A big data-based financial management method and system

Abstract

The application relates to the technical field of data processing, in particular to a financial management method and system based on big data, which comprises the following steps: S1, performing semantic-visual fusion cleaning and standardization processing on unstructured financial data of a small and medium-sized enterprise, and outputting a standardized financial data matrix; and S2, based on the standardized financial data matrix, combining ERP data and external supply chain credit data, and performing cash flow health degree prediction and supply chain payment strategy linkage; when the application is used, semantic-visual fusion cleaning is used to reduce the workload of manual arrangement of unstructured data, improve the data accuracy, and solve the pain points of difficult data entry, many errors and omissions of the small and medium-sized enterprise; through real-time cash flow health degree prediction, the small and medium-sized enterprise can identify the fund gap risk 7-30 days in advance, and passive supply interruption can be avoided; through a dynamic payment strategy, the overdue payment probability is reduced, the cooperation stability of the core supplier is improved, an advance payment discount can be obtained, and the procurement cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically to a financial management method and system based on big data. Background Technology

[0002] Financial management is the core management activity of an enterprise or organization to systematically plan, control, make decisions, and supervise the movement of funds. It aims to optimize resource allocation, control risks, and maximize value. Big data-based financial management refers to the use of big data technologies, such as data mining, machine learning, and artificial intelligence, to collect, integrate, analyze, and visualize massive amounts of multi-source financial and non-financial data. It also involves the deep integration of big data technology with traditional financial management processes to achieve data-driven and intelligent financial decision-making, thereby improving financial management efficiency and risk control.

[0003] Patent publication number CN201810382074.4 describes in its specification "A financial analysis management system and method based on big data provided in this application embodiment, the system comprising: a financial statement acquisition module for acquiring original forms of various financial statements of a target enterprise; a financial item extraction module for extracting predetermined types of financial items and corresponding financial amounts from the original forms of the various financial statements; a financial vector generation module for generating financial vectors; a financial vector recognition module for matching the financial vectors with a pre-trained financial vector recognition model; and a financial type output module for outputting the financial type corresponding to the target enterprise after determining the type of the financial vectors." Service type. The big data-based financial analysis management system of this application can comprehensively extract financial indicators from financial statements and perform comprehensive analysis based on machine learning, fully explore the regular characteristics of the analyzed objects in terms of financial information, make the analysis results more comprehensive, and save human resources costs. Although the above technology has the advantage of breaking through the depth and efficiency bottlenecks of traditional analysis by upgrading financial analysis from human experience-driven to data intelligence-driven, the unstructured financial data of SMEs has the problems of fragmentation, low quality and difficulty in correlation. Traditional single OCR or text parsing technology cannot achieve standardized processing, resulting in low data utilization. At the same time, existing cash flow management is mostly a post-event response and lacks real-time forecasting capabilities.

[0004] In conclusion, developing a financial management method and system based on big data remains a key issue that urgently needs to be addressed in the field of data processing technology. Summary of the Invention

[0005] The purpose of this invention is to address the problems of fragmented, low-quality, and difficult-to-connect unstructured financial data of small and medium-sized enterprises in the existing technology. Traditional single OCR or text parsing technologies cannot achieve standardized processing, resulting in low data utilization. At the same time, existing cash flow management is mostly reactive and lacks real-time forecasting capabilities. This invention provides a financial management method and system based on big data.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention provides a big data-based financial management method, comprising:

[0008] S1. Perform semantic-visual fusion cleaning and standardization on unstructured financial data of SMEs, and output a standardized financial data matrix;

[0009] S2. Based on the standardized financial data matrix, combined with ERP data and external supply chain credit data, perform cash flow health prediction and link supply chain payment strategies.

[0010] Furthermore, in step S1, the method for performing semantic-visual fusion cleaning and standardization on the unstructured financial data of SMEs and outputting a standardized financial data matrix is ​​as follows:

[0011] The method employs a BERT model pre-trained in the financial domain. It inputs unstructured text data from unstructured financial data of SMEs and then uses entity recognition algorithms to extract five core entities: the person requesting reimbursement, the reason for reimbursement, the amount, the time of occurrence, and the associated business document number. The pre-trained BERT model in the financial domain includes collecting 500,000 textual corpora from the financial domain of SMEs, including but not limited to reimbursement dialogues, financial notices, and reconciliation emails, to construct a corpus. in, For each text The word segmentation process uses Jieba word segmentation combined with an expanded financial terminology database to generate word sequences. Pre-train the BERT base model in the financial domain, optimizing the objective function as bidirectional language model loss + financial entity masking loss, expressed as:

[0012]

[0013] In the formula, For traditional masked language model loss, For the number of mask words, For non-masked word sequences, To predict probabilities for traditional masked language models, This is the financial entity masking loss, used to perform 100% masking of the five core entities already labeled in the corpus. For a collection of entities, To obtain the text sequence after entity removal, the pre-training process was performed for 1500 iterations, using a linear decay strategy for the learning rate. ,final When the perplexity of the validation set is ≤18, pre-training is stopped, and a pre-trained BERT model for the financial domain is obtained. .

[0014] Furthermore, in step S1, the method for performing semantic-visual fusion cleaning and standardization on the unstructured financial data of SMEs and outputting a standardized financial data matrix is ​​as follows:

[0015] Using the BERT model pre-trained in the financial domain Based on this, a conditional random field layer is introduced to extract five types of core entities, using unstructured text data as input. The last hidden layer outputs the feature vector of each word. ,in, The position of a word in the text; Input to a linear layer to obtain 11 emission score matrices for each word. ,in, Given the text length, a conditional random field layer calculates the globally optimal path for the label sequence, aiming to maximize the joint probability of the label sequence:

[0016]

[0017] In the formula, the optimal label sequence is solved using the Viterbi algorithm. Extract the corresponding 5 types of core entities ,in, The person seeking reimbursement It is a reason for reimbursement. It's the amount. It is the time of occurrence. It is associated with the business document number, and the entity identification confidence score is calculated:

[0018]

[0019] In the formula, For entities Word length, The set of locations of entities in the text. For entities The corresponding tag set, For the Sigmoid function, in When the time is right, mark it as a low-confidence entity.

[0020] Furthermore, in step S1, the method for performing semantic-visual fusion cleaning and standardization on the unstructured financial data of SMEs and outputting a standardized financial data matrix is ​​as follows:

[0021] Furthermore, an improved YOLOv8 object detection algorithm is employed to locate key regions in image-based unstructured financial data of SMEs. These key regions include, but are not limited to, amount boxes, invoice code areas, and seal areas. The improved YOLOv8 object detection algorithm involves collecting 50,000 financial images of SMEs, including but not limited to images taken from different angles, with varying lighting intensities, and at different resolutions. LabelImg is used to annotate three types of key regions, generating bounding boxes. ,in, For the amount box, For the invoice code area, For the seal area, construct a financial image dataset. Based on the K-means++ algorithm Clustering of annotation boxes generates anchor box sizes adapted to the financial area:

[0022] Amount frame anchor frame ,

[0023] Invoice code area anchor frame ,

[0024] Seal area anchor frame Furthermore, a category weighting factor is introduced based on the original CIoU loss to highlight the positioning of key financial areas. The expression is:

[0025]

[0026] In the formula, The number of detection layers, The number of anchor frames per layer. This is the actual annotation box. To improve the predicted bounding boxes of the YOLOv8 model, The length of the diagonal of the smallest rectangle enclosing the two frames. The probability that a real target exists. To improve the target confidence predicted by the YOLOv8 model, Category weighting factors are set based on the importance of financial regions: Amount Frame Invoice code area Seal area ; Weights are lost based on the target confidence level; the AdamW optimizer is used with weight decay. Training is performed for 1000 epochs. Training stops when the Average Precision on the validation set is ≥ 95%. Input the image to be processed. The model outputs predicted bounding boxes for three types of key regions. ,in and location reliability ,exist When the confidence level is less than 0.9, it is marked as a low-confidence region. Then, combined with a customized OCR model for financial scenarios, text information within the key regions is extracted. The customized OCR model has optimized the logic for recognizing handwritten cursive characters and fuzzy handwriting. The customized OCR model includes collecting 20,000 text samples from key financial regions, including but not limited to handwritten amounts, fuzzy invoice codes, and seal text, annotating the text content and character positions, and generating sample pairs. ,in, Image patches representing key regions. To correspond to real text, a dedicated dataset for financial OCR was constructed. The customized OCR model adds an attention mechanism layer after the CNN feature extraction module to enhance the feature response of handwritten cursive characters and blurred characters. The attention weight calculation expression is as follows:

[0027]

[0028] In the formula, The feature map output by the CNN. This is the attention weight matrix. For the query vector dimension; for numbers and special characters (including but not limited to ¥ and -) in financial text, a character-level language model constraint is introduced during CTC decoding, with the decoding probability being:

[0029]

[0030] In the formula, For the CTC decoding probability of the CRNN model, For text length, For 3-gram language model probabilities, As a weighting factor;

[0031] It is important to note that the extracted text information is validated for consistency and relevance using a semantic-visual cross-validation rule base to remove invalid data; and the cleaned data is mapped to structured fields that conform to accounting standards to generate a standardized financial data matrix.

[0032] The consistency check includes comparing the core entities extracted by the entity recognition algorithm with the text information extracted by the OCR model. If the deviation is greater than 5%, a manual review reminder is automatically triggered.

[0033] In addition, the correlation verification includes: by collecting historical financial data of enterprises, constructing a correlation rule model, marking isolated data of unrelated business documents as invalid data and removing them.

[0034] Furthermore, in step S2, the method for linking cash flow health forecasting with supply chain payment strategies based on the standardized financial data matrix, combined with ERP data and external supply chain credit data, is as follows:

[0035] The cash flow health prediction includes constructing an LSTM model with a time-series attention mechanism; the input normalization features of the LSTM model... This includes core entities, textual information, ERP data, and external supply chain credit data; temporal features are extracted using a two-layer LSTM, with the first LSTM layer calculating the hidden state and introducing a temporal attention mechanism to calculate attention weights; finally, the future output is passed through a fully connected layer. Daily cash flow forecast :

[0036]

[0037] In the formula, To predict the weight matrix, For bias; a weighted average absolute error is used to assign higher weights to the cash flow forecasts for days 1-7; the Adam optimizer is used with a learning rate of The training process is repeated 500 times, with validation every 20 times. The process is stopped early if the loss on the validation set does not decrease for 10 consecutive times. The LSTM model outputs a cash flow prediction curve for the next 7-30 days and a three-level health index of "safety-early warning-risk".

[0038] Furthermore, in step S2, the method for linking cash flow health forecasting with supply chain payment strategies based on the standardized financial data matrix, combined with ERP data and external supply chain credit data, is as follows:

[0039] Furthermore, based on the aforementioned three-level health index,

[0040] Security level: ;

[0041] Warning level: ;

[0042] Risk level: The three-level health index is converted into a quantitative score. :

[0043]

[0044] In the formula, , For security level, It is at the warning level. The risk level is set at 60%. A supply chain payment priority strategy is dynamically generated using an integer programming algorithm. The actual data after the supply chain payment priority strategy is executed is fed back to the LSTM model, and the weights of the LSTM model are updated using a gradient descent algorithm. The model is optimized once per quarter.

[0045] Furthermore, in step S2, the method for linking cash flow health forecasting with supply chain payment strategies based on the standardized financial data matrix, combined with ERP data and external supply chain credit data, is as follows:

[0046] It is important to note that the three-tiered health index includes a safe level, a warning level, and a risk level. The supply chain payment priority strategy is based on the three-tiered health index. For the safe level, priority is given to paying suppliers with payment terms of ≤30 days and a credit rating of A or above, triggering applications for early payment discounts. The credit rating is based on external credit data. For the warning level, with the constraint that core suppliers are not overdue, priority is given to paying suppliers of essential raw materials for production, followed by suppliers of non-core auxiliary materials. For the risk level, an emergency mechanism is triggered, automatically extracting information on assets that can be quickly converted into cash, and generating payment extension plans for finance personnel to negotiate with suppliers.

[0047] On the other hand, the present invention also provides a big data-based financial management system, comprising:

[0048] The data standardization processing module performs semantic-visual fusion cleaning and standardization processing on unstructured financial data of SMEs, and outputs a standardized financial data matrix.

[0049] The forecasting and strategy linkage module, based on the standardized financial data matrix and combined with ERP data and external supply chain credit data, performs cash flow health forecasting and linkage with supply chain payment strategies.

[0050] Furthermore, the operation process of the data standardization processing module includes:

[0051] The method employs a BERT model pre-trained in the financial domain. It inputs unstructured text data from unstructured financial data of SMEs and then uses entity recognition algorithms to extract five core entities: the person requesting reimbursement, the reason for reimbursement, the amount, the time of occurrence, and the associated business document number. The pre-trained BERT model in the financial domain includes collecting 500,000 textual corpora from the financial domain of SMEs, including but not limited to reimbursement dialogues, financial notices, and reconciliation emails, to construct a corpus. in, For each text The word segmentation process uses Jieba word segmentation combined with an expanded financial terminology database to generate word sequences. Pre-train the BERT base model in the financial domain, optimizing the objective function as bidirectional language model loss + financial entity masking loss, expressed as:

[0052]

[0053] In the formula, For traditional masked language model loss, For the number of mask words, For non-masked word sequences, To predict probabilities for traditional masked language models, This is the financial entity masking loss, used to perform 100% masking of the five core entities already labeled in the corpus. For a collection of entities, To obtain the text sequence after entity removal, the pre-training process was performed for 1500 iterations, using a linear decay strategy for the learning rate. ,final When the perplexity of the validation set is ≤18, pre-training is stopped, and a pre-trained BERT model for the financial domain is obtained. .

[0054] Using the BERT model pre-trained in the financial domain Based on this, a conditional random field layer is introduced to extract five types of core entities, using unstructured text data as input. The last hidden layer outputs the feature vector of each word. ,in, The position of a word in the text; Input to a linear layer to obtain 11 emission score matrices for each word. ,in, Given the text length, a conditional random field layer calculates the globally optimal path for the label sequence, aiming to maximize the joint probability of the label sequence:

[0055]

[0056] In the formula, the optimal label sequence is solved using the Viterbi algorithm. Extract the corresponding 5 types of core entities ,in, The person seeking reimbursement It is a reason for reimbursement. It's the amount. It is the time of occurrence. It is associated with the business document number, and the entity identification confidence score is calculated:

[0057]

[0058] In the formula, For entities Word length, The set of locations of entities in the text. For entities The corresponding tag set, For the Sigmoid function, in When the time is right, mark it as a low-confidence entity.

[0059] Furthermore, an improved YOLOv8 object detection algorithm is employed to locate key regions in image-based unstructured financial data of SMEs. These key regions include, but are not limited to, amount boxes, invoice code areas, and seal areas. The improved YOLOv8 object detection algorithm involves collecting 50,000 financial images of SMEs, including but not limited to images taken from different angles, with varying lighting intensities, and at different resolutions. LabelImg is used to annotate three types of key regions, generating bounding boxes. ,in, For the amount box, For the invoice code area, For the seal area, construct a financial image dataset. Based on the K-means++ algorithm Clustering of annotation boxes generates anchor box sizes adapted to the financial area:

[0060] Amount frame anchor frame ,

[0061] Invoice code area anchor frame ,

[0062] Seal area anchor frame Furthermore, a category weighting factor is introduced based on the original CIoU loss to highlight the positioning of key financial areas. The expression is:

[0063]

[0064] In the formula, The number of detection layers, The number of anchor frames per layer. This is the actual annotation box. To improve the predicted bounding boxes of the YOLOv8 model, The length of the diagonal of the smallest rectangle enclosing the two frames. The probability that a real target exists. To improve the target confidence predicted by the YOLOv8 model, Category weighting factors are set based on the importance of financial regions: Amount Frame Invoice code area Seal area ; Weights are lost based on the target confidence level; the AdamW optimizer is used with weight decay. Training is performed for 1000 epochs. Training stops when the Average Precision on the validation set is ≥ 95%. Input the image to be processed. The model outputs predicted bounding boxes for three types of key regions. ,in and location reliability ,exist When the confidence level is less than 0.9, it is marked as a low-confidence region. Then, combined with a customized OCR model for financial scenarios, text information within the key regions is extracted. The customized OCR model has optimized the logic for recognizing handwritten cursive characters and fuzzy handwriting. The customized OCR model includes collecting 20,000 text samples from key financial regions, including but not limited to handwritten amounts, fuzzy invoice codes, and seal text, annotating the text content and character positions, and generating sample pairs. ,in, Image patches representing key regions. To correspond to real text, a dedicated dataset for financial OCR was constructed. The customized OCR model adds an attention mechanism layer after the CNN feature extraction module to enhance the feature response of handwritten cursive characters and blurred characters. The attention weight calculation expression is as follows:

[0065]

[0066] In the formula, The feature map output by the CNN. This is the attention weight matrix. For the query vector dimension; for numbers and special characters (including but not limited to ¥ and -) in financial text, a character-level language model constraint is introduced during CTC decoding, with the decoding probability being:

[0067]

[0068] In the formula, For the CTC decoding probability of the CRNN model, For text length, For 3-gram language model probabilities, As a weighting factor;

[0069] It is important to note that the extracted text information is validated for consistency and relevance using a semantic-visual cross-validation rule base to remove invalid data; and the cleaned data is mapped to structured fields that conform to accounting standards to generate a standardized financial data matrix.

[0070] The consistency check includes comparing the core entities extracted by the entity recognition algorithm with the text information extracted by the OCR model. If the deviation is greater than 5%, a manual review reminder is automatically triggered.

[0071] In addition, the correlation verification includes: by collecting historical financial data of enterprises, constructing a correlation rule model, marking isolated data of unrelated business documents as invalid data and removing them.

[0072] Furthermore, the operational process of the prediction and strategy linkage module includes:

[0073] The cash flow health prediction includes constructing an LSTM model with a time-series attention mechanism; the input normalization features of the LSTM model... This includes core entities, textual information, ERP data, and external supply chain credit data; temporal features are extracted using a two-layer LSTM, with the first LSTM layer calculating the hidden state and introducing a temporal attention mechanism to calculate attention weights; finally, the future output is passed through a fully connected layer. Daily cash flow forecast :

[0074]

[0075] In the formula, To predict the weight matrix, For bias; a weighted average absolute error is used to assign higher weights to the cash flow forecasts for days 1-7; the Adam optimizer is used with a learning rate of The training process is repeated 500 times, with validation every 20 times. The process is stopped early if the loss on the validation set does not decrease for 10 consecutive times. The LSTM model outputs a cash flow prediction curve for the next 7-30 days and a three-level health index of "safety-early warning-risk".

[0076] Furthermore, based on the aforementioned three-level health index,

[0077] Security level: ;

[0078] Warning level: ;

[0079] Risk level: The three-level health index is converted into a quantitative score. :

[0080]

[0081] In the formula, , For security level, It is at the warning level. The risk level is set at 60%. A supply chain payment priority strategy is dynamically generated using an integer programming algorithm. The actual data after the supply chain payment priority strategy is executed is fed back to the LSTM model, and the weights of the LSTM model are updated using a gradient descent algorithm. The model is optimized once per quarter.

[0082] It is important to note that the three-tiered health index includes a safe level, a warning level, and a risk level. The supply chain payment priority strategy is based on the three-tiered health index. For the safe level, priority is given to paying suppliers with payment terms of ≤30 days and a credit rating of A or above, triggering applications for early payment discounts. The credit rating is based on external credit data. For the warning level, with the constraint that core suppliers are not overdue, priority is given to paying suppliers of essential raw materials for production, followed by suppliers of non-core auxiliary materials. For the risk level, an emergency mechanism is triggered, automatically extracting information on assets that can be quickly converted into cash, and generating payment extension plans for finance personnel to negotiate with suppliers.

[0083] Beneficial effects

[0084] Compared with known public technologies, the technical solution provided by this invention has the following beneficial effects:

[0085] In use, this invention reduces the workload of manual processing of unstructured data through semantic-visual fusion cleaning, improves data accuracy, and solves the pain points of difficult data entry and numerous errors for SMEs, thus facilitating improved data processing efficiency. Through real-time cash flow health prediction, SMEs can identify funding gap risks 7-30 days in advance, avoiding passive supply disruptions and enhancing cash flow forecasting capabilities. Dynamic payment strategies reduce the probability of overdue payments, improving the stability of cooperation with core suppliers and enabling early payment discounts, reducing procurement costs and optimizing supply chain decision-making accuracy. By achieving a closed loop of unstructured data standardization, cash flow prediction, and supply chain decision-making for the first time, it differs from existing technical solutions that only focus on data integration or single decision-making, exhibiting non-obviousness. Attached Figure Description

[0086] Figure 1 This is a flowchart illustrating a big data-based financial management method according to the present invention.

[0087] Figure 2 This is a system diagram of a big data-based financial management system according to the present invention. Detailed Implementation

[0088] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0089] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0090] The present invention will now be described in further detail with reference to the accompanying drawings:

[0091] Example 1:

[0092] like Figure 1 As shown, this invention provides a big data-based financial management method, including:

[0093] S1. Perform semantic-visual fusion cleaning and standardization on unstructured financial data of SMEs, and output a standardized financial data matrix;

[0094] Furthermore, in step S1, the method for performing semantic-visual fusion cleaning and standardization on the unstructured financial data of SMEs and outputting a standardized financial data matrix is ​​as follows:

[0095] The method employs a BERT model pre-trained in the financial domain. It inputs unstructured text data from unstructured financial data of SMEs and then uses entity recognition algorithms to extract five core entities: the person requesting reimbursement, the reason for reimbursement, the amount, the time of occurrence, and the associated business document number. The pre-trained BERT model in the financial domain includes collecting 500,000 textual corpora from the financial domain of SMEs, including but not limited to reimbursement dialogues, financial notices, and reconciliation emails, to construct a corpus. in, For each text The word segmentation process uses Jieba word segmentation combined with an expanded financial terminology database to generate word sequences. Pre-train the BERT base model in the financial domain, optimizing the objective function as bidirectional language model loss + financial entity masking loss, expressed as:

[0096]

[0097] In the formula, For traditional masked language model loss, For the number of mask words, For non-masked word sequences, To predict probabilities for traditional masked language models, This is the financial entity masking loss, used to perform 100% masking of the five core entities already labeled in the corpus. For a collection of entities, To obtain the text sequence after entity removal, the pre-training process was performed for 1500 iterations, using a linear decay strategy for the learning rate. ,final When the perplexity of the validation set is ≤18, pre-training is stopped, and a pre-trained BERT model for the financial domain is obtained. .

[0098] Furthermore, in step S1, the method for performing semantic-visual fusion cleaning and standardization on the unstructured financial data of SMEs and outputting a standardized financial data matrix is ​​as follows:

[0099] Using the BERT model pre-trained in the financial domain Based on this, a conditional random field layer is introduced to extract five types of core entities, using unstructured text data as input. The last hidden layer outputs the feature vector of each word. ,in, The position of a word in the text; Input to a linear layer to obtain 11 emission score matrices for each word. ,in, Given the text length, a conditional random field layer calculates the globally optimal path for the label sequence, aiming to maximize the joint probability of the label sequence:

[0100]

[0101] In the formula, the optimal label sequence is solved using the Viterbi algorithm. Extract the corresponding 5 types of core entities ,in, The person seeking reimbursement It is a reason for reimbursement. It's the amount. It is the time of occurrence. It is associated with the business document number, and the entity identification confidence score is calculated:

[0102]

[0103] In the formula, For entities Word length, The set of locations of entities in the text. For entities The corresponding tag set, For the Sigmoid function, in When the time is right, mark it as a low-confidence entity.

[0104] Furthermore, in step S1, the method for performing semantic-visual fusion cleaning and standardization on the unstructured financial data of SMEs and outputting a standardized financial data matrix is ​​as follows:

[0105] Furthermore, an improved YOLOv8 object detection algorithm is employed to locate key regions in image-based unstructured financial data of SMEs. These key regions include, but are not limited to, amount boxes, invoice code areas, and seal areas. The improved YOLOv8 object detection algorithm involves collecting 50,000 financial images of SMEs, including but not limited to images taken from different angles, with varying lighting intensities, and at different resolutions. LabelImg is used to annotate three types of key regions, generating bounding boxes. ,in, For the amount box, For the invoice code area, For the seal area, construct a financial image dataset. Based on the K-means++ algorithm Clustering of annotation boxes generates anchor box sizes adapted to the financial area:

[0106] Amount frame anchor frame ,

[0107] Invoice code area anchor frame ,

[0108] Seal area anchor frame Furthermore, a category weighting factor is introduced based on the original CIoU loss to highlight the positioning of key financial areas. The expression is:

[0109]

[0110] In the formula, The number of detection layers, The number of anchor frames per layer. This is the actual annotation box. To improve the predicted bounding boxes of the YOLOv8 model, The length of the diagonal of the smallest rectangle enclosing the two frames. The probability that a real target exists. To improve the target confidence predicted by the YOLOv8 model, Category weighting factors are set based on the importance of financial regions: Amount Frame Invoice code area Seal area ; Weights are lost based on the target confidence level; the AdamW optimizer is used with weight decay. Training is performed for 1000 epochs. Training stops when the Average Precision on the validation set is ≥ 95%. Input the image to be processed. The model outputs predicted bounding boxes for three types of key regions. ,in and location reliability ,exist When the confidence level is less than 0.9, it is marked as a low-confidence region. Then, combined with a customized OCR model for financial scenarios, text information within the key regions is extracted. The customized OCR model has optimized the logic for recognizing handwritten cursive characters and fuzzy handwriting. The customized OCR model includes collecting 20,000 text samples from key financial regions, including but not limited to handwritten amounts, fuzzy invoice codes, and seal text, annotating the text content and character positions, and generating sample pairs. ,in, Image patches representing key regions. To correspond to real text, a dedicated dataset for financial OCR was constructed. The customized OCR model adds an attention mechanism layer after the CNN feature extraction module to enhance the feature response of handwritten cursive characters and blurred characters. The attention weight calculation expression is as follows:

[0111]

[0112] In the formula, The feature map output by the CNN. This is the attention weight matrix. For the query vector dimension; for numbers and special characters (including but not limited to ¥ and -) in financial text, a character-level language model constraint is introduced during CTC decoding, with the decoding probability being:

[0113]

[0114] In the formula, For the CTC decoding probability of the CRNN model, For text length, For 3-gram language model probabilities, As a weighting factor;

[0115] It is important to note that the extracted text information is validated for consistency and relevance using a semantic-visual cross-validation rule base to remove invalid data; and the cleaned data is mapped to structured fields that conform to accounting standards to generate a standardized financial data matrix.

[0116] The consistency check includes comparing the core entities extracted by the entity recognition algorithm with the text information extracted by the OCR model. If the deviation is greater than 5%, a manual review reminder is automatically triggered.

[0117] In addition, the correlation verification includes: by collecting historical financial data of enterprises, constructing a correlation rule model, marking isolated data of unrelated business documents as invalid data and removing them.

[0118] In this embodiment, data processing efficiency is significantly improved, the time spent manually organizing unstructured financial data is reduced, the workload of staff is decreased, and the labor intensity of staff is reduced; the deviation rate of tax declaration is reduced, the accuracy of data is improved, and tax risks caused by data errors and omissions are avoided; the standardized data matrix can be directly connected to the enterprise's ERP system to support core financial work such as cash flow forecasting and cost accounting, laying the foundation for subsequent financial management, helping enterprises improve the efficiency of completing monthly financial closing, and improving the speed of decision response.

[0119] S2. Based on the standardized financial data matrix, combined with ERP data and external supply chain credit data, perform cash flow health prediction and supply chain payment strategy linkage.

[0120] Furthermore, in step S2, the method for linking cash flow health forecasting with supply chain payment strategies based on the standardized financial data matrix, combined with ERP data and external supply chain credit data, is as follows:

[0121] The cash flow health prediction includes constructing an LSTM model with a time-series attention mechanism; the input normalization features of the LSTM model... This includes core entities, textual information, ERP data, and external supply chain credit data; temporal features are extracted using a two-layer LSTM, with the first LSTM layer calculating the hidden state and introducing a temporal attention mechanism to calculate attention weights; finally, the future output is passed through a fully connected layer. Daily cash flow forecast :

[0122]

[0123] In the formula, To predict the weight matrix, For bias; a weighted average absolute error is used to assign higher weights to the cash flow forecasts for days 1-7; the Adam optimizer is used with a learning rate of The training process is repeated 500 times, with validation every 20 times. The process is stopped early if the loss on the validation set does not decrease for 10 consecutive times. The LSTM model outputs a cash flow prediction curve for the next 7-30 days and a three-level health index of "safety-early warning-risk".

[0124] Furthermore, in step S2, the method for linking cash flow health forecasting with supply chain payment strategies based on the standardized financial data matrix, combined with ERP data and external supply chain credit data, is as follows:

[0125] Furthermore, based on the aforementioned three-level health index,

[0126] Security level: ;

[0127] Warning level: ;

[0128] Risk level: The three-level health index is converted into a quantitative score. :

[0129]

[0130] In the formula, , For security level, It is at the warning level. The risk level is set at 60%. A supply chain payment priority strategy is dynamically generated using an integer programming algorithm. The actual data after the supply chain payment priority strategy is executed is fed back to the LSTM model, and the weights of the LSTM model are updated using a gradient descent algorithm. The model is optimized once per quarter.

[0131] Furthermore, in step S2, the method for linking cash flow health forecasting with supply chain payment strategies based on the standardized financial data matrix, combined with ERP data and external supply chain credit data, is as follows:

[0132] It is important to note that the three-tiered health index includes a safe level, a warning level, and a risk level. The supply chain payment priority strategy is based on the three-tiered health index. For the safe level, priority is given to paying suppliers with payment terms of ≤30 days and a credit rating of A or above, triggering applications for early payment discounts. The credit rating is based on external credit data. For the warning level, with the constraint that core suppliers are not overdue, priority is given to paying suppliers of essential raw materials for production, followed by suppliers of non-core auxiliary materials. For the risk level, an emergency mechanism is triggered, automatically extracting information on assets that can be quickly converted into cash, and generating payment extension plans for finance personnel to negotiate with suppliers.

[0133] In this embodiment, the accuracy of identifying funding gaps within 7 to 30 days is significantly improved, while supply chain payment efficiency and revenue are optimized, the overdue payment rate of core suppliers is reduced, and annual revenue can be increased through discounts obtained through advance payments. When assessing the risk level, compliant payment extension plans can be generated quickly, improving the negotiation success rate with suppliers and avoiding losses due to production line shutdowns caused by supply disruptions. Overall, the intelligence and risk resistance capabilities of enterprise financial management are improved.

[0134] Example 2:

[0135] like Figure 2 As shown, Embodiment 2 provides a big data-based financial management system, including:

[0136] The data standardization processing module performs semantic-visual fusion cleaning and standardization processing on unstructured financial data of SMEs, and outputs a standardized financial data matrix.

[0137] The forecasting and strategy linkage module, based on the standardized financial data matrix and combined with ERP data and external supply chain credit data, performs cash flow health forecasting and linkage with supply chain payment strategies.

[0138] Furthermore, the operation process of the data standardization processing module includes:

[0139] The method employs a BERT model pre-trained in the financial domain. It inputs unstructured text data from unstructured financial data of SMEs and then uses entity recognition algorithms to extract five core entities: the person requesting reimbursement, the reason for reimbursement, the amount, the time of occurrence, and the associated business document number. The pre-trained BERT model in the financial domain includes collecting 500,000 textual corpora from the financial domain of SMEs, including but not limited to reimbursement dialogues, financial notices, and reconciliation emails, to construct a corpus. in, For each text The word segmentation process uses Jieba word segmentation combined with an expanded financial terminology database to generate word sequences. Pre-train the BERT base model in the financial domain, optimizing the objective function as bidirectional language model loss + financial entity masking loss, expressed as:

[0140]

[0141] In the formula, For traditional masked language model loss, For the number of mask words, For non-masked word sequences, To predict probabilities for traditional masked language models, This is the financial entity masking loss, used to perform 100% masking of the five core entities already labeled in the corpus. For a collection of entities, To obtain the text sequence after entity removal, the pre-training process was performed for 1500 iterations, using a linear decay strategy for the learning rate. ,final When the perplexity of the validation set is ≤18, pre-training is stopped, and a pre-trained BERT model for the financial domain is obtained. .

[0142] Using the BERT model pre-trained in the financial domain Based on this, a conditional random field layer is introduced to extract five types of core entities, using unstructured text data as input. The last hidden layer outputs the feature vector of each word. ,in, The position of a word in the text; Input to a linear layer to obtain 11 emission score matrices for each word. ,in, Given the text length, a conditional random field layer calculates the globally optimal path for the label sequence, aiming to maximize the joint probability of the label sequence:

[0143]

[0144] In the formula, the optimal label sequence is solved using the Viterbi algorithm. Extract the corresponding 5 types of core entities ,in, The person seeking reimbursement It is a reason for reimbursement. It's the amount. It is the time of occurrence. It is associated with the business document number, and the entity identification confidence score is calculated:

[0145]

[0146] In the formula, For entities Word length, The set of locations of entities in the text. For entities The corresponding tag set, For the Sigmoid function, in When the time is right, mark it as a low-confidence entity.

[0147] Furthermore, an improved YOLOv8 object detection algorithm is employed to locate key regions in image-based unstructured financial data of SMEs. These key regions include, but are not limited to, amount boxes, invoice code areas, and seal areas. The improved YOLOv8 object detection algorithm involves collecting 50,000 financial images of SMEs, including but not limited to images taken from different angles, with varying lighting intensities, and at different resolutions. LabelImg is used to annotate three types of key regions, generating bounding boxes. ,in, For the amount box, For the invoice code area, For the seal area, construct a financial image dataset. Based on the K-means++ algorithm Clustering of annotation boxes generates anchor box sizes adapted to the financial area:

[0148] Amount frame anchor frame ,

[0149] Invoice code area anchor frame ,

[0150] Seal area anchor frame Furthermore, a category weighting factor is introduced based on the original CIoU loss to highlight the positioning of key financial areas. The expression is:

[0151]

[0152] In the formula, The number of detection layers, The number of anchor frames per layer. This is the actual annotation box. To improve the predicted bounding boxes of the YOLOv8 model, The length of the diagonal of the smallest rectangle enclosing the two frames. The probability that a real target exists. To improve the target confidence predicted by the YOLOv8 model, Category weighting factors are set based on the importance of financial regions: Amount Frame Invoice code area Seal area ; Weights are lost based on the target confidence level; the AdamW optimizer is used with weight decay. Training is performed for 1000 epochs. Training stops when the Average Precision on the validation set is ≥ 95%. Input the image to be processed. The model outputs predicted bounding boxes for three types of key regions. ,in and location reliability ,exist When the confidence level is less than 0.9, it is marked as a low-confidence region. Then, combined with a customized OCR model for financial scenarios, text information within the key regions is extracted. The customized OCR model has optimized the logic for recognizing handwritten cursive characters and fuzzy handwriting. The customized OCR model includes collecting 20,000 text samples from key financial regions, including but not limited to handwritten amounts, fuzzy invoice codes, and seal text, annotating the text content and character positions, and generating sample pairs. ,in, Image patches representing key regions. To correspond to real text, a dedicated dataset for financial OCR was constructed. The customized OCR model adds an attention mechanism layer after the CNN feature extraction module to enhance the feature response of handwritten cursive characters and blurred characters. The attention weight calculation expression is as follows:

[0153]

[0154] In the formula, The feature map output by the CNN. This is the attention weight matrix. For the query vector dimension; for numbers and special characters (including but not limited to ¥ and -) in financial text, a character-level language model constraint is introduced during CTC decoding, with the decoding probability being:

[0155]

[0156] In the formula, For the CTC decoding probability of the CRNN model, For text length, For 3-gram language model probabilities, As a weighting factor;

[0157] It is important to note that the extracted text information is validated for consistency and relevance using a semantic-visual cross-validation rule base to remove invalid data; and the cleaned data is mapped to structured fields that conform to accounting standards to generate a standardized financial data matrix.

[0158] The consistency check includes comparing the core entities extracted by the entity recognition algorithm with the text information extracted by the OCR model. If the deviation is greater than 5%, a manual review reminder is automatically triggered.

[0159] In addition, the correlation verification includes: by collecting historical financial data of enterprises, constructing a correlation rule model, marking isolated data of unrelated business documents as invalid data and removing them.

[0160] Furthermore, the operational process of the prediction and strategy linkage module includes:

[0161] The cash flow health prediction includes constructing an LSTM model with a time-series attention mechanism; the input normalization features of the LSTM model... This includes core entities, textual information, ERP data, and external supply chain credit data; temporal features are extracted using a two-layer LSTM, with the first LSTM layer calculating the hidden state and introducing a temporal attention mechanism to calculate attention weights; finally, the future output is passed through a fully connected layer. Daily cash flow forecast :

[0162]

[0163] In the formula, To predict the weight matrix, For bias; a weighted average absolute error is used to assign higher weights to the cash flow forecasts for days 1-7; the Adam optimizer is used with a learning rate of The training process is repeated 500 times, with validation every 20 times. The process is stopped early if the loss on the validation set does not decrease for 10 consecutive times. The LSTM model outputs a cash flow prediction curve for the next 7-30 days and a three-level health index of "safety-early warning-risk".

[0164] Furthermore, based on the aforementioned three-level health index,

[0165] Security level: ;

[0166] Warning level: ;

[0167] Risk level: The three-level health index is converted into a quantitative score. :

[0168]

[0169] In the formula, , For security level, It is at the warning level. The risk level is set at 60%. A supply chain payment priority strategy is dynamically generated using an integer programming algorithm. The actual data after the supply chain payment priority strategy is executed is fed back to the LSTM model, and the weights of the LSTM model are updated using a gradient descent algorithm. The model is optimized once per quarter.

[0170] It is important to note that the three-tiered health index includes a safe level, a warning level, and a risk level. The supply chain payment priority strategy is based on the three-tiered health index. For the safe level, priority is given to paying suppliers with payment terms of ≤30 days and a credit rating of A or above, triggering applications for early payment discounts. The credit rating is based on external credit data. For the warning level, with the constraint that core suppliers are not overdue, priority is given to paying suppliers of essential raw materials for production, followed by suppliers of non-core auxiliary materials. For the risk level, an emergency mechanism is triggered, automatically extracting information on assets that can be quickly converted into cash, and generating payment extension plans for finance personnel to negotiate with suppliers.

[0171] In this embodiment, a small to medium-sized electronic component enterprise (50 employees, annual revenue of 20 million yuan) is used as an example:

[0172] (a) Implementation preparation

[0173] We obtain WeChat expense reimbursement chat records (average 500 messages per month), photos of handwritten expense reimbursement forms (average 300 photos per month), and bank statement screenshots (average 200 photos per month) through enterprise financial software interfaces. We also connect to the ERP system to obtain procurement / sales data and access third-party credit reporting platforms to obtain supplier credit rating data.

[0174] The BERT parsing model, the improved YOLOv8 model, and the temporal attention LSTM model of this invention are deployed to an enterprise private cloud server to achieve data interoperability with the enterprise financial software.

[0175] (II) Step S1 Implementation: Unstructured Data Processing

[0176] Multimodal analysis: For the text “WeChat reimbursement - Employee Kan San - Travel expenses - 2024.10.16 - High-speed rail ticket 380 yuan”, the BERT model extracts the entity “Reimbursement person: Kan San, Reason: Travel expenses, Amount: 380 yuan, Time: 2024.10.16”; for the photograph of the high-speed rail ticket, the YOLOv8 is improved to locate the amount area, and the OCR extracts “Total price including tax: 380 yuan”.

[0177] Cross-validation: The text amount matches the OCR amount (0% deviation) and is associated with "transportation invoice", thus it is determined to be valid data;

[0178] Standardized mapping: Mapped to "Management expenses - Travel expenses - Transportation expenses - Kan San - October 2024", generating structured data and importing it into the ERP system.

[0179] (III) Implementation of Step S2: Cash Flow Forecasting and Payment Decision

[0180] Cash flow forecast: On November 1, 2024, the LSTM model was input with standardized data from October (receipts of RMB 800,000 and expenses of RMB 300,000), ERP receivables of RMB 500,000 and payables of RMB 400,000, and the industry's Q4 collection cycle of 25 days. The output was the cash flow forecast curve for November: There was a shortfall of RMB 100,000 on November 15, and the health index was "warning level".

[0181] Payment strategy generation: The integer programming algorithm determines that "core suppliers (chip supplier A, with outstanding payment of 200,000 yuan) should be paid first, and non-core suppliers (packaging supplier B, with outstanding payment of 200,000 yuan) should apply for an extension of the payment period to December", and automatically generates a payment period application template and pushes it to the packaging supplier;

[0182] Feedback optimization: On November 20, packaging supplier B agreed to extend the payment period, and the actual payment received was 850,000 yuan (50,000 yuan more than predicted). This data was fed back to the model to update the LSTM weights, reducing the prediction error for December to 3%.

[0183] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A financial management method based on big data, characterized in that, include: S1. Perform semantic-visual fusion cleaning and standardization on unstructured financial data of SMEs, and output a standardized financial data matrix; S2. Based on the standardized financial data matrix, combined with ERP data and external supply chain credit data, perform cash flow health prediction and link supply chain payment strategies.

2. The financial management method based on big data according to claim 1, characterized in that, In step S1, the method for performing semantic-visual fusion cleaning and standardization on the unstructured financial data of SMEs and outputting a standardized financial data matrix is ​​as follows: The method employs a BERT model pre-trained in the financial domain. It inputs unstructured text data from unstructured financial data of SMEs and then uses entity recognition algorithms to extract five core entities: the person requesting reimbursement, the reason for reimbursement, the amount, the time of occurrence, and the associated business document number. The pre-trained BERT model in the financial domain includes collecting 500,000 textual corpora from the financial domain of SMEs, including but not limited to reimbursement dialogues, financial notices, and reconciliation emails, to construct a corpus. in, For each text The word segmentation process uses Jieba word segmentation combined with an expanded financial terminology database to generate word sequences. Pre-train the BERT base model in the financial domain, optimizing the objective function as bidirectional language model loss + financial entity masking loss, expressed as: In the formula, For traditional masked language model loss, For the number of mask words, For non-masked word sequences, To predict probabilities for traditional masked language models, This is the financial entity masking loss, used to perform 100% masking of the five core entities already labeled in the corpus. For a collection of entities, To obtain the text sequence after entity removal, the pre-training process was performed for 1500 iterations, using a linear decay strategy for the learning rate. ,final When the perplexity of the validation set is ≤18, pre-training is stopped, and a pre-trained BERT model for the financial domain is obtained. .

3. The financial management method based on big data according to claim 2, characterized in that, In step S1, the method for performing semantic-visual fusion cleaning and standardization on the unstructured financial data of SMEs and outputting a standardized financial data matrix is ​​as follows: Using the BERT model pre-trained in the financial domain Based on this, a conditional random field layer is introduced to extract five types of core entities, using unstructured text data as input. The last hidden layer outputs the feature vector of each word. ,in, The position of a word in the text; Input to a linear layer to obtain 11 emission score matrices for each word. ,in, Given the text length, a conditional random field layer calculates the globally optimal path for the label sequence, aiming to maximize the joint probability of the label sequence: In the formula, the optimal label sequence is solved using the Viterbi algorithm. Extract the corresponding 5 types of core entities ,in, The person seeking reimbursement It is a reason for reimbursement. It's the amount. It is the time of occurrence. It is associated with the business document number, and the entity identification confidence score is calculated: In the formula, For entities Word length, The set of locations of entities in the text. For entities The corresponding tag set, For the Sigmoid function, in When the time is right, mark it as a low-confidence entity.

4. The financial management method based on big data according to claim 3, characterized in that, In step S1, the method for performing semantic-visual fusion cleaning and standardization on the unstructured financial data of SMEs and outputting a standardized financial data matrix is ​​as follows: Furthermore, an improved YOLOv8 object detection algorithm is employed to locate key regions in image-based unstructured financial data of SMEs. These key regions include, but are not limited to, amount boxes, invoice code areas, and seal areas. The improved YOLOv8 object detection algorithm involves collecting 50,000 financial images of SMEs, including but not limited to images taken from different angles, with varying lighting intensities, and at different resolutions. LabelImg is used to annotate three types of key regions, generating bounding boxes. ,in, For the amount box, For the invoice code area, For the seal area, construct a financial image dataset. Based on the K-means++ algorithm Clustering of annotation boxes generates anchor box sizes adapted to the financial area: Amount frame anchor frame , Invoice code area anchor frame , Seal area anchor frame Furthermore, a category weighting factor is introduced based on the original CIoU loss to highlight the positioning of key financial areas. The expression is: In the formula, The number of detection layers, The number of anchor frames per layer. This is the actual annotation box. To improve the predicted bounding boxes of the YOLOv8 model, The length of the diagonal of the smallest rectangle enclosing the two frames. The probability that a real target exists. To improve the target confidence predicted by the YOLOv8 model, Category weighting factors are set based on the importance of financial regions: Amount Frame Invoice code area Seal area ; Weights are lost based on the target confidence level; the AdamW optimizer is used with weight decay. Training is performed for 1000 epochs. Training stops when the Average Precision on the validation set is ≥ 95%. Input the image to be processed. The model outputs predicted bounding boxes for three types of key regions. ,in and location reliability ,exist When the confidence level is less than 0.9, it is marked as a low-confidence region. Then, combined with a customized OCR model for financial scenarios, text information within the key regions is extracted. The customized OCR model has optimized the logic for recognizing handwritten cursive characters and fuzzy handwriting. The customized OCR model includes collecting 20,000 text samples from key financial regions, including but not limited to handwritten amounts, fuzzy invoice codes, and seal text, annotating the text content and character positions, and generating sample pairs. ,in, Image patches representing key regions. To correspond to real text, a dedicated dataset for financial OCR was constructed. The customized OCR model adds an attention mechanism layer after the CNN feature extraction module to enhance the feature response of handwritten cursive characters and blurred characters. The attention weight calculation expression is as follows: In the formula, The feature map output by the CNN. This is the attention weight matrix. For the query vector dimension; for numbers and special characters (including but not limited to ¥ and -) in financial text, a character-level language model constraint is introduced during CTC decoding, with the decoding probability being: In the formula, For the CTC decoding probability of the CRNN model, For text length, For 3-gram language model probabilities, As a weighting factor; It is important to note that the extracted text information is validated for consistency and relevance using a semantic-visual cross-validation rule base to remove invalid data; and the cleaned data is mapped to structured fields that conform to accounting standards to generate a standardized financial data matrix. The consistency check includes comparing the core entities extracted by the entity recognition algorithm with the text information extracted by the OCR model. If the deviation is greater than 5%, a manual review reminder is automatically triggered. In addition, the correlation verification includes: by collecting historical financial data of enterprises, constructing a correlation rule model, marking isolated data of unrelated business documents as invalid data and removing them.

5. A financial management method based on big data according to claim 4, characterized in that, In step S2, the method for linking cash flow health forecasting with supply chain payment strategies based on the standardized financial data matrix, combined with ERP data and external supply chain credit data, is as follows: The cash flow health prediction includes constructing an LSTM model with a time-series attention mechanism; the input normalization features of the LSTM model... This includes core entities, textual information, ERP data, and external supply chain credit data; temporal features are extracted using a two-layer LSTM, with the first LSTM layer calculating the hidden state and introducing a temporal attention mechanism to calculate attention weights; finally, the future output is passed through a fully connected layer. Daily cash flow forecast : In the formula, To predict the weight matrix, For bias; a weighted average absolute error is used to assign higher weights to the cash flow forecasts for days 1-7; the Adam optimizer is used with a learning rate of The training process is repeated 500 times, with validation every 20 times. The process is stopped early if the loss on the validation set does not decrease for 10 consecutive times. The LSTM model outputs a cash flow prediction curve for the next 7-30 days and a three-level health index of "safety-early warning-risk".

6. The financial management method based on big data according to claim 5, characterized in that, In step S2, the method for linking cash flow health forecasting with supply chain payment strategies based on the standardized financial data matrix, combined with ERP data and external supply chain credit data, is as follows: Furthermore, based on the aforementioned three-level health index, Security level: ; Warning level: ; Risk level: ; The three-level health index is converted into a quantitative score. : In the formula, , For security level, It is at the warning level. Risk level; supply chain payment priority strategy is dynamically generated using integer programming algorithm; The actual data after the execution of the supply chain payment priority strategy is fed back to the LSTM model, and the weights of the LSTM model are updated by the gradient descent algorithm, and optimized once per quarter.

7. A financial management method based on big data according to claim 6, characterized in that, In step S2, the method for linking cash flow health forecasting with supply chain payment strategies based on the standardized financial data matrix, combined with ERP data and external supply chain credit data, is as follows: It should be noted that the three-level health index includes a safe level, an early warning level, and a risk level; the supply chain payment priority strategy is based on the three-level health index. The safe level prioritizes payments to suppliers with a payment period of ≤30 days and a credit rating of A or above, triggering applications for early payment discounts. The credit rating is based on external credit data. Warning level: With core suppliers not overdue, priority is given to paying suppliers of essential raw materials for production, followed by suppliers of non-core auxiliary materials. Risk level: Triggers an emergency mechanism, automatically extracts information on assets that can be quickly converted into cash, and generates payment extension plans for finance personnel to negotiate with suppliers.

8. A big data-based financial management system, based on any one of claims 1-7, characterized in that, include: The data standardization processing module performs semantic-visual fusion cleaning and standardization processing on unstructured financial data of SMEs, and outputs a standardized financial data matrix. The forecasting and strategy linkage module, based on the standardized financial data matrix and combined with ERP data and external supply chain credit data, performs cash flow health forecasting and linkage with supply chain payment strategies.

9. A big data-based financial management system according to claim 8, characterized in that, Operation process of data standardization processing module include: The method employs a BERT model pre-trained in the financial domain. It inputs unstructured text data from unstructured financial data of SMEs and then uses entity recognition algorithms to extract five core entities: the person requesting reimbursement, the reason for reimbursement, the amount, the time of occurrence, and the associated business document number. The pre-trained BERT model in the financial domain includes collecting 500,000 textual corpora from the financial domain of SMEs, including but not limited to reimbursement dialogues, financial notices, and reconciliation emails, to construct a corpus. in, For each text The word segmentation process uses Jieba word segmentation combined with an expanded financial terminology database to generate word sequences. Pre-train the BERT base model in the financial domain, optimizing the objective function as bidirectional language model loss + financial entity masking loss, expressed as: In the formula, For traditional masked language model loss, For the number of mask words, For non-masked word sequences, To predict probabilities for traditional masked language models, This is the financial entity masking loss, used to perform 100% masking of the five core entities already labeled in the corpus. For a collection of entities, To obtain the text sequence after entity removal, the pre-training process was performed for 1500 iterations, using a linear decay strategy for the learning rate. ,final When the perplexity of the validation set is ≤18, pre-training is stopped, and a pre-trained BERT model for the financial domain is obtained. . Using the BERT model pre-trained in the financial domain Based on this, a conditional random field layer is introduced to extract five types of core entities, using unstructured text data as input. The last hidden layer outputs the feature vector of each word. ,in, The position of a word in the text; Input to a linear layer to obtain 11 emission score matrices for each word. ,in, Given the text length, a conditional random field layer calculates the globally optimal path for the label sequence, aiming to maximize the joint probability of the label sequence: In the formula, the optimal label sequence is solved using the Viterbi algorithm. Extract the corresponding 5 types of core entities ,in, The person seeking reimbursement It is a reason for reimbursement. It's the amount. It is the time of occurrence. It is associated with the business document number, and the entity identification confidence score is calculated: In the formula, For entities Word length, The set of locations of entities in the text. For entities The corresponding tag set, For the Sigmoid function, in When the time is right, mark it as a low-confidence entity. Furthermore, an improved YOLOv8 object detection algorithm is employed to locate key regions in image-based unstructured financial data of SMEs. These key regions include, but are not limited to, amount boxes, invoice code areas, and seal areas. The improved YOLOv8 object detection algorithm involves collecting 50,000 financial images of SMEs, including but not limited to images taken from different angles, with varying lighting intensities, and at different resolutions. LabelImg is used to annotate three types of key regions, generating bounding boxes. ,in, For the amount box, For the invoice code area, For the seal area, construct a financial image dataset. Based on the K-means++ algorithm Clustering of annotation boxes generates anchor box sizes adapted to the financial area: Amount frame anchor frame , Invoice code area anchor frame , Seal area anchor frame Furthermore, a category weighting factor is introduced based on the original CIoU loss to highlight the positioning of key financial areas. The expression is: In the formula, The number of detection layers, The number of anchor frames per layer. This is the actual annotation box. To improve the predicted bounding boxes of the YOLOv8 model, The length of the diagonal of the smallest rectangle enclosing the two frames. The probability that a real target exists. To improve the target confidence predicted by the YOLOv8 model, Category weighting factors are set based on the importance of financial regions: Amount Frame Invoice code area Seal area ; Weights are lost based on the target confidence level; the AdamW optimizer is used with weight decay. Training is performed for 1000 epochs. Training stops when the Average Precision on the validation set is ≥ 95%. Input the image to be processed. The model outputs predicted bounding boxes for three types of key regions. ,in and location reliability ,exist When the confidence level is less than 0.9, it is marked as a low-confidence region. Then, combined with a customized OCR model for financial scenarios, text information within the key regions is extracted. The customized OCR model has optimized the logic for recognizing handwritten cursive characters and fuzzy handwriting. The customized OCR model includes collecting 20,000 text samples from key financial regions, including but not limited to handwritten amounts, fuzzy invoice codes, and seal text, annotating the text content and character positions, and generating sample pairs. ,in, Image patches representing key regions. To correspond to real text, a dedicated dataset for financial OCR was constructed. The customized OCR model adds an attention mechanism layer after the CNN feature extraction module to enhance the feature response of handwritten cursive characters and blurred characters. The attention weight calculation expression is as follows: In the formula, The feature map output by the CNN. This is the attention weight matrix. For the query vector dimension; for numbers and special characters (including but not limited to ¥ and -) in financial text, a character-level language model constraint is introduced during CTC decoding, with the decoding probability being: In the formula, For the CTC decoding probability of the CRNN model, For text length, For 3-gram language model probabilities, As a weighting factor; It is important to note that the extracted text information is validated for consistency and relevance using a semantic-visual cross-validation rule base to remove invalid data; and the cleaned data is mapped to structured fields that conform to accounting standards to generate a standardized financial data matrix. The consistency check includes comparing the core entities extracted by the entity recognition algorithm with the text information extracted by the OCR model. If the deviation is greater than 5%, a manual review reminder is automatically triggered. In addition, the correlation verification includes: by collecting historical financial data of enterprises, constructing a correlation rule model, marking isolated data of unrelated business documents as invalid data and removing them.

10. A big data-based financial management system according to claim 9, characterized in that, Operational flow of the prediction and strategy linkage module include: The cash flow health prediction includes constructing an LSTM model with a time-series attention mechanism; the input normalization features of the LSTM model... This includes core entities, textual information, ERP data, and external supply chain credit data; temporal features are extracted using a two-layer LSTM, with the first LSTM layer calculating the hidden state and introducing a temporal attention mechanism to calculate attention weights; finally, the future output is passed through a fully connected layer. Daily cash flow forecast : In the formula, To predict the weight matrix, For bias; a weighted average absolute error is used to assign higher weights to the cash flow forecasts for days 1-7; the Adam optimizer is used with a learning rate of The training process is repeated 500 times, with validation every 20 times. The process is stopped early if the loss on the validation set does not decrease for 10 consecutive times. The LSTM model outputs a cash flow prediction curve for the next 7-30 days and a three-level health index of "safety-early warning-risk". Furthermore, based on the aforementioned three-level health index, Security level: ; Warning level: ; Risk level: The three-level health index is converted into a quantitative score. : In the formula, , For security level, It is at the warning level. The risk level is set at 60%. A supply chain payment priority strategy is dynamically generated using an integer programming algorithm. The actual data after the supply chain payment priority strategy is executed is fed back to the LSTM model, and the weights of the LSTM model are updated using a gradient descent algorithm. The model is optimized once per quarter. It is important to note that the three-tiered health index includes a safe level, a warning level, and a risk level. The supply chain payment priority strategy is based on the three-tiered health index. For the safe level, priority is given to paying suppliers with payment terms of ≤30 days and a credit rating of A or above, triggering applications for early payment discounts. The credit rating is based on external credit data. For the warning level, with the constraint that core suppliers are not overdue, priority is given to paying suppliers of essential raw materials for production, followed by suppliers of non-core auxiliary materials. For the risk level, an emergency mechanism is triggered, automatically extracting information on assets that can be quickly converted into cash, and generating payment extension plans for finance personnel to negotiate with suppliers.

Images