Enterprise financial consultation platform based on financial data analysis

Abstract

The invention discloses an enterprise financial consultation platform based on financial data analysis, and relates to the technical field of enterprise financial data analysis, and the platform comprises a storage medium health monitoring module which collects physical health indexes of storage equipment in real time, constructs a dynamic portrait of the health degree of a storage medium, marks a high-risk storage area, and stores the physical health indexes; generating a dynamically updated storage medium health thermodynamic diagram; according to the method, the physical health indexes of the storage equipment are collected in real time, and the dynamic portrait of the health degree of the storage medium is constructed, so that the aging risk of the storage medium is accurately identified, invisible hardware aging is converted into quantifiable risk information, the risk identification capability is improved, dual-mode verification of a bit layer and a semantic layer is adopted, and the risk identification efficiency is improved. The functions of identifying high-risk silent data errors and capturing local damage of the storage medium are realized; and through a hierarchical repair strategy, the document repair effect is improved by utilizing an erasure code decoder and a FinDoc-RepairGAN processor.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of enterprise financial data analysis, and particularly relates to an enterprise financial consulting platform based on financial data analysis. BACKGROUND

[0002] With the acceleration of the digitalization process of enterprises, in the enterprise financial consulting platform, a large number of scanned paper financial vouchers such as invoices, contracts, bank receipts, etc. need to be processed, and these document images are stored in a mixed storage pool composed of different batches and different manufacturers of hard disks / SSDs.

[0003] The physical storage medium has a non-uniform aging phenomenon, which can cause data degradation and thus cause problems such as OCR misrecognition, thereby directly affecting core businesses such as enterprise cost accounting and tax audit. For example, the bit error rate (BER) of some hardware sectors / Blocks increases due to frequent writing or congenital defects; when the financial document image is stored in a physical area that is severely aged, the underlying data of the file may flip bits (BitFlip) in silence, causing damage to the local pixel block of the image file; and the OCR engine is prone to misrecognition of key financial information such as amounts, dates, and account numbers when processing such damaged images, resulting in digital distortion and error concealment.

[0004] The prior art has the following deficiencies: Application No. CN202310205100.7 discloses a comprehensive financial audit system based on big data, which includes a big data platform and a comprehensive financial audit system. The big data platform is used for storage and integration of past financial data, and the comprehensive financial audit system is used for comprehensive audit of various financial data, realizing retrieval of past financial data, improving data processing convenience, and accurately, quickly and safely auditing assets, liabilities, profits and losses, and producing an audit report. However, it ignores the aging of physical storage media, lacks detection of local pixel block damage of images, and lacks a response mechanism for physical risks; Application No. CN202410629096.1 discloses a method, system, device and medium for profit table data extraction, which designs a model for intelligent filtering of interference data for data extraction, accurately and quickly locates errors, and incorporates artificial correction data into machine training to realize self-evolution of the profit table data extraction method and improve the accuracy of profit table data extraction. However, it also ignores the impact of physical storage medium aging on financial document image data and the problem of OCR misrecognition.

[0005] In summary, the current prior art ignores the financial document image data degradation and OCR misrecognition problems caused by the non-uniform storage medium aging, which cannot meet the needs of enterprises in actual financial consulting and data processing, and will bring serious challenges to the accuracy and security of enterprise financial data, therefore, it is urgent to propose an enterprise financial consulting platform based on financial data analysis to solve the above problems.

[0006] The above information disclosed in the background section is only for the purpose of enhancing the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0007] The purpose of the present application is to provide an enterprise financial consulting platform and method based on financial data analysis, which accurately identifies the aging risk of storage media through a storage media health monitoring module, identifies high-risk silent data errors and captures local damage to storage media through a silent data error detection module, repairs documents through a document repair and OCR module, and executes file data migration decisions through a storage optimization decision module to solve the problems in the background art.

[0008] In order to achieve the above-mentioned purpose, the present application provides the following technical scheme: an enterprise financial consulting platform based on financial data analysis, comprising a storage media health monitoring module: real-time acquisition of physical health indicators of storage devices, construction of storage media health dynamic portrait, marking of high-risk storage areas, and generation of dynamically updated storage media health heat map, transmission to the silent data error detection module; The silent data error detection module: according to the storage media health heat map, all financial document image files stored in the media area are retrieved, and the financial document images stored in the high-risk area are double-checked, the silent data errors are identified, and the suspicious file list is output to the document repair and OCR module; The document repair and OCR module: based on the suspicious file list of silent data errors, a hierarchical repair strategy is enabled, and context-aware OCR credibility enhancement is performed, to obtain repaired documents, OCR results and credibility reports, which are transmitted to the storage optimization decision module; The storage optimization decision module: receiving the storage media health dynamic portrait, the repaired documents, the OCR results and the credibility reports analyzes the document value and risk level, and performs migration cost-risk trade-off and intelligent migration, generates migration decisions, and feedbacks adjustment to dynamically optimize the storage strategy.

[0009] Optionally, the storage media health monitoring module comprises a multi-source data acquisition interface, an online training engine and a heat map generator, wherein the multi-source data acquisition interface: is used to connect the storage device controller to obtain real-time converged storage device SMART data; The online training engine is used to continuously update the parameters of the Gradient Boosting Tree (GBDT) model. The heatmap generator is used to visualize the distribution of high-risk storage areas. The physical health indicators are obtained by connecting to the storage device controller through a multi-source data acquisition interface. These indicators include real-time aggregated storage device SMART parameters, controller ECC error correction count, physical bad block mapping table, historical read / write error logs, and environmental temperature and humidity data. The real-time aggregated storage device SMART parameters include the number of remapped sectors, the number of P / E cycles, and the temperature.

[0010] Optionally, the steps for constructing the dynamic health profile of the storage medium are as follows: Input physical health indicators and perform preprocessing operations such as cleaning and normalization to generate preprocessed data; Feature data is obtained by extracting features that reflect the health status of the storage medium from the preprocessed data. Feature data is input into the Gradient Boosting Tree (GBDT) model of an online training engine for training, resulting in a prediction model for storage medium health to predict risk values. The formula for calculating the risk value using the GBDT model is as follows: ,and In the formula, Represented as an eigenvector, This represents the raw risk value output by the GBDT model predicting the health of the storage medium. Represented as the total number of trees, Represented as the first input feature vector after... The prediction function of a decision tree. Represented as a tree structure space of GBDT; Based on the predicted risk value, a risk score is quantified, a decision rule for high-risk marking is set, and high-risk storage areas are marked. The quantification formula for the risk score is as follows: In the formula, This is represented as the final risk score, and , Represented as a natural constant, This is represented as the slope parameter controlling the steepness of the kurtosis function. This is represented as the offset parameter for adjusting the threshold center; The decision-making rules for high-risk marking are to set risk score thresholds and sudden bad block thresholds; The marking criteria for high-risk storage areas must meet the following condition: number of newly added bad blocks. ,or ; Based on the predicted risk value and combined with the physical structure information of the storage device, a dynamic profile of the storage medium's health can be constructed. Extract features reflecting the health status of the storage medium from the preprocessed data to obtain feature data.

[0011] Optionally, the step of generating the storage medium health heat map is as follows: Divide the storage space grid: discretize the physical space of the physical storage device into a three-dimensional grid, including the first dimension of the logical block address (LBA) range of data storage in the storage device, the second dimension of the physical Die position in the solid state drive (SSD), and the third dimension of the sliding window size of a 24-hour time window; Calculate the grid heat value: calculate the comprehensive risk indicator according to the risk score and the access frequency of the high-risk storage area; Color mapping encoding: use HSV color space conversion and output RGB values; Storage medium health heat map output: according to the storage area division and color mapping relationship, use the heat map generator to draw the storage medium health heat map, and intuitively display the distribution of high-risk storage areas.

[0012] Optionally, the step of double-mode verification is as follows: A1. Based on the high-risk areas marked by the storage medium health heat map, retrieve all financial document image files stored in the high-risk areas; A2. Use the bit layer verification model to verify the data integrity of the financial document image files, and use the semantic layer verification model to verify the business logic of the financial document image files; The verification steps of the bit layer verification model are as follows: B1. Use the BLAKE3 algorithm to process the financial document image files in blocks, and calculate the BLAKE3 hash value for each data block to generate the current enhanced block hash value of the file; B2. Compare the current enhanced block hash value of the file with the reference hash value recorded when the file was initially stored to determine the difference block; The verification steps of the semantic layer verification model are as follows: C1. Use a lightweight financial document structure parser to locate and extract the pixel distribution features of the amount and date box regions in the financial document image files; C2. Compare the extracted pixel distribution features with the reference model of similar documents, calculate the abnormal score, and identify the abnormal structure distortion; A3. If either the bit layer verification model or the semantic layer verification model fails, it is determined that the financial document image file has a high-risk silent error; A4. According to the high-risk silent error result determined by the double-mode verification, analyze the bit error pattern and semantic anomaly mapping to infer the error type of the silent data.

[0013] Optionally, the generation logic of the suspicious file list is as follows: According to the financial document image file with high-risk silent errors, record the unique identifier of the file, the risk level, the error type and the damaged area prediction coordinates, and generate a suspicious file list; Quantitatively calculate the risk level and set the damaged area prediction rule to generate the damaged area prediction coordinates; Formulate error type fusion rules based on the type combination of the mapping table, and output the data structure of the suspicious file list.

[0014] Optionally, the document repair and OCR module includes an erasure code decoder, a FinDoc-RepairGAN processor and a credibility fusion engine, wherein the erasure code decoder is used to reconstruct and repair the bit error file to recover the original data block: The FinDoc-RepairGAN processor: loads a pre-trained invoice / contract special repair model, which is used to start the repair model for the structure abnormal file determined in the semantic layer verification; The credibility fusion engine: used to integrate the OCR confidence and the business rule verification result, and generate a field-level credibility score; The processing steps of the hierarchical repair strategy are as follows: According to the suspicious file list, the files are divided into two categories, one is the hash verification failed file with bit layer error, and the other is the structure abnormal file with semantic layer error; For the hash verification failed file with bit layer error, the erasure code decoder uses the characteristics of Reed-Solomon erasure code to reconstruct the original data block to repair the bit layer; For the structure abnormal file with semantic layer error, use the FinDoc-RepairGAN processor for repair work until the repair meets the constraints of the discriminator.

[0015] Optionally, the execution steps of the context-aware OCR credibility enhancement are as follows: The repaired image is used to extract text information using a financial OCR engine, map local confidence, and generate attention heat maps and character-level confidence by using the attention mechanism inside the OCR; Match the recognition result of the OCR with the financial knowledge base to verify the business rules, and in the matching process, quantitatively calculate the rule confidence and mark the recognition items that contradict the rules; Use the credibility fusion engine to integrate the local confidence and the results of the business rule verification to generate a field-level credibility score; According to the credibility fusion result credibility score calculation storage risk attenuation process, reflecting the credibility of each field recognition result, and output the weighted OCR result and credibility report.

[0016] Optionally, the migration cost-risk trade-off processing logic is as follows: Collecting storage media health dynamic portrait, repaired documents, OCR results and credibility report analyzing document value and risk level, quantifying file importance, comprehensive risk index and migration cost indicators; According to the quantified indicators, a migration value model is constructed; Using multi-objective optimization algorithm NSGA-II to solve the migration value model, and obtaining a Pareto optimal solution set; The migration decision generation step is as follows: According to the Pareto optimal solution set, a file migration priority list is generated; According to the file migration priority and value / risk attributes, the files are classified into different categories, and different migration strategies are adopted, including immediately migrating to a healthy storage pool for high value-high risk files, and performing cold / hot tiered storage; For low value-high risk files, downgrading storage, which can be transferred to a high redundancy but low performance area or only keep the repaired copy.

[0017] A computer device, comprising: a memory, a processor and a computer readable storage medium; the memory and the computer readable storage medium store a computer program, and the processor implements the enterprise financial consulting platform based on financial data analysis when executing the computer program.

[0018] In the above technical solution, the technical effects and advantages provided by the present application are: The present application realizes accurate identification of storage media aging risk by real-time collection of physical health indicators of storage devices and construction of storage media health dynamic portrait, converts invisible hardware aging into quantifiable risk information, improves risk identification capability, adopts "bit layer_semantic layer" dual-mode verification to realize identification of high-risk silent data errors, and captures the function of local damage of storage media; Through the hierarchical repair strategy, the effect of document repair is improved by using erasure code decoder and FinDoc-RepairGAN processor; Finally, based on the correlation between technical risk and business value, file data migration decision is made, which not only reduces the migration cost, but also improves the performance and reliability of the storage medium, and combined with the feedback adjustment of the storage strategy, it expands to adapt to the business needs and medium state of different enterprises, effectively safeguards the safety and accuracy of financial data, and provides strong support for the financial management of enterprises. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only represent some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0020] Fig. 1 A module block diagram of the enterprise financial consulting platform based on financial data analysis of the present application.

[0021] Fig. 2 A flowchart of the dual-mode verification of the present application.

[0022] Fig. 3 A flowchart of the grading repair of the suspicious file of the present application. DETAILED DESCRIPTION

[0023] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these example implementations are provided so that this disclosure will be thorough and complete, and will fully convey the scope of example implementations to those skilled in the art.

[0024] EMBODIMENTS The present application provides an enterprise financial consulting platform based on financial data analysis as shown in Figs. 1-3 The enterprise financial consulting platform based on financial data analysis comprises a storage medium health monitoring module: real-time collection of physical health indicators of a storage device, construction of a storage medium health degree dynamic portrait, marking of a high-risk storage area, and generation of a dynamically updated storage medium health heat map, transmission to a silent data error detection module. Specifically, the storage medium health monitoring module comprises a multi-source data acquisition interface, an online training engine and a heat map generator, wherein the multi-source data acquisition interface is used to connect a storage device controller to obtain real-time converged storage device SMART data. The online training engine is used to continuously update gradient boosting tree GBDT model parameters. The heat map generator is used to visualize the distribution of high-risk storage areas.

[0025] Specifically, the physical health indicators are real-time converged storage device SMART parameters, controller ECC error correction count, physical bad block mapping table, historical read / write error logs and environmental temperature and humidity data obtained by connecting the storage device controller through the multi-source data acquisition interface, wherein the real-time converged storage device SMART parameters include the number of remapping sectors, the number of P / E cycles and the temperature.

[0026] Specifically, the construction steps of the storage medium health degree dynamic portrait are as follows: The input physical health indicators are preprocessed by cleaning and normalization to remove noise and outliers, and make the data comparable, to generate preprocessed data; Features reflecting the health status of the storage medium are extracted from the preprocessed data to obtain feature data, wherein the feature data includes time series derived features and cross features, and the expression of the time series derived features is , wherein represents the ECC correction increment per unit time, represents the ECC correction count increment in the time window, represents the time window, represents the current time, represents the ECC correction count; , wherein represents the bad block density, represents the total number of bad blocks counted from the bad block mapping table, represents the total number of blocks of the storage medium; The expression of the cross feature is , wherein represents the temperature stress, represents the maximum value function, represents the current temperature, represents the temperature threshold, represents the logarithm of the number of P / E cycles, represents the number of P / E cycles; The vector expression of the feature data is , wherein represents the feature vector; The feature data is input into a gradient boosting decision tree (GBDT) model of an online training engine for training to obtain a prediction model of the health degree of the storage medium to predict the risk value, wherein the calculation formula of the GBDT model to predict the risk value is , and , wherein represents the original risk value output by the GBDT model to predict the health degree of the storage medium, represents the total number of trees, represents the prediction function of the th decision tree after inputting the feature vector, represents the tree structure space of the GBDT; The risk score is quantified according to the predicted risk value, a decision rule for setting a high-risk mark is set, and a high-risk storage area is marked, wherein the quantification formula of the risk score is , wherein represents the final risk score, and , is expressed as a natural constant, is expressed as a slope parameter of the control function steepness, is expressed as an offset parameter of the adjustment threshold center; The decision rule of the high-risk mark is to set the risk score threshold and the sudden bad block threshold, wherein the risk score threshold is 0.85, and the sudden bad block threshold is 5 per hour; The marking condition of the high-risk storage area needs to meet the number of newly added bad blocks , or ; According to the predicted risk value and in combination with the physical structure information of the storage device, a storage medium health dynamic portrait is constructed.

[0027] Specifically, the generation logic of the storage medium health heat map is as follows: Divide the storage space grid: discretize the physical space of the physical storage device into a three-dimensional grid, including the first dimension of the logical block address (LBA) range of the data storage in the storage device, the second dimension of the physical Die position in the solid state disk (SSD), and the third dimension of the sliding window size of the 24-hour time window. The index expression of the grid in the LBA dimension is , wherein is expressed as the index of the grid in the LBA dimension, is expressed as the index of the grid in the LBA dimension, is expressed as the minimum logical block address, is expressed as the number of span blocks in the LBA direction of each grid, is expressed as a floor function; Calculate the grid heat value: calculate the comprehensive risk indicator according to the risk score and the access frequency of the high-risk storage area, wherein the calculation formula of the comprehensive risk indicator is , wherein is expressed as the heat value of the grid in the space and time coordinates , is expressed as the weight coefficient of the grid risk score, is expressed as the risk score of the grid in the space and time coordinates , is expressed as the weight coefficient of the grid risk score, is expressed as the risk score of the grid in the space and time coordinates , is expressed as the weight coefficient of the corresponding access rate, is expressed as the area access frequency of the grid in the time window, is expressed as the global maximum access number, is expressed as the spatial coordinate of the grid; Color mapping encoding: using HSV color space conversion and outputting RGB values, wherein the expression of the HSV color space conversion is , and and set the saturation fixed at 100% and the lightness fixed at 90%, wherein is expressed as a hue, shown as a gradient from green to red, for indicating the transition process from low risk to high risk, and is red when 0, is green when 1; The expression of the output RGB value is , wherein is expressed as an RGB value, is expressed as an RGB conversion function; Storage medium health heat map output: According to the storage area division and the color mapping relationship, the storage medium health heat map is drawn using the heat map generator, and the distribution of high-risk storage areas is intuitively displayed.

[0028] Silent data error detection module: According to the storage medium health heat map, all financial document image files stored in the medium area are retrieved, and the financial document images stored in the high-risk area are subjected to double-mode verification to identify silent data errors, and the suspicious file list is output to the document repair and OCR module; Specifically, the steps of double-mode verification are as follows: A1, based on the high-risk area marked by the storage medium health heat map, all financial document image files such as invoices and contract scans stored in the high-risk area are retrieved; A2, using a bit layer verification model to verify the data integrity of the financial document image file, and using a semantic layer verification model to verify the business logic of the financial document image file; The verification steps of the bit layer verification model are as follows: B1, using BLAKE3 algorithm to block process the financial document image file, and calculating BLAKE3 hash value for each data block respectively, generating the current enhanced block hash value of the file, wherein the expression of block processing is , and , , and , wherein represents the total number of blocks after the financial document image file is segmented, represents the total size of the original financial document image file, represents the size of each data block, and , represents the upward rounding function, represents the content of the th data block, represents the financial document image file reading function, represents the unique identifier of the financial document image file,​ Represented as the first The starting offset of each data block in the file. Represented as a data block index; The formula for calculating the BLAKE3 hash value is as follows: In the formula, Represented as the first The hash value currently calculated for each data block. Represented as the BLAKE3 cryptographic hash function, Represented as the first The base hash value initially stored for each data block. This is represented as the baseline hash lookup function, used to retrieve the initial hash value from trusted storage; B2. Compare the current enhanced block hash value of the file with the base hash value recorded when the file was initially stored to determine the difference blocks. The conditions for determining the difference blocks are as follows: In the formula, Represented as the first Error flags for each data block. The result is a binary value; a result of 1 indicates an error, and a result of 0 indicates a normal result. Represented as the Hamming distance function, Represented as calculating the first The hash value currently calculated for each data block With the The base hash value initially stored for each data block The number of bits that differ between the two hash values; The validation steps for the semantic layer validation model are as follows: C1. Using a lightweight financial document structure parser, locate and extract the pixel distribution features of the amount and date box regions in the financial document image file. The expression for the pixel distribution features is: ,and ,as well as In the formula, It is represented as a vector of pixel distribution features. Represented as the mean of the red channel. Represented as the color channel mean. This is represented as the average value of the green channel. This is expressed as the variance of the edge gradient after Canny edge detection. This represents the total number of pixels in the ROI region. Represented as pixels within a region, This refers to extracting the amount and date frame area from a financial document image file. Represented as pixels The value in the R channel, Represented as pixels the value of the G channel, the total number of edge pixels in the ROI, the value of the B channel, the total number of edge pixels in the ROI, the gradient amplitude of the pixel point , the average gradient amplitude, C2, comparing the extracted pixel distribution features with the baseline model of the same type of document, identifying the structural distortion with abnormal score through calculation, wherein the calculation formula of the abnormal score is , wherein, the abnormal score, the Euclidean distance, the baseline feature vector of the same type of document cluster, the total area of the ROI region, the total area of the financial document image file, the area weight, A3, when any one of the bit layer verification model or the semantic layer verification model fails, it is determined that the financial document image file has a high-risk silent error; A4, according to the high-risk silent error result determined by the dual-mode verification, the error type of the silent data is inferred by analyzing the bit error pattern and the semantic abnormal mapping, wherein the expression of the bit error pattern is , and , wherein, the error type of the silent data, large area damage, local damage, continuous error threshold, total number of error blocks, The content of semantic abnormal mapping is large area damage corresponding to the amount / date digital area damage of the financial document, the reason is that the loss of continuous data blocks causes large area pixel distortion, and local damage corresponding to the table line fracture / stamp defect of the financial document, the reason is that isolated data block error causes local pixel anomaly.

[0029] Specifically, the generation logic of the suspicious file list is as follows: According to the financial document image file with high-risk silent error, record the unique identifier , risk level, error type and damaged area prediction coordinates, generate a suspicious file list, wherein the expression of the suspicious file list is , wherein, the suspicious file list, a unique identifier of the financial document image file, a risk level, an error type represented as silent data, a damaged area prediction coordinate represented as a damaged area prediction coordinate; The risk level is quantitatively calculated, and the damaged area prediction rule is set to generate the damaged area prediction coordinate, wherein the quantitative calculation formula of the risk level is , wherein a bit error weight coefficient, a semantic anomaly weight coefficient, a bit error intensity; The expression of the damaged area prediction coordinate is , wherein a pixel distribution feature vector of the amount area, a pixel distribution feature vector of the date area, a feature deviation threshold value; An error type fusion rule is formulated based on the type combination of the mapping table, and a data structure of the suspicious file list is output, wherein the expression of the error type fusion rule is , wherein a fused error type, a bit layer silent data error type, a semantic layer damaged area prediction coordinate, a type combination operator.

[0030] Document repair and OCR module: based on the silent data error of the suspicious file list, a hierarchical repair strategy is enabled, and a context-aware OCR credibility enhancement is performed, to obtain a repaired document, an OCR result and a credibility report, which are transmitted to the storage optimization decision module; Specifically, the document repair and OCR module includes an erasure code decoder, a FinDoc-RepairGAN processor and a credibility fusion engine, wherein the erasure code decoder is used to reconstruct and repair the bit error file to recover the original data block: The FinDoc-RepairGAN processor loads a pre-trained invoice / contract special repair model, which is used to start the repair model for the structure abnormal file determined in the semantic layer check, and the FinDoc-RepairGAN processor includes a generator and a discriminator, wherein the generator receives a damaged image and a document type label; The discriminator has a built-in financial rule constraint unit to ensure that the repaired area conforms to the visual rules of financial vouchers, and the financial rule constraint unit includes a digital font consistency detector, a table line continuity evaluator and a seal integrity verifier; The credibility fusion engine is used to integrate OCR confidence scores with business rule verification results and generate field-level credibility scores.

[0031] Specifically, the processing logic of the graded repair strategy is as follows: Based on the list of suspicious files, the files are divided into two categories: one is files with failed hash verification due to bit-level errors, and the other is files with structural anomalies due to semantic-level errors. For files with failed hash verification due to bit layer errors, the erasure coding decoder is called to reconstruct the original data blocks using the characteristics of Reed-Solomon erasure coding in order to repair the bit layer. For structurally abnormal files with semantic layer errors, the FinDoc-RepairGAN processor is used for repair until the repair meets the discriminator's constraints. The generator takes the damaged image and document type label as input to reconstruct and generate the repaired image. The discriminator then checks the generated repaired area according to the built-in financial rule constraint unit to ensure that the repaired image conforms to the visual rules of financial documents, such as consistency of digital fonts, continuity of table lines, and integrity of seals.

[0032] Specifically, the execution steps for context-aware OCR credibility enhancement are as follows: The repaired image is then processed using a financial OCR engine to extract text information and map local confidence scores. By leveraging the OCR's internal attention mechanism, an attention heatmap and character-level confidence scores are generated. The expression for the attention heatmap is as follows: ,and , In the formula, Represented as coordinates Attention weights Represented as pixel coordinates for OCR recognition. Represented as a natural exponential function, This represents the summation of positions across the entire graph. This is represented as the semantic vector of the current character recognized by the decoder query. Represented as Visual feature image key vector, Represented as feature dimension, This is represented as a vector transpose. Represented as a scaling factor; The expression for character-level confidence is: In the formula, Represented as characters Confidence level, Represented as a character index, This represents the character determined by the OCR detection box. The bounding box area, This is represented by the maximum value sign; The OCR results are matched against a financial knowledge base to verify business rules. During the matching process, rule confidence is quantified, and contradictory entries are flagged. The financial knowledge base includes various financial business rules, such as whether the total amount (including tax) of an invoice equals the sum of all other amounts, whether the date logic is reasonable, and whether the account has passed Luhn verification. The expression for rule confidence is as follows: In the formula, Represented as a field Rule confidence, Represented as a field; The credibility fusion engine integrates the results of local confidence and business rule validation to generate field-level credibility scores. The expression for the credibility score after credibility fusion is as follows: In the formula, Represented as a field The base confidence level, Represented as the corresponding character The confidence level weighting coefficient This is represented by the minimum value symbol. Represented as the corresponding field The weighting coefficients of the rule confidence; The storage risk attenuation process is calculated based on the credibility score of the credibility fusion result, reflecting the credibility level of the recognition result for each field, and outputs weighted OCR results and a credibility report. The formula for calculating storage risk attenuation is as follows: In the formula, Represented as a field The ultimate credibility, Represented as a sensitive factor, Represented as a risk attenuation factor, It is represented as a natural exponential decay function.

[0033] Storage optimization decision module: Receives dynamic health profiles of storage media, repaired documents, OCR results and credibility reports, analyzes document value and risk level, performs migration cost-risk trade-offs and executes intelligent migration, generates migration decisions, and provides feedback to adjust dynamic optimization storage strategies.

[0034] Specifically, the logic for handling the migration cost-risk trade-off is as follows: This involves collecting dynamic health profiles of storage media, analyzing repaired documents, OCR results, and credibility reports to determine document value and risk levels. It also quantifies file importance, comprehensive risk index, and migration costs. The expression for file importance is: In the formula, Indicated as the importance of the document, a weight coefficient representing the access frequency, a weight coefficient representing the access frequency of the file, a weight coefficient representing the platform business audit and legality, a weight coefficient representing the platform business audit and legality, the expression of the comprehensive risk index is , wherein, a weight coefficient representing the platform business audit and legality, a weight coefficient representing the platform business audit and legality, a weight coefficient representing the OCR credibility score, the expression of the migration cost is , wherein, a weight coefficient representing the file migration cost, a weight coefficient representing the bandwidth cost, a weight coefficient representing the size of the file to be migrated, a weight coefficient representing the available bandwidth, a weight coefficient representing the IOPS cost, a weight coefficient representing the IOPS required for migration, a weight coefficient representing the available migration IOPS, a migration value model is constructed according to the quantitative indicators, wherein the expression of the migration value model is , wherein, a weight coefficient representing the migration value, a weight coefficient representing the smoothing constant to prevent the denominator from being 0, a multi-objective optimization algorithm NSGA-II is used to solve the migration value model to obtain a Pareto optimal solution set.

[0035] Specifically, the generation steps of the migration decision are as follows: According to the Pareto optimal solution set, a migration priority list of the file is generated; According to the migration priority and value / risk attributes of the file, the file is classified into different categories, and different migration strategies are adopted, including immediately migrating the high value-high risk file to the healthy storage pool and performing cold / hot tiered storage; For low value-high risk files, degrade the storage, and can be stored in a high redundancy but low performance area or only keep a repaired copy.

[0036] Specifically, the logic of dynamically optimizing the storage strategy is as follows: According to the migration effect and the health trend of the storage medium, the effectiveness of the current storage strategy is evaluated, and an evaluation result is generated; According to the evaluation result, the health threshold is dynamically adjusted, the inspection frequency is optimized, and the redundancy strategy is upgraded; The above effect evaluation and adjustment optimization process is repeatedly performed to achieve continuous dynamic optimization of the storage strategy.

[0037] The above formulas are dimensionless values calculated, and the formulas are obtained by collecting a large amount of data to simulate a formula of the most recent real situation. The preset parameters in the formula are set by a person skilled in the art according to the actual situation.

[0038] The above embodiments can be realized wholly or partially by software, hardware, firmware, or any other combination. When realized by software, the above embodiments can be realized wholly or partially in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are wholly or partially generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another, for example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center through a wired or wireless (such as infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. containing one or more available medium collections. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state disk.

[0039] It should be understood that the size of the sequence number of each process described above in various embodiments of the present application does not mean the order of execution, and the execution order of each process should be determined by its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0040] Those of ordinary skill in the art can realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be realized in electronic hardware, or a combination of computer software and electronic hardware. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0041] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A business financial advisory platform based on financial data analysis, characterized in that, The storage medium health monitoring module includes a multi-source data acquisition interface, an online training engine, and a heat map generator. The multi-source data acquisition interface is used to connect a storage device controller to obtain real-time converged storage device SMART data. The online training engine is used to continuously update gradient boosting decision tree (GBDT) model parameters. The heat map generator is used to visualize high-risk storage area distribution. The physical health indicators are real-time converged storage device SMART parameters, controller ECC error correction count, physical bad block mapping table, historical read / write error logs, and environmental temperature and humidity data obtained by connecting the storage device controller through the multi-source data acquisition interface. The real-time converged storage device SMART parameters include the number of remapping sectors, the number of P / E cycles, and the temperature.

2. The enterprise financial advisory platform based on financial data analysis as claimed in claim 1, wherein, The construction steps of the storage medium health dynamic portrait are as follows: Input the physical health indicators and perform preprocessing operations such as cleaning and normalization to generate preprocessed data. Extract features that reflect the health status of the storage medium from the preprocessed data to obtain feature data. The decision rule for high-risk labeling is to set risk score thresholds and sudden bad block thresholds.

3. The enterprise financial advisory platform based on financial data analysis as claimed in claim 1 wherein, According to the predicted risk value and the physical structure information of the storage device, the storage medium health dynamic portrait is constructed. Extract features that reflect the health status of the storage medium from the preprocessed data to obtain feature data. The generation steps of the storage medium health heat map are as follows: The feature data is input to a gradient boosting decision tree (GBDT) model of an online training engine for training to obtain a prediction model of the health degree of the storage medium to predict a risk value, wherein a calculation formula of the GBDT model for predicting the risk value is , and , wherein is represented as a feature vector, is represented as an original risk value output by the GBDT model for predicting the health degree of the storage medium, is represented as a total number of trees, is represented as a prediction function of the th decision tree after inputting the feature vector, is represented as a tree structure space of the GBDT. According to the predicted risk value, a risk score is quantified, a decision rule for setting a high-risk mark is set, and a high-risk storage area is marked, wherein a quantification formula of the risk score is , wherein, is expressed as a final risk score, and , is expressed as a natural constant, is expressed as a slope parameter of steepness of a control function, is expressed as an offset parameter for adjusting a center of a threshold. Divide the storage space grid: discretize the physical space of the physical storage device into a three-dimensional grid, including the first dimension of the logical block address (LBA) range of the data storage in the storage device, the second dimension of the physical Die position in the solid state drive (SSD), and the third dimension of the sliding window size of the 24-hour time window. The marking condition of the high-risk storage area needs to meet the number of newly added bad blocks , or ; Calculate the grid heat value: calculate the comprehensive risk indicator according to the risk score and the access frequency of the high-risk storage area. Color mapping encoding: use HSV color space conversion and output RGB values.

4. The enterprise financial advisory platform based on financial data analysis as claimed in claim 3 wherein, ​ ​ ​ ​ Storage media health heat map output: according to the storage area division and the color mapping relationship, a storage media health heat map is drawn using a heat map generator, and the distribution of high-risk storage areas is intuitively displayed.

5. The enterprise financial advisory platform based on financial data analysis as claimed in claim 1 wherein, The steps of the double-mode verification are as follows: A1, based on the high-risk area marked by the storage media health heat map, all financial document image files stored in the high-risk area are retrieved; A2, the data integrity of the financial document image files is verified using a bit layer verification model, and the business logic of the financial document image files is verified using a semantic layer verification model; The verification steps of the bit layer verification model are as follows: B1, using BLAKE3 algorithm to process the financial document image file, and calculating BLAKE3 hash value for each data block respectively, generating the current enhanced block hash value of the file; B2, comparing the current enhanced block hash value of the file with the baseline hash value recorded when the file is initially stored to determine the difference block; The verification steps of the semantic layer verification model are as follows: C1, using a lightweight financial document structure parser to locate and extract the pixel distribution features of the amount and date box area in the financial document image file; C2, comparing and analyzing the extracted pixel distribution features with the baseline model of the same type of document, identifying the abnormal structure distortion by calculating the abnormal score; A3, if any one of the bit layer verification model or the semantic layer verification model fails, it is determined that the financial document image file has high-risk silent errors; A4, according to the high-risk silent error result determined by the double-mode verification, the error type of the silent data is inferred by analyzing the bit error mode and the semantic abnormal mapping.

6. The enterprise financial advisory platform based on financial data analysis as claimed in claim 5 wherein, The generation logic of the suspicious file list is as follows: According to the financial document image file with high-risk silent errors, the unique identifier, risk level, error type and damaged area prediction coordinates of the file are recorded to generate a suspicious file list; Quantitative calculation is performed on the risk level, and damaged area prediction coordinates are generated according to the damaged area prediction rules; Based on the type combination of the mapping table, the error type fusion rules are formulated, and the data structure of the suspicious file list is output.

7. The enterprise financial advisory platform based on financial data analysis as claimed in claim 6 wherein, The document repair and OCR module includes an erasure code decoder, a FinDoc-RepairGAN processor and a credibility fusion engine, wherein the erasure code decoder is used to reconstruct and repair the bit error file to recover the original data block: The FinDoc-RepairGAN processor loads a pre-trained invoice / contract special repair model to start the repair model for the structure abnormal file determined in the semantic layer verification; The credibility fusion engine is used to integrate the OCR confidence and the business rule verification result, and generate a field-level credibility score; The processing steps of the hierarchical repair strategy are as follows: According to the suspicious file list, the files are divided into two categories, one is the hash verification failed file of the bit layer error, and the other is the structure abnormal file of the semantic layer error; For the hash verification failed file of the bit layer error, the Reed-Solomon erasure code decoder is called to reconstruct the original data block to repair the bit layer; For the structural abnormal files with semantic layer errors, the FinDoc-RepairGAN processor is used for repair work until the repair meets the constraints of the discriminator.

8. The enterprise financial advisory platform based on financial data analysis as claimed in claim 7 wherein, The execution steps of the context-aware OCR confidence enhancement are as follows: The repaired image is used to extract text information using a financial OCR engine, map local confidence, and generate attention heat maps and character-level confidence by utilizing the attention mechanism inside the OCR. The recognition results of the OCR are matched with a financial knowledge base to verify business rules, and during the matching process, the rule confidence is quantified, and the recognized items that contradict the rules are marked. The confidence fusion engine is used to integrate the local confidence and the results of business rule verification to generate field-level confidence scores. According to the confidence fusion results, a risk decay process is calculated based on the confidence score, reflecting the degree of confidence of each field recognition result, and the weighted OCR results and confidence report are output.

9. The enterprise financial advisory platform based on financial data analysis as claimed in claim 8, wherein, The processing logic of the migration cost-risk trade-off is as follows: The storage medium health dynamic image, repaired document, OCR result, and confidence report are collected to analyze the document value and risk level, quantify the indicators of file importance, comprehensive risk index, and migration cost; A migration value model is constructed according to the quantified indicators; The multi-objective optimization algorithm NSGA-II is used to solve the migration value model to obtain a Pareto optimal solution set; The generation steps of the migration decision are as follows: According to the Pareto optimal solution set, a migration priority list of the file is generated; According to the migration priority and value / risk attributes of the file, the file is classified into different categories, and different migration strategies are adopted, including immediate migration to a healthy storage pool for high-value-high-risk files; For low-value-high-risk files, the storage is downgraded.

10. A computer device comprising: A memory, a processor, and a computer-readable storage medium; the memory and the computer-readable storage medium store a computer program, characterized in that: when the processor executes the computer program, the enterprise financial consulting platform based on financial data analysis of any one of claims 1-9 is realized.

Images