An intelligent operation and maintenance system based on AI big model and RPA

Abstract

The present invention relates to an intelligent operation and maintenance system based on an AI big model and RPA, which relates to the field of data processing. The system receives user input information in response to an RPA robot, performs semantic matching through an AI semantic model, and obtains consulting questions; processes the consulting questions through an answer material library, and matches a number of historical answers; performs centralized weight distribution on the several historical answers to obtain a number of centralized weights; processes the consulting questions through a question and answer log, and performs answer frequency weight distribution on the several historical answers to obtain a number of frequency weights; optimizes the several historical answers according to the several centralized weights and the several frequency weights, obtains a target answer, and feeds back the target answer to the RPA robot to reply to the user, thereby solving the technical problems that the intelligent operation and maintenance system is difficult to effectively handle complex and non-standardized user questions, the answer matching algorithm has low adaptability, and the multimedia information processing capability is limited.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to an intelligent operation and maintenance system based on AI big models and RPA. Background Art

[0002] With the rapid development of information technology, the complexity and workload of enterprise operations and maintenance are increasing, and the requirements for operation and maintenance efficiency and quality are becoming increasingly stringent. Traditional operation and maintenance methods mainly rely on manual operations and empirical judgment, which is not only inefficient but also prone to errors. To improve operation and maintenance efficiency and accuracy, intelligent operation and maintenance systems have emerged. However, most existing intelligent operation and maintenance systems use intelligent question-and-answer methods based on rules or templates, which can only handle simple, standardized user questions. For complex and non-standardized user questions, existing systems are often unable to provide accurate and satisfactory answers, and often require users to ask repeatedly to meet their needs. This results in low operation and maintenance efficiency and may even require manual intervention, which undoubtedly reduces the system's practicality and user experience. Summary of the Invention

[0003] The present invention addresses the technical problems in the existing technology that intelligent operation and maintenance systems are difficult to effectively handle complex and non-standardized user questions, the answer matching algorithm has low adaptability, and the multimedia information processing capabilities are limited. An intelligent operation and maintenance system based on AI big models and RPA is provided to solve these problems.

[0004] The technical solution of the present invention to solve the above technical problems is as follows:

[0005] The present invention provides an intelligent operation and maintenance system based on an AI big model and RPA, and the execution steps include: in response to the RPA robot receiving user input information, performing semantic matching through the AI ​​semantic model to obtain consulting questions; processing the consulting questions through the answer material library to match a number of historical answers; performing centralized weight distribution on the several historical answers to obtain a number of centralized weights; processing the consulting questions through the question and answer log, performing answer frequency weight distribution on the several historical answers to obtain a number of frequency weights; optimizing the several historical answers according to the several centralized weights and the several frequency weights, obtaining target answers and feeding back to the RPA robot to reply to the user.

[0006] The beneficial effect of the present invention is that by combining the AI ​​big model with RPA technology, an efficient intelligent operation and maintenance system is realized. The system can accurately understand user questions and intelligently select the best answers from a large number of historical answers for reply, significantly improving operation and maintenance efficiency and user experience, while enhancing the system's multimedia information processing capabilities and the accuracy of answer matching. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1This is a flowchart of the execution steps of an intelligent operation and maintenance system based on AI big model and RPA provided by the present invention.

[0008] Figure 2 This is a flowchart of the execution steps for semantic matching in an intelligent operation and maintenance system based on an AI big model and RPA provided by the present invention. DETAILED DESCRIPTION

[0009] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0010] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the specified features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0011] In the description of the present invention, the term "for example" is used to mean "used as an example, illustration or illustration". Any embodiment of the present invention described as "for example" is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is given to enable any person skilled in the art to implement and use the present invention. In the following description, details are listed for the purpose of explanation. It should be understood that a person of ordinary skill in the art can recognize that the present invention can be implemented without using these specific details. In other examples, well-known structures and processes are not elaborated in detail to avoid obscuring the description of the present invention with unnecessary details. Therefore, the present invention is not intended to be limited to the embodiments shown, but is consistent with the widest scope consistent with the principles and features disclosed herein.

[0012] Example:

[0013] like Figure 1 As shown, an embodiment of the present invention provides an intelligent operation and maintenance system based on an AI big model and RPA, and the execution steps include:

[0014] S10: In response to the RPA robot receiving user input information, semantic matching is performed through the AI ​​semantic model to obtain the consulting question.

[0015] S20: Process the consultation question through the answer material library and match several historical answers.

[0016] S30: Perform centralized weight distribution on the plurality of historical answers to obtain a plurality of centralized weights.

[0017] S40: Processing the consulting question through the question-and-answer log, performing answer frequency weight distribution on the several historical answers, and obtaining several frequency weights.

[0018] S50: Optimize the historical answers according to the multiple concentration weights and the multiple frequency weights, obtain a target answer, and feed it back to the RPA robot to reply to the user.

[0019] For example, intelligent operations and maintenance ("AOM") refers to a series of steps and procedures followed in actual O&M work, typically encompassing multiple stages, including problem reception, problem analysis, problem resolution, and result feedback. In intelligent O&M systems, these processes can be optimized and accelerated through automation and intelligentization. AI big models refer to mathematical models constructed using artificial intelligence technologies, such as machine learning and deep learning. In intelligent O&M systems, AI big models can process and analyze large amounts of O&M data, extracting useful information and patterns to support O&M decision-making and problem-solving. AI big models here specifically refer to models with powerful processing capabilities and extensive knowledge coverage, such as large language models. Robotic Process Automation (RPA) is a technology that uses software robots to simulate human computer operations to automate business processes. In intelligent O&M systems, RPA can perform repetitive, standardized O&M tasks, such as data entry and report generation, thereby reducing the workload of O&M personnel and improving O&M efficiency. RPA robots can be deployed on either the server or client side, depending on the system architecture and O&M requirements. They automate operational tasks by simulating user actions (such as clicks and inputs) without requiring human intervention. The combination of AI models and RPA can further enhance the intelligence level of the system and improve operational efficiency and accuracy.

[0020] Specifically, when an RPA robot is designed for use in an intelligent operation and maintenance system, it can perform a series of automated tasks to respond to and process user input. Within an intelligent operation and maintenance system, the RPA robot is configured to receive user input, which can come from various channels, such as questions submitted through the user interface, work orders imported from other systems, or consultation requests sent via email, chat applications, and other means. Upon receiving this information, the RPA robot immediately initiates a pre-defined process to process it. During this process, the RPA robot passes the user input to an AI semantic model for further processing. An AI semantic model, built using artificial intelligence technology, deeply understands and analyzes textual information to extract key information and intent. The AI ​​semantic model is specifically trained to understand and match user inquiries. When the RPA robot passes user input to the AI ​​semantic model, the model analyzes the information and attempts to match it with entries in a known question library or knowledge base. This process requires a deep understanding of the text's semantics, including analysis of vocabulary, grammar, and context. Through matching with the AI ​​semantic model, the system can accurately identify the specific inquiry being addressed by the user. Once a problem is identified, the intelligent operation and maintenance system can automatically provide users with solutions, answers, or guide them to the next step based on pre-set rules and processes. In this way, the combination of RPA robots and AI semantic models not only improves the automation level of the operation and maintenance system, but also significantly enhances the efficiency and accuracy of user consultation processing.

[0021] The RPA robot then compares the identified inquiry question with its answer library. This library is a vast database containing a variety of past user inquiries and their corresponding answers or solutions. These historical answers undergo rigorous review and organization to ensure their accuracy and effectiveness. During this comparison process, the RPA robot leverages advanced algorithms and technologies, such as natural language processing (NLP) and machine learning, to identify the historical answer that best matches the current inquiry question. This matching is based not only on superficial text similarity but, more importantly, on understanding the intent and context behind the question to find the answer that best meets the user's needs. Through this comparison and matching process, the RPA robot may identify several historical answers relevant to the inquiry question. These answers may vary in content, format, and level of detail, but they all address the core question of the user's inquiry.

[0022] Next, to more accurately select the best answer or provide comprehensive recommendations, the system applies a centralized weight distribution to these historical answers. First, the system performs a preliminary evaluation of each matched historical answer, typically based on multiple dimensions such as accuracy, completeness, currency, and user satisfaction. These dimensions are pre-defined and are intended to comprehensively measure the quality and applicability of answers. Next, based on these evaluation dimensions, each historical answer is assigned one or more weights. These weights reflect the answer's performance across each dimension and are typically expressed numerically, with higher values ​​representing better performance in that dimension. In assigning weights, the system may employ various algorithms and techniques, such as weighted averaging and fuzzy comprehensive evaluation, to ensure objective and appropriate weighting. These algorithms and techniques comprehensively consider information from multiple dimensions, avoiding the biased evaluation of answers based on a single metric. Through this centralized weight distribution process, the system ultimately obtains several centralized weights, one for each historical answer. These weights not only reflect the answer's overall performance across various dimensions but also provide important information for subsequent answer selection or comprehensive recommendations. In short, by comprehensively evaluating multiple dimensions of the answers, the system is provided with a more comprehensive and objective basis for selection.

[0023] Optionally, to further optimize answer selection strategies and enhance user experience, the system uses Q&A logs to conduct in-depth analysis of consultation questions and their historical answers. Q&A logs are a crucial data source for recording user consultation history, answer provision, and user feedback. They contain information such as questions asked, answers provided, user acceptance of these answers (e.g., clicks, likes, comments, and other interactive behaviors), and whether the answers ultimately resolved the user's questions. After the RPA robot matches several historical answers related to the current consultation question through the answer library, the system further searches the Q&A logs to analyze how often these answers have been used to answer similar questions. This frequency reflects the popularity, practicality, and effectiveness of the answers and is a key indicator of answer quality. To quantify this frequency, the system collects and analyzes data from the Q&A logs, calculates the number of times each historical answer has been used to answer similar questions, and converts this into a frequency weight. A higher frequency weight indicates that the answer has been used more frequently to solve similar questions in the past, and is therefore likely to be more practical and accurate. When calculating the frequency weight, the system may also consider other factors, such as the time since the answer was provided and the level of positive user feedback, to further refine the weighting. These factors help the system more comprehensively evaluate the quality of answers, avoiding misleading responses due to frequently used but ineffective answers. Ultimately, the system assigns a frequency weight to each historical answer based on the analysis of the Q&A log. These weights provide an important reference for subsequent answer selection, allowing the system to prioritize answers that have proven effective and popular in the past. This approach not only improves the accuracy of answer selection and user satisfaction, but also continuously optimizes the answer library and recommendation strategies based on user feedback and behavioral habits, providing users with a more personalized and intelligent service experience.

[0024] Finally, the system has matched several historical answers related to the inquiry using the answer library and calculated their concentration weights and frequency weights. The concentration weight reflects the answer's comprehensive performance across multiple evaluation dimensions, while the frequency weight reflects the frequency and popularity of the answer in solving similar problems in the past. Next, the system considers these two weights to optimize the historical answers. This process involves complex algorithms and logic, aiming to balance the quality, practicality, and historical performance of the answers to find the answer that best meets the current inquiry. During the optimization process, the system combines the concentration weights and frequency weights by weighted summation or other methods to generate a comprehensive score. This comprehensive score reflects the comprehensive performance of each historical answer across multiple dimensions and serves as an important basis for the system's target answer selection. The highest-scoring answer from the historical answers is then selected as the target answer based on the comprehensive score. This target answer not only provides high-quality content but also undergoes rigorous system screening and optimization to best meet user needs. Once the target answer is determined, the RPA robot immediately provides feedback to the user. This process may occur via text message, email, chat, or other communication channels, depending on the user's preferences and system configuration. After receiving the answer, users can perform further operations or ask new questions based on their needs. In this way, the system can fully utilize the information of concentration weight and frequency weight to accurately optimize historical answers, thereby providing users with the highest quality and most personalized service experience, improving question processing efficiency and matching.

[0025] In a preferred embodiment, Figure 2 As shown, in response to the RPA robot receiving user input information, semantic matching is performed through the AI ​​semantic model to obtain consulting questions, including: constructing a question grammatical structure, wherein the question grammatical structure includes a subject and an execution action; configuring a user input label and a question identifier that conforms to the question grammatical structure, wherein the user input label is text information including the subject and the execution action; based on the user input label and the question identifier, supervised training of N front-end AI semantic models through N machine learning models with different topological structures; based on the outputs of the N front-end AI semantic models and the question identifier, supervised training of the post-end AI semantic model through the LSTM-DSSM neural network; and fusing the N output branches of the N front-end AI semantic models with the input layer of the post-end AI semantic model to obtain the AI ​​semantic model.

[0026] For example, an RPA robot first receives user input, typically a text or voice message. It then uses an AI semantic model to perform semantic matching on the input to extract the core elements of the consultation question. To accurately understand the user's question, the system first constructs the grammatical structure of the question, identifying the subject (typically the user or the role they represent) and the action (the action the user wants to perform or the information they request). Next, based on the identified subject and action, the system assigns user input labels. These labels are textual information containing the subject and action, and are used for subsequent model training and prediction. The system also generates a unique question identifier for each question, which is used to distinguish different question instances during training. To capture the diversity and complexity of questions, the system uses N machine learning models with different topologies as the front-end AI semantic models. These models can be based on different algorithms, such as decision trees, support vector machines, and random forests. The system then uses the user input labels and question identifiers to perform supervised training on these front-end models. During training, the models learn how to predict the semantic content of the question based on the input labels and question identifiers. After obtaining the output of the pre-trained AI semantic model, the system uses the LSTM-DSSM (Long Short-Term Memory-Deep Structured Semantic Model) neural network as the post-trained AI semantic model. The LSTM-DSSM model can capture long-term dependencies in sequential data and is particularly effective for understanding complex semantic structures. The post-trained model undergoes supervised training based on the pre-trained model's output (i.e., N output branches) and the question identifier. This step aims to further refine and integrate the pre-trained model's predictions to improve the accuracy of semantic understanding. This training process is known as the stacking training method, which integrates the output branches with the post-trained model's input layer. The output branches of the N pre-trained AI semantic models serve as part of the post-trained AI semantic model's input layer. This fusion approach allows the post-trained model to learn richer information from the predictions of multiple pre-trained models, thereby improving its ability to discern complex semantics. This system integrates the strengths of different models, mitigates the limitations of a single model, and achieves a more accurate understanding of complex semantic questions. The training and optimization steps described above ultimately yield a highly optimized AI semantic model that accurately understands user inquiries and provides precise semantic matching results for the RPA robot. Through this process, the RPA robot leverages the AI ​​semantic model to precisely understand user input. In particular, the stacking training method integrates multiple front-end models with different topologies into a powerful back-end model, significantly improving its ability to identify complex semantic questions.

[0027] In a preferred embodiment, in response to the RPA robot receiving user input information, semantic matching is performed through the AI ​​semantic model to obtain consulting questions, including: when the user input information is text, sending it to the AI ​​semantic model; when the user input information is a picture, activating the OCR text recognizer, processing the picture, obtaining the text recognition result and sending it to the AI ​​semantic model.

[0028] Specifically, the RPA robot first determines the type of user input. This information can be in text form, such as text entered through the chat interface, or in image form, such as screenshots, photos, or document scans uploaded by the user. The robot has an internal mechanism to identify the type of user input. If the input is text, it proceeds directly to the next step; if the input is an image, a specific processing flow is triggered. If the user enters text, the RPA robot directly sends the text to the AI ​​semantic model. The AI ​​semantic model uses advanced natural language processing (NLP) and deep learning algorithms to parse and understand the text, extracting the specific question the user wants to consult. For image input, the RPA robot first activates the Optical Character Recognition (OCR) text recognizer. OCR technology recognizes text content in images and converts it into editable text. The robot uses the OCR text recognition results as input and sends them to the AI ​​semantic model for further semantic matching. Whether the input is text or converted from an image, the AI ​​semantic model conducts in-depth analysis and understanding. The model leverages extensive training data and sophisticated algorithms to identify the core elements of user inquiries, such as the topic, keywords, and intent. After processing using the AI ​​semantic model, the RPA robot accurately extracts the user's specific inquiry and presents it in a structured format, facilitating subsequent querying, analysis, or responses. Through this process, the RPA robot can flexibly process user-entered text or image information and accurately match it with the AI ​​semantic model. This capability not only enhances the robot's understanding of user needs but also significantly improves the efficiency and accuracy of automated services. Whether handling simple text inquiries or complex image information, the RPA robot can flexibly and efficiently and intelligently provide satisfactory responses to users.

[0029] In a preferred embodiment, the several historical answers are subjected to centralized weight distribution to obtain several centralized weights, including: performing semantic similarity analysis on the several historical answers to obtain a semantic similarity set; traversing the several historical answers to perform distribution density analysis based on the semantic similarity set to obtain several distribution densities; calculating the ratio of the several distribution densities to the distribution sum density to obtain the several centralized weights.

[0030] Optionally, to more effectively utilize historical answers to respond to user inquiries, the system analyzes the centralized weight distribution of these answers. First, several historical answers relevant to the current inquiry are collected. Next, semantic analysis techniques (such as word embedding and sentence embedding in natural language processing (NLP)) are used to perform semantic similarity analysis on these answers. Semantic similarity analysis aims to measure the semantic proximity between each historical answer and the inquiry question, as well as the similarity between each historical answer. The analysis results in a semantic similarity set, which contains the similarity score of each historical answer to the inquiry question or other historical answers. Next, the system iterates through these historical answers and performs distribution density analysis. This analysis assesses the distribution density of each historical answer in the semantic space—that is, the distribution of its similarity scores with other answers. If a historical answer has high similarity with multiple other answers, its distribution density in the semantic space will be relatively high. Through this step, the system calculates a distribution density value for each historical answer. After obtaining several distribution densities, the system calculates the ratio of these distribution densities to the sum of the distribution densities to determine the centralized weight. The summed distribution density refers to the sum of the distribution densities of all historical answers, which reflects the overall distribution of the entire set of historical answers in the semantic space. For each historical answer, its centralized weight is equal to the ratio of its distribution density to the summed distribution density. This ratio reflects the relative importance of the answer in the overall distribution. Through calculation, the system will obtain a centralized weight value for each historical answer, and these weight values ​​will be used in the subsequent answer selection or sorting process. Through the above process, the system can analyze the centralized weight distribution of several historical answers to obtain the relative importance of each answer in the semantic space. This analysis not only helps the system to understand the semantic relationship between historical answers more accurately, but also provides strong support for subsequent answer selection or sorting.

[0031] In a preferred embodiment, according to the semantic similarity set, the several historical answers are traversed to perform distribution density analysis to obtain several distribution densities, including: extracting the first historical answer; extracting k neighboring semantic similarities with the semantic similarity of the first historical answer from the semantic similarity set from large to small, where 0.2M≤k value≤0.5M, M represents the number of historical answers; calculating the mean of the k neighboring semantic similarities, setting it as the first distribution density, and adding it to the several distribution densities.

[0032] Furthermore, when it's necessary to analyze the distribution density of historical answers, this is typically done based on their semantic similarity. At this point, the system has already collected a series of historical answers and calculated their semantic similarity using natural language processing techniques, forming a semantic similarity set. In this set, each historical answer shares a certain semantic similarity score with the others. To begin the analysis, the system first randomly or sequentially selects the first historical answer from all historical answers as a starting point. Next, the semantic similarity scores of the k most similar neighboring answers to this first historical answer are extracted from the semantic similarity set. The value of k is a key parameter, determining the number of most similar answers to consider. Based on given conditions, the value of k is set between 20% and 50% of the number of historical answers M, i.e., 0.2M ≤ k ≤ 0.5M. This range is chosen to ensure computational efficiency while capturing a sufficient number of similar answers to reflect the distribution of the first historical answer in the semantic space. After obtaining k neighboring semantic similarity scores, the system will average them to obtain the first distribution density. This mean reflects the average semantic proximity between the first historical answer and its k most similar answers, and thus, to a certain extent, reflects its distribution density in the semantic space. Finally, this calculated first distribution density will be added to a distribution density set maintained by the system. This set will be used to store the distribution density values ​​of all historical answers for subsequent analysis or sorting operations. Through the above steps, the system can calculate a distribution density value for each historical answer, which reflects the relative distribution of the answer in the semantic space. This method not only helps to understand the semantic relationship between historical answers, but also provides useful information for subsequent analysis or processing.

[0033] In a preferred embodiment, the consulting questions are processed through the question and answer log, and the answer frequency weight distribution of the several historical answers is performed to obtain several frequency weights, including: according to the question and answer log, the several historical answers are sorted according to the answer time sequence to obtain the historical answer sorting result; the several latest answer times of the historical answer sorting result are counted; the historical answers whose time interval with the current time is greater than a preset time length among the several latest answer times are deleted to obtain several updated historical answers; the answer frequency weight distribution of the several updated historical answers is performed through the question and answer log to obtain several frequency weights.

[0034] Specifically, when processing consultation questions, using Q&A logs to analyze the frequency weight distribution of historical answers is an effective method for assessing which answers are most popular or frequently used. First, the system extracts several historical answers related to the current consultation question from the Q&A log. Next, these answers are ranked based on the order in which they appeared in the Q&A log—that is, the order in which they were used to answer similar questions. The resulting ranking is an ordered list that reflects the temporal distribution of the historical answers. After obtaining the ranked historical answers, the system further calculates the latest answer time for each answer in the ranked list. The latest answer time refers to the most recent time when the answer was used to answer a similar question in the Q&A log. This step helps identify which answers have been frequently used recently. Next, the system examines the time interval between these latest answer times and the current time. If the time interval between the latest answer time and the current time for a historical answer exceeds a preset time interval, the answer is considered outdated and removed from the updated historical answer set. This preset time interval can be adjusted based on actual conditions to reflect the system's requirements for answer timeliness. This step ensures that the historical answers used for subsequent analysis are active and relevant in the recent period. After deleting outdated answers, the system will calculate the answer frequency weight distribution for the remaining updated historical answers. Answer frequency refers to the number of times each answer is used to answer similar questions in the Q&A log. The weight distribution is determined based on the answer frequency, that is, the more times an answer is answered, the greater its weight. This step can result in a set of several frequency weights that reflect the relative importance of each updated historical answer in answering similar questions. Through the above process, the system can use the Q&A log to analyze the answer frequency weight distribution of historical answers. This method not only helps to identify which answers have been frequently used in the recent period, but also provides strong support for subsequent answer selection or sorting.

[0035] In a preferred embodiment, optimizing the multiple historical answers according to the multiple concentration weights and the multiple frequency weights, obtaining a target answer and feeding it back to the RPA robot to reply to the user, includes: constructing an optimization fitness function:

[0036] ;

[0037] ;

[0038] in, Characterizes the fitness of any historical answer, Characterizes the centralized weight bias parameter, Characterize the frequency weight bias parameter, Characterizes the natural constants, Characterize the centralized weight, Characterize the frequency weight, represents the convergence probability, Characterizes the fitness threshold, Representing the number of iterations; optimizing the several historical answers according to the optimization fitness function, and when the convergence probability is met, outputting the optimal historical answer and setting it as the target answer.

[0039] Furthermore, in order to select the best answer from several historical answers, a fitness function needs to be constructed to evaluate the quality of each answer. This fitness function takes into account multiple factors, including concentration weight, frequency weight, and some adjustment parameters to balance these factors. The specific formula of the optimization fitness function is:

[0040] ;

[0041] ;

[0042] in, Characterizes the fitness of any historical answer, which is based on the centralized weight and frequency weights To calculate; centralized weight bias parameter Used to adjust the influence of centralized weight on fitness; frequency weight bias parameter Used to adjust the influence of frequency weight on fitness; Characterizing natural constants, convergence probability Used to determine whether the optimization process has reached a stable state, that is, whether a good enough answer has been found; fitness threshold It is used to set a standard. When the fitness of an answer exceeds this threshold, it can be considered a good enough answer. Indicates how many iterations the optimization process has been through. With the fitness function, we can start optimizing several historical answers. The optimization process involves multiple iterations, and each iteration will evaluate each answer according to the fitness function and adjust the order or weight of the answer based on the evaluation results. During the iteration process, the system will continuously check the convergence probability. To determine whether a good enough answer has been found, that is, the fitness exceeds the threshold The optimization process ends when the convergence probability is met, and the optimal historical answer is output as the target answer. Once the target answer is found, the system feeds it back to the RPA robot, which then uses it to respond to the user's question. Through the above process, the system combines concentration and frequency weighting to optimize several historical answers, ultimately finding the optimal answer and feeding it back to the RPA robot for response. This approach not only considers the importance and frequency of the answer but also ensures the accuracy and efficiency of the optimization process through mechanisms such as fitness functions and convergence probability.

[0043] In a preferred embodiment, a semantic similarity analysis is performed on the several historical answers to obtain a semantic similarity set, including: obtaining a first historical answer and a second historical answer; performing semantic analysis on the video of the first historical answer to obtain first video semantics, performing semantic analysis on the picture of the first historical answer to obtain first picture semantics, and extracting first text semantics of the first historical answer; performing semantic analysis on the video of the second historical answer to obtain second video semantics, performing semantic analysis on the picture of the second historical answer to obtain second picture semantics, and extracting second text semantics of the second historical answer; calculating the video semantic similarity of the first video semantics and the second video semantics, calculating the picture semantic similarity of the first picture semantics and the second picture semantics, and calculating the text semantic similarity of the first text semantics and the second text semantics; performing mean calculation on the video semantic similarity, the picture semantic similarity, and the text semantic similarity to obtain a first semantic similarity, and adding it to the semantic similarity set.

[0044] Specifically, two historical answers are selected for comparison, with the first and second historical answers being used as examples. Semantic parsing is performed on the first historical answer. If it contains a video, the video content is semantically parsed to extract key information, forming the first video semantics. If it contains an image, the image content is semantically parsed, similarly extracting key information to form the first image semantics. The text content in the first historical answer is extracted and semantically parsed to obtain the first text semantics. A similar semantic parsing is performed on the second historical answer to obtain the second video semantics, the second image semantics, and the second text semantics. Next, the video semantic similarity between the first video semantics and the second video semantics is calculated. This involves extracting and comparing features of the video content to determine the degree of similarity between the two. Similarly, the image semantic similarity between the first image semantics and the second image semantics is calculated. This involves extracting and comparing features such as objects, colors, and textures in the images. The text semantic similarity between the first text semantics and the second text semantics is calculated. This can be achieved by comparing the vocabulary, sentence structure, and themes in the texts. After obtaining the video semantic similarity, image semantic similarity, and text semantic similarity, the three similarity values ​​are averaged to obtain a comprehensive semantic similarity value, namely the first semantic similarity. This first semantic similarity value is added to the semantic similarity set, which will contain the semantic similarity values ​​between all historical answer pairs. The above process can be repeated to compare all possible historical answer pairs to construct a complete semantic similarity set. Through the above process, semantic similarity analysis can be performed on several historical answers, and a set containing the semantic similarities between all historical answer pairs can be obtained. This method not only considers the similarity between text content, but also the similarity between non-text content such as videos and images, thereby enabling a more comprehensive assessment of the content relevance between historical answers.

[0045] The embodiments of the present invention provide an intelligent operation and maintenance system based on an AI big model and RPA, which has at least the following technical effects:

[0046] 1. The AI ​​semantic model performs semantic matching on user input, enabling rapid and accurate identification of user inquiries, regardless of whether the input is text or image. For text input, the AI ​​semantic model is used directly for matching; for image input, the input is first converted to text using an OCR text recognizer before matching. This design enables the system to flexibly handle a variety of user input methods, improving the efficiency and accuracy of question answering.

[0047] 2. Historical answers are analyzed using both concentration and frequency weights. The quality and relevance of historical answers are comprehensively assessed through methods such as semantic similarity analysis, distribution density analysis, and answer frequency weight distribution. Then, a fitness function is used to optimize multiple historical answers, finding the optimal answer and feeding it back to the RPA robot for response. This process considers not only the accuracy and relevance of the answers, but also their timeliness and frequency of use, thereby improving answer quality and user experience.

[0048] 3. By continuously updating the answer library and Q&A log, we maintain real-time tracking of historical answers and their accuracy. In the Q&A log, the system sorts historical answers based on response time and deletes outdated or no longer relevant answers to ensure the accuracy and timeliness of updated historical answers. This dynamic update mechanism enables the system to continuously learn and optimize, continuously improving the accuracy and efficiency of question answering.

[0049] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An intelligent operation and maintenance system based on AI big model and RPA, characterized by: The execution steps include: In response to the RPA robot receiving user input information, it performs semantic matching through the AI ​​semantic model to obtain the consultation question; Process the consultation question through the answer material library and match several historical answers; Performing centralized weight distribution on the plurality of historical answers to obtain a plurality of centralized weights; Processing the consulting question through the question-and-answer log, performing answer frequency weight distribution on the several historical answers, and obtaining several frequency weights; Optimizing the historical answers based on the multiple concentration weights and the multiple frequency weights to obtain a target answer and feeding it back to the RPA robot to reply to the user; The steps of performing centralized weight distribution on the plurality of historical answers to obtain a plurality of centralized weights include: Performing semantic similarity analysis on the plurality of historical answers to obtain a semantic similarity set; According to the semantic similarity set, traverse the plurality of historical answers to perform distribution density analysis to obtain a plurality of distribution densities; Calculating the ratio of the plurality of distribution densities to the distribution sum density to obtain the plurality of concentration weights; The consultation question is processed through the question-and-answer log, and the frequency weight distribution of the historical answers is performed to obtain a number of frequency weights. The execution steps include: According to the question and answer log, the plurality of historical answers are sorted according to the reply time sequence to obtain a historical answer sorting result; Counting the multiple historical answers at the multiple latest reply times of the historical answer sorting results; Deleting the historical answers whose time intervals between the latest reply time and the current time are greater than a preset time length from the multiple latest reply time points, to obtain multiple updated historical answers; Performing answer frequency weight distribution on the several updated historical answers through the question and answer log to obtain several frequency weights; The steps of optimizing the historical answers based on the multiple concentration weights and the multiple frequency weights, obtaining a target answer and feeding it back to the RPA robot to reply to the user include: Construct an optimization fitness function: , , in, Characterizes the fitness of any historical answer, Characterizes the centralized weight bias parameter, Characterize the frequency weight bias parameter, Characterizes the natural constants, Characterize the centralized weight, Characterize the frequency weight, represents the convergence probability, Characterizes the fitness threshold, Characterize the number of iterations; The plurality of historical answers are optimized according to the optimization fitness function, and when the convergence probability is satisfied, the optimal historical answer is output and set as the target answer.

2. The system according to claim 1, wherein In response to the RPA robot receiving user input information, it performs semantic matching through the AI ​​semantic model to obtain the consultation question. The execution steps include: Constructing a question grammatical structure, wherein the question grammatical structure includes a subject and an action; Configuring a user input tag and a question identifier that conforms to the question grammatical structure, wherein the user input tag is text information including the subject and the execution action; Based on the user input label and the question identifier, supervise the training of N front-end AI semantic models through N machine learning models with different topological structures; Based on the outputs of the N front-end AI semantic models and the problem identifier, supervise the training of the back-end AI semantic model through the LSTM-DSSM neural network; The N output branches of the N front-end AI semantic models are fused with the input layer of the rear-end AI semantic model to obtain the AI ​​semantic model.

3. The system according to claim 1, wherein: In response to the RPA robot receiving user input information, it performs semantic matching through the AI ​​semantic model to obtain the consultation question. The execution steps include: When the user input information is text, it is sent to the AI ​​semantic model; When the user input information is a picture, the OCR text recognizer is activated, and after processing the picture, the text recognition result is obtained and sent to the AI ​​semantic model.

4. The system according to claim 1, wherein: According to the semantic similarity set, traversing the plurality of historical answers to perform distribution density analysis to obtain a plurality of distribution densities, the execution steps include: Extract the first history answer; Extracting k neighboring semantic similarities with the first historical answer from the semantic similarity set, in descending order of semantic similarity, where 0.2M≤k≤0.5M, and M represents the number of historical answers; The average of the k adjacent semantic similarities is calculated, set as the first distribution density, and added to the plurality of distribution densities.

5. The system according to claim 1, wherein: Performing semantic similarity analysis on the plurality of historical answers to obtain a semantic similarity set, the execution steps include: Get the first history answer and the second history answer; Performing semantic analysis on the video of the first historical answer to obtain first video semantics, performing semantic analysis on the picture of the first historical answer to obtain first picture semantics, and extracting first text semantics of the first historical answer; Performing semantic analysis on the video of the second historical answer to obtain second video semantics, performing semantic analysis on the picture of the second historical answer to obtain second picture semantics, and extracting second text semantics of the second historical answer; Calculating the video semantic similarity between the first video semantics and the second video semantics, calculating the picture semantic similarity between the first picture semantics and the second picture semantics, and calculating the text semantic similarity between the first text semantics and the second text semantics; Performing mean calculation on the video semantic similarity, the image semantic similarity, and the text semantic similarity to obtain a first semantic similarity, and adding the first semantic similarity to the semantic similarity set.

Images