described in claim 1, characterized in that, The method for determining whether there is a risk of data fraud is as follows: Construct a dataset {x1, x2, ..., x} from the financial data of the same type. n }; Extract the first non-zero digit of each data point in the dataset; The number of times the non-zero digits appear is counted, and the ratio of the number of times the non-zero digits appear to n is calculated. When the difference between the ratio and the preset expected value is not less than the preset threshold, there is a risk of data fraud.

3. The RPA-based financial data verification method as described in claim 2, characterized in that, When the ratio is less than a preset threshold, the statistic M is calculated: ; When the statistic M is greater than a preset threshold, there is a risk of data fraud. Among them, O d E represents the number of times the number d is counted. d Let d be the expected number of times. .

4. The RPA-based financial data verification method as described in claim 1, characterized in that, The method for determining whether a single field extreme value exists is as follows: Construct a dataset {x1, x2, ..., x} from the financial data of the same type. n }; Calculate the maximum absolute deviation of the dataset. ; If G > G c The critical value, then the x i For single-field extreme values; The G c The method for calculating the critical value is as follows: ； Where n is the number of data points, t α / (2n),n 2 is the critical value of the t-distribution. The significance level is denoted by s, and the standard deviation of the dataset is s. is the mean of the dataset.

5. The RPA-based financial data verification method as described in claim 1, characterized in that, The method for determining whether there is an anomaly in the combination of multiple fields is as follows: Extract several features from the financial data to form a feature vector group; The feature vector set is input into the trained self-learning model to calculate the anomaly score; When the abnormal score is greater than a preset threshold, the financial data has multiple field combination anomalies.

6. The RPA-based financial data verification method as described in claim 5, characterized in that, The self-learning model has T data sets, and the training method for the T data sets is as follows: Obtain historical financial data from the financial system; S1: Randomly select m data points from the historical financial data to form a data group; S2: Randomly select one feature from the features of the financial data; S3: Randomly select a segmentation value point for one of the features; the value of the segmentation value point is between the minimum and maximum values ​​in the data of the current feature; S4: Based on the values ​​of the segmentation points, divide the current data group into two sub-data groups, left and right; S5: Repeat steps S2 to S4 until the set conditions are met, then the training of the data set is complete; Repeat steps S1 to S5 until T data sets are obtained; The conditions are as follows: The number of sub-data groups reaches the maximum limit value ceil(log2(m)); where ceil represents rounding up; Or the current subgroup contains only one piece of data; Or all feature values ​​of all data in the current subgroup are exactly the same.

7. The RPA-based financial data verification method as described in claim 6, characterized in that, The abnormal score is calculated as follows: Abnormal scores , Where h(x) is the number of times the value x is split from the beginning of a certain data group in the self-learning model until the data is isolated; E(h(x)) represents the average number of splits of data point x in T data groups; and c(p) represents the expected value of the average number of splits given sample p.

8. The RPA-based financial data verification method as described in claim 1, characterized in that, After generating the verification report and before submitting it for subsequent process execution, it is determined whether the financial data is text data. If so, a text comparison method is used to verify the financial data.

9. The RPA-based financial data verification method as described in claim 8, characterized in that, The text comparison method is as follows: The text data is compared with preset text data. When the similarity between the text data and the preset text data is greater than a threshold, the text data of the financial data is adjusted to the preset text data and the financial data is marked.

10. The RPA-based financial data verification method as described in claim 9, characterized in that, The method for calculating the similarity of the text data is as follows: Generate a character binary grammar set by combining the text data with the preset text data; Construct a union vector space from the character binary grammar sets of both; Create frequency vectors for each of the two strings respectively; Calculate the cosine similarity between the frequency vectors of the two vectors.

Images