Financial analysis method and system based on big data

Abstract

The invention provides a financial analysis method and system based on big data, and the method comprises the steps: analyzing the dependency relationship between a non-standard code combination and a fuzzy text mode in a material purchasing scene based on a credible rule library, extracting a new attribute combination feature, and forming an optimization rule feature set if the new attribute combination feature can improve the matching accuracy; performing rule-based key field and attribute set verification matching on the real-time non-standard transaction data according to the optimization rule feature set in combination with a dependency relationship between a non-standard code combination and a fuzzy text mode, and outputting a matched transaction record set; and comparing the difference between the matched transaction record set and the historical manual matching record, extracting and judging whether the newly added non-standard attribute combination is applicable to the non-standard transaction data which is not covered or not after a matching rule is adjusted by adopting an incremental learning algorithm, and if so, forming a dynamic matching table.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a financial analysis method and system based on big data. Background Art

[0002] As a core component ensuring transparency in capital flows and the accuracy of operational decisions, corporate financial reconciliation is of irreplaceable importance in the modern business environment. With the rapid evolution of supply chain collaboration platforms and the diversification of payment methods between customers and suppliers, the need to process non-standardized transaction data has become increasingly prominent, making it a key area of digital financial management. Traditional reconciliation methods rely on standardized purchase order numbers or invoice serial numbers for matching. However, these methods often struggle with the irregular data generated by new electronic invoicing or payment systems. Existing solutions typically employ pre-set rules or manual intervention. While these solutions can address this issue to a certain extent, they suffer from poor adaptability, inefficiency, and insufficient recognition of non-standard data patterns. This makes rapid response and adjustment difficult, especially in scenarios involving supply chain upgrades or frequent external system changes. Specifically, the core challenge facing current reconciliation systems lies in effectively handling non-standardized transaction data generated by external system changes. This data often lacks standard identifiers, such as purchase order numbers or invoice serial numbers, or contains ambiguous text patterns and non-standard code combinations, making it difficult for the system to directly match it to the corresponding business document or cost center. Financial personnel are forced to manually categorize these vouchers into the correct business records one by one through repeated operations. This process is not only time-consuming and labor-intensive, but also lacks an automated learning mechanism, making it impossible to transform manual experience into system rules. The resulting technical difficulty lies in the system's difficulty in extracting patterns from the matching behavior of financial personnel, nor in dynamically generating new matching rules based on non-standard attribute combinations, resulting in limited reconciliation efficiency and accuracy. Therefore, how to continuously monitor the manual matching operations of financial personnel to capture the implicit non-standard code combinations or fuzzy text patterns, and how to use AI models to autonomously learn and extract new matching patterns applicable to non-standard transaction data has become a key issue in improving the adaptability and intelligence of corporate financial reconciliation. Summary of the Invention

[0003] The present invention provides a financial analysis method based on big data, which mainly includes:

[0004] Obtain payment records and electronic vouchers received by the enterprise ERP system, filter out abnormal transaction records associated with business documents, analyze the account, cost center, and supplier code fields, identify and mark the commonly occurring non-standard parts, and form a non-standard data set;

[0005] Group non-standard code combinations and fuzzy text patterns with a frequency above a threshold in the non-standard data set, extract business document samples that match historical payment records / electronic vouchers, calculate the matching probability, and based on the rules with matching probabilities above the threshold, determine whether the transaction instances that meet the rules in the current non-standard transaction data are consistent with the business documents and cost center allocations. If they are consistent, mark the rule as a trusted rule, and build a verified trusted rule library;

[0006] Based on the trusted rule base, new attribute combination features are extracted in the procurement material scenario. If the new attribute combination features can improve the matching accuracy, an optimized rule feature set is formed;

[0007] Based on the optimized rule feature set, the real-time non-standard transaction data is verified and matched with the key fields and attribute sets based on the rules, and the matching transaction record set is output;

[0008] Compare the differences between the matched transaction record set and the historical manually matched records, and use the incremental learning algorithm to adjust the matching rules to form a dynamic matching table;

[0009] When the enterprise ERP system, supply chain collaboration platform or payment system is changed or upgraded, the change characteristics and trends are extracted from the newly generated non-standard transaction data, and the dynamic matching table is adaptively optimized to obtain a rule set that adapts to the new system environment. The unmatched non-standard transaction data is processed in batches. If uncovered non-standard code combinations or fuzzy text patterns are found, they are added to the matching model and the final matching solution is output.

[0010] Furthermore, the payment and receipt records and electronic vouchers received by the enterprise ERP system are obtained, and abnormal transaction records associated with business documents are screened out. The account, cost center, and supplier code fields are analyzed, and commonly occurring non-standard components are identified and marked to form a non-standard data set. This includes: obtaining the transaction amount, transaction time, payment method, general ledger account code, supplier code, and cost center code from the enterprise payment and receipt records; and obtaining the invoice serial number, purchase order number, and voucher summary from the electronic voucher to generate the original record set. A supplier transaction mapping table is established based on the general ledger account codes and their corresponding supplier codes described in the original record set. Each record in the supplier transaction mapping table is standardized and verified according to the code rules in the preset supplier file. If the combination of supplier code and general ledger account code does not comply with the rules, a preliminary abnormal record set is obtained. A supplier behavior feature table is established based on the frequency of occurrence of the supplier code in the general ledger account code described in the preliminary abnormal record set and the integrity of the associated documents. The occurrence pattern of the supplier code is determined through frequency counting. If the missing rate of the associated documents exceeds a preset threshold of 0.2 and the number of repeated occurrences exceeds a preset threshold, a high-risk code group is identified. Based on the non-standard cost center codes described in the high-risk code group, key fields are extracted to construct a string feature vector. The cosine similarity algorithm is used to perform clustering calculations on the string feature vectors. When the string similarity exceeds the preset threshold of 0.8, a text anomaly pattern group is obtained. A correspondence table is established between the string feature vectors described in the text anomaly pattern group and the standard cost center codes. The string edit distance is calculated using the minimum edit distance algorithm. When the edit distance is less than the preset threshold, a standard code mapping set is obtained. Based on the mapping relationship described in the standard code mapping set, the non-standard codes in the preliminary abnormal record set are replaced, and the records that have completed standardization processing are eliminated using a rule filtering method to obtain a non-standard data set.

[0011] Furthermore, non-standard code combinations and fuzzy text patterns with a frequency higher than a threshold in the non-standard data set are grouped, business document samples matching historical payment records / electronic vouchers are extracted, and the matching probability is calculated. Based on the rule with a matching probability higher than the threshold, the transaction instances that meet the rule in the current non-standard transaction data are judged to be consistent with the business document and cost center allocation. If consistent, the rule is marked as a trusted rule, and a verified trusted rule base is constructed, including: obtaining non-standard code combinations from the non-standard data set, establishing a frequency statistics table according to the general ledger account code, supplier code, and cost center code described in the code combination, using a counting method to count the number of occurrences of each code combination within half a year, and obtaining a high-frequency code table when the number of occurrences is greater than a preset benchmark value. According to the code matching rules described in the high-frequency code table, the corresponding voucher summary text is extracted from the non-standard data set, the voucher summary text is segmented to obtain a keyword sequence, and a text feature vector is constructed using a word frequency statistics method, from which a key text group is obtained whose word frequency weight exceeds a preset benchmark value. According to the text content described in the key text group, matching records are retrieved from historical payment records and electronic vouchers. A business rule library is constructed based on standard business document samples. The matching degree between each record and the rule library is calculated using a text similarity algorithm. When the similarity exceeds the preset benchmark value, a preliminary rule table is obtained. The rule content is extracted from the preliminary rule table, and transaction records that meet the rules are screened from the current non-standard transaction data. The consistency of the transaction records and business documents is verified using the document field comparison method. If the transaction records and business documents are consistent, a rule verification table is obtained. The cost center allocation is checked based on the rule content described in the rule verification table. The cost center code is verified for relevance using the field mapping method. When the relevance verification passes, a compliance rule group is obtained. A rule scoring table is constructed based on the compliance rule group. The rule application scope, matching accuracy, and verification pass rate are quantitatively evaluated using a weighted calculation method. When the scoring result exceeds the preset benchmark value, a trusted rule library is obtained.

[0012] Furthermore, based on the trusted rule base, new attribute combination features are extracted in the procurement material scenario. If the new attribute combination features can improve the matching accuracy, an optimized rule feature set is formed, including: obtaining non-standard code combinations and fuzzy texts in the procurement material scenario from the trusted rule base, establishing an association matrix according to the general ledger account code, supplier code, cost center code, material code, and voucher summary, calculating the support between fields through the association rule mining algorithm, and obtaining a code dependency table when the support is greater than the preset threshold of 0.6. A text feature mapping table is constructed based on the code combinations and voucher summaries described in the code dependency table, a word frequency statistical method is used to extract text keywords, and the correlation strength between keywords and code combinations is calculated through mutual information entropy. From the calculation results, a text feature group is obtained when the correlation strength exceeds the preset threshold of 0.7. Cross-field combination rules are extracted from the feature content of the text feature group, a combination frequency statistics table is established based on the field combination rules, and the frequency of occurrence of field combinations is calculated through the frequent pattern tree algorithm. When the frequency value exceeds the preset threshold of 0.8, a high-frequency combination set is obtained. A feature evaluation table is constructed according to the combination rules described in the high-frequency combination set. Scoring criteria are established based on three dimensions: inter-field correlation, text similarity, and combination stability. Feature scores are calculated using numerical normalization methods. A candidate feature set is obtained from the scoring results when the score exceeds the preset threshold of 0.85. A matching test is conducted on the validation dataset based on the feature content described in the candidate feature set. The feature matching accuracy is calculated using cross-validation methods. An optimized feature table is obtained from the verification results when the accuracy improvement exceeds the preset baseline value of 0.1. Feature application verification is conducted in the material procurement scenario based on the feature content described in the optimized feature table. A multidimensional feature matrix is constructed using feature combination methods. An optimized feature set is obtained when the feature application effect exceeds the preset threshold of 0.9.

[0013] Furthermore, a keyword set is generated based on the fuzzy text of the purchase order. Based on the code-text mapping relationship in the rule base, the code features are associated and matched with the keyword set. If the confidence level of the associated match meets the requirements, a code-text joint feature vector is generated. High-frequency code-text combinations are extracted from the vector and added to the rule base as new rule features. This includes: obtaining incomplete order numbers and material code fragments from the purchase order, establishing a character distribution table based on the code character composition pattern, extracting code prefix features and length features using character statistics, and obtaining missing character positions and missing lengths exceeding a preset baseline from the character distribution table to obtain a code missing table. Based on the code fragments in the code missing table, the corresponding product name and specification descriptions are extracted from the fuzzy text field. A word segmentation algorithm is used to segment the product name and specification descriptions. Similar product descriptions are replaced using a domain dictionary. From the segmentation results, word frequency weights exceeding a preset baseline are obtained to obtain a specification word table. The code-text mapping relationship is obtained from the rule base, and an association matrix is established based on the keywords in the specification word table and the code features in the code missing table. A collaborative filtering algorithm based on user-item ratings is used to calculate the matching confidence level. When the confidence level exceeds a preset baseline, a feature mapping table is obtained. A field dependency network is established based on the feature content described in the feature mapping table. The node weight calculation method based on the item association is used to assign values to the connection strengths between the code snippets and the specification keywords. The frequent subgraph mining algorithm is used to extract the node group whose connection weight exceeds the preset benchmark value. The node group with co-occurrence frequency exceeding the preset benchmark value is obtained from the node group to obtain a combined feature table. A feature vector matrix is constructed according to the feature content described in the combined feature table. A reference vector set is established based on the rule base entries. The feature similarity is calculated using the cosine similarity algorithm. The calculation results are used to obtain the new feature table with a similarity lower than the preset benchmark value. A rule verification set is constructed based on the feature content described in the new feature table. The cross-validation method is used to verify the applicability of the features. When the verification result exceeds the preset benchmark value, a rule update table is obtained.

[0014] Furthermore, based on the optimized rule feature set, the real-time non-standard transaction data is matched with the key fields and attribute sets based on the rules, and a matching transaction record set is output, including: obtaining the code combination rules and text pattern rules from the optimized rule feature set, establishing a field rule table according to the general ledger account code, supplier code, cost center code, and voucher summary, using the field attribute extraction method to obtain the attribute values of the three dimensions of field length, value range, and format specification from the field rule table, and obtaining a verification rule table when the attribute values meet the rule requirements. Transaction records are extracted from the real-time non-standard transaction data, and the fields of the transaction records are standardized according to the rule requirements described in the verification rule table. Invalid characters and outliers are removed through data cleaning methods, and a standard record table is obtained from the cleaning results whose integrity exceeds the preset threshold. A feature vector matrix is constructed based on the record content described in the standard record table, and the field similarity and text similarity are calculated for the code combination and text content respectively. The cosine similarity algorithm is used for feature matching calculation, and the similarity exceeds the preset threshold from the matching results to obtain a preliminary record table. According to the record contents described in the preliminary record table, the preset business documents and cost center information are retrieved, the field values are verified using the field mapping method, and the linear correlation coefficient between the fields is verified using the correlation calculation method. When the correlation coefficient exceeds the preset threshold, a verification record table is obtained. The three dimensional features of code matching, text similarity, and field correlation are extracted from the record contents described in the verification record table, and the feature values are standardized using the normalization method. When the standardized feature values exceed the preset threshold, a feature vector table is obtained. A confidence score table is constructed based on the feature contents described in the feature vector table, and the record confidence score is calculated using the weighted average method based on the feature weight. The matching record set is obtained from the calculation results with the confidence exceeding the preset threshold.

[0015] Furthermore, by comparing the differences between the matched transaction record set and the historical manual matching records, the incremental learning algorithm is used to adjust the matching rules to form a dynamic matching table, including: obtaining field attribute values from the matched transaction record set and the historical manual matching records, establishing a field difference table according to the general ledger account code, supplier code, cost center code, and voucher summary, calculating the field value distribution difference using the statistical variance method, and obtaining the difference degree exceeding the preset threshold from the calculation result to obtain the adjustment rule table. According to the rule content described in the adjustment rule table, non-standard attribute combinations are extracted, the attribute combinations are clustered using the text classification method, the cluster center offset is calculated using the incremental learning algorithm, and the change amount exceeding the preset threshold is obtained from the offset calculation result to obtain the rule optimization table. A new rule verification set is constructed from the rule content described in the rule optimization table, and the rule application scope is calculated according to the code mapping rule, field association rule, and text description rule. The rule parameters are determined using the grid search method, and the parameter optimization table is obtained when the parameter verification result exceeds the preset threshold. Rule validation is performed on unmatched non-standard data using the parameter values described in the parameter optimization table. A validation dataset is constructed using a sample partitioning method. The rule accuracy is calculated using a cross-validation method. A validation evaluation table is obtained from the validation results when the accuracy exceeds a preset threshold. A rule update set is constructed based on the evaluation results described in the validation evaluation table. The rule update step size is calculated using an adaptive adjustment method. A dynamic rule table is obtained from the update results when the stability index exceeds a preset threshold. A rule integration set is constructed based on the rule content described in the dynamic rule table. The new and old rules are merged using a rule merging method. Rule correspondence is established using a feature mapping method. A dynamic matching table is obtained when the completeness of the correspondence exceeds a preset threshold.

[0016] Furthermore, when the enterprise ERP system, supply chain collaboration platform or payment system is changed or upgraded, the change characteristics and trends are extracted from the newly generated non-standard transaction data, the dynamic matching table is adaptively optimized, and a rule set adapted to the new system environment is obtained. The unmatched non-standard transaction data is batch processed. If an uncovered non-standard code combination or fuzzy text pattern is found, the matching model is added and the final matching solution is output, including: obtaining the general ledger account code, supplier code, cost center code, and voucher summary from the newly generated non-standard transaction data, establishing a change mapping table according to the field characteristics, using the time series pattern mining method to count the field format and value changes, obtaining the change amplitude exceeding the preset threshold from the change mapping table to obtain the feature change table. A rule adaptation table is constructed based on the change content described in the feature change table, and the parameters of the field format rule, value range rule, and association logic rule are optimized through the adaptive adjustment method. The incremental learning algorithm is used to calculate the rule update value, and the update value obtained from the calculation result exceeds the preset threshold to obtain the adaptation rule table. A validation dataset is constructed from the rules described in the adaptation rule table. A stratified sampling method is used to group unmatched non-standard transaction data. Rule coverage is calculated using a cross-validation method. The data table to be supplemented is obtained from the validation results with coverage below a preset threshold. Non-standard code combinations and text patterns are extracted according to the data content described in the data table to be supplemented. A feature extraction table is established based on the code mapping rules and text association rules. Features are matched using a similarity calculation method. A supplementary rule table is obtained from the matching results with similarities below a preset threshold. A rule validation set is constructed based on the rule content described in the supplementary rule table. A grid search method is used to optimize rule parameters. Rule accuracy is calculated using a cross-validation method. A validation rule table is obtained from the validation results with accuracy exceeding a preset threshold. A rule integration set is constructed based on the rule content described in the validation rule table. A weighted rule merging method is used to fuse the original rules and supplementary rules. Rule coverage is calculated using an integrity assessment method. The final matching solution is obtained from the assessment results with coverage exceeding a preset threshold.

[0017] The present invention provides a financial analysis system based on big data, which mainly includes:

[0018] The data acquisition and screening module is used to obtain transaction records from the enterprise ERP system, screen out abnormal transaction records associated with business documents, identify and mark non-standard parts in the account, cost center and supplier code fields, and form a non-standard data set;

[0019] The non-standard data analysis and rule building module is used to group non-standard portions of the non-standard data set whose frequency exceeds a threshold, extract business document samples that match historical transaction records, calculate the matching probability, and, based on the rules whose matching probability exceeds the threshold, determine whether the transaction instances that meet the rules in the current transaction data that does not conform to the standard format or encoding rules are consistently assigned to business documents and cost centers. If so, the rule is marked as a trusted rule, thus building a verified trusted rule library.

[0020] The rule optimization and feature extraction module is used to extract new attribute combination features in the procurement material scenario based on the trusted rule base. If the new attribute combination features can improve the matching accuracy, an optimized rule feature set is formed;

[0021] The real-time matching and verification module is used to verify and match the key fields and attribute sets of real-time non-standard transaction data based on the optimized rule feature set, and output the matching transaction record set;

[0022] The incremental learning and dynamic matching module is used to compare the differences between the matched transaction record set and the historical manual matching records, and to form a dynamic matching table after adjusting the matching rules using the incremental learning algorithm.

[0023] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0024] The present invention discloses a financial analysis method based on big data. The method analyzes the payment records and electronic vouchers in the enterprise ERP system, identifies non-standard code combinations and fuzzy text patterns in abnormal transactions, and builds a trusted rule base. The present invention uses historical data to verify matching rules, extracts new attribute combination features, and optimizes matching accuracy. Real-time non-standard transaction data is subjected to rule-based verification and matching, and a set of matching transaction records is output. Through an incremental learning algorithm, the present invention can dynamically update the matching model to adapt to new non-standard attribute combinations. When changes occur in related systems, the present invention can adaptively optimize the matching model and process unmatched non-standard transaction data. This method significantly improves the matching efficiency and accuracy of non-standard transaction data, providing strong support for enterprise financial management. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 The figure is a flow chart of a financial analysis method based on big data of the present invention.

[0026] Figure 2 This is a block diagram of a financial analysis system based on big data of the present invention. DETAILED DESCRIPTION

[0027] To further understand the content of the present invention, the present invention is described in detail with reference to the accompanying drawings and examples. The present application is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are intended only to illustrate the relevant invention and are not intended to limit the invention. It should also be noted that, for ease of description, only portions relevant to the invention are shown in the accompanying drawings.

[0028] like Figure 1 In this embodiment, a financial analysis method based on big data may specifically include:

[0029] S101 obtains transaction records from the enterprise ERP system, filters out abnormal transaction records associated with business documents, identifies and marks non-standard parts appearing in the account, cost center, and supplier code fields, and forms a non-standard data set.

[0030] In an embodiment of the present invention, when an enterprise ERP system receives externally input financial data, the system automatically triggers an analysis process. The financial analysis system first generates an initial data set by capturing fields such as the transaction amount, payment method, general ledger account code, and supplier code from payment and receipt records, as well as the invoice serial number and voucher summary from electronic vouchers. The system then checks data integrity based on pre-set rules. If a record is found to be missing a purchase order number or invoice serial number, it is classified as an abnormal transaction record.

[0031] The S1011 financial analysis system extracts key information such as transaction amount, transaction time, and general ledger account code from payment records, and obtains voucher summaries and supplier codes from electronic vouchers. It generates an original record set by merging fields, and then constructs a supplier transaction mapping table based on the general ledger account code and supplier code. The mapping table is standardized and verified. If it is found that the code combination does not meet the preset rules, it is marked as a preliminary abnormal record set.

[0032] In this embodiment of the present invention, the system uses field extraction and mapping table construction to initially screen out abnormal data. For example, if the general ledger account code corresponding to a supplier code does not match the business scope in the supplier's profile, the system will record this as an anomaly. This approach ensures basic identification of non-standard data.

[0033] The S1012 financial analysis system analyzes the frequency of occurrence of supplier codes in general ledger accounts and document integrity based on the preliminary abnormal record set, and generates a supplier behavior feature table. If the document missing rate exceeds the preset threshold and the number of repeated occurrences is high, it is identified as a high-risk code group. Subsequently, the cosine similarity algorithm is used to cluster the non-standard cost center codes in the high-risk code group, and the minimum edit distance algorithm is combined with the standard code for comparison. If the edit distance is less than the threshold, a non-standard data set is formed.

[0034] In this embodiment of the present invention, the system further analyzes the regularity of abnormal records. For example, if a supplier code frequently appears under a specific general ledger account and lacks associated documents, the system will mark it as a high-risk code group. Subsequently, through clustering and edit distance calculations, the system can identify pattern differences between "FIN2022A" and "FIN2022," and normalize the non-standard codes to generate a non-standard dataset containing abnormal patterns.

[0035] In this embodiment of the present invention, the dataset generation process fully considers the characteristics of non-standard code combinations and ambiguous text. For example, in the transaction records of a supplier at Baotai Co., Ltd., the supplier used the general ledger account code 660102 to record office supply expenses. However, the supplier's file was listed as a construction materials supplier, which did not match the expected account 220102 and lacked a purchase order number. Analysis revealed that this code appeared 15 times within six months, far exceeding the normal frequency, and the "office supplies" in the voucher summary did not match the actual items purchased, such as "cement." The system uses cosine similarity calculations to confirm the similarity of cost center codes and generates an accurate anomaly dataset through standardized replacement.

[0036] It is understood that the embodiments of the present invention do not impose excessive restrictions on the specific algorithms or threshold settings for field extraction, which can be adjusted by technical personnel based on actual scenarios. This method ensures the accuracy and scalability of the reconciliation process by identifying implicit patterns in non-standard data.

[0037] S102 groups the non-standard parts in the non-standard data set whose frequency of occurrence is higher than the threshold, extracts business document samples that match historical transaction records, calculates the matching probability, and based on the rules with matching probabilities higher than the threshold, determines whether the transaction instances that meet the rules in the current transaction data that does not meet the standard format or coding rules are consistently allocated to the business documents and cost centers. If they are consistent, the rule is marked as a trusted rule to build a verified trusted rule library.

[0038] In this embodiment of the present invention, the system first analyzes a non-standard dataset to ensure comprehensive rule extraction. This dataset includes code combinations and text patterns found in unusual transaction records. The system generates trusted rules through multi-dimensional processing, providing intelligent support for subsequent financial reconciliation.

[0039] The S1021 financial analysis system extracts non-standard code combinations from non-standard data sets, constructs a frequency statistics table based on general ledger account codes, supplier codes, and cost center codes, and uses counting methods to count the number of occurrences of various code combinations within half a year. If the number exceeds the preset threshold, a high-frequency code table is generated. Then, based on the code matching relationship in the high-frequency code table, the corresponding voucher summary text is extracted. Natural language processing technology is used to segment the text and extract keywords, generate keyword sequences, and calculate word frequency weights. If the weight exceeds the preset threshold, a key text group is formed.

[0040] In this embodiment of the present invention, a frequency statistics table traverses every transaction record in the non-standard dataset, recording the number of occurrences of each code combination. For example, if a supplier code and general ledger account code combination appears more than 20 times within six months, it will be included in the high-frequency code table. The processing of voucher summaries relies on word segmentation technology, breaking the text into individual words and quantifying the importance of each word through word frequency statistics to ensure that key text groups reflect the true characteristics of the transaction.

[0041] The S1022 financial analysis system retrieves matching records from historical payment records and electronic vouchers based on key text groups, establishes an initial business rule base based on standard business document samples, and uses the cosine similarity algorithm to calculate the similarity between the matching records and the rule base. If the similarity is higher than the preset threshold, a preliminary rule table is generated. Subsequently, transaction records that meet the rules are screened from the preliminary rule table, and the field mapping method is used to verify the consistency of the transaction records and business documents and the relevance of the cost center code. If the verification passes, a compliance rule group is formed.

[0042] In this embodiment of the present invention, the cosine similarity algorithm quantifies the similarity between text feature vectors and sample vectors in the rule base. For example, if the similarity between a transaction record's keyword sequence and a historical sample exceeds 0.9, it is considered to conform to the underlying rule. The field mapping method ensures the applicability of the rule by comparing the cost center code in the transaction record with the expected code in the business document.

[0043] The S1023 financial analysis system constructs a rule scoring table based on the compliance rule group, and comprehensively evaluates the scope of application, matching accuracy and verification pass rate of the rules through a weighted calculation method. If the scoring result exceeds the preset threshold, a trusted rule set is generated and stored as a trusted rule library for matching processing of non-standard transaction data.

[0044] In this embodiment of the present invention, the rule scoring table uses a weighted calculation method, for example, the scope of application accounts for 40%, the matching accuracy accounts for 35%, and the verification pass rate accounts for 25%. Through quantitative evaluation, efficient rules are screened. The construction of a trusted rule base ensures the high reliability and practicality of the rules.

[0045] For example, within a certain company's financial data, transaction records for a supplier, a trading company, showed 87 instances of the cost center code ADM2023 and general ledger account code 660305 combined within six months, far exceeding the normal business threshold of 20 instances. This combination is used to record advertising expenses, but the supplier actually sells building materials, indicating an anomaly in code usage. The system extracted keywords such as "advertising design fees" and "promotional materials" from the voucher summary through word segmentation, with frequency weights exceeding 0.35. However, the business documents indicated that the purchases were for cement and steel, which was inconsistent with the text description. A comparison of historical data revealed that a similar pattern recurred multiple times in early 2023, with a similarity as high as 0.92, reflecting typical anomalies. Further verification revealed that ADM2023 is a variant of the standard code ADM2022, frequently used for transactions with incomplete documents or large amounts, involving over 200 records, suggesting a tendency to evade monitoring. The rule scoring shows that the scope of application of this rule covers 90% of building materials procurement business, with a matching accuracy of 0.95 and a verification pass rate of 0.88, which meets the trustworthy rule standards.

[0046] In an embodiment of the present invention, this method systematically extracts underlying patterns in non-standard data through frequency statistics, text analysis, and similarity calculations. Compared to traditional processing methods that rely on fixed rules, this method can dynamically adapt to changes in abnormal patterns. For example, when a supplier evades monitoring through code variants, the system can identify and incorporate new rules through cross-field verification and historical sample comparison. The resulting rule base not only improves the efficiency of identifying abnormal transactions, but also ensures the reliability and timeliness of the rules through quantitative assessment, providing strong technical support for corporate compliance management.

[0047] It is understood that the embodiments of the present invention do not impose excessive restrictions on the parameters or threshold settings of the specific algorithms, which can be adjusted by technical personnel based on actual scenarios. For example, the threshold for cosine similarity can be flexibly configured based on the data scale. This flexibility ensures the applicability of the method in different enterprise environments, while also improving the operability and accuracy of rule generation through multi-dimensional verification.

[0048] S103 extracts new attribute combination features in the procurement material scenario based on the trusted rule base. If the new attribute combination features can improve the matching accuracy, an optimized rule feature set is formed.

[0049] In this embodiment of the present invention, the system uses correlation analysis and feature extraction to deeply explore potential patterns in non-standard data. This approach not only improves matching accuracy but also dynamically adapts to changing procurement scenarios. For example, when processing unusual transactions for a supplier, a certain trading company, the system discovered patterns where the code did not match the actual business, and optimized the rules accordingly.

[0050] The S1031 financial analysis system extracts non-standard code combinations and fuzzy text data in the procurement material scenario from the trusted rule base, constructs an association matrix based on the general ledger account code, supplier code, cost center code, material code and voucher summary, and uses the association rule mining algorithm to calculate the support between fields. If the support exceeds the preset threshold, a code dependency table is generated. Subsequently, a text feature mapping table is constructed based on the table. Keywords are extracted through word frequency statistics and the mutual information entropy method is used to calculate the association strength between them and the code combination. If the strength exceeds the threshold, a text feature group is formed.

[0051] In one embodiment, the association rule mining algorithm generates a code dependency table with high support by analyzing the co-occurrence frequency between fields. For example, the combination support of supplier code SP10086 and general ledger account code 660204 reaches 0.72, indicating that they are highly correlated. Mutual information entropy ensures that the text feature group accurately reflects the transaction characteristics by quantifying the degree of dependence between keywords such as "office equipment procurement" and codes. The S1032 financial analysis system extracts cross-field combination rules for the text feature group, establishes a combination frequency statistics table, and calculates the frequency of occurrence of field combinations through the frequent pattern tree algorithm. If the frequency exceeds the threshold, a high-frequency combination set is generated. The matching effect of these combinations is then tested on the verification data set, and the accuracy improvement is calculated using the cross-validation method. If the improvement meets the standard, an optimized feature table is formed, and then its application effect is verified through a multidimensional feature matrix. If the effect is better than expected, an optimized feature set is generated.

[0052] For example, in transactions for a certain trading company, the code combination SP10086, 660204, ADM2023, and material code M20230518 frequently appeared, with a frequency of 0.85. The correlation strength between the voucher summary keyword "computer parts replacement" and this combination was 0.82, but the transaction actually involved the procurement of construction materials. Cross-validation showed that the new feature significantly improved the matching accuracy from 0.82 to 0.94.

[0053] In purchase order processing, the system also extracts code prefix and length distribution features for incomplete order numbers such as PO2023X and material code M20230518Y, and discovers their missing patterns. For example, PO2023X should be 12 digits but is only 8 digits long, with the last digit missing. By segmenting the specification description "304 stainless steel plate 5 mm thickness", the keywords "304", "stainless steel", and "5mm" are generated, and the domain dictionary is used to replace "plate" with "steel plate". The matching confidence of these keywords and the code prefix "M2023" is 0.87, and the similarity between the constructed feature vector and the rule base is only 0.45, indicating its uniqueness. The new feature identified 187 abnormal transactions in 500 orders with an accuracy rate of 0.92.

[0054] In this embodiment of the present invention, this method uses field dependency networks and frequent pattern mining to capture the deep connections between code and text. For example, the connection weight between "M2023" and "stainless steel" is 0.92, reflecting a stable mapping relationship. Compared to traditional single-field matching, this method incorporates multidimensional features, significantly improving its ability to handle non-standard data. Technicians can adjust the threshold according to the scenario to ensure the method's flexibility.

[0055] In another embodiment, an association rule mining algorithm, such as the Apriori algorithm, calculates support based on this. The support reflects the prevalence of field combinations in a data set. For example, the support of "M123" and "S001" is 0.9, which means that 90% of procurement transactions contain this combination. If the preset support threshold is 0.8, the combination meets the conditions and enters the code dependency table. This process ensures that high-frequency and reliable field relationships are mined, providing a structured basis for subsequent analysis. For example, the generation of the code dependency table is based on the support screening results. In one possible implementation, the dependency table may record "M123→S001" to represent the dependency relationship between the material code and the supplier code, and also contain other field combinations, such as "5101→

[0056] C101". This table not only clearly shows the strong correlation between fields, but also provides a reference for data governance. For example, enterprises can use this to optimize supplier selection and reduce the arbitrariness of non-standard material procurement. It should be noted that the dynamic update of the dependency table can adapt to changes in procurement scenarios. For example, when a new supplier joins, the dependency relationship can be adjusted in a timely manner to maintain the timeliness of the analysis. In one embodiment, the construction of the text feature mapping table is based on the code dependency table to further mine the fuzzy text data in the voucher summary. For example, the summary text "Bolt Procurement in March 2025" may contain the keywords "bolt" and "procurement". Through word frequency statistics, assuming that the frequency of "bolt" is 95% and "procurement" is 80%, it shows that these words are high-frequency features. Preferably, the mutual information entropy method is used to calculate the correlation strength of the keyword and code combination. For example, the mutual information entropy value of "bolt" and the material code "M123" is high, indicating that the two are strongly correlated. If the strength threshold is set to 0.7, "bolt" is included in the text feature group.

[0057] In actual applications, the optimized feature set successfully identified 152 suspected anomalous transactions, 143 of which were confirmed to be violations. This accuracy stems from a comprehensive analysis of incomplete code patterns and ambiguous text, enabling the system to infer complete information and reveal hidden anomalous patterns, providing efficient support for procurement management.

[0058] S104 verifies and matches the key fields and attribute sets of the real-time non-standard transaction data based on the optimized rule feature set, and outputs a matching transaction record set.

[0059] In this embodiment of the present invention, the system ensures the accuracy of matching real-time transaction data through multi-level field processing and feature analysis. For example, in a corporate procurement scenario, the system verified a transaction record involving office supplies, demonstrating its efficiency and reliability.

[0060] The S1041 financial analysis system extracts code combination rules and text pattern rules from the optimization rule feature set, builds a field rule table based on the general ledger account code, supplier code, cost center code and voucher summary, obtains the attribute values of field length, value range and format specification through the field attribute extraction method and generates a validation rule table, then extracts transaction records from real-time non-standard transaction data, performs field standardization according to the requirements of the validation rule table, and uses data cleaning methods to eliminate invalid characters and abnormal values. If the field integrity exceeds the preset threshold, a standard record table is generated.

[0061] For example, a transaction record has the general ledger account code 660102, the vendor code SP1024ABC, the cost center code ADM2023, and the voucher summary "Office Supplies Purchase." The field rule table requires the general ledger account code to be 6 digits with the first digit being 6, the vendor code to be 10 digits and beginning with SP, and the cost center code to be a 7-digit alphanumeric combination. The system normalizes lowercase characters to uppercase and completes the cost center code to ADM2023A. After cleaning, the integrity reaches 0.96, exceeding the 0.85 threshold.

[0062] The S1042 financial analysis system constructs a feature vector matrix based on the standard record table, calculates field similarity and text similarity for code combinations and text content respectively, and uses the cosine similarity algorithm for feature matching. If the similarity exceeds the preset threshold, a preliminary record table is generated, and then the preset business documents and cost center information are retrieved. The field values are verified through field mapping and the linear correlation coefficient between fields is calculated. If the correlation coefficient exceeds the threshold, a verification record table is formed.

[0063] In this embodiment of the present invention, the cosine similarity algorithm quantifies the degree of match by comparing the transaction record's feature vector with the standard vector in the rule base. The code combination similarity for this record is 0.89, the text similarity is 0.92, and the overall match reaches 0.90, exceeding the 0.75 threshold. Field mapping confirms that 660102 corresponds to office expenses, and SP1024ABC is associated with office supply suppliers. The correlation coefficient is 0.87, exceeding the 0.80 threshold.

[0064] During the verification process, the system found that the supplier's business documents were consistent with the administrative department's cost center, confirming the legitimacy of the transaction. Compared to traditional manual verification, this automated matching significantly improves efficiency and reduces human error.

[0065] The S1043 financial analysis system extracts three-dimensional features of code matching, text similarity and field correlation from the verification record table, standardizes the feature values through normalization method and constructs a confidence score table, and uses weighted average method to calculate the confidence score of the record. If the score exceeds the preset threshold, the matching record set is output.

[0066] In this embodiment of the present invention, the system assigns weights to the three-dimensional features: code matching at 0.4, text similarity at 0.3, and field relevance at 0.3. The feature values for this record are 0.89, 0.92, and 0.87, respectively, resulting in a calculated confidence level of 0.893, exceeding the 0.85 threshold, and successfully included in the matching record set. This weighting method balances the contributions of each dimension, ensuring comprehensiveness of the results.

[0067] In practical applications, this method not only verifies field standardization but also confirms the rationality of business logic through correlation analysis. For example, another abnormal transaction was eliminated due to an incorrect cost center attribution, demonstrating the system's ability to identify anomalies. It is understood that thresholds and weights can be adjusted according to enterprise needs to meet the matching requirements of different scenarios.

[0068] S105 compares the differences between the matched transaction record set and the historical manual matching records, and forms a dynamic matching table after adjusting the matching rules using an incremental learning algorithm.

[0069] The S1051 financial analysis system extracts field attribute values from the matched transaction record set and historical manual matching records, constructs a field difference table based on the general ledger account code, supplier code, cost center code and voucher summary, and uses the statistical variance method to calculate the difference in field value distribution. If the difference exceeds the preset threshold, an adjustment rule table is generated. Subsequently, non-standard attribute combinations are extracted based on the adjustment rule table, clustered using the text classification method, and the offset of the cluster center is calculated using the incremental learning algorithm. If the offset exceeds the threshold, a rule optimization table is formed.

[0070] For example, historical records for a certain supplier, a technology company, show that it used general ledger account code 660204 to record office equipment purchases, while the current record set frequently uses 660304 for R&D equipment purchases. Statistical variance calculations show that the account code difference reaches 0.82, exceeding the 0.75 threshold. Text classification revealed that the voucher summary changed from "office computer equipment" to "computing equipment for R&D." The cluster center offset was 0.79, exceeding the 0.7 threshold, reflecting a significant change in business attributes.

[0071] The S1052 financial analysis system constructs a new rule verification set from the rule optimization table, calculates the applicable scope based on the code mapping rules, field association rules and text description rules respectively, and uses the grid search method to optimize the rule parameters. If the verification result exceeds the threshold, a parameter optimization table is generated, and then the rules are verified on the unmatched non-standard data. The verification data set is constructed through sample division and the accuracy is calculated using the cross-validation method. If the accuracy meets the standard, a verification evaluation table is formed.

[0072] In validating the new rules, the system set parameters for equipment procurement transactions: a match between the account code and the supplier's business scope greater than 0.8, consistency between the cost center and the equipment-using department greater than 0.85, and a relevance between the voucher summary and the business greater than 0.9. A grid search confirmed these parameters with an accuracy of 0.93 on the validation set. Verification on 2,000 unmatched data sets showed a cross-validation accuracy of 0.91, demonstrating the rule's strong generalization capabilities.

[0073] In this embodiment of the present invention, this method uses incremental learning to capture business changes. For example, after adjusting the recognition rules for R&D equipment procurement, they successfully cover 95% of relevant transactions, demonstrating greater flexibility than traditional fixed rules. The parameter adjustment process also considers dynamic correlations between fields, ensuring rule adaptability.

[0074] The S1053 financial analysis system constructs a rule update set based on the verification evaluation table, calculates the rule update step and evaluates the stability through the adaptive adjustment method, generates a dynamic rule table if the stability index exceeds the threshold, then integrates the new and old rules and establishes a corresponding relationship through the feature mapping method, and outputs a dynamic matching table if the completeness meets the standard.

[0075] During rule updates, the adaptive adjustment method dynamically adjusts the step size based on changes in accuracy, stopping optimization when the stability index reaches 0.89. After integrating the new rules with the existing ones, the completeness reaches 0.94, covering both office and R&D equipment procurement scenarios. This integration not only preserves the effectiveness of the existing rules but also expands support for new businesses.

[0076] In practical applications, this mechanism significantly improves the system's responsiveness to business transformations. For example, when a company shifts from purchasing office equipment to R&D equipment, the rule base is updated promptly and identifies 95% of relevant transactions, reducing the need for manual intervention. It's understood that thresholds and parameters can be adjusted based on specific scenarios to ensure the method's practicality.

[0077] S106 When the enterprise ERP system, supply chain collaboration platform or payment system is changed or upgraded, the change characteristics and trends are extracted from the newly generated non-standard transaction data, the dynamic matching table is adaptively optimized, and a rule set that adapts to the new system environment is obtained. The unmatched non-standard transaction data is processed in batches. If uncovered non-standard code combinations or fuzzy text patterns are found, they are added to the matching model and the final matching solution is output.

[0078] The S1061 financial analysis system extracts general ledger account codes, supplier codes, cost center codes, and voucher summaries from newly generated non-standard transaction data, constructs a change mapping table, and uses a time series pattern mining method to analyze the changing trends of field formats and values. If the magnitude of the change exceeds the preset threshold, a feature change table is generated. Subsequently, a rule adaptation table is constructed based on the feature change table. The parameters of the field format rules, value range rules, and association logic rules are optimized through an adaptive adjustment method. The rule update value is calculated using an incremental learning algorithm. If the updated value exceeds the threshold, an adaptation rule table is formed.

[0079] For example, after a supply chain system upgrade, supplier codes changed from a 10-digit prefix "SP" to a 12-digit prefix "SUP." Temporal pattern mining revealed that 80% of these codes changed within three months, with a change of 0.85, exceeding the 0.8 threshold. The system adjusted the length rule to 12 digits and supported prefixes like "SP|SUP," achieving an update value of 0.92, significantly improving adaptability.

[0080] In this embodiment of the present invention, time series pattern mining analyzes the time series of field values to identify regularities in format changes. Compared to static analysis, this approach can better capture the dynamic impact of system upgrades and ensure forward-looking rule adjustments.

[0081] S1062 In an embodiment of the present invention, the financial analysis system constructs a verification data set from the adaptation rule table, adopts a stratified sampling method to group the unmatched non-standard transaction data and calculates the rule coverage through cross-validation. If the coverage is lower than the threshold, a data table to be supplemented is generated, and then the non-standard code combination and text pattern are extracted and a feature extraction table is constructed. The features are matched through a similarity calculation method. If the similarity is lower than the threshold, a supplementary rule table is generated, and then the grid search method is used to optimize the parameters and verify the accuracy. If the accuracy meets the standard, a verification rule table is formed.

[0082] A stratified sampling of 5,000 unmatched transactions revealed a coverage rate of 0.95 for large transactions, 0.87 for medium transactions, and only 0.62 for small transactions, all below the 0.8 threshold. Temporary supplier codes prefixed with "RD" and the text pattern "Temporary Purchase - Office Supplies" were found in small transactions, with a similarity of only 0.45 to the existing rules. After grid search optimization, the supplementary rules achieved an accuracy rate of 0.92, a temporary code recognition rate of 0.95, and a text matching rate of 0.89.

[0083] The S1063 financial analysis system constructs a rule integration set based on the verification rule table, adopts a weight-based rule merging method to integrate the original rules and supplementary rules, and calculates the coverage through the integrity assessment method. If the coverage exceeds the threshold, a final rule table is generated and applied to unmatched data processing.

[0084] The original rule weight is 0.7, and the supplementary rule weight is 0.3, resulting in a combined coverage of 0.96, fully covering both new and existing transaction models. This approach not only adapts to format changes but also expands support for scenarios such as temporary purchases, significantly improving the robustness of the rules.

[0085] In practice, this mechanism has effectively addressed the challenges presented by system upgrades. For example, supplementary rules have been implemented to address blind spots in small transactions, improving processing efficiency by approximately 30%. Technicians can adjust thresholds and weights based on business needs, ensuring the flexibility and practicality of the method.

[0086] like Figure 2 The present invention provides a financial analysis system based on big data, which mainly includes:

[0087] The data acquisition and screening module is used to obtain payment records and electronic vouchers received by the enterprise ERP system, filter out abnormal transaction records associated with business documents, analyze the account, cost center, and supplier code fields, identify and mark commonly occurring non-standard code combinations and fuzzy text patterns, and form a non-standard data set;

[0088] The non-standard data analysis and rule building module is used to group non-standard code combinations and fuzzy text patterns with a frequency above a threshold in non-standard data sets, extract business document samples that match historical payment records / electronic vouchers, calculate the matching probability, and based on the rules with matching probabilities above the threshold, determine whether the transaction instances that meet the rules in the current non-standard transaction data are consistent with the business documents or cost center allocations. If they are consistent, the rule is marked as a trusted rule, thus building a verified trusted rule library;

[0089] The rule optimization and feature extraction module is used to analyze the dependency between non-standard code combinations and fuzzy text patterns in procurement material scenarios based on a trusted rule base, extract new attribute combination features, and form an optimized rule feature set if the new attribute combination features can improve matching accuracy.

[0090] The real-time matching and verification module is used to verify and match the key fields and attribute sets of real-time non-standard transaction data based on the optimized rule feature set, combined with the dependency relationship between non-standard code combinations and fuzzy text patterns, and output the matching transaction record set;

[0091] The incremental learning and dynamic matching module compares the differences between the matching transaction record set and the historical manually matched records. After adjusting the matching rules using the incremental learning algorithm, it extracts and determines whether the newly added non-standard attribute combination is applicable to the uncovered non-standard transaction data. If so, a dynamic matching table is formed.

[0092] The system adaptive optimization module is used to extract change characteristics and trends from newly generated non-standard transaction data when the enterprise ERP system, supply chain collaboration platform or payment system is changed or upgraded, and adaptively optimize the dynamic matching table to obtain a rule set that adapts to the new system environment. It batch processes unmatched non-standard transaction data and, if uncovered non-standard code combinations or fuzzy text patterns are found, adds the matching model to output the final matching solution.

[0093] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A financial analysis method based on big data, characterized in that: The method comprises: Obtain transaction records from the enterprise ERP system, filter out abnormal transaction records associated with business documents, identify and mark non-standard parts in the account, cost center, and supplier code fields, and form a non-standard data set; The non-standard parts of the non-standard data set with a frequency higher than a threshold are grouped, and business document samples that match historical transaction records are extracted. The matching probability is calculated. Based on the rules with matching probabilities higher than the threshold, it is determined whether the transaction instances that meet the rules in the current transaction data that does not meet the standard format or encoding rules are consistently assigned to business documents and cost centers. If they are, the rule is marked as a trusted rule, and a verified trusted rule library is constructed. Based on the trusted rule base, new attribute combination features are extracted in the procurement material scenario. If the new attribute combination features can improve the matching accuracy, an optimized rule feature set is formed; Based on the optimized rule feature set, the real-time non-standard transaction data is verified and matched with key fields and attribute sets, and the matching transaction record set is output; By comparing the differences between the matched transaction record set and the historical manual matching records, the matching rules are adjusted using the incremental learning algorithm to form a dynamic matching table.

2. The method according to claim 1, characterized in that The process of obtaining transaction records from the enterprise ERP system, screening out abnormal transaction records associated with business documents, identifying and marking non-standard parts appearing in the account, cost center, and supplier code fields, and forming a non-standard data set includes: Obtain the general ledger account code and supplier code from the enterprise transaction record, obtain the invoice serial number and purchase order number from the electronic voucher, and obtain the original record set; Establishing a supplier transaction mapping table based on the general ledger account code and supplier code in the original record set, performing standardized verification on the supplier transaction mapping table using preset supplier file code rules, and forming a preliminary abnormal record set based on the supplier transaction mapping tables that fail the standardized verification; A supplier behavior characteristic table is established based on the frequency of supplier codes appearing in general ledger account codes and the integrity of associated business documents in the preliminary abnormal record set. The invoice serial number and purchase order number in the supplier behavior characteristic table are used to determine the missing rate and repeated occurrence count of associated business documents. If the missing rate of associated business documents exceeds a missing rate threshold and the repeated occurrence count of associated business documents exceeds a repeated occurrence threshold, a high-risk code group is determined. The cosine similarity algorithm is used to extract key fields of the cost center code in the high-risk code group to construct string feature vectors for clustering. The minimum edit distance algorithm is used to calculate the edit distance with the cost center code. If the edit distance is less than the preset threshold, a non-standard data set is obtained.

3. The method according to claim 1, characterized in that The non-standard parts of the non-standard data set with a frequency higher than a threshold are grouped, business document samples matching historical transaction records are extracted, and matching probabilities are calculated. Based on the rules with matching probabilities higher than the threshold, it is determined whether the transaction instances that meet the rules in the current transaction data that does not meet the standard format or encoding rules are consistently assigned to the business documents and cost centers. If so, the rule is marked as a trusted rule, and a verified trusted rule library is constructed, including: Obtain code combinations from non-standard data sets, use a counting method to count the number of occurrences of the code combinations within a preset time, and obtain a high-frequency code table if the number of occurrences exceeds a preset benchmark value; Extract the corresponding voucher summary text according to the high-frequency code table, obtain a keyword sequence by word segmentation processing of the summary text, construct a text feature vector for the keyword sequence using a word frequency statistics method, and obtain a key text group when the word frequency weight obtained from the text feature vector exceeds a preset benchmark value; Retrieving matching records from historical transaction records according to the key text group, calculating the similarity between the matching records and the trusted rules using a text similarity algorithm, and obtaining a preliminary rule table when the similarity exceeds a preset benchmark value; Filter out transaction records that meet the preliminary selection rules in the preliminary selection rule table, verify the consistency of the transaction records with the business documents through the document field comparison method, and obtain a rule verification table if the transaction records are consistent with the business documents; Check the cost center allocation according to the rules in the rule verification table, use the field mapping method to verify the relevance of the cost center code, and obtain a verified and trusted rule base when the relevance verification passes.

4. The method according to claim 1, wherein Based on the trusted rule base, new attribute combination features are extracted in the procurement material scenario. If the new attribute combination features can improve the matching accuracy, an optimized rule feature set is formed, including: Obtain non-standard code combinations and fuzzy text in the procurement material scenario from the trusted rule library, calculate the support between fields through the association rule mining algorithm, and obtain a code dependency table if the support is greater than a preset threshold; Constructing a text feature mapping table according to the code dependency table, extracting text keywords using a word frequency statistics method, calculating the correlation strength between the text keywords and the non-standard code combination using mutual information entropy, and obtaining a text feature group when the correlation strength exceeds a preset threshold; Field combination rules are extracted for the text feature group, a combination frequency statistics table is established based on the field combination rules, the frequency of occurrence of field combinations is calculated by the frequent pattern tree algorithm, a feature matching test is performed on the part where the frequency of occurrence of field combinations exceeds a preset threshold, and the feature matching accuracy is calculated using a cross-validation method. When the improvement in the accuracy exceeds a preset benchmark value, an optimized rule feature set is obtained.

5. The method according to claim 1, characterized in that Based on the trusted rule base, new attribute combination features are extracted in the procurement material scenario. If the new attribute combination features can improve the matching accuracy, an optimized rule feature set is formed, including: generating a keyword set based on the fuzzy text of the purchase order, associating and matching code features with the keyword set based on the code and text mapping relationship in the verified trusted rule base, if the confidence of the associated matching meets the standard, generating a code and text joint feature vector, extracting a combination of code and text co-occurrences from it, calculating the similarity between the combination of code and text co-occurrences and the rules in the verified trusted rule base, and if the similarity is lower than the similarity threshold, adding the combination of code and text co-occurrences as a new rule feature to the rule base.

6. The method according to claim 1, characterized in that The method of matching key fields and attribute sets of real-time non-standard transaction data based on the optimized rule feature set and outputting a matching transaction record set includes: A field rule table is established according to the code combination rule and the text pattern rule, wherein the field rule table includes field attribute values of the general ledger account code, supplier code, cost center code, and voucher summary; The transaction records in the field rule table are processed using a field standardization method, invalid characters and outliers are removed using a data cleaning method, and code combinations whose integrity exceeds a preset threshold are obtained. The code combinations and text content are matched with each other using a cosine similarity algorithm, and a preliminary record table is obtained when the similarity exceeds the preset threshold. Retrieve preset business documents and cost center information according to the record contents of the preliminary record table, verify the field values using the field mapping method, form a verification record table based on the verified field values, calculate the record confidence score using the weighted average method, and obtain the matching transaction record set whose confidence exceeds the preset threshold from the calculation results.

7. The method according to claim 1, characterized in that The differences between the transaction record set and the historical manual matching records are compared and matched, and the matching rules are adjusted using an incremental learning algorithm to form a dynamic matching table, including: Using a statistical variance method to calculate the distribution difference between the fields of the transaction record set and the historical manually matched records, obtaining the difference between the distribution differences exceeding a preset threshold from the difference distribution differences to obtain an adjustment rule table; Extracting non-standard attribute combinations based on the content of the adjustment rule table, performing clustering operations on the non-standard attribute combinations using a text classification method, and calculating the center offset of the clustering operation using an incremental learning algorithm to obtain a rule optimization table; Constructing a rule verification set from the content of the rule optimization table, wherein the rule verification set uses a grid search method to determine the rule parameter value, and obtaining a parameter optimization table when the parameter value verification result exceeds a preset threshold; The unmatched non-standard data are validated according to the parameter values in the parameter optimization table. The unmatched non-standard data are used to construct a validation data set through a sample partitioning method, and the rule accuracy is calculated using a cross-validation method to obtain a validation evaluation table. A rule update set is constructed based on the evaluation results of the verification evaluation table, the rule update step is calculated through the adaptive adjustment method, and the stability index obtained from the update results exceeds the preset threshold to obtain a dynamic matching table.

8. The method according to claim 1, characterized in that The method includes: when an enterprise ERP system, supply chain collaboration platform or payment system is changed or upgraded, extracting change characteristics and trends from newly generated non-standard transaction data, adaptively optimizing the dynamic matching table to obtain a rule set that adapts to the new system environment, batch processing unmatched non-standard transaction data, and if uncovered non-standard code combinations or fuzzy text patterns are found, adding them to the matching model and outputting the final matching solution.

9. The method according to claim 8, characterized in that When the enterprise ERP system, supply chain collaboration platform or payment system is changed or upgraded, the change characteristics and trends are extracted from the newly generated non-standard transaction data, and the dynamic matching table is adaptively optimized to obtain a rule set that adapts to the new system environment. The unmatched non-standard transaction data is processed in batches. If uncovered non-standard code combinations or fuzzy text patterns are found, they are added to the matching model and the final matching solution is output, including: Obtain general ledger account codes and supplier codes from non-standard transaction data, perform statistical analysis on the codes using time series pattern mining methods, and obtain a feature change table based on change amplitude thresholds. Constructing a rule adaptation table based on the feature change table, optimizing parameters of the field format rule using an adaptive adjustment method, and obtaining an adaptation rule table by calculating using an incremental learning algorithm; A stratified sampling method is used to group data in the adaptation rule table, and a rule coverage rate of the data grouping is calculated by a cross-validation method. If the rule coverage rate is lower than a preset threshold, a data table to be supplemented is obtained; A rule integration set is established based on the data table to be supplemented, and the rule integration set is merged using a weight-based rule merging method. The rule coverage is calculated using an integrity assessment method. If the rule coverage exceeds a preset threshold, the final matching solution is obtained.

10. A financial analysis system based on big data, characterized in that: The system comprises: The data acquisition and screening module is used to obtain transaction records from the enterprise ERP system, screen out abnormal transaction records associated with business documents, identify and mark non-standard parts in the account, cost center and supplier code fields, and form a non-standard data set; The non-standard data analysis and rule building module is used to group non-standard portions of the non-standard data set whose frequency exceeds a threshold, extract business document samples that match historical transaction records, calculate the matching probability, and, based on the rules whose matching probability exceeds the threshold, determine whether the transaction instances that meet the rules in the current transaction data that does not conform to the standard format or encoding rules are consistently assigned to business documents and cost centers. If so, the rule is marked as a trusted rule, thus building a verified trusted rule library. The rule optimization and feature extraction module is used to extract new attribute combination features in the procurement material scenario based on the trusted rule base. If the new attribute combination features can improve the matching accuracy, an optimized rule feature set is formed; The real-time matching and verification module is used to verify and match the key fields and attribute sets of real-time non-standard transaction data based on the optimized rule feature set, and output the matching transaction record set; The incremental learning and dynamic matching module is used to compare the differences between the matched transaction record set and the historical manual matching records, and to form a dynamic matching table after adjusting the matching rules using the incremental learning algorithm.

Images