Digital office method, device and equipment based on RPA and medium

Abstract

The invention relates to a digital office method, device and equipment based on RPA and a medium, and belongs to the technical field of artificial intelligence. The method comprises the steps that a starting condition is preset, a controller controls an RPA to operate according to the starting condition, the RPA conducts data collection by logging in a micronucleus terminal to obtain a cross-border token matrix, character recognition is conducted through a text recognition algorithm according to the cross-border token matrix to obtain extracted content data, and a filling template is preset; the RPA fills the extracted content data according to the filling template to obtain real-time filling data; acquiring historical filling data, calculating through a comparison model according to the historical filling data and the real-time filling data to obtain a comparison result, acquiring a verification API, transmitting the comparison result to the data verification terminal by the RPA according to the verification API, and acquiring a confirmation signal; the RPA performs a classification operation according to the acknowledgement signal. The automation of digital office is realized, and the office efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of artificial intelligence, and particularly relates to a digital office method and device based on RPA, equipment and medium. BACKGROUND

[0002] In the field of digital office, especially in the field of import and export customs and related business processes, traditional technical means face many challenges.

[0003] In the data collection link, in the past, manual data collection from various platforms and systems, such as the customs single window platform, trade document processing system and mailbox, is not only inefficient, but also prone to omissions and errors. At the same time, the traditional automatic data collection method has poor flexibility, and it is difficult to achieve accurate and timely data acquisition according to diversified business scenarios and complex start conditions.

[0004] For text recognition, traditional OCR technology has obvious defects. When processing cross-border document matrix related files such as customs declaration and invoice, due to the diversity of file formats and the complexity of content, there are a large number of irregular fonts, blurred images and special symbols, etc., and the recognition accuracy of traditional OCR technology is low. This leads to subsequent data processing based on incorrect or incomplete recognition results, which seriously affects the accuracy and reliability of the business process.

[0005] In terms of data comparison, the data comparison between the customs declaration and the accompanying single has always been a difficulty. The traditional method relies on manual comparison, which consumes a lot of manpower and time, and it is difficult to guarantee the comprehensiveness and accuracy of the comparison. Even if some simple automatic comparison tools are used, they cannot effectively handle complex business logic and semantic association, and cannot accurately identify the subtle differences and potential problems between data.

[0006] In addition, for price reasonableness judgment, the traditional method lacks effective prediction models. Often only according to experience or simple price fluctuation range to judge whether the real-time price is reasonable, cannot fully utilize historical price data for in-depth analysis and accurate prediction, and is difficult to adapt to the dynamic changes of market price and complex economic environment. SUMMARY

[0007] To solve the above problems in the prior art, the present application provides a digital office method, device, equipment and medium based on RPA, The purpose of the present application can be achieved by the following technical solutions: A preset starting condition, the controller controls the RPA to run according to the starting condition, the RPA collects data through a login microkernel terminal to obtain a cross-border pass matrix, and obtains extraction content data through a text recognition algorithm according to the cross-border pass matrix, a preset filling template, and the RPA fills the extraction content data according to the filling template to obtain real-time filling data; Obtain historical filling data, calculate the comparison result through a comparison model according to the historical filling data and the real-time filling data, obtain a verification API, and the RPA transmits the comparison result to a data verification terminal according to the verification API and obtains a confirmation signal; The RPA performs a classification operation according to the confirmation signal; The classification operation includes: calculating a filling data type through a classification model according to the real-time filling data, obtaining a classification system API, and the RPA transmits the real-time filling data to a wisdom kernel coordination hub through the classification system API according to the filling data type.

[0008] Specifically, the starting condition includes but is not limited to a time-based starting condition, an event-based starting condition, a data-based starting condition, and an external instruction-based starting condition.

[0009] Specifically, the microkernel terminal includes but is not limited to a customs single window platform, various trade document processing systems, and a mailbox.

[0010] Specifically, the text recognition algorithm includes the following steps: According to the cross-border pass matrix, image segmentation is performed through a high-resolution segmentation network to obtain segmented image data; According to the segmented image data, text spatiotemporal sequence information is calculated through an improved LSTM model; According to the text spatiotemporal sequence information, a probability graph is calculated through a character mask module, a preset probability threshold is set, the probability graph is binarized according to the probability threshold to obtain a binary graph, an outermost rectangle of an outer contour is calculated through a contour detection algorithm according to the binary graph, a channel probability value is calculated according to the outermost rectangle of the outer contour, the channel probability value is sorted to obtain a maximum channel probability value, a text prediction result is obtained according to the maximum channel probability value, a preset sorting rule is set, and the extraction content data is obtained by sorting the text prediction result according to the sorting rule; The expression calculated by the character mask module is: , Wherein, F0 is a probability graph, Fi is text spatiotemporal sequence information, k is an experience value, and p is a character mask output feature.

[0011] Specifically, the improved LSTM model expression is: , wherein h t is the hidden layer state at time t, that is, the text space sequence information, f t is the forgetting gate at time t, i t is the memory gate at time t, a t is the temporary state information at time t, c t is the cell state at time t, o t is the output gate at time t, δ is the Leaky Relu function, w f , w i , w c , w o are weight coefficients, h t-1 is the hidden layer state at t-1, X t is the segmented image data, b f , b i , b t , b o are bias terms, and t represents time.

[0012] Specifically, the RPA fills the extracted content data according to the filling template to obtain real-time filling data, and the method specifically comprises the following steps: obtaining the filling template type, confirming the parsing method according to the filling template type, and parsing the filling template according to the parsing method to obtain cell position and table header information; calculating a similarity value according to the table header information and the extracted data through a cosine similarity algorithm, presetting a similarity threshold, and screening the table header information according to the similarity threshold and the similarity value to obtain a corresponding table header; filling the extracted content according to the cell position of the corresponding table header to obtain the real-time filling data.

[0013] Specifically, the comparison model comprises a cross comparison model and a prediction comparison model.

[0014] Specifically, the cross comparison model specifically comprises the following steps: The real-time filling data comprises customs declaration form filling data and accompanying single filling data; presetting a retrieval semantic sequence, obtaining a retrieval semantic by traversing the retrieval semantic sequence, and retrieving the customs declaration form filling data and the accompanying single filling data according to the retrieval semantic to obtain a comparison data pair; According to the comparison data, the comparison similarity calculated by the Levenshtein distance algorithm is obtained, a preset comparison similarity threshold is set, the comparison similarity is judged according to the comparison similarity threshold, if the comparison similarity is not less than the comparison similarity threshold, no processing is performed; if the comparison similarity is less than the comparison similarity threshold, the comparison data pair is marked to obtain the comparison result.

[0015] Specifically, the prediction comparison model specifically includes the following calculation steps: A preset price retrieval semantic is set, the historical filling data and the real-time filling data are retrieved according to the price retrieval semantic to obtain historical price data and real-time price data, and the historical price data is stored to obtain a historical price sequence; According to the historical price sequence, a sequence frequency average amplitude is calculated, a maximum frequency amplitude is calculated according to the sequence frequency average amplitude, a cycle length is calculated according to the maximum frequency amplitude, and an average pooling operation is performed on the historical price sequence according to the cycle length to obtain a down-sampling sequence, and the down-sampling sequence is spliced to obtain a multi-scale sequence; The sequence frequency average amplitude calculation expression is: , Wherein, A is the sequence frequency average amplitude, Amp represents the calculation amplitude, Avg represents the average operation along the variable dimension, FFT represents the Fourier transform, and X represents the historical price sequence; The maximum frequency amplitude calculation expression is: , Wherein, f is the maximum frequency amplitude, and A represents the sequence frequency average amplitude; The cycle length calculation expression is: , Wherein, p is the cycle length, L is the length of the historical price sequence, and f is the maximum frequency amplitude; The down-sampling sequence calculation expression is: , Wherein, Y is the down-sampling sequence, X is the historical price sequence, and p is the cycle length; According to the multi-scale sequence, a trend item and a seasonal item are obtained by time series decomposition, a preset time window is set, the seasonal item is intercepted according to the time window to obtain an intercepted seasonal item, the trend item and the intercepted seasonal item are respectively predicted to obtain a trend prediction value and a seasonal prediction value, and the trend prediction value and the seasonal prediction value are added to obtain a price prediction value; The trend item calculation expression is: X1 is a trend item, n is the number of influencing factors, i is a variable parameter, and x is a multi-scale sequence. The seasonal term calculation expression is: X2 is a seasonal term, X is a multi-scale sequence, and X1 is a trend item. The trend prediction value expression is: X3 is a trend prediction value, X1 is a trend item, and Liner represents a linear layer. The seasonal prediction value expression is: X4 is a seasonal prediction value, X5 is a truncated seasonal term, and Liner represents a linear layer. The difference between the price prediction value and the real-time price data is calculated, a preset difference threshold is set, it is judged whether the difference is greater than the difference threshold, if yes, the real-time price data is marked to the comparison result; if not, no processing is performed.

[0016] A digital office system based on RPA, comprising: a preprocessing module, a verification module, a classification module; The preprocessing module is used for presetting a starting condition, and a controller controls RPA to run according to the starting condition; the RPA performs data acquisition by logging in a microkernel terminal to obtain a cross-border pass matrix, performs character recognition on the cross-border pass matrix by a text recognition algorithm to obtain extracted content data, and fills in a preset filling template; and the RPA fills in the extracted content data according to the filling template to obtain real-time filling data. The verification module is used for obtaining historical filling data, calculating a comparison result according to the historical filling data and the real-time filling data by a comparison model, obtaining a verification API, and transmitting the comparison result to a data verification terminal and obtaining a confirmation signal according to the verification API. The classification module is used for the RPA to perform a classification operation according to the confirmation signal. The classification operation includes: calculating a filling data type according to the real-time filling data by a classification model, obtaining a classification system API, and transmitting the real-time filling data to a wisdom kernel collaborative hub according to the filling data type by the classification system API.

[0017] The present application has the following advantages: ​​​(1) Set up a text recognition algorithm, use a high-resolution segmentation network for image segmentation, combine an improved LSTM model to calculate text spatio-temporal sequence information, and through a series of operations such as a character mask module, effectively solve the problem of inaccurate recognition of traditional OCR technology. When processing complex format and content cross-border token matrix files, it can accurately identify the text and obtain high-quality extracted content data, providing a reliable foundation for subsequent data filling and processing, and improving the accuracy and stability of the entire business process.

[0018] (2) The comparison model includes a cross-comparison model and a prediction comparison model. The cross-comparison model can comprehensively and accurately compare the data filled in the customs declaration and the data filled in the accompanying single by traversing the semantic sequence and using the levenshtein distance algorithm, effectively identifying data differences and marking abnormal data pairs. The prediction comparison model accurately predicts price trends by deep analysis of historical price data, including calculating sequence frequency average amplitude, maximum frequency, and cycle length, constructing multi-scale sequences, and performing time series decomposition. By comparing with real-time price data, it can accurately judge whether the real-time price is reasonable, providing strong support for business decision-making and avoiding economic losses caused by unreasonable prices.

[0019] (3) Based on accurate data collection, recognition and comparison results, RPA can perform classification operations according to confirmation signals, calculate the type of filled data through a classification model, and accurately transmit the data to the intelligence core collaborative hub. This process realizes the automation and intelligentization of business processes, optimizes data flow and resource allocation, improves overall office efficiency, reduces labor costs, and enhances the competitiveness of enterprises in a digital office environment. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to facilitate the understanding of those skilled in the art, the present application will be further described below with reference to the accompanying drawings.

[0021] Figure 1 The flowchart of a digital office method based on RPA. DETAILED DESCRIPTION

[0022] In order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined invention purpose, the specific implementation, structure, features and effects of the present application will be described in detail below with reference to the preferred embodiments and the accompanying drawings.

[0023] Please refer to Figure 1 , a digital office method, device, equipment and medium based on RPA; A preset starting condition, a controller controls the RPA to run according to the starting condition, the RPA collects data through a login microkernel terminal to obtain a cross-border pass matrix, and obtains extraction content data through a text recognition algorithm according to the cross-border pass matrix, a preset filling template, and the RPA fills the extraction content data according to the filling template to obtain real-time filling data; Obtain historical filling data, calculate the comparison result through a comparison model according to the historical filling data and the real-time filling data, obtain a verification API, and the RPA transmits the comparison result to a data verification terminal according to the verification API and obtains a confirmation signal; The RPA performs a classification operation according to the confirmation signal; The classification operation includes: calculating a filling data type through a classification model according to the real-time filling data, obtaining a classification system API, and the RPA transmits the real-time filling data to a wisdom kernel coordination hub through the classification system API according to the filling data type.

[0024] Specifically, the starting condition includes but is not limited to a time-based starting condition, an event-based starting condition, a data-based starting condition, and an external instruction-based starting condition.

[0025] Specifically, the microkernel terminal includes but is not limited to a customs single window platform, various trade document processing systems, and a mailbox.

[0026] Specifically, the text recognition algorithm includes the following steps: According to the cross-border pass matrix, image segmentation is performed through a high-resolution segmentation network to obtain segmented image data; According to the segmented image data, text spatiotemporal sequence information is calculated through an improved LSTM model; According to the text spatiotemporal sequence information, a probability graph is calculated through a character mask module, a probability threshold is preset, the probability graph is binarized according to the probability threshold to obtain a binary graph, an outermost rectangle of an outer contour is calculated through a contour detection algorithm according to the binary graph, a channel probability value is calculated according to the outermost rectangle of the outer contour, the channel probability value is sorted to obtain a maximum channel probability value, a character prediction result is obtained according to the maximum channel probability value, a sorting rule is preset, and the extraction content data is obtained by sorting the character prediction result according to the sorting rule; The expression calculated by the character mask module is: , Wherein, F0 is a probability graph, Fi is text spatiotemporal sequence information, k is an experience value, and p is a character mask output feature.

[0027] In this embodiment, the segmented image data is obtained by dividing the original image into multiple subregions with similar features using an image segmentation algorithm. Each subregion corresponds to a specific category or object. In high-resolution image segmentation, these subregions can be pixel-level segmentation results or higher-level objects or regions. The contour detection algorithm includes but is not limited to the findcontours function, the minAreaRect function, and the convexHull function.

[0028] Specifically, the improved LSTM model expression is: , Among them, h t is the hidden state at time t, which is the spatiotemporal sequence information of the text, f t is the forget gate at time t, i t is the memory gate at time t, a t is the temporary status information at time t, c t is the cell state at time t, o t is the output gate at time t, δ is the Leaky Relu function, w f 、w i 、w c 、w o is the weight coefficient, h t-1 is the hidden state at time t-1, X t is the segmented image data, b f 、b i 、b t 、b o is the bias term, and t represents time.

[0029] In this embodiment, the improved LSTM model can directly process multidimensional data by changing the fully connected layer in the LSTM model into a convolution operation, and modifying the activation function to the ELU function and the Leaky Relu function can alleviate the gradient disappearance problem and the problem of complete failure in the negative value area.

[0030] Specifically, the RPA fills the extracted content data according to the filling template to obtain the filling data, which specifically includes: Obtaining the filling template type, determining a parsing method according to the filling template type, and parsing the filling template according to the parsing method to obtain cell positions and header information; Calculating a similarity value based on the header information and the extracted data using a cosine similarity algorithm, presetting a similarity threshold, and filtering the header information based on the similarity threshold and the similarity value to obtain a corresponding header; The filling data is obtained by filling the extracted content according to the cell position of the corresponding table header.

[0031] Specifically, the comparison model includes a cross comparison model and a prediction comparison model, and the specific calculation steps of the cross comparison model include: The real-time filling data includes customs declaration filling data and accompanying single filling data. A preset retrieval semantic sequence is traversed to obtain a retrieval semantic, and the customs declaration filling data and the accompanying filling data are retrieved according to the retrieval semantic to obtain a comparison data pair. A comparison similarity is calculated according to the comparison data pair through a levenshtein distance algorithm, a comparison similarity threshold is preset, the comparison similarity is judged according to the comparison similarity threshold, if the comparison similarity is not less than the comparison similarity threshold, no processing is performed, and if the comparison similarity is less than the comparison similarity threshold, the comparison data pair is labeled to obtain the comparison result.

[0032] Specifically, the specific calculation steps of the prediction comparison model include: A price retrieval semantic is preset, the historical filling data and the real-time filling data are retrieved according to the price retrieval semantic to obtain historical price data and real-time price data, and the historical price data is stored to obtain a historical price sequence. A sequence frequency average amplitude is calculated according to the historical price sequence, a maximum frequency of amplitude is calculated according to the sequence frequency average amplitude, a cycle length is calculated according to the maximum frequency of amplitude, an average pooling operation is performed on the historical price sequence according to the cycle length to obtain a down-sampling sequence, and the down-sampling sequence is spliced to obtain a multi-scale sequence. The sequence frequency average amplitude calculation expression is: , Wherein, A is a sequence frequency average amplitude, Amp represents a calculation amplitude, Avg represents an average operation along a variable dimension, FFT represents a Fourier transform, and X represents a historical price sequence. The maximum frequency of amplitude calculation expression is: , Wherein, f is a maximum frequency of amplitude, and A represents a sequence frequency average amplitude. The cycle length calculation expression is: , Wherein, p is a cycle length, L is a historical price sequence length, and f is a maximum frequency of amplitude. The down-sampling sequence calculation expression is: , Wherein, Y is a down-sampling sequence, X is a historical price sequence, and p is a cycle length. According to the multi-scale sequence, a trend item and a seasonal item are obtained through time series decomposition, a preset time window is set, the seasonal item is intercepted according to the time window to obtain an intercepted seasonal item, the trend item and the intercepted seasonal item are respectively predicted to obtain a trend prediction value and a seasonal prediction value, and the trend prediction value and the seasonal prediction value are added to obtain a price prediction value. The trend item calculation expression is: , Wherein, X1 is a trend item, n is the number of influencing factors, i is a variable parameter, and x is a multi-scale sequence. The seasonal item calculation expression is: , Wherein, X2 is a seasonal item, X is a multi-scale sequence, and X1 is a trend item. The trend prediction value expression is: , Wherein, X3 is a trend prediction value, X1 is a trend item, and Liner represents a linear layer. The seasonal prediction value expression is: , Wherein, X4 is a seasonal prediction value, X5 is an intercepted seasonal item, and Liner represents a linear layer. The difference between the price prediction value and the real-time price data is calculated, a preset difference threshold is set, it is judged whether the difference is greater than the difference threshold, if yes, the real-time price data is marked to the comparison result; if not, no processing is performed.

[0033] Specifically, the classification model specifically calculates the steps comprising: According to the filling data, a word embedding vector is calculated through a K-BERT model, a feature extraction vector is obtained through a TextCNN model for feature extraction according to the word embedding vector, and the filling data type is obtained through a Softmax classification layer for classification according to the feature extraction vector.

[0034] A digital office system based on RPA, comprising: a preprocessing module, a verification module, and a classification module. The preprocessing module is used for presetting a starting condition, and a controller controls an RPA to run according to the starting condition, the RPA collects data through login of a microkernel terminal to obtain a cross-border pass matrix, performs character recognition on the cross-border pass matrix through a text recognition algorithm to obtain extracted content data, and a preset filling template is filled in the extracted content data by the RPA to obtain real-time filling data; The verification module is used for obtaining historical filling data, calculating a comparison result through a comparison model according to the historical filling data and the real-time filling data, obtaining a verification API, and transmitting the comparison result to a data verification terminal and obtaining a confirmation signal according to the verification API by the RPA; The classification module is used for the RPA to perform a classification operation according to the confirmation signal. The classification operation includes: calculating a filling data type through a classification model according to the real-time filling data, obtaining a classification system API, and transmitting the real-time filling data to a wisdom kernel coordination hub through the classification system API according to the filling data type by the RPA.

[0035] An electronic device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor implements the RPA-based digital office method as described above when executing the program.

[0036] A storage medium containing computer executable instructions for executing the RPA-based digital office method as described above when executed by a computer processor.

[0037] The computer storage medium of the embodiment of the application can adopt any combination of one or more computer readable media. The computer readable medium can be a computer readable signal medium or a computer readable storage medium. The computer readable storage medium may, for example, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or component, or any combination thereof. More specific examples (non-exhaustive list) of computer readable storage media include: electrical connections with one or more conductive wires, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this document, the computer readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device or component.

[0038] A computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, which carries computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0039] The program code included in the computer-readable medium can be transmitted with any appropriate medium, including but not limited to wireless, electric wire, optical cable, RF or the like, or any suitable combination thereof. The computer program code for performing the operation of the present invention can be written in one or more programming languages ​​or a combination thereof, and the programming language includes an object-oriented programming language such as Java, Smalltalk, C++, and also includes a conventional procedural programming language such as "C" language or similar programming language. The program code can be executed completely on the user's computer, partially on the user's computer, as an independent software package, partially on the user's computer and partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (for example, utilizing an Internet service provider to connect through the Internet).

[0040] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to equivalent embodiments using the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A digital office method based on RPA, characterized in that: include: A startup condition is preset, and the controller controls the RPA to run according to the startup condition. The RPA logs in to the micro-core terminal to collect data to obtain a cross-border token matrix, and performs text recognition using a text recognition algorithm based on the cross-border token matrix to obtain extracted content data. A filling template is preset, and the RPA fills the extracted content data according to the filling template to obtain real-time filling data; Obtain historical filling data, calculate a comparison result based on the historical filling data and the real-time filling data using a comparison model, obtain a verification API, and the RPA transmits the comparison result to a data verification terminal according to the verification API and obtains a confirmation signal; The RPA performs a classification operation according to the confirmation signal; The classification operation includes: calculating the filling data type through the classification model according to the real-time filling data, obtaining the classification system API, and the RPA transmitting the real-time filling data to the smart core collaborative center through the classification system API according to the filling data type.

2. The RPA-based digital office method according to claim 1, characterized in that: The start condition includes but is not limited to a time-based start condition, an event-based start condition, a data-based start condition, and an external instruction-based start condition.

3. The RPA-based digital office method according to claim 1, characterized in that: The micro-core terminal includes but is not limited to the customs single window platform, various trade document processing systems, and mailboxes.

4. The RPA-based digital office method according to claim 1, characterized in that: The specific calculation steps of the text recognition algorithm include: Perform image segmentation using a high-resolution segmentation network according to the cross-border token matrix to obtain segmented image data; The text spatiotemporal sequence information is calculated by improving the LSTM model according to the segmented image data; A probability map is calculated by a character mask module based on the spatiotemporal sequence information of the text, a probability threshold is preset, the probability map is binarized based on the probability threshold to obtain a binary map, a minimum circumscribed rectangle of an outer contour is calculated based on the binary map using a contour detection algorithm, channel probability values ​​are calculated based on the minimum circumscribed rectangle of the outer contour, the channel probability values ​​are sorted to obtain a maximum channel probability value, a text prediction result is obtained by indexing based on the maximum channel probability value, a sorting rule is preset, and the text prediction results are sorted according to the sorting rule to obtain the extracted content data; The character mask module calculation expression is: ， Among them, F0 is the probability map, Fi is the spatiotemporal sequence information of the text, k is the empirical value, and p is the character mask output feature.

5. The RPA-based digital office method according to claim 4, characterized in that: The improved LSTM model expression is: ， Among them, h t is the hidden state at time t, which is the spatiotemporal sequence information of the text, f t is the forget gate at time t, i t is the memory gate at time t, a t is the temporary status information at time t, c t is the cell state at time t, o t is the output gate at time t, δ is the Leaky Relu function, w f 、w i 、w c 、w o is the weight coefficient, h t-1 is the hidden state at time t-1, X t is the segmented image data, b f 、b i 、b t 、b o is the bias term, and t represents time.

6. The RPA-based digital office method according to claim 1, characterized in that: The RPA fills the extracted content data according to the filling template to obtain real-time filling data specifically including: Obtaining the filling template type, determining a parsing method according to the filling template type, and parsing the filling template according to the parsing method to obtain cell positions and header information; Calculating a similarity value based on the header information and the extracted data using a cosine similarity algorithm, presetting a similarity threshold, and filtering the header information based on the similarity threshold and the similarity value to obtain a corresponding header; The extracted content is filled according to the cell position of the corresponding table header to obtain the real-time filling data.

7. The RPA-based digital office method according to claim 1, characterized in that: The alignment model includes a cross alignment model and a predictive alignment model.

8. The RPA-based digital office method according to claim 7, characterized in that: The specific calculation steps of the cross comparison model include: The real-time filling data includes customs declaration filling data and accompanying document filling data; Preset a search semantic sequence, traverse the search semantic sequence to obtain search semantics, and search the customs declaration form filling data and the accompanying filling data according to the search semantics to obtain a comparison data pair; A comparison similarity is calculated based on the comparison data pair using a Levenshtein distance algorithm, a comparison similarity threshold is preset, and the comparison similarity is judged based on the comparison similarity threshold. If the comparison similarity is not less than the comparison similarity threshold, no processing is performed; if the comparison similarity is less than the comparison similarity threshold, the comparison data pair is labeled to obtain the comparison result.

9. The RPA-based digital office method according to claim 6, characterized in that: The specific calculation steps of the prediction comparison model include: Preset price retrieval semantics, retrieve the historical filling data and the real-time filling data according to the price retrieval semantics to obtain historical price data and real-time price data, and store the historical price data to obtain a historical price sequence; The average amplitude of the sequence frequency is calculated based on the historical price sequence, the maximum amplitude frequency is calculated based on the average amplitude of the sequence frequency, the period length is calculated based on the maximum amplitude frequency, the historical price sequence is average pooled according to the period length to obtain a down-sampled sequence, and the down-sampled sequences are spliced ​​to obtain a multi-scale sequence; The calculation expression of the average amplitude of the sequence frequency is: , Where A is the average amplitude of the sequence frequency, Amp represents the calculated amplitude, Avg represents the average operation along the variable dimension, FFT represents Fourier transform, and X represents the historical price sequence; The calculation expression of the maximum frequency of the amplitude is: ， Among them, f is the maximum amplitude frequency, A represents the average amplitude of the sequence frequency; The cycle length calculation expression is: ， Where p is the cycle length, L is the length of the historical price series, and f is the frequency with the maximum amplitude; The downsampling sequence calculation expression is: ， Where Y is the downsampled sequence, X is the historical price sequence, and p is the cycle length; Decomposing the multi-scale sequence by time series to obtain a trend term and a seasonal term, presetting a time window, intercepting the seasonal term according to the time window to obtain an intercepted seasonal term, predicting the trend term and the intercepted seasonal term respectively to obtain a trend forecast value and a seasonal forecast value, and adding the trend forecast value and the seasonal forecast value to obtain a price forecast value; The trend item calculation expression is: ， Among them, X1 is the trend term, n is the number of influencing factors, i is the variable parameter, and x is the multi-scale sequence; The seasonal term calculation expression is: ， Among them, X2 is the seasonal term, X is the multi-scale sequence, and X1 is the trend term; The trend prediction value expression is: ， Among them, X3 is the trend prediction value, X1 is the trend term, and Liner represents the linear layer; The seasonal prediction value expression is: Among them, X4 is the seasonal prediction value, X5 is the intercepted seasonal term, and Liner represents the linear layer; Calculate the difference between the price prediction value and the real-time price data, preset a difference threshold, and determine whether the difference is greater than the difference threshold. If so, match the real-time price data to the comparison result; if not, do nothing.

10. A digital office system based on RPA, characterized in that: include: Preprocessing module, verification module, classification module; The pre-processing module is used to preset startup conditions. The controller controls the RPA to run according to the startup conditions. The RPA logs in to the micro-core terminal to collect data to obtain a cross-border token matrix. The text recognition algorithm is used to perform text recognition based on the cross-border token matrix to obtain extracted content data. A filling template is preset. The RPA fills the extracted content data according to the filling template to obtain real-time filling data. The verification module is configured to obtain historical filling data, calculate a comparison result based on the historical filling data and the real-time filling data using a comparison model, obtain a verification API, and transmit the comparison result to a data verification terminal according to the verification API and obtain a confirmation signal. The classification module is configured for the RPA to perform a classification operation according to the confirmation signal; The classification operation includes: calculating the filling data type through the classification model according to the real-time filling data, obtaining the classification system API, and the RPA transmitting the real-time filling data to the smart core collaborative center through the classification system API according to the filling data type.

Images