CN117875906B - Electronic bill auditing method based on artificial intelligence - Google Patents

Abstract

The invention provides an electronic bill auditing method based on artificial intelligence, which comprises the following steps: acquiring an electronic bill image, and identifying the electronic bill image to obtain bill information; matching the bill information to an order database to obtain bill auditing information; based on the product information and the client information, determining a plurality of raw materials and corresponding raw material demand of the product; predicting raw material supply amounts of the plurality of raw materials based on the warehouse information, the current season information and the position of the processing plant; judging whether the raw material supply amount of each raw material is larger than the corresponding raw material demand amount or not respectively; if the first signing time and the second signing time are predicted based on the supply chain information, the processing information and the transportation information; respectively judging whether the first signing time is smaller than or equal to a first preset time and whether the second signing time is smaller than or equal to a second preset time; if the electronic bill is smaller than the electronic bill, the electronic bill is checked and approved; so as to reduce the loss rate of the prefabricated dishes and improve the quality of the dishes.

Description

Electronic bill auditing method based on artificial intelligence

Technical Field

The invention relates to the technical field of bill auditing, in particular to an electronic bill auditing method based on artificial intelligence.

Background

Cold chain logistics refers to a logistics transportation mode for maintaining a certain temperature of goods all the time in the links of processing, storage, transportation, distribution, retail and the like of the products for maintaining the quality of the foods or the efficacy of other products and reducing transportation loss. The current flow of food cold chain logistics is to pre-cool agricultural products in an original product producing area, transport the pre-cooled agricultural products to a processing factory for cold chain processing, transport the processed frozen products or fresh-keeping products to a cold chain storage for storage, and finally deliver the frozen products to a distributor for sale in a cold chain transportation mode. Because the objects of the food cold chain logistics are mostly food with longer shelf life such as vegetables, fruits, frozen products and the like, the transportation aging requirement on the food is loose. Because the shelf life of the prefabricated dish is shorter, usually about 5 days, the conventional cold chain logistics cannot well meet the transportation requirement of the prefabricated dish, and because the error of time control in the process is larger, the loss rate of the prefabricated dish in the transportation process is higher. And traditional cold chain logistics are used for conveying raw materials to various places without finishing food; however, the population tastes vary from region to region, simply marketing the prepared dish to each location does not fit well into the well-suited eating habits. Therefore, there is a need to optimize the conventional food cold chain stream to accommodate the transportation of prepared vegetables.

In view of the above, the invention provides an electronic bill auditing method based on artificial intelligence, which controls the raw material transportation time and the product transportation time to realize the overall grasp of the product transportation time, so as to reduce the loss rate of the prefabricated dishes and improve the quality of the dishes.

Disclosure of Invention

The invention aims to provide an electronic bill auditing method based on artificial intelligence, which comprises the following steps: acquiring an electronic bill image, and identifying the electronic bill image to obtain bill information; the bill information comprises an order number, a product name, the number of products, a client name, a client address, a transportation mode and a transportation temperature; matching the bill information to an order database to obtain bill auditing information; the bill auditing information comprises order information, product information, supply chain information, warehouse information, processing information, customer information and transportation information; determining a plurality of raw materials and corresponding raw material demand of the product obtained by processing based on the product information and the client information; predicting raw material supply amounts of the plurality of raw materials based on the warehouse information, the current season information, and the process plant location; judging whether the raw material supply amount of each raw material is larger than the corresponding raw material demand amount or not respectively; if any raw material supply quantity is smaller than the corresponding raw material demand quantity, the electronic bill verification is not passed; if the raw material supply amount of each raw material is greater than or equal to the corresponding raw material demand amount, predicting a first signing time and a second signing time based on the supply chain information, the processing information and the transportation information; respectively judging whether the first signing time is smaller than or equal to a first preset time and whether the second signing time is smaller than or equal to a second preset time; if the first signing time is longer than the first preset time or the second signing time is longer than the second preset time, the electronic bill verification is not passed; the first preset time is related to a shelf life of the plurality of raw materials; the second preset time is related to a shelf life of the product and an aging requirement of a customer; if the first signing time is smaller than or equal to the first preset time and the second signing time is smaller than or equal to the second preset time, the electronic bill is checked and approved.

Further, before the raw materials are loaded and transported to the processing plant, the method further comprises the following steps: acquiring a raw material picture and raw material electronic bill information; identifying the raw material picture to obtain a raw material identification result; matching the raw material identification result with the raw material name in the raw material electronic receipt information, and if the matching is not passed, checking the raw material electronic receipt information is not passed; if the matching is passed, the raw material electronic bill information is checked and passed, the raw material electronic bill information is printed and posted on the corresponding raw material.

Further, before the product is loaded and transported to the customer, the method further comprises: acquiring a product picture and product electronic bill information; identifying the product picture to obtain a product identification result; matching the product identification result with the product name in the product electronic bill information, and if the matching is not passed, checking the product electronic bill information is not passed; if the matching is passed, the product electronic bill information is checked and passed, the product electronic bill information is printed and posted on the corresponding product.

Further, the acquiring the electronic document image and identifying the electronic document image to obtain document information includes: determining a plurality of colors in the electronic bill image, and obtaining a plurality of color channels according to the colors; channel selection is carried out on the electronic document image through the plurality of color channels, so that a plurality of channel images are obtained; performing contrast processing on the plurality of channel images, and performing differential processing on the processed channel images to obtain a plurality of differential images; carrying out gray scale processing on the plurality of differential images to obtain a plurality of gray scale images; respectively carrying out expansion and contraction treatment on the plurality of gray level images to obtain a plurality of treated gray level images; and inputting the plurality of processed gray images into an image recognition model, and outputting the bill information by the image recognition model.

Further, the image recognition model is obtained by training an initial image recognition model, and the initial image recognition model comprises an input module, a convolution module, a full connection module, an attention module, an output module and a recognition module; the input module is used for receiving the plurality of processed gray images and transmitting the plurality of processed gray images to the plurality of convolution modules; the convolution module is used for respectively carrying out convolution processing on the plurality of processed gray images to obtain a plurality of first feature images; the full-connection module is used for receiving the first feature map and fully connecting the first feature map to obtain a full-connection map; the attention module is used for receiving the first feature map and extracting the first feature map to obtain a plurality of groups of attention force diagrams; the output module is used for receiving the full-connection image and the multiple groups of attention force diagrams, processing and identifying the full-connection image and the multiple groups of attention force diagrams, and obtaining the bill information.

Further, the expression of the loss function for training the initial image recognition model is:

，/> ; wherein/> Representing a loss function; /(I)Representing a fourth variable, R representing the total number of neurons in the output module; /(I)Representing the actual output of the training samples, i.e., the labels; /(I)An output representing an initial image recognition model; /(I)Representing an excitation function of neurons within the output module; /(I)Representing weights between the full connection module to the output module neurons; /(I)Representing an excitation function of the fully connected module; /(I)Representing a second variable, i=j; /(I)Representing the total number of neurons within the convolution module; /(I)Representing the weights of the convolution module neurons to the fully connected modules; /(I)Representing an excitation function of neurons within an input module; /(I)Representing a first variable; /(I)Representing the total number of neurons in the input module; /(I)Representing weights between the input module and the convolution module; /(I)Representing an input variable; /(I)A threshold value representing a j-th neuron in the input module; /(I)A threshold value representing a fully connected module; /(I)A threshold value representing neurons connected to the fully connected module within the output module; /(I)Represents a third variable, k=l; /(I)Representing the total number of neurons within the attention module; /(I)Representing the weight between the attention module neurons to the output module neurons; /(I)Representing an excitation function of the attention module; /(I)Representing the weight of the convolution module neurons to the attention module neurons; /(I)A threshold value representing a kth neuron within the attention module; /(I)Representing the threshold value of the first neuron in the output module connected to the attention module.

Further, the attention module comprises a plurality of attention units, and each attention unit is respectively used for processing a different first characteristic diagram; the attention unit comprises a plurality of anti-attention models, a first fusion model, a second fusion model and an attention model; the attention model is used for extracting attention characteristics of the current first characteristic diagram to obtain an attention characteristic diagram; the anti-attention model is used for carrying out anti-attention feature extraction on the residual first feature images to obtain a plurality of anti-attention feature images; the first fusion model is used for fusing the plurality of anti-attention feature images to obtain a fused feature image; and the second fusion model is used for fusing the attention characteristic diagram and the fusion characteristic diagram to obtain the current attention diagram.

Further, determining, based on the product information and the customer information, a plurality of raw materials and corresponding raw material demands for processing the product, including: establishing a combined customer portrait and an independent customer portrait set according to the historical customer information; extracting a current customer portrait from the independent customer portrait set according to the customer information; processing the combined customer portraits according to the customer information to obtain a plurality of relevant customer portraits; inputting the current customer representation, the plurality of related customer representations, and the product information into a raw material demand prediction model, the raw material demand prediction model outputting the predicted raw material demand.

Further, the predicting the raw material supply amounts of the plurality of raw materials based on the warehouse information, the current season information, and the process plant location includes: acquiring a historical order number according to a time sequence, and matching the historical order number to an order database to obtain historical bill auditing information; extracting the historical document audit information according to the historical order number to obtain a plurality of groups of historical warehouse information, historical season information, historical processing plant positions, historical raw material supply quantity and supply scores; the supply score is related to the remaining amount of raw material or the shortage amount of raw material; taking the set of the historical warehouse information, the historical season information and the historical processing plant position as a first training sample, and taking the corresponding set of the historical raw material supply quantity and the supply score as a first label; inputting the first training sample into an initial raw material supply model and determining a loss function based on an output of the initial raw material supply model and a difference of the first label; repeating training the initial raw material supply model based on the first training sample and the first label until the value of the first loss function is less than a first preset threshold; taking the initial raw material supply quantity model when the value of the first loss function is smaller than the first preset threshold value as a trained raw material supply quantity model; the warehouse information, the current season information, and the process plant location are input to the raw material supply quantity model, which outputs the raw material supply quantity.

Further, the method further comprises the step of judging whether the transportation temperature is within a preset temperature range or not based on the transportation mode and the transportation distance, and if not, the electronic bill verification is not passed.

The technical scheme of the embodiment of the invention has at least the following advantages and beneficial effects:

according to the invention, the raw material transportation time and the product transportation time are controlled, so that the overall grasp of the product transportation time is realized, the loss rate of the prefabricated dishes is reduced, and the quality of the dishes is improved.

The invention can avoid the problem of bill dragging caused by insufficient raw materials by auditing the raw material supply quantity, and improves the satisfaction rate of customers.

According to the invention, the electronic bill is checked before the raw materials and products are loaded, so that the situation that the plates are not in stock can be avoided, and the situation of error in supply is effectively reduced.

According to the invention, the bill is processed and identified through the color of the bill image, so that the interference information can be reduced, and the accuracy of identifying the bill information can be improved.

According to the invention, the rest first feature images are subjected to attention extraction through the anti-attention mechanism to obtain the anti-attention feature images, and then are fused with the attention feature images of the extracted current first feature images to obtain the final attention map of the current first attention feature images, so that the accuracy of attention identification is improved, and the images of the segmented bill information are more accurate.

Drawings

FIG. 1 is an exemplary flow chart of an electronic document auditing method based on artificial intelligence according to some embodiments of the present invention.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments of the present invention. The components of the embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations.

FIG. 1 is an exemplary flow chart of an electronic document auditing method based on artificial intelligence according to some embodiments of the present invention. As shown in fig. 1, the electronic document auditing method provided by the invention comprises the following steps:

acquiring an electronic bill image, and identifying the electronic bill image to obtain bill information; the bill information comprises an order number, a product name, a product number, a client name, a client address, a transportation mode and a transportation temperature.

The electronic document image may be a document image obtained after a customer places a pre-formed menu. And after the electronic bill image verification is passed, the electronic bill image verification can be printed into a paper bill for later process use. The electronic document image may include document information, information category, document expiration, dividing frame, and generation time. Various information in the electronic document image may be identified in different colors to distinguish. The bill information may refer to information related to a current bill, and may include an order number, a product name, a product number, a client name, a client address, a transportation mode, a transportation temperature, and the like. The information category may refer to a hint word for indicating the document information category. For example, "order number 202402267895" may be considered as an information category, and "202402267895" may be considered as document information. Document expiration may refer to the document being valid as a credential, e.g., the document expiration may be valid for the day. The division frame may refer to a frame for dividing document information by category to facilitate recognition. The transportation means includes a means of transporting the raw material of the product and a means of transporting the product, and for example, the transportation means may include automobile transportation, air transportation, railway transportation, and the like. The transportation temperature may refer to the temperature of the raw materials of the transportation product and the temperature of the transportation product.

In some embodiments, the acquiring the electronic document image and identifying the electronic document image to obtain document information includes:

And determining a plurality of colors in the electronic document image, and obtaining a plurality of color channels according to the colors.

And selecting the channels of the electronic document image through the plurality of color channels to obtain a plurality of channel images. For example, the electronic document image includes three colors of red, yellow and blue, and a red channel image, a yellow channel image and a blue channel image can be obtained according to the three colors of red, yellow and blue, respectively.

And carrying out contrast processing on the plurality of channel images, and carrying out differential processing on the processed channel images to obtain a plurality of differential images.

Carrying out gray scale processing on the plurality of differential images to obtain a plurality of gray scale images;

and respectively carrying out expansion and contraction treatment on the plurality of gray level images to obtain a plurality of treated gray level images.

And inputting the plurality of processed gray images into an image recognition model, and outputting the bill information by the image recognition model.

In some embodiments, the image recognition model is derived by training an initial image recognition model that includes an input module, a convolution module, a full connection module, an attention module, an output module, and a recognition module; the input module is used for receiving the plurality of processed gray images and transmitting the plurality of processed gray images to the plurality of convolution modules; the convolution module is used for respectively carrying out convolution processing on the plurality of processed gray images to obtain a plurality of first feature images; the full-connection module is used for receiving the first feature map and fully connecting the first feature map to obtain a full-connection map; the attention module is used for receiving the first feature map and extracting the first feature map to obtain a plurality of groups of attention force diagrams; the output module is used for receiving the full-connection image and the multiple groups of attention force diagrams, processing and identifying the full-connection image and the multiple groups of attention force diagrams, and obtaining the bill information. The expression of the loss function for training the initial image recognition model is as follows:

，/> ; wherein/> Representing a loss function; /(I)Representing a fourth variable, R representing the total number of neurons in the output module; /(I)Representing the actual output of the training samples, i.e., the labels; /(I)An output representing an initial image recognition model; /(I)Representing an excitation function of neurons within the output module; /(I)Representing weights between the full connection module to the output module neurons; /(I)Representing an excitation function of the fully connected module; /(I)Representing a second variable, i=j; /(I)Representing the total number of neurons within the convolution module; /(I)Representing the weights of the convolution module neurons to the fully connected modules; /(I)Representing an excitation function of neurons within an input module; /(I)Representing a first variable; /(I)Representing the total number of neurons in the input module; /(I)Representing weights between the input module and the convolution module; /(I)Representing an input variable; /(I)A threshold value representing a j-th neuron in the input module; /(I)A threshold value representing a fully connected module; /(I)A threshold value representing neurons connected to the fully connected module within the output module; /(I)Represents a third variable, k=l; /(I)Representing the total number of neurons within the attention module; /(I)Representing the weight between the attention module neurons to the output module neurons; /(I)Representing an excitation function of the attention module; /(I)Representing the weight of the convolution module neurons to the attention module neurons; /(I)A threshold value representing a kth neuron within the attention module; /(I)Representing the threshold value of the first neuron in the output module connected to the attention module.

In some embodiments, the attention module includes a plurality of attention units, each for processing a different first feature map; the attention unit comprises a plurality of anti-attention models, a first fusion model, a second fusion model and an attention model; the attention model is used for extracting attention characteristics of the current first characteristic diagram to obtain an attention characteristic diagram. The current first feature map may refer to a first feature map processed by the attention unit. The anti-attention model is used for carrying out anti-attention feature extraction on the residual first feature images to obtain a plurality of anti-attention feature images. The remaining first feature map refers to the first feature map other than the current first feature map. The first fusion model is used for fusing the plurality of anti-attention feature images to obtain a fused feature image. And the second fusion model is used for fusing the attention characteristic diagram and the fusion characteristic diagram to obtain the current attention diagram. The current attention profile may refer to an attention profile of the current first feature map. The attention model may be CBAM (Convolutional Block Attention Module) or DANet (Dual Attention Network for Scene Segmentation) model, or the like.

Matching the bill information to an order database to obtain bill auditing information; the document audit information includes order information, product information, supply chain information, warehouse information, process information, customer information, and shipping information.

The order database refers to a database for storing bill information. The order database stores the bill information according to the order number. And expanding the bill information in an index mode. For example, for product a, the raw materials in the raw materials database may be obtained by indexing. For another example, the warehouse data in the warehouse database is obtained by indexing. The bill auditing information refers to information for auditing the bill information. Order information may refer to information related to a pre-prepared order. For example, the order information may include order number, time of order, aging requirements, order status, and the like. The product information may refer to information related to the prepared dish. For example, the product information may include product name, product quantity, product raw material, processing style, raw material shelf life, product shelf life, and the like. Supply chain information refers to information related to product production and product transportation. For example, the supply chain information may be the entire supply chain process of transporting agricultural products from a raw warehouse (warehouse) to a processing plant for processing to produce a product, and then transporting the product to a customer. Warehouse information may refer to information related to a raw warehouse. For example, the warehouse information may include warehouse location, raw material residuals, warehouse temperature, warehouse type, and the like. The process information refers to information related to the process plant. For example, the process information may include process plant location, process recipe, dish type, process preferences, and the like. Customer information may refer to information related to a customer. Such as customer name, customer address, customer preferences, etc. Transportation information may refer to information related to transportation. For example, the transportation information may include transportation mode, transportation temperature, transportation equipment, and the like. Specifically, the order number can be matched to an order database to obtain an order data packet corresponding to the order number, and order information in the order data packet is extracted; then matching the product name to an order data packet to obtain product information; and matching the client name to the order data packet to obtain supply chain information, warehouse information, processing information, client information and corresponding transportation information. In other embodiments, the document audit information may also be stored in a specific data format, and when relevant data needs to be extracted, the data is extracted according to the data format.

And determining a plurality of raw materials and corresponding raw material demand of the product obtained by processing based on the product information and the client information.

The raw materials refer to agricultural products required for producing the dishes. For example, the product may be diced chicken, and the raw materials may include chicken, peanuts, lettuce, peppers, oils, salt, and the like. For the package, the prepared dish can also comprise rice and/or egg. The raw material demand may refer to the amount of raw material required to produce the product in the product quantity.

In some embodiments, a set of federated and independent customer portraits may be established based on historical customer information; the history client information refers to information of clients after documents have been completed. A federated customer representation may refer to an overall representation established for all collaborated customers. For example, the federated customer representation may be in the form of a knowledge graph with nodes of the knowledge graph being customer names, product numbers, product names, customer addresses, etc., edges representing relationships between the nodes. An independent customer representation may refer to a representation of an individual customer. The individual client images may include the client's preferences. And extracting the current customer portrait from the independent customer portrait set according to the customer information. The current customer representation may refer to the customer representation of the customer to be predicted, and for the first-time collaboration customer, an independent customer representation of the historical customer whose similarity exceeds a similarity threshold may be taken as the current customer representation. The similarity threshold value refers to a preset similarity threshold value. And processing the combined customer portraits according to the customer information to obtain a plurality of relevant customer portraits. The relevant customer portrayal may refer to a portrayal of a historical customer having some association with the current customer. Wherein the association may include customer addresses that are similar (e.g., same province or distance less than a distance threshold, etc.), desired products that are the same (e.g., same as a pap chicken) and types that are similar (e.g., same as a customer belonging to one type, which may include a customer who is playing a body-building, a customer who is losing weight, and a customer who is paying attention to taste, a customer who is paying attention to product quality, etc.). Inputting the current customer representation, the plurality of related customer representations, and the product information into a raw material demand prediction model, the raw material demand prediction model outputting the predicted raw material demand. The raw material demand prediction model can be obtained by training an initial raw material demand prediction model through an artificial intelligence training method. The initial raw material demand prediction model can be a neural network model, a semantic matching model and the like. Specifically, the raw material demand prediction model can perform feature extraction on the current customer representation to obtain the current customer feature; the current customer characteristics may be related to the type, address, product name, number of products, preferences, etc. of the current customer. Extracting features of the relevant customer portraits to obtain a plurality of relevant customer features; the relevant customer characteristics may be related to the type, address, product name, etc. of the current customer. And extracting the association between the current client characteristics and the related client characteristics to obtain association characteristics. And splicing the current client features and the associated features to obtain spliced features. And processing the splicing characteristics to obtain the raw material demand.

And predicting the raw material supply amounts of the plurality of raw materials based on the warehouse information, the current season information and the process plant location.

The current season information may refer to a season when the product is processed. The process plant location refers to the location of the process plant that processed the product. The location of the process plant may be selected based on the location of the customer and the customer's preferences to take care of subsequent product transportation and customer taste requirements. The raw material supply amount may refer to the amount of raw material supplied.

In some embodiments, the historical order numbers may be obtained in a time series and matched to an order database to obtain historical document audit information. The history order number refers to an order number of history document information. The historical bill auditing information refers to bill auditing information of a bill. Extracting the historical bill auditing information according to the historical order number to obtain a plurality of groups of historical warehouse information, historical season information, historical processing plant positions, historical raw material supply quantity and supply scores; the supply score is related to the remaining amount of raw material or the shortage amount of raw material. The history repository information refers to repository information when the raw material of the order number is supplied in history. The history season information is information of seasons when the history order number is supplied. The historic process plant location refers to the location of the process plant that processed the order number product. The historical raw material supply amount refers to the amount of raw material actually supplied to the processing plant for product processing. The supply score may be determined based on the end product and the degree of completion of its quantity. And taking the set of the historical warehouse information, the historical season information and the historical processing plant position as a first training sample, and taking the corresponding set of the historical raw material supply quantity and the supply score as a first label. The first training sample is input into an initial stock supply model and a loss function is determined based on an output of the initial stock supply model and a difference of the first label. Wherein the initial raw material supply amount model may be various machine learning models in consideration of time. For example, LSTM model, etc. Repeating training the initial raw material supply model based on the first training sample and the first label until the value of the first loss function is less than a first preset threshold. And taking the initial raw material supply quantity model when the value of the first loss function is smaller than a first preset threshold value as a trained raw material supply quantity model. The warehouse information, the current season information, and the process plant location are input to the raw material supply quantity model, which outputs the raw material supply quantity.

Judging whether the raw material supply amount of each raw material is larger than the corresponding raw material demand amount or not respectively; if any raw material supply quantity is smaller than the corresponding raw material demand quantity, the electronic bill verification is not passed.

If the raw material supply amount of each raw material is greater than or equal to the corresponding raw material demand amount, the first signing time and the second signing time are predicted based on the supply chain information, the processing information and the transportation information.

The first time of issuance may refer to the time the process plant signed up for the last material. The second time of receipt may refer to the time at which the customer signed the product. In some embodiments, the first time of receipt may be determined from historical supply chain information; and determining a second signing time according to the historical processing information and the historical transportation information. For example, a machine learning model for predicting the first signing time and the second signing time may be trained by historical processing information, historical transportation information, and historical supply chain information, and then the supply chain information, the processing information, and the transportation information are processed according to the trained machine learning model to predict the first signing time and the second signing time. The method for predicting the first signing time and the second signing time by the machine learning model is similar to that for predicting the raw material supply amount, and will not be described here again.

Respectively judging whether the first signing time is smaller than or equal to a first preset time and whether the second signing time is smaller than or equal to a second preset time; if the first signing time is longer than the first preset time or the second signing time is longer than the second preset time, the electronic bill verification is not passed; the first preset time is related to a shelf life of the plurality of raw materials; the second preset time is related to the shelf life of the product and the aging requirements of the customer.

If the first signing time is smaller than or equal to the first preset time and the second signing time is smaller than or equal to the second preset time, the electronic bill is checked and approved.

In some embodiments, prior to shipping the feedstock to the process plant, further comprising: and acquiring the raw material picture and the raw material electronic bill information. The raw material picture is a picture of raw materials taken during loading. Raw material electronic bill information may refer to bill information in electronic form that is related to raw materials. The raw material electronic bill information may include information such as order number, raw material name, raw material amount, raw material transportation mode, and raw material transportation temperature. And identifying the raw material picture to obtain a raw material identification result. The raw material identification result may be correlated with a raw material name. For example, the raw material identification result may be a raw material name. And matching the raw material identification result with the raw material name in the raw material electronic receipt information, and if the matching is not passed, checking the raw material electronic receipt information is not passed. If the matching is passed, the raw material electronic bill information is checked and passed, the raw material electronic bill information is printed and posted on the corresponding raw material.

In some embodiments, prior to shipping the product to the customer, further comprising: and obtaining the product picture and the product electronic bill information. The product picture refers to a picture of a product taken when the product is loaded. Product electronic document information may refer to document information in electronic form that is related to a product. The product electronic bill information may include order number, product name, product quantity, product transportation mode, product transportation temperature, and the like. And identifying the product picture to obtain a product identification result. The product identification result may be related to the product name. For example, the product identification result may be a product name. Matching the product identification result with the product name in the product electronic bill information, and if the matching is not passed, checking the product electronic bill information is not passed; if the matching is passed, the product electronic bill information is checked and passed, the product electronic bill information is printed and posted on the corresponding product.

In some embodiments, the method further includes determining whether the transportation temperature is within a preset temperature range based on the transportation mode and the transportation distance, and if not, the electronic document is not approved.

The above is only a preferred embodiment of the present invention, and is not intended to limit the present invention, but various modifications and variations can be made to the present invention by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (2)

1. An electronic document auditing method based on artificial intelligence is characterized by comprising the following steps:

acquiring an electronic bill image, and identifying the electronic bill image to obtain bill information; the bill information comprises an order number, a product name, the number of products, a client name, a client address, a transportation mode and a transportation temperature; the method for acquiring the electronic bill image and identifying the electronic bill image to obtain bill information comprises the following steps:

Determining a plurality of colors in the electronic bill image, and obtaining a plurality of color channels according to the colors;

Channel selection is carried out on the electronic document image through the plurality of color channels, so that a plurality of channel images are obtained;

performing contrast processing on the plurality of channel images, and performing differential processing on the processed channel images to obtain a plurality of differential images;

Carrying out gray scale processing on the plurality of differential images to obtain a plurality of gray scale images;

respectively carrying out expansion and contraction treatment on the plurality of gray level images to obtain a plurality of treated gray level images;

Inputting the plurality of processed gray level images into an image recognition model, and outputting the bill information by the image recognition model; the image recognition model is obtained by training an initial image recognition model, and the initial image recognition model comprises an input module, a convolution module, a full connection module, an attention module, an output module and a recognition module;

The input module is used for receiving the plurality of processed gray images and transmitting the plurality of processed gray images to the plurality of convolution modules;

The convolution module is used for respectively carrying out convolution processing on the plurality of processed gray images to obtain a plurality of first feature images;

The full-connection module is used for receiving the first feature map and fully connecting the first feature map to obtain a full-connection map;

The attention module is used for receiving the first feature map and extracting the first feature map to obtain a plurality of groups of attention force diagrams; the attention module comprises a plurality of attention units, and each attention unit is respectively used for processing a different first characteristic diagram; the attention unit comprises a plurality of anti-attention models, a first fusion model, a second fusion model and an attention model;

the attention model is used for extracting attention characteristics of the current first characteristic diagram to obtain an attention characteristic diagram;

the anti-attention model is used for carrying out anti-attention feature extraction on the residual first feature images to obtain a plurality of anti-attention feature images;

The first fusion model is used for fusing the plurality of anti-attention feature images to obtain a fused feature image;

The second fusion model is used for fusing the attention characteristic diagram and the fusion characteristic diagram to obtain a current attention diagram;

the output module is used for receiving the full-connection image and the multiple groups of attention force diagrams, processing and identifying the full-connection image and the multiple groups of attention force diagrams, and obtaining the bill information;

the expression of the loss function for training the initial image recognition model is as follows:

，

Wherein/> Representing a loss function; /(I)Representing the fourth variable,/>Representing the total number of neurons in the output module; /(I)Representing the actual output of the training samples, i.e., the labels; /(I)An output representing an initial image recognition model; /(I)Representing an excitation function of neurons within the output module; /(I)Representing weights between the full connection module to the output module neurons; /(I)Representing an excitation function of the fully connected module; /(I)Representing the second variable,/>；/>Representing the total number of neurons within the convolution module; /(I)Representing the weights of the convolution module neurons to the fully connected modules; /(I)Representing an excitation function of neurons within an input module; /(I)Representing a first variable; /(I)Representing the total number of neurons in the input module; /(I)Representing weights between the input module and the convolution module; /(I)Representing an input variable; /(I)Representing the/>, within an input moduleA threshold for individual neurons; /(I)A threshold value representing a fully connected module; /(I)A threshold value representing neurons connected to the fully connected module within the output module; /(I)Represents a third variable, k=l; /(I)Representing the total number of neurons within the attention module; /(I)Representing the weight between the attention module neurons to the output module neurons; /(I)Representing an excitation function of the attention module; /(I)Representing the weight of the convolution module neurons to the attention module neurons; /(I)A threshold value representing a kth neuron within the attention module; /(I)A threshold value representing a first neuron in the output module connected to the attention module;

Matching the bill information to an order database to obtain bill auditing information; the bill auditing information comprises order information, product information, supply chain information, warehouse information, processing information, customer information and transportation information;

Determining a plurality of raw materials and corresponding raw material demand of the product obtained by processing based on the product information and the client information; wherein, based on the product information and the customer information, determining a plurality of raw materials and corresponding raw material demands for processing the product, comprising:

establishing a combined customer portrait and an independent customer portrait set according to the historical customer information;

extracting a current customer portrait from the independent customer portrait set according to the customer information;

processing the combined customer portraits according to the customer information to obtain a plurality of relevant customer portraits;

inputting the current customer representation, the plurality of related customer representations, and the product information into a raw material demand prediction model, the raw material demand prediction model outputting the predicted raw material demand;

predicting raw material supply amounts of the plurality of raw materials based on the warehouse information, the current season information, and the process plant location; wherein the predicting the raw material supply amounts of the plurality of raw materials based on the warehouse information, the current season information, and the process plant location includes:

Acquiring a historical order number according to a time sequence, and matching the historical order number to an order database to obtain historical bill auditing information;

Extracting the historical document audit information according to the historical order number to obtain a plurality of groups of historical warehouse information, historical season information, historical processing plant positions, historical raw material supply quantity and supply scores; the supply score is related to the remaining amount of raw material or the shortage amount of raw material;

taking the set of the historical warehouse information, the historical season information and the historical processing plant position as a first training sample, and taking the corresponding set of the historical raw material supply quantity and the supply score as a first label;

inputting the first training sample into an initial raw material supply model and determining a first loss function based on an output of the initial raw material supply model and a difference of the first label;

Repeating training the initial raw material supply model based on the first training sample and the first label until the value of the first loss function is less than a first preset threshold;

Taking the initial raw material supply quantity model when the value of the first loss function is smaller than the first preset threshold value as a trained raw material supply quantity model;

inputting the warehouse information, the current season information and the process plant location into the raw material supply model, the raw material supply model outputting the raw material supply;

Judging whether the raw material supply amount of each raw material is larger than the corresponding raw material demand amount or not respectively; if any raw material supply quantity is smaller than the corresponding raw material demand quantity, the electronic bill verification is not passed;

If the raw material supply amount of each raw material is greater than or equal to the corresponding raw material demand amount, predicting a first signing time and a second signing time based on the supply chain information, the processing information and the transportation information;

Respectively judging whether the first signing time is smaller than or equal to a first preset time and whether the second signing time is smaller than or equal to a second preset time; if the first signing time is longer than the first preset time or the second signing time is longer than the second preset time, the electronic bill verification is not passed; the first preset time is related to a shelf life of the plurality of raw materials; the second preset time is related to a shelf life of the product and an aging requirement of a customer;

if the first signing time is smaller than or equal to the first preset time and the second signing time is smaller than or equal to the second preset time, the electronic bill is checked and approved.

2. The electronic bill auditing method based on artificial intelligence according to claim 1, further comprising judging whether the transportation temperature is within a preset temperature range based on the transportation mode and the transportation distance, and if not, failing the electronic bill auditing.