Production and operation cost allocation system based on large model

Abstract

The invention relates to a production and operation cost allocation system based on a large model, and the system comprises a rule algorithm self-maintenance module which is used for monitoring event changes, and carrying out the dynamic maintenance marking rule and dynamic marking of generated production and operation income and expenditure data through employing a decision tree and a first star large model; and the real-time revenue and expenditure collection module is used for acquiring real-time revenue and expenditure data, processing the real-time revenue and expenditure data by adopting a distance calculation method and a text processing large model according to a dynamic marking result and performing mutual verification so as to judge the belonging product type of the real-time revenue and expenditure data and realize cost allocation of the real-time revenue and expenditure data. Compared with the prior art, the system has the advantages of realizing intelligent allocation of production and operation costs and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of cost sharing, and in particular to a production and operation cost sharing system based on a large model. Background Art

[0002] At present, it has become a necessary requirement for all enterprises to establish a financial report analysis system that integrates business and finance, and to establish a multi-dimensional indicator system that runs through the entire business management chain of the enterprise and covers all products, markets, and projects of the enterprise. Existing enterprises have issued a document on the evaluation and analysis system for business and finance integration. The document puts forward clear requirements for establishing a management closed loop with product lines as the chain, penetrating revenue, cost, investment to cash flow, and incorporates this work into the performance improvement assessment plan. The document requires professional companies to classify and implement policies based on the relevance of their business to the main business of the communications industry, and to pilot the evaluation and analysis system for products within the unit.

[0003] Matching the company's operating costs, investments, and R&D inputs with its products to form a closed-loop management system is crucial for the company's sustainable development. However, in actual operations, whether it is business-finance integration, cost allocation to products, or motivation association, there are difficulties that need to be solved one by one:

[0004] 1. When using traditional revenue and expenditure statistics methods to support the development and operation of multiple products at the same time, it is often difficult to accurately allocate the expenditure costs between products. This may lead to uneven resource allocation, where some products may bear unreasonable costs, while other products may receive undue resource investment.

[0005] 2. Currently, most of the operating costs of enterprises are indirectly allocated to products, and there is no real input-output relationship between resources. This is especially true for large state-owned enterprises, which have complex historical situations, and the financial model structure and business structure have not changed for a long time. The business and financial departments often use different data and indicators to measure performance and costs, making it difficult to calculate real performance. Therefore, it is urgent to find the optimal evolution path to adapt to business development.

[0006] 3. In closed-loop management, the links of income reporting, expenditure application, and expense reimbursement need to be linked, but the traditional method must rely on manual operations to link expenditures and income. Not only is it time-consuming and labor-intensive, it also increases additional investment, making it difficult to maintain and operate in the long term, and it is impossible to manage motivation, and the goal of business and financial integration cannot be achieved. Summary of the invention

[0007] The purpose of the present invention is to provide a production and operation cost allocation system based on a large model for realizing intelligent allocation of production and operation costs.

[0008] The purpose of the present invention can be achieved by the following technical solutions:

[0009] A production and operation cost allocation system based on a large model, comprising:

[0010] Rule algorithm self-maintenance module: used to monitor event changes. For the revenue and expenditure data of production and operation that have been generated, decision trees and the First Star big model are used to dynamically maintain marking rules and dynamically mark.

[0011] Real-time income and expenditure collection module: used to obtain real-time income and expenditure data, and according to the dynamic labeling results, respectively use the distance calculation method and the text processing model to process and verify each other, so as to determine the product category to which the real-time income and expenditure data belongs, and realize the cost sharing of the real-time income and expenditure data.

[0012] Furthermore, the construction process of the decision tree is represented as a generation process of labeling rules, and the construction process of the decision tree includes:

[0013] 1) Normalize the features of the generated income and expenditure data;

[0014] 2) Based on the feature normalization results and in combination with existing business rules, different weight factors are assigned to each feature data;

[0015] 3) Take the feature normalization result as the entire data set and calculate the entropy of the entire data set;

[0016] 4) Based on the weight factor and entropy, the information gain of each feature data is calculated, and the feature data with the largest information gain is selected as the segmentation feature of the current node;

[0017] 5) Based on the remaining data sets, repeat steps 4)-5) until a leaf node is constructed to obtain a constructed decision tree.

[0018] Furthermore, the normalized expression is:

[0019]

[0020] In the formula, x is the original data, min(x) is the smallest value in the data set, which is used to adjust the lower bound of the data to 0, max(x) is the largest value in the data set, which is used to adjust the upper bound of the data to 1, and x′ is the normalized data.

[0021] Furthermore, the calculation expression of the information gain is:

[0022]

[0023] In the formula, p1 represents the number of samples in the first category of the data set, p2 represents the number of samples in the second category of the data set, and p1+p2 represents the total number of samples in the data set. Indicates the proportion of the first category in the entire data set, represents the proportion of the second category in the entire data set, Entropy(D) represents the entropy of the entire data set D, the degree of uncertainty in the data set, Gain(D,A) represents the information gain of feature A in data set D, |D v ∣ means that in the data set D, the value of feature A is a subset D of v v The number of samples in the dataset D, |D| represents the total number of samples in the dataset D, Entropy(D v ): indicates that when the value of feature A is v, the corresponding subset D v The entropy of Values(A) represents the set of feature A values, w A Represents the weight factor of feature A, which is used to adjust the priority of the feature.

[0024] Furthermore, the training steps of the first large star model include:

[0025] Obtain historical production and operation cost-related data, including rules and regulations data and historical financial data;

[0026] The first star big model is retrieved, and the historical production and operation cost related data is used for tuning to realize model localization and obtain the trained first star big model.

[0027] Furthermore, the step of using the distance calculation method for processing includes:

[0028] Based on the real-time income and expenditure data, the unstructured data therein is identified by using the OCR mode to obtain the key structured data therein, and the mean vector of the historical income and expenditure data of each product is calculated by cosine distance together with the structured data in the real-time income and expenditure data. The product category to which the real-time income and expenditure data belongs is determined according to the cosine distance calculation result, wherein the cosine distance calculation expression is:

[0029]

[0030] Where A represents the vector corresponding to the real-time income and expenditure data, B represents the mean vector, and both vectors A and B have n elements, which are represented by A=[A1,A2,...,A i ,...A n ] and B=[B1,B2,...,B i ,...B n ], cos(θ) represents the cosine of the angle θ between vectors A and B;

[0031] Based on the real-time income and expenditure data, the Euclidean distance between the real-time income and expenditure data and the cluster center of all products in different time periods is calculated, and the product category to which the real-time income and expenditure data belongs is determined according to the Euclidean distance calculation result, wherein the cluster center is obtained by performing k-means clustering processing on the historical income and expenditure data of each product, and the calculation expression of the Euclidean distance is:

[0032]

[0033] Where d(x,y) represents the Euclidean distance between vector x and vector y, x represents the vector corresponding to the real-time income and expenditure data, y represents the vector corresponding to the cluster center, and both vectors x and y contain n elements, which are respectively represented as x=(x1,x2...,x i ,...x n ) and y=(y1,y2,...,y i ,...y n );

[0034] Determine whether the judgment result corresponding to the cosine distance calculation result and the judgment result corresponding to the Euclidean distance calculation result are the same. If they are not the same, the real-time income and expenditure data is manually judged to belong to the product category to achieve the allocation of the real-time income and expenditure data. If they are the same, the judgment result corresponding to the cosine distance calculation result or the judgment result corresponding to the Euclidean distance calculation result is used as the final result to achieve the allocation of the real-time income and expenditure data.

[0035] Furthermore, the steps of using the text processing model for processing include:

[0036] Preprocess the real-time income and expenditure data to generate feature vectors, and use a pre-trained text processing model to perform predictions, and output the product categories and motivation results of the real-time income and expenditure data;

[0037] The training steps of the text processing large model include:

[0038] Based on the historical business and financial revenue and expenditure data of each product, the OCR mode is used to identify the unstructured data, obtain the key structured data, and vectorize it together with the structured data to form a feature vector. The historical revenue and expenditure data contains two labels, namely the motivation label and the product label.

[0039] Building a large text processing model and setting training parameters, wherein the large text processing model includes a shared encoder, a motivation identification decoder, and a product classification decoder;

[0040] The feature vector is segmented, the label is encoded, and the two are jointly output to the encoder for processing. Then, a motivation identification decoder and a product classification decoder are used to identify motivation labels and product labels respectively. Through continuous training, a well-trained text processing large model is obtained.

[0041] Furthermore, it also includes a trend prediction and auxiliary decision-making module, and the execution steps of the trend prediction and auxiliary decision-making module include:

[0042] Based on the real-time revenue and expenditure data allocation results and the historical revenue and expenditure data of each product, preprocessing is performed to obtain input data in the form of time series;

[0043] Inputting the input data in the form of time series into a pre-trained time series model for prediction, obtaining and outputting a first trend analysis result for assisting decision making;

[0044] The input data in the form of time series is input into a pre-trained second star model for prediction, and a second trend analysis result is obtained and output for auxiliary decision-making, wherein the second star model is constructed using an LSTM model based on a multi-head attention mechanism, and the mean square error is used as the loss function. A Dropout layer is provided in the multi-head attention mechanism, which is used to randomly discard some neurons in each iterative training.

[0045] Furthermore, the pre-processing step comprises:

[0046] Formatting the real-time income and expenditure data allocation results and the historical income and expenditure data of each product into a time series form to obtain input data in an initial time series form;

[0047] The input data in the initial time series form is processed for outliers and missing values, and is normalized to obtain the final input data in the time series form.

[0048] Furthermore, in the rule algorithm self-maintenance module, when the marking results of the decision tree and the first star large model are inconsistent, manual marking is adopted; in the real-time income and expenditure aggregation module, when the processing results of the distance calculation method and the text processing large model are inconsistent, manual judgment is adopted.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] (1) The present invention uses a rule algorithm self-maintenance module to perform rule self-maintenance and dynamic labeling through a decision tree and the first star big model. On this basis, the real-time income and expenditure collection module classifies and distinguishes real-time income and expenditure data by adopting a distance calculation method and a text processing big model, and allocates each income and expenditure data to the corresponding product, forming a relationship between expenses and products, and realizing intelligent allocation of production and operation expenses.

[0051] (2) The rule algorithm self-maintenance module and the real-time income and expenditure collection module of the present invention both use two methods to verify each other for expense identification, thereby improving the accuracy of expense allocation. The distance calculation method and the text processing large model in the real-time income and expenditure collection module respectively allocate expenses from the perspective of similarity calculation and text processing, thereby further improving the accuracy of expense allocation.

[0052] (3) The present invention sets up trend prediction and decision-making support modules to perform periodic analysis and emergency event detection, adopts time series model and second star big model to make predictions, conducts trend predictions on products and business conditions, optimizes resource allocation, and assists decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a schematic diagram of the system structure of the present invention;

[0054] Figure 2 A schematic diagram of the system structure of an embodiment of the present invention;

[0055] Figure 3 A schematic diagram of marking according to an embodiment of the present invention;

[0056] Figure 4 A backward cost table of an embodiment of the present invention;

[0057] Figure 5 This is a cost allocation report according to an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0059] This embodiment provides a production and operation cost allocation system based on a large model, such as Figure 1 As shown, the system includes a rule algorithm self-maintenance module 1, a real-time income and expenditure collection module 2, and a trend prediction and auxiliary decision-making module 3. Figure 1 , the method comprises the following steps:

[0060] 1. Rule Algorithm Self-Maintenance Module 1

[0061] This module is used to classify and label basic data, build decision tree algorithms, and train large models. Using this module, you can obtain various cost allocation models based on the company's specific needs and business model.

[0062] This module uses two methods to perform pre-classification and labeling: 1. Construct a decision tree algorithm and use the traditional method of manual classification and labeling. 2. Use the First Star model for pre-training, dynamic labeling and allocation. If the final labeling results of the two methods are different, manual intervention will be used for judgment.

[0063] Let's first introduce the first traditional method. The initialization of such rules requires "labeling" assets and personnel based on current finances and policies, and then configuring various rules based on the "labeling" results. The so-called "labeling" is to "label" such "sources of cost generation" as fixed assets, intangible assets, employees, cloud resources, communication lines, etc., so that the subsequent costs of the labeled "sources of cost generation" can be allocated to a certain product in proportion. For example, after manual labeling, wages will need to be paid every month, and fixed assets will be depreciated every month.

[0064] We first use the decision tree algorithm to analyze historical data and existing rules, and then label the provided data to improve the accuracy and efficiency of labeling.

[0065] The specific algorithm steps are as follows:

[0066] First, data preprocessing and basic parameter preparation are performed. That is, because most features in financial data have different magnitudes, such as income and cost, car fare reimbursement, software development expenses, cloud resource costs, AI computing power fees, etc., the relevant features need to be normalized and the feature values ​​​​scaled to the range of 0 to 1 so that they are unified to the same magnitude when calculating information gain.

[0067] Normalization formula:

[0068]

[0069] In the formula, x is the original data, min(x) is the smallest value in the data set, which is used to adjust the lower bound of the data to 0, max(x) is the largest value in the data set, which is used to adjust the upper bound of the data to 1, and x′ is the normalized data.

[0070] At the same time, in the process of building the decision tree, the segmentation of certain features is manually adjusted in combination with business rules and the knowledge of financial experts. And based on the opinions of experts, weight factors w are added to certain features. A , so as to reflect the importance of this feature when calculating the information gain.

[0071] Then use the following formula to calculate the entropy of the entire data set.

[0072]

[0073] Then according to the previously determined weight factor w A , use the following formula to calculate the information gain of each feature, and then select the feature with the largest information gain as the segmentation feature of the current node.

[0074]

[0075] In the formula, p1 represents the number of samples in the first category of the data set, p2 represents the number of samples in the second category of the data set, and p1+p2 represents the total number of samples in the data set. Indicates the proportion of the first category in the entire data set, represents the proportion of the second category in the entire data set, Entropy(D) represents the entropy of the entire data set D, the degree of uncertainty in the data set, Gain(D,A) represents the information gain of feature A in data set D, |D v ∣ means that in the data set D, the value of feature A is a subset D of v v The number of samples in the dataset D, |D| represents the total number of samples in the dataset D, Entropy(D v ): indicates that when the value of feature A is v, the corresponding subset D v The entropy of Values(A) represents the set of feature A values, w A Represents the weight factor of feature A, which is used to adjust the priority of the feature.

[0076] Repeat the above steps for the segmented sub-datasets to construct sub-nodes. If the sub-dataset is pure (i.e., all samples belong to the same category), mark the node as a leaf node and end the recursion. After the decision tree is built, the current basic data (i.e., personnel, fixed assets) can be "labeled" using the decision tree, and the labeling results can be temporarily stored for subsequent applications.

[0077] The second method is to use the First Star big model for auxiliary judgment of "marking". First, the big model is localized, and then the unit is optimized in a targeted manner. The input data is divided into rules and regulations data and historical financial data.

[0078] First, let's explain the rules and regulations data, that is, the accounting standards, the People's Republic of China Invoice Management Measures, the People's Republic of China Enterprise Income Tax Implementation Regulations and other currently implemented regulations are sorted according to the requirements of the big model. Such regulations are basically implemented stably and rarely adjusted, and can be used as training data. The regulations within the telecommunications system and the relevant management regulations and financial requirements of the unit are relatively more variable, so they can be used as external resources of the big model and are also sorted according to the requirements of the big model so that the big model can make logical judgments based on these latest specifications when needed.

[0079] The historical financial data includes the historical data in the relevant financial system of the unit and related financial statements, business data, invoice data, etc. After vectorization, this type of data is also used as external resource data for the large model.

[0080] After completing the relevant adjustments, you can "label" the assets and personnel.

[0081] After obtaining the marking results by the above two methods, "dynamic marking" is also required for each subsequent event change. That is, when events such as asset transfer and personnel transfer to other positions or departments occur, the marking rules are dynamically maintained through decision trees and large models. Or if personnel participate in the research and development of a certain product, or because a certain product purchases a certain "atomic capability" interface, marking must be re-run. If the marking results are eventually changed, it can be considered that the relevant expenses incurred by such assets and personnel are re-attributed to the new product to keep the results "dynamic" and accurate.

[0082] In the process of using large models to implement such rules and "labeling", the results can be compared with those produced by the first traditional method. If the comparison results are inconsistent, they will be manually identified and ultimately all such "sources of cost generation" will be "labeled" so that the various costs they generate can be directly attributed to the cost of the product.

[0083] 2. Real-time income and expenditure collection module 2

[0084] This module is responsible for collecting business and financial data from various departments and systems of the company. This data may include product sales data, various cost and revenue data, single-time data of outsourced labor and advertising expenses, and travel and conference expense data. This module needs to collect data while ensuring the integrity and accuracy of the data, and use algorithms to identify the relationship between expenses and products, so as to achieve the purpose of directly allocating various expenditures.

[0085] In this example, two methods are also used to identify expenses. The traditional method is divided into three steps. The first step is to organize historical financial data, that is, according to the classified business and financial data of each product (usually mainly income and expenditure data, which will be referred to as income and expenditure data below), through the associated data of structured income and expenditure data in the historical financial database, the specific associated data include: reimbursement type, associated contract, associated customers and suppliers, transaction association type, transaction details, whether to collect payment, invoice information (tax rate, invoicing item, amount, invoice type) cost and profit center, economic item code, budget indicator, budget responsibility center and other structured data. After vectorizing these data, calculate the mean vector of all income and expenditure data under this product (that is, assuming there are only three vectors, add the values ​​of the three vectors in each dimension, and then divide by the number of vectors to get the mean vector. Assume that the three vectors are A, B, and C: M = (A+B+C) / 3), so as to obtain the mean vector of all products.

[0086] The second step is to use cosine similarity combined with Euclidean distance to determine which product the current income and expenditure data belongs to when the enterprise has income and expenditure and generates income and expenditure data. The specific method is to use the relevant associated data of the income and expenditure data submitted this time, specifically based on the values ​​of these structured data of this data (Note, according to the income and expenditure type and income and expenditure situation, only some data actually have values, accounting type, associated contracts, associated customers and suppliers, transaction association type, transaction details, whether to collect payment, invoice information (tax rate, invoicing items, amount, invoice type) cost and profit center, economic matter code, budget indicator, budget responsibility center, etc.), as well as the unstructured data of related attachment images such as accounting forms, settlement forms, invoices, travel and transportation (tickets, itinerary sheets). After gathering all the data of this income and expenditure, first identify the unstructured data through the OCR mode to obtain the key structured data, and then calculate the cosine distance of the mean vector of the income and expenditure data of each product in the historical accounting together with the current structured data. The calculation formula is:

[0087]

[0088] Where A represents the vector corresponding to the real-time income and expenditure data, B represents the mean vector, and both vectors A and B have n elements, which are represented by A=[A1,A2,...,A i ,...A n ] and B=[B1,B2,...,B i ,...B n ], cos(θ) represents the cosine of the angle θ between vectors A and B.

[0089] The ones with cosine values ​​closer to 1 are the most similar, and we can preliminarily conclude that the income and expenditure data belong to the most similar products.

[0090] The third step is to use Euclidean distance to further identify the judgment results, that is, determine the K value from the dimensions of time, amount segment, and income and expenditure type, and then use kmeans clustering to classify the historical product income and expenditure data to obtain the cluster centers of all products in different time periods, and then calculate the corresponding Euclidean distance based on the values ​​of this income and expenditure data and all cluster centers. The Euclidean distance formula is:

[0091]

[0092] Where d(x,y) represents the Euclidean distance between vector x and vector y, x represents the vector corresponding to the real-time income and expenditure data, y represents the vector corresponding to the cluster center, and both vectors x and y contain n elements, which are respectively represented as x=(x1,x2...,x i ,...x n ) and y=(y1,y2,...,y i ,...y n ).

[0093] If the calculation result has a large deviation, that is, it is not close to the Euclidean distance of the product corresponding time period calculated in the second step, but is closer to the cluster center of other product time periods, and exceeds the predetermined threshold, then this data needs to be manually judged as belonging to the product.

[0094] While using traditional methods for identification, we also use large models for identification. The identification using large models is also divided into three steps. The first step is to organize historical financial data, extract information from unstructured data through OCR, and then vectorize it with structured information. All income and expenditure data form feature vectors. All data are labeled. There are two forms of labels: one is associated with known motivations, which is partly for training large models to identify motivations, and the other is product labels, which are for the purpose of grouping products. The second step is to select a relevant model that is relatively suitable for processing text data, such as BERT, and use the model's tokenizer to encode the feature vector as the input of the model, and then use a shared encoder (such as a Transformer encoder) to encode the input text and convert the text into a context-related representation. Then design an independent decoder for each task: the motivation identification decoder is responsible for identifying motivation labels. The product classification decoder is responsible for identifying product classification labels. The labels (products to which the income and expenditure data belongs) are encoded and converted into a form that the model can process. Then define the training parameters, such as learning rate, batch size, number of training rounds, etc., and then use the training data to train the model. By adjusting the parameters, optimize the model performance, and finally ensure the model convergence and complete the training. The third step is to pre-process the new income and expenditure data, generate feature vectors, and then use the trained large model to predict, output the product category and motivation results to which the data belongs, and judge according to the confidence of the model output. If it exceeds the threshold, it will be directly classified as a product, otherwise it will be marked as data that requires manual judgment.

[0095] When the above two classification products are consistent, it can be considered that the classification products of this income and expenditure data are correct, and no manual intervention is required. If the judgment is inconsistent, manual judgment will be intervened when the income and expenditure data is recorded.

[0096] 3. Trend Forecasting and Decision Support Module 3

[0097] This module is based on the product allocation data generated by the first two modules, and timely identifies and predicts from the product perspective, and conducts trend analysis, periodic analysis, and emergency detection. That is, first format these data together with historical data into a time series form. During processing, it is necessary to ensure that the timestamps and values ​​of the data are accurate, and that the data from different data sources are correctly aligned. It is also necessary to pay attention to seasonal factors and the time characteristics of financial data. Then deal with missing values ​​and outliers, and perform normalization. Finally, it is also necessary to pay attention to the company's financial characteristics, and introduce parameters such as cash flow, profit margin, and debt-to-asset ratio in the company's financial report data.

[0098] After the data is prepared, two methods are also used for prediction. The first traditional method is to build a time series model for each product data, select appropriate parameters (p, d, q parameters of the ARIMA model), and then use the product's historical data for model training and fitting of the time series model. Then use the model to perform trend analysis, periodicity analysis, and emergency event detection. It is necessary to take into account that financial data often has significant seasonal and periodic characteristics, such as financial settlement at the end of the quarter, annual budget, etc. The long-term trend elimination method can be used to eliminate seasonal factors, and then abnormal trends and abnormal fluctuations that do not conform to the periodicity, as well as sudden cost surges or sharp drops in sales, etc., are found respectively. These events may cause product cost overruns, waste of resources, and reduced product profitability. In short, possible financial risks are identified, and the identified risks are evaluated to determine their severity and possible impact, and then alarms are triggered according to the set risk alarm thresholds.

[0099] The second method is to use the Second Star model, which relies on the LSTM and multi-head attention mechanisms to handle complex time series forecasting tasks, and then predict financial data to improve the accuracy of predictions by capturing long-term and short-term dependencies and processing multiple input features.

[0100] First of all, data preprocessing is required, that is, the collected data must be cleaned, missing values ​​must be processed, outliers must be identified and processed, and then normalized, and date features (year, quarter, month, week, day, weekday / weekend) must be generated. Feature engineering is then performed, such as time features, financial features, and other features. Time features require the inclusion of special holidays, long vacations, and financial settlement and account opening and closing dates. Financial features include year-on-year change rate, month-on-month change rate, and moving average. Other features are mainly external factors, such as financial policies, regulations, and macroeconomic indicators, industry data, etc.

[0101] After completing the above steps, fine-tune the second star model, such as adjusting the number of attention heads, adding financial domain knowledge, etc., and then you can perform training and evaluation, select the mean square error (MSE) as the loss function, and use Dropout to prevent overfitting. Specifically, add a Dropout layer to the multi-head attention mechanism and specify the dropout rate. In this way, in each training iteration, the Dropout layer will randomly drop some neurons, and the dropped neurons will not participate in the forward and backward propagation of the training. In the inference (prediction) stage, use the complete neural network (no neuron dropout), and multiply the output of all neurons by (1-p) to scale, thereby effectively preventing the model from overfitting.

[0102] After adjustment and training, the model can be deployed and an API can be designed for access, so that subsequent predictions can be effectively made through the financial data generated in real time.

[0103] After completing the predictions of the two methods, the prediction values ​​of both methods are finally presented to the user. The prediction results may include cost overruns, waste of resources, decreased profitability, etc., allowing users to make decisions based on the prediction results.

[0104] According to the above system, according to Figure 2 As shown, the steps in the actual application scenario are:

[0105] Step S1: Configure detailed account matching cost and expense rules and use algorithms to complete various "labeling" tasks.

[0106] First, by configuring the account code, profit center, cost center, supplier, and functional scope, determine the matching relationship between the detailed account data and the cost expenses, match these data with the cost expenses, and set the matching rules for rental property, intangible asset amortization, fixed asset depreciation and other expenses, organize various data and attributes (examples of item attributes: application date, purchase date, price, type, item name, usage record, approval record), as the input data of the decision tree algorithm and the First Star Big Model, as described in the previous article, build a decision tree model and provide it to the First Star Big Model as external data. In actual applications, according to the configuration of the account and different profit center combinations, the subsequent detailed account data on SAP can be associated with the three types of costs: business management fees, capability service unit management fees, or comprehensive support unit management fees.

[0107] The rent allocation rules for the big data AI center department are configured according to the corresponding product-related income and the indirect allocation ratio. The actual cost of the corresponding rent can be calculated based on the personnel area, and then the program can automatically allocate it.

[0108] like Figure 3 As shown, you can select an asset detail on the left, click the [Mark] button on the left, and select a product for marking. Marking means calling two methods to perform calculations and determine the product to which this asset belongs. After marking, this asset is the asset used by the corresponding product. If the two methods calculate and determine differently, you can set the allocation manually.

[0109] The above is the initial setting. After running correctly, the "labeling" is completed according to the asset application, purchase record, and other asset attributes in combination with two algorithms. For the newly added data, "labeling" is performed every time it enters.

[0110] Since the decision tree model has good interpretability, coupled with the assistance of the First Star large model, and can be calculated at any time, it can replace manual labeling and save manual operation and maintenance.

[0111] Step S2: Collect data from the forward and backward systems.

[0112] The backward system uses the SAP system, that is, the detailed account is imported through the data interface of the SAP system to ensure the integrity and accuracy of the data. The imported data needs to be verified to ensure that the data format is consistent with the preset rules to facilitate subsequent matching and allocation operations. The forward system is a variety of reimbursement systems or business domain systems (when a customer generates an order through a business touchpoint, the income is calculated immediately, and the amount obtained by the corresponding merchant participating in the account sharing is calculated as the corresponding expenditure).

[0113] Step S3: Match and generate apportioned cost and expense data.

[0114] After importing the data, the system will automatically calculate the detailed account data and cost expenses using two algorithms based on the data collected in the previous step S2, that is, obtain the final allocated cost and expense data, and associate the products. In the event of inconsistency between the two algorithms, it will be marked in red during allocation and wait for manual selection.

[0115] Step S4: After generating the cost and expense data to be allocated, the large model system will identify and mark a variety of possible driver data for manual selection and processing.

[0116] Since financial data is monopolized within the enterprise, data collection is not a problem, but it is often difficult to collect motivations. Therefore, it is necessary to collect relevant motivations while collecting data, or to train historical data through large models, and finally determine and save the motivations by obtaining various keyword information and relevant quantity information related to income and expenditure.

[0117] Step S5: Configure cost sharing rules.

[0118] According to the labeling results of the driver data, configure the allocation rules for different costs and expenses. Configure the department, allocation type, detailed rules, product, allocation method, allocation ratio and other rules to allocate the costs and expenses to each product or department.

[0119] Cloud resources (including AI resources) are indirectly allocated to all Haobai products, that is, the proportion can be set manually, or it can be automatically allocated according to revenue or profit through a program.

[0120] For one-time related expenses, such as advertising fees, marketing incentives and certain R&D project fees, a time series analysis method (ARIMA model) algorithm is used to evaluate the impact on the product and then allocate the expenses.

[0121] Step S6: Initiate a cost sharing form.

[0122] The user initiates a cost allocation form, selects the allocation date, and the system automatically extracts the corresponding SAP detailed account data. Apply the configured allocation rules to generate allocation data.

[0123] like Figure 4 The Backward Cost form shown.

[0124] Step S7: Configure product report retrieval rules.

[0125] Configure the mapping relationship between group products and No. 100 products in the allocated data in the product report, and configure the mapping relationship between group key costs and expenses and No. 100 costs and expenses.

[0126] Step S8: Generate a product-by-product cost allocation report.

[0127] Based on the configured data retrieval rules, the system generates a product-by-product cost allocation report.

[0128] The final cost allocation report is as follows Figure 5 shown.

[0129] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0130] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A production and operation cost allocation system based on a large model, characterized in that: include: Rule algorithm self-maintenance module (1): used to monitor event changes. For the revenue and expenditure data of production and operation that have been generated, decision trees and the first star big model are used to dynamically maintain the marking rules and dynamically mark them; Real-time income and expenditure collection module (2): used to obtain real-time income and expenditure data, and based on the dynamic labeling results, respectively use the distance calculation method and the text processing model to process and verify each other, so as to determine the product category to which the real-time income and expenditure data belongs, and realize the cost sharing of the real-time income and expenditure data.

2. A large-model-based production and operation cost sharing system according to claim 1, characterized in that: The construction process of the decision tree is represented as a generation process of labeling rules, and the construction process of the decision tree includes: 1) Normalize the features of the generated income and expenditure data; 2) Based on the feature normalization results and in combination with existing business rules, different weight factors are assigned to each feature data; 3) Take the feature normalization result as the entire data set and calculate the entropy of the entire data set; 4) Based on the weight factor and entropy, the information gain of each feature data is calculated, and the feature data with the largest information gain is selected as the segmentation feature of the current node; 5) Based on the remaining data sets, repeat steps 4)-5) until a leaf node is constructed to obtain a constructed decision tree.

3. A large-model-based production and operation cost sharing system according to claim 2, characterized in that: The normalized expression is: In the formula, x is the original data, min(x) is the smallest value in the data set, which is used to adjust the lower bound of the data to 0, max(x) is the largest value in the data set, which is used to adjust the upper bound of the data to 1, and x′ is the normalized data.

4. A large-model-based production and operation cost sharing system according to claim 2, characterized in that: The calculation expression of the information gain is: In the formula, p1 represents the number of samples in the first category of the data set, p2 represents the number of samples in the second category of the data set, and p1+p2 represents the total number of samples in the data set. Indicates the proportion of the first category in the entire data set, represents the proportion of the second category in the entire data set, Entropy(D) represents the entropy of the entire data set D, the degree of uncertainty in the data set, Gain(D,A) represents the information gain of feature A in data set D, |D v ∣ means that in the data set D, the value of feature A is a subset D of v v The number of samples in the dataset D, |D| represents the total number of samples in the dataset D, Entropy(D v ): indicates that when the value of feature A is v, the corresponding subset D v The entropy of Values(A) represents the set of feature A values, w A Represents the weight factor of feature A, which is used to adjust the priority of the feature.

5. The large-model-based production and operation cost sharing system according to claim 1 is characterized in that: The training steps of the first star model include: Obtain historical production and operation cost-related data, including rules and regulations data and historical financial data; The first star big model is retrieved, and the historical production and operation cost related data is used for tuning to realize model localization and obtain the trained first star big model.

6. A large-model-based production and operation cost allocation system according to claim 1, characterized in that: The step of using the distance calculation method for processing includes: Based on the real-time income and expenditure data, the unstructured data therein is identified by using the OCR mode to obtain the key structured data therein, and the mean vector of the historical income and expenditure data of each product is calculated by cosine distance together with the structured data in the real-time income and expenditure data. The product category to which the real-time income and expenditure data belongs is determined according to the cosine distance calculation result, wherein the cosine distance calculation expression is: Where A represents the vector corresponding to the real-time income and expenditure data, B represents the mean vector, and both vectors A and B have n elements, which are represented by A=[A1,A2,...,A i ,...A n ] and B=[B1,B2,...,B i ,...B n ], cos(θ) represents the cosine of the angle θ between vectors A and B; Based on the real-time income and expenditure data, the Euclidean distance between the real-time income and expenditure data and the cluster center of all products in different time periods is calculated, and the product category to which the real-time income and expenditure data belongs is determined according to the Euclidean distance calculation result, wherein the cluster center is obtained by performing k-means clustering processing on the historical income and expenditure data of each product, and the calculation expression of the Euclidean distance is: Where d(x,y) represents the Euclidean distance between vector x and vector y, x represents the vector corresponding to the real-time income and expenditure data, y represents the vector corresponding to the cluster center, and both vectors x and y contain n elements, which are respectively represented as x=(x1,x2...,x i ,...x n ) and y=(y1,y2,...,y i ,...y n ); Determine whether the judgment result corresponding to the cosine distance calculation result and the judgment result corresponding to the Euclidean distance calculation result are the same. If they are not the same, the real-time income and expenditure data is manually judged to belong to the product category to achieve the allocation of the real-time income and expenditure data. If they are the same, the judgment result corresponding to the cosine distance calculation result or the judgment result corresponding to the Euclidean distance calculation result is used as the final result to achieve the allocation of the real-time income and expenditure data.

7. The large-model-based production and operation cost sharing system according to claim 1 is characterized in that: The steps of using the large text processing model include: Preprocess the real-time income and expenditure data to generate feature vectors, and use a pre-trained text processing model to perform predictions, and output the product categories and motivation results of the real-time income and expenditure data; The training steps of the text processing large model include: Based on the historical business and financial revenue and expenditure data of each product, the OCR mode is used to identify the unstructured data, obtain the key structured data, and vectorize it together with the structured data to form a feature vector. The historical revenue and expenditure data contains two labels, namely the motivation label and the product label. Building a large text processing model and setting training parameters, wherein the large text processing model includes a shared encoder, a motivation identification decoder, and a product classification decoder; The feature vector is segmented, the label is encoded, and the two are jointly output to the encoder for processing. Then, a motivation identification decoder and a product classification decoder are used to identify motivation labels and product labels respectively. Through continuous training, a well-trained text processing large model is obtained.

8. The large-model-based production and operation cost allocation system according to claim 1 is characterized in that: It also includes a trend prediction and decision-making assistance module (3), wherein the execution steps of the trend prediction and decision-making assistance module (3) include: Based on the real-time revenue and expenditure data allocation results and the historical revenue and expenditure data of each product, preprocessing is performed to obtain input data in the form of time series; Inputting the input data in the form of time series into a pre-trained time series model for prediction, obtaining and outputting a first trend analysis result for assisting decision making; The input data in the form of time series is input into a pre-trained second star model for prediction, and a second trend analysis result is obtained and output for auxiliary decision-making, wherein the second star model is constructed using an LSTM model based on a multi-head attention mechanism, and the mean square error is used as the loss function. A Dropout layer is provided in the multi-head attention mechanism, which is used to randomly discard some neurons in each iterative training.

9. A large-model-based production and operation cost sharing system according to claim 8, characterized in that: The pre-processing step comprises: Formatting the real-time income and expenditure data allocation results and the historical income and expenditure data of each product into a time series form to obtain input data in an initial time series form; The input data in the initial time series form is processed for outliers and missing values, and is normalized to obtain the final input data in the time series form.

10. The large model-based production and operation cost allocation system according to claim 1, characterized in that: In the rule algorithm self-maintenance module, when the marking results of the decision tree and the first star large model are inconsistent, manual marking is adopted; in the real-time income and expenditure collection module, when the processing results of the distance calculation method and the text processing large model are inconsistent, manual judgment is adopted.

Images