Power field office processing method and system fusing RPA and AI algorithms

Abstract

The invention discloses an electric power field office processing method and system fusing RPA and AI algorithms, and the method comprises the steps: collecting the content of a document, extracting a text in the document, and classifying the document according to the text content; analyzing the extracted text, completing the classification of text topics, and generating an abstract for the text content at the same time; and automatically reviewing the content of the document by using an RPA technology, and feeding back according to a review result. The office automation level of the power industry is improved, and economic and social values are brought to enterprises.

Description

Technical Field

[0001] The present invention relates to the technical field of office document processing, and specifically to an office processing method and system in the electric power field that integrates RPA and AI algorithms. Background Art

[0002] With the rapid development of the power industry, the demand for document processing and management is increasing, and traditional office methods can no longer meet the requirements of high efficiency and high accuracy. At present, the development of technologies such as artificial intelligence, machine learning, and RPA (Robotic Process Automation) provides a technical basis for automated office. The power industry needs a system that can integrate multiple office functions, which can not only integrate multiple office functions, but also automate office processes, thereby reducing labor costs and significantly improving work efficiency. Against this background, the patent of this invention came into being, aiming to create a comprehensive document processing and management solution through the integration of innovative technologies. In the wave of digital transformation, ensuring the security of document data and the compliance of approval processes is crucial to the power industry. The design of this system fully considers these needs, and ensures the security of the approval process and the confidentiality of document data by using advanced handwriting recognition technology for identity authentication and implementing strict data security measures. Summary of the invention

[0003] In view of the above-mentioned problems, the present invention is proposed.

[0004] The present invention provides the following technical solution: an office processing method in the electric power field integrating RPA and AI algorithms, comprising:

[0005] Collect the content of documents, extract the text in the documents, and classify the documents according to the text content;

[0006] Analyze the extracted text, classify the text topics, and generate a summary of the text content;

[0007] Use RPA technology to automatically review document content and provide feedback based on the review results.

[0008] As a preferred solution of the power field office processing method integrating RPA and AI algorithms described in the present invention, wherein: the text extraction in the document includes, if the document is plain text, extracting text from the scanned paper document or PDF file by OCR technology; using semantic reasoning and semantic analysis to complete the character information of the unrecognized part for the characters not recognized by OCR;

[0009] If the document is a picture, after pre-processing the picture by removing noise, adjusting contrast, and sharpening, analyze the image content and distinguish between text areas and non-text areas; apply OCR to the image of the text part to extract the text in the picture;

[0010] If the document is voice data, the voice content is converted into text using automatic speech recognition technology;

[0011] The speech recognition process is as follows:

[0012] Preprocess the input speech signal, including: noise reduction, echo cancellation and sound enhancement;

[0013] Through MFCC, features are extracted from the preprocessed sound and the sound signal is converted into the form of feature vector;

[0014] Use the deep learning model bidirectional long short-term memory network combined with the attention mechanism to process the MFCC features and output the text sequence;

[0015] c t =Bi-LSTM(MFCC t )

[0016] y t =Attention(c t )

[0017] Where t=1,2,...,T, t represents time, T represents time length; MFCC t is the MFCC feature at time step t; c t is the hidden state of Bi-LSTM; y t is the attention-weighted output, representing the converted text; Attention represents the calculation of the attention mechanism.

[0018] As a preferred solution of the power field office processing method integrating RPA and AI algorithms described in the present invention, wherein: the extracting text from the document also includes, combining NLP and machine learning technology, performing semantic analysis and understanding on the extracted text content;

[0019] Use contextual information to relate text content from different sources;

[0020] For texts from three different sources, a pre-trained word embedding model is used to represent the text as word vectors and vector combinations;

[0021]

[0022]

[0023]

[0024] in, The nth word vector representing the document source; The mth word vector representing the source of the image; The kth word vector representing the source of speech; V d A set of word vectors representing the source of the document; V p A set of word vectors representing the source of the image; V s A collection of word vectors representing the source of speech;

[0025] Through natural language processing technology, the coherent word vectors in the text are combined to obtain the combined word vector:

[0026]

[0027]

[0028] Among them, C d A collection of combined word vectors representing the source of the document; C p A collection of combined word vectors representing the source of the image; C s Represents the set of combined word vectors of speech sources; N represents the number of combined word vectors of document sources; M represents the number of combined word vectors of image sources; J represents the number of combined word vectors of speech sources; The mth combined word vector representing the source of the document is the result of the accumulation of the vectors from the rth word vector to the r+jth word vector in the order of the text content. The next combined word vector is accumulated again starting from the r+n+1th word vector. The mth combined word vector representing the source of the image is the result of the accumulation of the vectors from the uth word vector to the u+oth word vector in the order of the text content. The next combined word vector is accumulated again starting from the u+o+1th word vector. The mth combined word vector representing the source of speech is the result of the accumulation of the vectors from the eth word vector to the e+tth word vector in the order of text content. The next combined word vector is accumulated again starting from the e+t+1th word vector.

[0029] The combined word vector set and the word vector set are combined to obtain a comprehensive vector set;

[0030] Z d =V d ∪C d

[0031] Z p =V p ∪C p

[0032] Z s =Vs ∪C s

[0033] Use cosine similarity to calculate the similarity between word vectors and vector combinations:

[0034]

[0035] Wherein, x represents a vector selected from the comprehensive vector set; y represents any vector selected from the comprehensive vector set that has a different source from x;

[0036] Use the Bayesian algorithm to calculate the probability association between word vectors and other content:

[0037]

[0038] Among them, g represents any vector different from x in the comprehensive vector set; P(g|x) represents the conditional probability of vector g given vector x; P(x|g) represents the conditional probability of vector x given vector g; P(g) represents the prior probability of vector g; P(x) represents the prior probability of vector x;

[0039] Using similarity and probability association as association indicators, calculate the importance of each word vector and vector combination:

[0040]

[0041] Among them, α and β represent weight coefficients respectively; UU=Z d ∪Z p ∪Z s Represents all vectors of the comprehensive vector set from three sources, where x and y have different sources, g≠x; A represents the number of optional vectors for y; B represents the number of optional vectors for g.

[0042] As a preferred solution of the power field office processing method integrating RPA and AI algorithms described in the present invention, the document classification includes: extracting features of the document content through the CNN model; detecting relevant features in the document information through the convolution layer, and capturing information between different sentences by applying the convolution kernel to the input sequence, realizing the detection of different feature patterns in the sequence, and capturing the dynamic changes of the sequence as the time window slides; reducing the dimension of the feature map through the pooling layer to retain important information;

[0043] Use support vector machines to train classification models and use the trained models to classify documents.

[0044] As a preferred solution of the office processing method in the power field integrating RPA and AI algorithms described in the present invention, the document classification using the trained model includes converting the text into a numerical feature vector using Word2Vec technology, performing machine learning using a support vector machine, and performing sentiment analysis or topic classification:

[0045] Emtion=SVM(D)

[0046] Topic = DeepNetwork (D)

[0047] Among them, D represents the feature vector for the entire text; Emtion and Topic are the results of sentiment analysis and topic classification respectively; DeepNetwork represents a deep learning model, which receives the text feature vector D as input and performs the task of topic classification.

[0048] As a preferred solution of the power field office processing method integrating RPA and AI algorithms described in the present invention, wherein: the generation of a summary of the text content includes using the Seq2Seq model in the natural language generation technology for generation; the Seq2sea model includes an encoder and a decoder, specifically:

[0049] Q = Encoder (q)

[0050] Summary(q)=Decoder(Q)

[0051] Among them, Encoder refers to the encoder, which is used to convert the input text into a vector representation of a fixed size; q represents the processing dimension of the encoder; Decoder(c) refers to the decoder, which receives the internal representation c of the encoder as input and generates output; Summary(q) receives the output of the encoder and generates a summary; Q represents the vector of the final state of the encoder;

[0052] The attention mechanism focuses on specific parts of the input document when the decoder generates a summary, and the attention weight a ij Indicates how much the decoder values ​​the i-th word in the input document when generating the j-th word;

[0053] a ij =γ·AttentionWeight(h i ,s j )+δ·I(x)

[0054] Among them, h i is the hidden state of the i-th word in the encoder; s jis the hidden state of the decoder when generating the j-th word; γ and δ are weight coefficients, and their sum is 1;

[0055] The decoder combines the attention weights and the context vector to generate the summary. At each step of generating a word, the model calculates the conditional probability distribution to select the most likely next word; which is expressed as:

[0056] p(w j |w<j,Q,a ij ) = DecoderWordProbability(|w<j,Q,a ij )

[0057] where w j represents the j-th word in the summary; |w<j represents the word sequence before word j; DecoderWordProbability represents the probability model for the decoder to generate the next word, which is used to calculate the likelihood of the next word under a specific context.

[0058] As a preferred solution of the office processing method in the power field that integrates RPA and AI algorithms according to the present invention, wherein: the automatic review includes automatically reviewing the document content using RPA technology, and the steps are as follows:

[0059] Structured information extraction: Using natural language processing technology to extract key information in the document;

[0060] Rule matching: According to the preset review rules, match them one by one with the extracted structured information;

[0061] Keyword check: Verify whether the document contains specific keywords or sensitive words, and whether the quantity and content of the words meet the regulations;

[0062] Time information verification: Check whether the time information in the document meets the preset time rules, including the accuracy and reasonableness of the time;

[0063] Quantity calculation verification: For documents involving amount or quantity calculations, automatically calculate the relevant values and verify their accuracy;

[0064] Feedback review results: According to the results of rule matching, generate a review report indicating whether it meets the standards;

[0065] Subsequent processing: For documents that do not meet the standards, automatically trigger an alarm and return them to the originator for correction;

[0066] Update and maintenance of review rules: Regularly analyze historical contract data to generate an updated draft of the review rules.

[0067] An office processing system in the electric power field that integrates RPA and AI algorithms using the method described in the present invention is characterized by:

[0068] Document receiving and preprocessing module: used for document uploading, automatic classification and content extraction: extracting text content from documents, pictures or voices to prepare for subsequent processing;

[0069] Text analysis and summarization module: used for text analysis and automatic summarization;

[0070] RPA automated review module: used for content review, judgment condition setting, and review result processing;

[0071] Information transmission and approval reminder module: used for approval reminder and information transmission: automatically transmit the reviewed documents to the approval system of the next stage;

[0072] Risk warning and electronic review module: used for risk warning and electronic review;

[0073] Identity verification and electronic signature module: used for signature verification, electronic signature, and signature storage;

[0074] Document approval and feedback module: used for approval process management and feedback to relevant departments;

[0075] Data security and confidentiality module: used for data encryption and access control;

[0076] Monitoring and auditing module: used for system monitoring and auditing logs;

[0077] User interface and interaction module: used for user operation and management.

[0078] A computer device comprises: a memory and a processor; the memory stores a computer program, wherein: the processor implements the steps of any one of the methods of the present invention when executing the computer program.

[0079] A computer-readable storage medium stores a computer program, wherein: when the computer program is executed by a processor, the steps of any one of the methods of the present invention are implemented.

[0080] Beneficial effects of the present invention: The office processing method in the power field that integrates RPA and AI algorithms provided by the present invention significantly reduces the need for manual operations and greatly improves the processing speed of office documents. Standardized review and content extraction using RPA technology and machine learning algorithms reduce human errors and improve the accuracy of document processing. By automating and optimizing office processes, the dependence on human resources is reduced, which helps to reduce the operating costs of enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0082] Figure 1 An overall flow chart of an office processing method in the power field that integrates RPA and AI algorithms provided in the first embodiment of the present invention;

[0083] Figure 2 A schematic diagram of the CNN algorithm principle in the power field office processing method integrating RPA and AI algorithms provided in the first embodiment of the present invention;

[0084] Figure 3 An RPA automatic review flow chart of an office processing method in the power field that integrates RPA and AI algorithms provided in the first embodiment of the present invention;

[0085] Figure 4 A diagram showing the key system steps and module structure of the power field office processing method integrating RPA and AI algorithms provided in the first embodiment of the present invention;

[0086] Figure 5 A system framework diagram of an office processing method in the power field that integrates RPA and AI algorithms, provided in the first embodiment of the present invention. DETAILED DESCRIPTION

[0087] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0088] Example 1, reference Figure 1 , which is an embodiment of the present invention, provides an office processing method in the power field integrating RPA and AI algorithms, including:

[0089] S1: Collect the content of the document, extract the text in the document, and classify the document according to the text content.

[0090] First, extract the feature vector x that represents the content of each document. Here, a convolutional neural network (CNN) is selected. The CNN model contains convolutional layers, pooling layers, and fully connected layers, which helps to extract key features of information content. The convolutional layer is used to detect relevant features in document information. By applying the convolution kernel to the input sequence, it can capture information between different sentences and detect different feature patterns in the sequence. As the time window slides, the CNN model effectively captures the dynamic changes of the sequence. The introduction of the pooling layer helps to reduce the dimension of the feature map and retain the most important information.

[0091] The core of convolutional neural network is convolution operation, such as Figure 2 , which is equivalent to the filter operation in image processing. For a convolution kernel of size m×n, W is an m×n matrix. Each weight w in the convolution kernel W can determine its corresponding weight according to the influence of different features on the final recognition result, and multiply it with the corresponding pixel x in X, and then sum it. The calculation formula is as follows:

[0092]

[0093] Use the extracted feature vectors to train a classification model. Use a support vector machine to train the model:

[0094] y=f(x)=sign(W T x+b)

[0095] Among them, W represents the convolution kernel, an m×n matrix used to extract the features of the input data, and w represents the weight in the convolution kernel. X represents the input data, which can be a local area of ​​the image or a feature map, and the corresponding x represents the pixel. k represents the kth element in the convolution kernel W; θ is the weight vector in the feature space, which is a model parameter used to map the feature space to the decision boundary; b is the bias term, which controls the translation of the classification decision boundary along the feature space; the sign function is used to determine the classification category. x represents the input feature vector, which contains the extracted features for classification. θT actually represents the transpose of the weight vector θ. This transposition allows θ to be dot-producted with the input feature vector x, that is, θTx represents the sum of the corresponding elements of θ and x multiplied, which is a way to calculate the distance between a point in the feature space and the decision boundary.

[0096] For a new document, the system first extracts its feature vector x' and then uses the trained model to classify it. The classification result y' represents the category to which the document is assigned.

[0097] The text extraction in the document includes, if the document is plain text, extracting text from a scanned paper document or PDF file by OCR technology; using semantic reasoning and semantic analysis to complete the character information of the unrecognized part for the characters not recognized by OCR. Intelligent character recognition is an artificial intelligence technology based on OCR with computer deep learning. Semantic reasoning and semantic analysis are used. For characters not recognized by OCR, the character information of the unrecognized part can be completed based on the context sentence information of the character and combined with the knowledge base of the semantic network of EAI. For documents in specific formats, such as tables or reports, more specialized recognition technology is used. Finally, documents in different formats (such as Word, Excel, PPT) are converted into a unified processing format (such as TXT or HTML).

[0098] If the document is a picture, after pre-processing the picture by denoising, adjusting the contrast, and sharpening, the image content is analyzed to distinguish between text areas and non-text areas; OCR is applied to the image of the text part to extract the text in the picture.

[0099] If the document is voice data, automatic speech recognition technology is used to convert the voice content into text.

[0100] The speech recognition process is as follows:

[0101] The speech recognition process first requires preprocessing, that is, preprocessing the input speech signal, including: noise reduction, echo cancellation, and sound enhancement. The goal of this stage is to extract clear and recognizable sound features. The next step is to extract features from the preprocessed sound. The commonly used feature extraction method is Mel Frequency Cepstral Coefficient (MFCC). MFCC converts the sound signal into a series of feature vectors by simulating the auditory characteristics of the human ear. Then, the deep learning model Bidirectional Long Short-Term Memory Network (Bi-LSTM) is used in combination with the attention mechanism to process the MFCC features and output a text sequence. Bi-LSTM is able to take into account the context before and after the speech sequence, while the attention mechanism helps the model focus on the key parts.

[0102] Among them, the deep learning model bidirectional long short-term memory network is combined with the attention mechanism to process the MFCC features, and the output text sequence is expressed as:

[0103] c t =Bi-LSTM(MFCC t )

[0104] y t =Attention(c t )

[0105] Where, t=1,2,...,T; MFCC tis the MFCC feature at time step t; c t is the hidden state of Bi-LSTM. In the speech recognition process mentioned above, c t represents the hidden state of the Bi-LSTM network at time step t. Specifically: c t It is the internal state representation calculated by the forward and backward LSTM layers of the Bi-LSTM network when processing the MFCC feature vector at time step t. This hidden state captures the sequence information up to the current time step t, including the past and future context information of the speech signal, because Bi-LSTM can process bidirectional information flow. t can be considered as a high-dimensional feature vector at time step t, which encodes the dynamic and static characteristics of the speech signal at that time point. ;y t is the attention-weighted output, representing the converted text; Attention represents the calculation process of the attention mechanism, which is a function used to determine the model in generating the output text sequence y t We should pay attention to the input feature c t Specifically: Attention(c t ) is a Bi-LSTM hidden state c based on the current time step t t This function calculates a weight for the hidden state of each time step in the sequence. These weights reflect the impact of each state on the current output y t The weights are calculated through a softmax layer to ensure that the weights of all time steps add up to 1, forming a probability distribution. Finally, these weights are combined with the corresponding hidden state c t Multiply and sum them to get a context vector that integrates the entire sequence information. This vector represents the weighted representation of the current time step. Therefore, y t =Attention(c t ) indicates that at time step t, the model focuses on the hidden state sequence c1, c2, ..., c of the Bi-LSTM through the attention mechanism. T to generate the output text sequence y t This process enables the model to dynamically focus on the features in the input speech signal that are most important for the current prediction, thereby improving the accuracy and robustness of recognition.

[0106] In the process of speech recognition, the role of the attention mechanism is to provide the model with the ability to focus on a specific part of the input sequence, so as to better capture and understand the key information in the speech signal. Specifically, the role of the attention mechanism in the above process can be broken down into the following aspects:

[0107] Focus on key information: In speech recognition, not all acoustic features are equally important. The attention mechanism allows the model to identify and focus on those parts that are most critical to understanding the speech signal, such as specific phonemes or intonation patterns.

[0108] Enhanced model performance: By focusing on the most informative parts of the input features, the model can generate more accurate and reliable outputs. This helps improve recognition accuracy, especially in complex situations such as noisy environments or speaker speed changes.

[0109] Improve long-distance dependency issues: In speech signals, important acoustic features may span a long time range. The attention mechanism can help the model capture these long-distance dependencies, which is crucial for understanding the semantics of the entire word or phrase.

[0110] Improve decoding efficiency: The attention mechanism can make the decoding process more direct and efficient. When generating the output text sequence, the model can use the attention weights to quickly locate the most relevant input parts instead of treating all inputs equally.

[0111] Improves model generalization: By learning to recognize and focus on different features in different speech inputs, the model can better generalize to new, unseen data.

[0112] Providing interpretability: Attention weights can provide an explanation for the model’s decision-making process, helping researchers and developers understand how the model makes a specific prediction, which is very useful for debugging and improving the model.

[0113] Combine NLP and machine learning technologies to perform semantic analysis and understanding on the extracted text content.

[0114] Leverage contextual information to correlate text content from different sources.

[0115] For texts from three different sources, a pre-trained word embedding model is used to represent the text as word vectors and vector combinations.

[0116]

[0117]

[0118]

[0119] in, The nth word vector representing the document source; The mth word vector representing the source of the image; The kth word vector representing the source of speech; V d A set of word vectors representing the source of the document; V pA set of word vectors representing the source of the image; V s A collection of word vectors representing the source of speech;

[0120] Through natural language processing technology, the coherent word vectors in the text are combined to obtain the combined word vector:

[0121]

[0122]

[0123] Among them, C d A collection of combined word vectors representing the source of the document; C p A collection of combined word vectors representing the source of the image; C s Represents the set of combined word vectors from speech sources; N represents the number of combined word vectors from document sources; M represents the number of combined word vectors from image sources; J represents the number of combined word vectors from speech sources; The mth combined word vector representing the source of the document is the result of the accumulation of the vectors from the rth word vector to the r+jth word vector in the order of the text content. The next combined word vector is accumulated again starting from the r+n+1th word vector. The mth combined word vector representing the source of the image is the result of the accumulation of the vectors from the uth word vector to the u+oth word vector in the order of the text content. The next combined word vector is accumulated again starting from the u+o+1th word vector. The mth combined word vector representing the source of the speech is the result of the accumulation of the vectors from the eth word vector to the e+tth word vector in the order of the text content. The next combined word vector is accumulated again starting from the e+t+1th word vector.

[0124] The combined word vector set and the word vector set are combined to obtain a comprehensive vector set.

[0125] Z d =V d ∪C d

[0126] Z p =V p ∪C p

[0127] Z s =V s ∪C s

[0128] Use cosine similarity to calculate the similarity between word vectors and vector combinations:

[0129]

[0130] Among them, x represents a vector selected from the comprehensive vector set; y represents an arbitrary vector selected from the comprehensive vector set that has a different source from x.

[0131] Use the Bayesian algorithm to calculate the probability association between word vectors and other content:

[0132]

[0133] Among them, g represents any vector different from x in the comprehensive vector set; P(g|x) represents the conditional probability of vector g given vector x; P(x|g) represents the conditional probability of vector x given vector g; P(g) represents the prior probability of vector g; P(x) represents the prior probability of vector x;

[0134] Using similarity and probability association as association indicators, calculate the importance of each word vector and vector combination:

[0135]

[0136] Among them, α and β represent weight coefficients respectively; UU=Z d ∪Z p ∪Z s Represents all vectors of the comprehensive vector set from three sources, where x and y have different sources, g≠x; A represents the number of optional vectors for y; B represents the number of optional vectors for g.

[0137] S2: Analyze the extracted text, classify the text topics, and generate a summary of the text content.

[0138] In the first stage of text analysis, the input text is standardized, including word segmentation, stop word removal, and stemming. This step is to simplify the text and make it more suitable for subsequent analysis.

[0139] Next, the text is converted into numerical feature vectors using Word2Vec technology. These feature vectors can capture the semantic information of the text and provide input for subsequent machine learning processing. Finally, support vector machines are used for machine learning to perform sentiment analysis or topic classification. For example, sentiment analysis may involve determining the emotional tendency of the text (positive, negative, or neutral), while topic classification is to classify the text into specific topic categories.

[0140] The CNN model is used to extract features from document content. The convolutional layer is used to detect relevant features in document information. By applying the convolution kernel to the input sequence, the information between different sentences can be captured, and the detection of different feature patterns in the sequence can be achieved. As the time window slides, the dynamic changes of the sequence can be captured. The pooling layer is used to reduce the dimension of the feature map and retain important information.

[0141] Use support vector machines to train classification models and use the trained models to classify documents. Use Word2Vec technology to convert text into numerical feature vectors and use support vector machines for machine learning to perform sentiment analysis or topic classification:

[0142] Emtion=SVM(D)

[0143] Topic = DeepNetwork (D)

[0144] Where D represents the feature vector for the entire text; Emtion and Topic are the results of sentiment analysis and topic classification, respectively. DeepNetwork represents a deep learning model that receives the text feature vector D as input and performs topic classification tasks.

[0145] S3: Use RPA technology to automatically review document content and provide feedback based on the review results.

[0146] The Seq2Seq model in natural language generation technology is used for generation; the Seq2sea model includes an encoder and a decoder, specifically:

[0147] Q = Encoder (q)

[0148] Summary(q)=Decoder(Q)

[0149] Among them, Encoder refers to the encoder, which is used to convert the input text into a vector representation of a fixed size; q represents the processing dimension of the encoder; Decoder(c) refers to the decoder, which receives the internal representation c of the encoder as input and generates output; Summary(q) receives the output of the encoder and generates a summary; Q represents the vector of the final state of the encoder.

[0150] The attention mechanism focuses on specific parts of the input document when the decoder generates a summary, and the attention weight a ij It indicates how much importance the decoder attaches to the i-th word in the input document when generating the j-th word.

[0151] a ij =γ·AttentionWeight(h i ,s j )+δ·I(x)

[0152] Among them, h i is the hidden state of the i-th word in the encoder; s j is the hidden state of the decoder when generating the jth word; γ and δ are weight coefficients, and the sum is 1.

[0153] The decoder combines attention weights and context vectors to generate a summary. When generating a word at each step, the model calculates the conditional probability distribution to select the most likely next word, which is expressed as:

[0154] p(w j |w<j,Q,a ij ) = DecoderWordProbability(|w<j,Q,a ij )

[0155] where w j represents the j-th word in the summary; |w<j represents the word sequence before word j. DecoderWordProbability represents the probability model for the decoder to generate the next word, which is used to calculate the likelihood of the next word in a specific context. It is usually implemented by a neural network model, such as a recurrent neural network (RNN), long short-term memory network (LSTM), or gated recurrent unit (GRU). Since different types of neural networks are involved, DecoderWordProbability is used here to represent this complex process. These models can capture the sequential dependencies in text data and generate text sequences that conform to the language model.

[0156] The steps to automatically review document content using RPA (Robotic Process Automation) technology can be summarized as follows:

[0157] Structured information extraction: Use natural language processing (NLP) technology to extract key information in the document, such as the contract parties, time, amount, etc., and convert it into structured data. As shown specifically Figure 3 below.

[0158] Rule matching: According to the pre-set review rules, such as keyword rules, time rules, quantity calculation formulas, etc., match them one by one with the extracted structured information.

[0159] Keyword check: Verify whether the document contains specific keywords or sensitive words, and whether the quantity and content of these words comply with the regulations.

[0160] Time information verification: Check whether the time information in the document complies with the pre-set time rules, including the accuracy and reasonableness of the time.

[0161] Quantity calculation verification: For documents involving amount or quantity calculations, automatically calculate the relevant values and verify their accuracy.

[0162] Feedback on review results: Based on the results of rule matching, a review report is generated to indicate which parts meet the standards and which parts are suspected of or do not meet the standards.

[0163] Subsequent processing: For documents that do not meet the standards, an alarm can be automatically triggered or they can be returned to the initiator for correction. For documents suspected of non-compliance, they can be submitted to manual for further review.

[0164] Update and maintenance of review rules: Regularly analyze historical contract data, generate updated drafts of review rules, compare existing rules, and extract differences as update suggestions to maintain the timeliness and accuracy of review rules.

[0165] Furthermore, the data is encrypted through data transmission and data encryption technology.

[0166] TCP (Transmission Control Protocol) combined with TLS (Transport Layer Security) is a common network security practice used to provide reliable, encrypted data transmission over the Internet.

[0167] TCP connection establishment: First, the client and server establish a reliable connection through the TCP three-way handshake. This process includes the exchange of SYN, SYN-ACK, and ACK messages.

[0168] TLS handshake process: After the TCP connection is established, the client and server start the TLS handshake process to negotiate encryption parameters and verify each other's identity. This process includes exchanging "Hello" messages, certificates, and keys.

[0169] Certificate verification: The client verifies the server's certificate chain to ensure the server's identity and check the validity of the certificate (for example, whether the certificate has expired and whether it was issued by a trusted certificate authority).

[0170] Key exchange: Using a key exchange algorithm such as Diffie-Hellman, the client and server negotiate a session key to encrypt and decrypt data.

[0171] Encrypted communication: Once the TLS handshake is complete, the client and server use the negotiated keys and cryptographic parameters (such as cipher suites) to encrypt and decrypt the data being sent.

[0172] Data transmission: Encrypted data is transmitted over a TCP connection. Since TCP is a stream protocol, the TLS layer usually uses the record layer to split the data into appropriate chunks and add a TLS header and encrypted payload data to each chunk.

[0173] Data integrity and authentication: TLS also provides message integrity checking and authentication, using MAC (Message Authentication Code) or digital signatures to ensure that data has not been tampered with during transmission.

[0174] Session End: When the communication ends, one or both parties can send a "close_notify" alert to securely close the TLS session, and the TCP connection will be closed.

[0175] Signature Verification Technology:

[0176] Signature verification is an identity verification technique that confirms a person's identity by analyzing their handwritten signature. Here are some of the key steps and techniques for signature verification:

[0177] Signature collection:

[0178] Digital signature collection: When the approver agrees with the content of the document, he needs to confirm the document and collect the signature directly on the electronic device through a touch screen or digital pen.

[0179] Image preprocessing:

[0180] Denoising: Remove noise from the image to improve the clarity of the signature.

[0181] Binarization: Convert the image into black and white to facilitate feature extraction.

[0182] Normalization: Adjust the image size to make it suitable for processing.

[0183] Feature extraction:

[0184] Geometric features: Extract the geometry of the signature, such as stroke length, angle, and direction.

[0185] Dynamic features: For online signatures, extract time series data during the writing process, such as pen pressure, speed, and acceleration.

[0186] Signature comparison:

[0187] Template matching: Compare the collected signature with a pre-stored signature template.

[0188] Similarity calculation: The difference between signatures is calculated using the similarity metric - Euclidean distance.

[0189] The Euclidean metric, also known as the Euclidean distance, is a commonly used distance definition in the field of machine learning. It refers to the real distance between two points in n-dimensional space, or the natural length of a vector (i.e. the distance from the point to the origin). The Euclidean distance in two-dimensional and three-dimensional space is the actual distance between two points. The formula for the Euclidean distance in n-dimensional space is as follows:

[0190]

[0191] For the signature verification problem, suppose the two feature vectors obtained by the feature extraction subnet for the input signature are X R =F W，b (S R ) reference signature (i.e. control group for comparison) and X T =F W，b (S T )Test signature.

[0192] Where X R =[x1 (R) ,x2 (R) …,x n (R) ] T ,X T =[x1 (T) ,x2 (T) …,x n (T) ] T , where n is the dimension of the feature vector and n∈R. At this time, X R With X T The Euclidean distance between them is as follows:

[0193]

[0194] If there are two points P1 and P2, their coordinates in n-dimensional space are (x1, x2, ..., x n ) and (y1,y2,...,y n ), for example, in 2D, n is 2, and in 3D, it is 3.

[0195] Since the closer the features between two signatures are, the smaller the Euclidean distance is, which means the greater the similarity between the two, the feature measurement of the two signatures needs to be the inverse of the Euclidean distance between the two, that is:

[0196] D(X R ,X T )=-d(X R ,X T )

[0197] Threshold judgment: judge whether the signatures are consistent based on the similarity calculation result and the preset threshold. In signature verification, a smaller Euclidean distance means that X R With X T If the distance is relatively similar, they may belong to the same person; a larger distance means that they are very different and may belong to different people. The signature verification system will set a threshold, and when the Euclidean distance is less than this threshold, the signature is considered valid.

[0198] On the other hand, this embodiment also provides an office processing system in the power field that integrates RPA and AI algorithms, such as Figure 4 As shown, it includes:

[0199] 1. Document reception and preprocessing module

[0200] Document upload: Users can upload documents, pictures or voice content through the system interface.

[0201] Automatic classification: Use machine learning algorithms to automatically classify and label uploaded documents.

[0202] Content extraction: Extract text content from documents, images, or voice for subsequent processing.

[0203] 2. Text Analysis and Summarization Module

[0204] Text analysis: Perform semantic analysis on the extracted text to identify key information and topics.

[0205] Automatic Summarization: Generates an automatic summary of the document content, providing a quick overview to facilitate initial understanding of the document content.

[0206] 3.RPA automated review module

[0207] Content review: Use RPA technology to automatically review document content and check whether it complies with preset standards and rules.

[0208] Judgment condition setting: Set specific review conditions, such as sensitive words, format specifications, etc.

[0209] Review result processing:

[0210] Reject request: If the document fails the review, the system automatically rejects the user's request and provides detailed reasons.

[0211] Pass review: If the document passes the review, the system sends the document to the next stage of approval.

[0212] 4. Information transmission and approval reminder module

[0213] Approval reminder: The system sends approval reminders to approvers through the internal network or mobile devices to ensure that approvers handle them in a timely manner.

[0214] Information transfer: Documents that have passed review will be automatically transferred to the approval system for the next stage.

[0215] 5. Risk Warning and Electronic Review Module

[0216] Risk warning: When the approver is approving, the system provides risk point prompts to assist them in making judgments.

[0217] Electronic Review: Approver can review electronically through the system interface, adding comments and opinions.

[0218] 6. Identity verification and electronic signature module

[0219] Signature verification: Use handwriting recognition technology to verify whether the electronic signature is signed by the approver himself.

[0220] Electronic signature: After confirming the identity of the approver, the system will manage the electronic signature to ensure the legality and validity of the document.

[0221] Signature storage: Store the signed documents on a secure server for easy subsequent review and management.

[0222] 7. Document approval and feedback module

[0223] Approval process management: Manage the entire approval process to ensure that documents pass through each approval link in order.

[0224] Feedback to relevant departments: After approval, the system will automatically feedback the document to relevant departments and users.

[0225] 8.Data security and confidentiality module

[0226] Data encryption: Encrypt transmitted and stored data to ensure data security.

[0227] Access Control: Implement strict access control mechanisms to ensure that only authorized users can access sensitive data.

[0228] 9. Monitoring and auditing module

[0229] System monitoring: Monitor the performance and status of the system to detect and resolve problems in a timely manner.

[0230] Audit log: A detailed log of all operations is recorded to facilitate post-audit and analysis.

[0231] 10. User interface and interaction module

[0232] User interface: Provides an intuitive and easy-to-use user interface for easy operation and management.

[0233] Interaction design: optimize user interaction process and improve user experience.

[0234] The system framework is built, such as Figure 5 As shown, including:

[0235] User Interface (Frontend): The front end uses the Vue.js framework combined with the Element UI component library to build a responsive and user interactive interface.

[0236] API Gateway: As an intermediary between front-end applications and back-end services, it provides a RESTful API for the front-end to send requests to the back-end and receive data.

[0237] Backend service:

[0238] SpringBoot: As the basic framework of the entire system, it provides functions such as automatic configuration, embedded server and health check.

[0239] Controllers & Services: Process requests from API Gateway and provide corresponding service logic.

[0240] Data Access Layer:

[0241] MyBatis Plus: As an ORM framework, it simplifies database operations and provides a database operation interface.

[0242] MySQL: A relational database that stores system data and works with MyBatis Plus.

[0243] CachingLayer:

[0244] Redis Cache: As a distributed cache system, it stores hot data and session information to reduce the load on the database and improve system performance.

[0245] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0246] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0247] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0248] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0249] Example 2 is an embodiment of the present invention, which provides an office processing method in the power field that integrates RPA and AI algorithms. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0250] Table 1 Comparison of system implementation effects

[0251]

[0252] Based on the data provided in the above table, we can make the following analysis:

[0253] Document processing time has been significantly reduced: Before the implementation of the system, it took an average of 5 days to process a document. After the implementation of the system, this time was reduced to 2 hours, an efficiency improvement of 96%. This shows that the automated process has greatly accelerated the processing of documents.

[0254] The error rate has been significantly reduced: Before the system was implemented, the error rate was 15 per thousand documents. After the system was implemented, the error rate was reduced to 2, a reduction of 87%. This is attributed to the reduction of human errors due to automation.

[0255] The approval efficiency has been greatly improved: before the implementation of the system, one person approved an average of 20 documents per day. After the implementation of the system, this number increased to 100, an increase of 400%, indicating that the automated approval process has significantly improved the work efficiency of the approvers.

[0256] Significant savings in labor costs: It is estimated that after the implementation of the system, labor costs can be saved by 1 million yuan per year. This shows that the automation system reduces the dependence on document processing personnel and reduces labor costs.

[0257] Security has been enhanced: the number of security incidents has dropped from 5 per quarter before implementation to 0, a 100% reduction, indicating that the system's security measures have effectively reduced the risk of data leakage and violations

[0258] Improved user satisfaction: User satisfaction increased from 3.5 to 4.7 (out of 5), an increase of 34%, reflecting users' acceptance and satisfaction with the new system.

[0259] Faster decision-making: The time for decision-making has been reduced from 3 days to 1 day, an improvement of 67%, thanks to the risk warning and rapid report generation functions provided by the system.

[0260] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An office processing method in the power field integrating RPA and AI algorithms, characterized in that: include: Collect the content of documents, extract the text in the documents, and classify the documents according to the text content; Analyze the extracted text, classify the text topics, and generate a summary of the text content; Use RPA technology to automatically review document content and provide feedback based on the review results.

2. The power field office processing method integrating RPA and AI algorithms as claimed in claim 1 is characterized by: The extracting of text from the document includes, if the document is plain text, extracting text from a scanned paper document or PDF file using OCR technology; using semantic reasoning and semantic analysis to complete the character information of the unrecognized part of the characters not recognized by OCR; If the document is a picture, after pre-processing the picture by removing noise, adjusting contrast, and sharpening, analyze the image content and distinguish between text areas and non-text areas; apply OCR to the image of the text part to extract the text in the picture; If the document is voice data, the voice content is converted into text using automatic speech recognition technology; The speech recognition process is as follows: Preprocess the input speech signal, including: noise reduction, echo cancellation and sound enhancement; Through MFCC, features are extracted from the preprocessed sound and the sound signal is converted into the form of feature vector; Use the deep learning model bidirectional long short-term memory network combined with the attention mechanism to process the MFCC features and output the text sequence; c t =Bi-LSTM(MFCC t ) y t =Attention(c t ) Where t=1,2,...,T, t represents time, T represents time length; MFCC t is the MFCC feature at time step t; c t is the hidden state of Bi-LSTM; y t is the attention-weighted output, representing the converted text; Attention represents the calculation of the attention mechanism.

3. The power field office processing method integrating RPA and AI algorithms as claimed in claim 2 is characterized by: Extracting text from the document also includes, combining NLP and machine learning techniques, performing semantic analysis and understanding on the extracted text content; Use contextual information to relate text content from different sources; For texts from three different sources, a pre-trained word embedding model is used to represent the text as word vectors and vector combinations; in, The nth word vector representing the document source; The mth word vector representing the source of the image; The kth word vector representing the source of speech; V d A set of word vectors representing the source of the document; V p A set of word vectors representing the source of the image; V s A collection of word vectors representing the source of speech; Through natural language processing technology, the coherent word vectors in the text are combined to obtain the combined word vector: Among them, C d A collection of combined word vectors representing the source of the document; C p A collection of combined word vectors representing the source of the image; C s Represents the set of combined word vectors of speech sources; N represents the number of combined word vectors of document sources; M represents the number of combined word vectors of image sources; J represents the number of combined word vectors of speech sources; The mth combined word vector representing the source of the document is the result of the accumulation of the vectors from the rth word vector to the r+jth word vector in the order of the text content. The next combined word vector is accumulated again starting from the r+n+1th word vector. The mth combined word vector representing the source of the image is the result of the accumulation of the vectors from the uth word vector to the u+oth word vector in the order of the text content. The next combined word vector is accumulated again starting from the u+o+1th word vector. The mth combined word vector representing the source of speech is the result of the accumulation of the vectors from the eth word vector to the e+tth word vector in the order of text content. The next combined word vector is accumulated again starting from the e+t+1th word vector. The combined word vector set and the word vector set are combined to obtain a comprehensive vector set; With d =V d ∪C d With p =V p ∪C p With s =V s ∪C s Use cosine similarity to calculate the similarity between word vectors and vector combinations: Wherein, x represents a vector selected from the comprehensive vector set; y represents any vector selected from the comprehensive vector set that has a different source from x; Use the Bayesian algorithm to calculate the probability association between word vectors and other content: Where g represents any vector different from x in the comprehensive vector set; P(g|x) represents the conditional probability of vector g given vector x; P(x|g) represents the conditional probability of vector x given vector g; P(g) represents the prior probability of vector g; P(x) represents the prior probability of vector x; Using similarity and probability association as association indicators, calculate the importance of each word vector and vector combination: Among them, α and β represent weight coefficients respectively; UU=Z d ∪Z p ∪Z s Represents all vectors of the comprehensive vector set from three sources, where x and y have different sources, g≠x; A represents the number of optional vectors for y; B represents the number of optional vectors for g.

4. The power field office processing method integrating RPA and AI algorithms as claimed in claim 3 is characterized by: The document classification includes extracting features from the document content through the CNN model; detecting relevant features in the document information through the convolution layer, and capturing information between different sentences by applying the convolution kernel to the input sequence, realizing the detection of different feature patterns in the sequence, and capturing the dynamic changes of the sequence as the time window slides; reducing the dimension of the feature map through the pooling layer to retain important information; Use support vector machines to train classification models and use the trained models to classify documents.

5. The power field office processing method integrating RPA and AI algorithms as claimed in claim 4 is characterized by: The document classification using the trained model includes converting the text into a numerical feature vector using the Word2Vec technology, performing machine learning using a support vector machine, and performing sentiment analysis or topic classification: Emtion=SVM(D) Topic = DeepNetwork (D) Among them, D represents the feature vector for the entire text; Emtion and Topic are the results of sentiment analysis and topic classification respectively; DeepNetwork represents a deep learning model, which receives the text feature vector D as input and performs the task of topic classification.

6. The power field office processing method integrating RPA and AI algorithms as claimed in claim 5 is characterized by: The generating summary of the text content includes using the Seq2Seq model in the natural language generation technology to generate the summary; the Seq2sea model includes an encoder and a decoder, specifically: Q = Encoder (q) Summary(q)=Decoder(Q) Among them, Encoder refers to the encoder, which is used to convert the input text into a vector representation of a fixed size; q represents the processing dimension of the encoder; Decoder(c) refers to the decoder, which receives the internal representation c of the encoder as input and generates output; Summary(q) receives the output of the encoder and generates a summary; Q represents the vector of the final state of the encoder; The attention mechanism focuses on specific parts of the input document when the decoder generates a summary, and the attention weight a ij Indicates how much the decoder values ​​the i-th word in the input document when generating the j-th word; a ij =γ·AttentionWeight(h i ,s j )+δ·I(x) Among them, h i is the hidden state of the i-th word in the encoder; s j is the hidden state of the decoder when generating the jth word; γ and δ are weight coefficients, and the sum is 1; The decoder combines the attention weights and the context vector to generate a summary. When generating words at each step, the model calculates the conditional probability distribution to select the most likely next word; expressed as: p(w j |w<j,Q,a ij )=DecoderWordProbability(|w<j,Q,a ij ) where, w j represents the j-th word in the abstract; |w<j represents the word sequence before word j; DecoderWordProbability represents the probability model for the decoder to generate the next word, which is used to calculate the likelihood of the next word in a specific context.

7. The power field office processing method integrating RPA and AI algorithms as claimed in claim 6 is characterized by: The automatic review includes automatically reviewing the document content using RPA technology, the steps are as follows: Structured information extraction: Use natural language processing technology to extract key information from documents; Rule matching: Match the extracted structured information one by one according to the pre-set review rules; Keyword check: Verify whether the document contains specific keywords or sensitive words, and whether the number and content of words comply with regulations; Time information verification: Check whether the time information in the document complies with the preset time rules, including the accuracy and rationality of the time; Quantity calculation verification: For documents involving amount or quantity calculation, automatically calculate the relevant values ​​and verify their accuracy; Feedback on review results: Generate a review report based on the results of rule matching, indicating whether it meets the standards; Subsequent processing: For documents that do not meet the standards, an alert is automatically triggered and they are returned to the initiator for correction; Update and maintenance of review rules: Regularly analyze historical contract data and generate updated drafts of review rules.

8. An office processing system in the electric power field that integrates RPA and AI algorithms using the method described in any one of claims 1 to 7, characterized in that: Document receiving and preprocessing module: used for document uploading, automatic classification and content extraction: extracting text content from documents, pictures or voices to prepare for subsequent processing; Text analysis and summarization module: used for text analysis and automatic summarization; RPA automated review module: used for content review, judgment condition setting, and review result processing; Information transmission and approval reminder module: used for approval reminder and information transmission: automatically transmit the reviewed documents to the approval system of the next stage; Risk warning and electronic review module: used for risk warning and electronic review; Identity verification and electronic signature module: used for signature verification, electronic signature, and signature storage; Document approval and feedback module: used for approval process management and feedback to relevant departments; Data security and confidentiality module: used for data encryption and access control; Monitoring and auditing module: used for system monitoring and auditing logs; User interface and interaction module: used for user operation and management.

9. A computer device comprising: A memory and a processor; the memory stores a computer program, characterized in that: when the processor executes the computer program, it implements the steps of an office processing method in the power field that integrates RPA and AI algorithms.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of an office processing method in the power field that integrates RPA and AI algorithms are implemented.

Images