Intelligent work order analysis method and device based on closed-loop self-evolution and medium

Abstract

The invention discloses a work order intelligent analysis method and device based on closed-loop self-evolution, and a medium, and the method comprises the steps: constructing a unified structured database, constructing a hot replaceable sub-network with work order appeal category recognition and urgency degree scoring capabilities based on semantic vector representation, and achieving the precise recall of millisecond-level related cases and policy provisions, thereby achieving the precise recall of the related cases and policy provisions. The problems of time consumption and large deviation of information query are solved, vectorization is performed on the current work order based on the formatted data, historical cases with the similarity higher than a threshold value are recalled in a historical work order library as decision reference data, and quasi-reply content with the quality reaching the standard is generated. According to the technical scheme, reply errors caused by personal understanding deviation are eradicated through auxiliary work order reply, the accuracy and speciality of the content are ensured before reply sending through integrating reply content quality real-time scoring and a closed-loop correction mechanism, so that the work order rehandling rate is greatly reduced, and the one-time handling rate and the operation efficiency of work order disposal are integrally improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a work order intelligent analysis method, device, and medium based on closed-loop self-evolution. Background Technology

[0002] The increasing demands for high-quality electricity services from society have placed enormous pressure and challenges on power supply companies in handling customer requests. Currently, the following technical and management bottlenecks urgently need to be addressed in the customer work order processing flow:

[0003] First, data silos abound, leading to inefficient information retrieval. Handling customer requests often requires accessing multi-dimensional data such as customer files, electricity usage information, and historical work orders. However, this data is typically scattered across multiple independent and heterogeneous systems, including marketing systems, next-generation electricity information collection systems, and power distribution management systems. Frontline staff are forced to repeatedly log in, switch between, and perform keyword searches across these systems—a cumbersome and time-consuming process. This fragmented approach not only results in slow response times but also makes the system highly susceptible to information bias due to incomplete information or synchronization delays, creating potential for subsequent misjudgments.

[0004] Secondly, the policies and regulations are complex, and knowledge relies heavily on personal experience. The power industry has numerous policies, regulations, and business procedures that are frequently updated. In handling work orders, especially policy consultation orders, there is a high degree of reliance on frontline staff's personal understanding, memory, and on-the-spot interpretation abilities regarding policies. If the policy is not properly grasped or misunderstood, it can easily lead to inaccurate or inconsistent responses to customer requests, resulting in customer doubts or even complaints, damaging the authority and credibility of power supply services.

[0005] Third, there is a lack of experience in handling such cases, and procedures lack standardization. For work orders arising from emerging businesses or complex malfunctions, due to their atypical nature and low incidence, frontline staff often lack relevant experience and are unfamiliar with established, standardized procedures. This leads to considerable arbitrariness in on-site handling, inconsistent quality of solutions, and difficulty in ensuring stable service quality, ultimately resulting in low customer satisfaction.

[0006] Fourth, the quality of response reports is inconsistent and lacks intelligent support. Work order responses are the final exit point in the processing flow, and their quality is crucial. Currently, work order specialists rely primarily on their individual business skills when drafting response reports, lacking accurate references to relevant historical cases and intelligent content generation technology. This not only leads to low efficiency in response report processing but also frequently results in inaccurate content, missing key points, or non-standard wording, making them highly susceptible to rejection by higher-level quality inspection departments. This leads to work orders being reprocessed, creating a vicious cycle of "processing-rejection-reprocessing," severely wasting human resources.

[0007] In summary, the existing power customer complaint handling model has significant shortcomings in data integration, knowledge application, process standardization, and quality control, making it difficult to meet the high-efficiency, accurate, and standardized requirements for quality power services under the new circumstances. Therefore, there is an urgent need for an integrated intelligent solution that can integrate multi-source data, internalize domain knowledge, empower frontline personnel, and ensure the quality of handling. Summary of the Invention

[0008] The technical problem this invention aims to solve is that the current handling of electricity customer complaints suffers from several bottlenecks, including data silos, reliance on personal experience for policy implementation, lack of standards for handling complex work orders, and low quality of response reports. The goal is to provide a closed-loop, self-evolving intelligent work order analysis method, device, and medium. By constructing a unified structured database and building a hot-swappable subnetwork based on semantic vector representation, capable of identifying work order complaint categories and scoring urgency, it achieves millisecond-level accurate retrieval of relevant cases and policy provisions, solving the problems of time-consuming and biased information queries. Based on the formatted data, the current work order is vectorized, and historical cases with similarity exceeding a threshold are retrieved from the historical work order database as decision-making reference data to generate quality-compliant proposed responses. This assists in work order responses by eliminating errors caused by individual misunderstandings. By integrating real-time scoring and closed-loop correction mechanisms for response content quality, the accuracy and professionalism of the content are ensured before the response is issued, thereby significantly reducing the work order reprocessing rate and improving the overall first-time completion rate and operational efficiency of work order handling.

[0009] This invention is achieved through the following technical solution:

[0010] The first aspect of this invention provides a work order intelligent analysis method based on closed-loop self-evolution, comprising the following specific steps:

[0011] Historical work order texts, policy and regulatory documents, and customer service dialogue records are obtained, preprocessed, and used to form a structured database;

[0012] A text encoder based on a pre-trained language model is constructed. Text content from a structured database is concatenated into a unified text input to the text encoder to obtain a semantic vector representation of the work order text. Based on the semantic vector representation, a hot-swappable subnetwork with the ability to identify work order request categories and score urgency is constructed.

[0013] Training data is extracted from the structured database, including text semantic features obtained from the work order vector database, knowledge graph features obtained from the policy and regulation database, and temporal dialogue features obtained from the customer service dialogue database. The hot-swappable subnetwork is then jointly trained to obtain the trained subnetwork.

[0014] Input the title and content of the customer request work order into the trained subnetwork, and output formatted data containing the request category and urgency score;

[0015] Based on the formatted data, the current work order is vectorized, and historical cases with similarity higher than the threshold are recalled from the historical work order database as decision reference data to generate a proposed response content that meets the quality standards.

[0016] The proposed response content is fed back to the structured database for periodic retraining of the subnetwork and updating of the graph to obtain the work order processing model.

[0017] Furthermore, the preprocessing of historical work order texts, policy and regulatory documents, and customer service dialogue records includes:

[0018] Historical work orders are divided into triples consisting of sentence + preceding sentence + following sentence, and the triple text is converted into vectors.

[0019] Natural language processing is used to identify the structure of policy and regulatory documents, resulting in a four-level heading group of articles, clauses, items, and subheadings. An anchor ID is created based on the subheading, and text is associated through the anchor ID to obtain a policy knowledge relational database that can be quickly queried and traversed.

[0020] The real-time voice dialogue stream in the customer service chat log is segmented and marked, and the segmented and marked voice segments are converted into corresponding text records.

[0021] Furthermore, the hot-swappable subnetwork constructed based on semantic vector representation, capable of identifying work order request categories and scoring urgency, specifically includes:

[0022] Build the initial model;

[0023] The title and content of the work order are concatenated into a unified text, which is then input into the initial model for encoding to obtain a multi-dimensional sentence vector.

[0024] Two feedforward branches, multi-label classification and linear regression, are connected in parallel to the multi-dimensional sentence vector. The two feedforward branches are used to output the multi-dimensional request category recognition of the work order and the urgency score, respectively.

[0025] The initial model is combined with the two branches to form a dual-task initial subnetwork that simultaneously has the ability to identify work order request categories and score urgency.

[0026] Furthermore, training data is extracted from the structured database, including textual semantic features obtained from the work order vector database, knowledge graph features obtained from the policy and regulation database, and temporal dialogue features obtained from the customer service dialogue database. The hot-swappable subnetwork is then jointly trained to obtain the trained subnetwork, specifically including:

[0027] Align text semantic features, knowledge graph features, and temporal dialogue features along the timeline to determine the temporal relationships between features;

[0028] Based on the attention mechanism, we analyze the logical relationships between semantic features, graph features, and dialogue features;

[0029] By using a feature fusion network to dynamically stitch together time-aligned and logically related features, a work order story flow containing a complete event timeline is generated.

[0030] The work order story flow is analyzed using a large language model to extract core structured fields, including issue summary tags and urgency scores.

[0031] A classification loss function is constructed, and the problem summary label and urgency score are used as supervision signals for training on the work order training data;

[0032] If the subnetwork's overall score on the reserved set exceeds the set threshold during verification, the training is considered complete, and the trained subnetwork is obtained.

[0033] Furthermore, when training on the work order training data using problem summary labels and urgency scores as supervisory signals, the process includes:

[0034] The problem summary tags are converted into query vectors, which are then searched in the policy knowledge relational database and the triple vector library to obtain the top K problem summary anchor points and top M historical solutions that are most relevant to the current work order.

[0035] The search results are used as external memory and vector-concatenated with the original work order text to form an enhanced input.

[0036] The enhanced input is fed into the hot-swappable subnetwork, and multi-label classification supervision is performed using problem summary labels, and regression supervision is performed using urgency scores, to conduct dual-task training with the assistance of external memory.

[0037] Furthermore, the step of inputting the title and content of the customer request work order into the trained subnetwork and outputting formatted data containing the request category and urgency score specifically includes:

[0038] The title and content of the customer request work order are input into the trained subnetwork, which has the ability to identify multi-label request categories and urgency regression.

[0039] The trained subnetwork outputs the raw score for the appeal category and the raw score for urgency.

[0040] The original urgency score is causally calibrated to obtain a calibrated urgency score;

[0041] Encapsulate the appeal category and the calibrated urgency score into formatted data and return it.

[0042] Furthermore, when the request category output by the trained subnetwork is policy consultation, the retrieval enhancement generation mechanism is triggered first before customer service processing. Relevant policy provisions are retrieved through vector similarity retrieval, and the relevant policy provisions are concatenated with the work order content and input into the subnetwork to automatically generate policy response text.

[0043] Furthermore, the generated proposed response content that meets the quality standards specifically includes:

[0044] The customer request work order is vectorized and the vector similarity is searched in the historical work order database. Historical work orders with a similarity higher than a set threshold and their processing results are recalled as decision reference data.

[0045] The trained subnetwork is invoked to score the quality of the customer service response content generated using decision reference data. If the score is not up to standard, a rewrite is prompted until a satisfactory response content is output.

[0046] A second aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a work order intelligent analysis method based on closed-loop self-evolution.

[0047] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a work order intelligent analysis method based on closed-loop self-evolution.

[0048] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0049] By constructing a unified structured database and building a hot-swappable subnetwork based on semantic vector representation, capable of identifying work order request categories and scoring urgency, millisecond-level accurate retrieval of relevant cases and policy provisions can be achieved. This solves the problems of time-consuming and biased information queries. Based on the formatted data, the current work order is vectorized, and historical cases with similarity exceeding a threshold are retrieved from the historical work order database as decision reference data to generate qualified response content. This assists in work order responses and eliminates errors in responses caused by personal misunderstandings. By integrating real-time scoring and closed-loop correction mechanisms for response content quality, the accuracy and professionalism of the content are ensured before the response is sent, thereby significantly reducing the work order reprocessing rate and improving the overall first-time completion rate and operational efficiency of work order handling. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0051] Figure 1 This is a work order intelligent analysis method based on closed-loop self-evolution in this embodiment of the invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0053] As one possible implementation method, such as Figure 1 As shown, this embodiment provides a work order intelligent analysis method based on closed-loop self-evolution, including the following specific steps:

[0054] Historical work order texts, policy and regulatory documents, and customer service dialogue records are obtained, preprocessed, and used to form a structured database;

[0055] A text encoder based on a pre-trained language model is constructed. Text content from a structured database is concatenated into a unified text input to the text encoder to obtain a semantic vector representation of the work order text. Based on the semantic vector representation, a hot-swappable subnetwork with the ability to identify work order request categories and score urgency is constructed.

[0056] Training data is extracted from a structured database, including text semantic features obtained from a work order vector database, knowledge graph features obtained from a policy and regulation database, and temporal dialogue features obtained from a customer service dialogue database. The hot-swappable subnetwork is then jointly trained to obtain the trained subnetwork.

[0057] Input the title and content of the customer request work order into the trained subnetwork, and output formatted data containing the request category and urgency score;

[0058] Based on the formatted data, the current work order is vectorized, and historical cases with similarity higher than the threshold are recalled from the historical work order database as decision reference data to generate a proposed response content that meets the quality standards.

[0059] The proposed response content is fed back to the structured database for periodic retraining of the subnetwork and updating of the graph to obtain the work order processing model.

[0060] This embodiment achieves end-to-end intelligent handling and self-evolution capabilities in the power grid customer service field through a five-stage closed loop consisting of preprocessing, dual-task sub-networks, vector recall, quality control, and data feedback, resulting in the following significant effects:

[0061] First, by uniformly preprocessing and structuring the scattered historical work orders, policies and regulations, and customer service dialogues into a database, the previously complex historical texts can be quickly retrieved. Building a structured database avoids repeated searches by agents, shortening work order processing time. A text encoder based on a pre-trained language model is constructed, concatenating the text content from the structured database into a unified text input to the encoder to obtain the semantic vector representation of the work order text. Based on this semantic vector representation, a dual-task, hot-swappable subnetwork capable of work order request category recognition and urgency scoring is constructed. Then, training data is extracted from the structured database, including textual semantic features from the work order vector database, knowledge graph features from the policy and regulations database, and temporal dialogue features from the customer service dialogue database. The system employs a joint training method for the hot-swappable subnetworks to obtain trained subnetworks. Based on these trained subnetworks, precise urgency scoring queue data processing is performed, automatically sorting work orders by category and urgency. This enables automatic classification, grading, and routing of work orders, eliminating the need for manual grading and directly processing orders based on their scores, thus preventing work order backlog. Formatted data vectorization is used to recall historical cases with similarity exceeding a set threshold, allowing for direct referencing of successful experiences. Simultaneously, the subnetwork performs quality scoring on proposed responses, forcing rewriting of responses below a set threshold to improve work order response quality and prevent invalid work order processing. The final processing results are automatically fed back daily, periodically triggering LoRA incremental fine-tuning and knowledge graph updates, achieving daily updates.

[0062] Furthermore, the hot-swappable subnetwork design allows the core algorithm model to be securely and seamlessly updated and upgraded without affecting online services, greatly improving the system's maintainability and stability. Managers can analyze accumulated system data (such as high-frequency issues, solution effectiveness, and quality score trends) to accurately identify bottlenecks, product defects, or knowledge gaps in the service process, providing quantitative data for operational optimization, product improvement, and training direction. From pre-process reference (retrieval) to in-process quality inspection (scoring), and then to post-process optimization (feedback and updates), the system achieves digital and intelligent management of the entire customer service process, transforming quality management from post-event spot checks to real-time, proactive intervention.

[0063] The process of preprocessing the obtained historical work order texts, policy and regulation documents, and customer service conversation records in this embodiment is uniformly completed with slicing - vectorization - anchoring - transcription before warehousing, significantly reducing subsequent retrieval and generation errors. The original and messy multi-source data is processed into high-quality and structured machine-readable data. By slicing the historical work orders into sentence + previous sentence + next sentence triples and immediately vectorizing them, retaining the context of multiple tokens before and after, the complete appeal scenario can still be restored during single-sentence recall, improving the hit rate of vector retrieval. Use natural language processing to automatically identify the four-level structure of article - section - item - sub-item, generate a unique anchor ID for each "sub-item", and then hang the main text, attachments, and associated articles to this anchor ID to form a policy knowledge relational database that can be quickly queried and traversed; the average latency of anchor retrieval is reduced, and the retrieval positioning basis time for the work order reply process is shortened; the real-time voice stream is segmented by energy detection + end-of-sentence silence, and the roles of the customer service / customer are marked synchronously, and then transcribed into text to eliminate the role misalignment caused by traditional whole-paragraph transcription; the transcription error rate is reduced; finally, the above three types of processed texts are uniformly output as JSON vector blocks with IDs and directly written into the same hybrid vector - relation library, achieving the integration of one text, one vector, and one anchor, reducing duplicate storage. The specific operation steps are as follows:

[0064] Processing of historical work orders: Clean the original work order text, including removing special characters, correcting typos, and unifying the entity formats such as dates / names. Use mature natural language processing tools to segment the work order text to obtain a list of sentences. In a sliding window manner, create the corresponding triple (previous sentence, current sentence, next sentence) for each sentence. Concatenate the three sentences in each triple into a long string using a delimiter. Use a text embedding model to convert this concatenated string into a high-dimensional vector. Finally, output a vector database that stores the vector corresponding to each triple and associates it back to the original work order ID and sentence position.

[0065] Processing of policies and regulations: Use a dedicated document parsing library to extract the original text and format information (such as font, size, position) from PDF or WORD files. Use rules or models to identify the hierarchical structure of the document, match the numbers through regular expressions, and combine the font and position information to determine the hierarchical relationship. Among them:

[0066] Chapter / article: Usually uses "Chapter 1", "Article 1" or Arabic numeral numbering, and the font is the largest, centered or bold.

[0067] Section: May not have a number, or uses "(1)", etc., and the font is slightly smaller.

[0068] Item: Uses numbering such as "1.", "(1)", etc.

[0069] Item: Use numbering such as “(a)”, “①”, etc., or as a specific paragraph under the item.

[0070] The document page is used as the image input model. The model predicts the label of each text block. That is, an anchor ID is created based on the target, text is associated through the anchor ID, and then a hierarchical tree is built based on the labels and position information of the associated text.

[0071] Processing of customer service chat logs:

[0072] First, the continuous audio stream is analyzed to detect speech signals and silence. Based on the length of the silence segments, the possible speaker switching points are initially determined as candidate points for segmentation. The entire audio file is received, and a series of segments are output. Each segment includes a start time, end time, and corresponding speaker ID, identifying who speaks and when, thus tagging the audio stream with speaker labels. Combining the results of VAD and SD, the audio stream is segmented into short, single-speaker audio segments at speaker switching points and long silence points. Each segmented audio segment is converted into text by an ASR engine. The text recognized by ASR is associated with its corresponding speaker ID and timestamp, and the text is integrated into the final dialogue record according to the time sequence.

[0073] Furthermore, a hot-swappable subnetwork with the ability to identify work order request categories and score urgency is constructed based on semantic vector representation, specifically including:

[0074] An initial model is constructed by selecting a suitable, open-source pre-trained large language model as the base. The structured information of the work order is transformed into input that the model can process, and its high-quality semantic representation is obtained. This involves concatenating the work order's title and content into a unified long text sequence using delimiters. This concatenated text sequence is then input into the initial model for encoding. Typically, the final hidden state corresponding to the CLS token is extracted, or the hidden states of all tokens are averaged and pooled to form a multi-dimensional sentence vector for the entire work order text. This vector is a high-dimensional, dense array of floating-point numbers, containing the semantic information of the entire work order.

[0075] Since a work order may belong to multiple request categories simultaneously, it is necessary to solve two different tasks, classification and regression, in parallel on the basis of shared semantic vectors. This allows a single model to complete both tasks simultaneously, achieving resource sharing and accelerating inference speed. Specifically, two feedforward branches, multi-label classification and linear regression, are connected in parallel to the multi-dimensional sentence vectors. These two feedforward branches are used to output the multi-dimensional request category recognition of the work order and the urgency score, respectively. The two branches are integrated with the initial model, and the model is trained to handle both tasks simultaneously through joint training. That is, the inputs of both branches are connected to the sentence vectors output by the initial model. The entire joint network is trained end-to-end using labeled data (each data point has a category label and an urgency score). The knowledge related to the two tasks is mutually reinforcing through the shared encoder, which often yields better results than training two independent models. This trained dual-task initial sub-network (including the initial model and the two branches) is packaged into an independent module or service, forming a dual-task initial sub-network with the ability to simultaneously identify work order request categories and score urgency.

[0076] A dual-task initial subnetwork with the ability to simultaneously identify work order request categories and score urgency was formed. A structured database was then used to train the hot-swappable subnetwork.

[0077] Training data is extracted from structured databases, including textual semantic features from a work order vector database, knowledge graph features from a policy and regulation database, and temporal dialogue features from a customer service dialogue database. These textual semantic features, knowledge graph features, and temporal dialogue features are aligned along a timeline to determine the temporal relationships between features. Logical relationships between semantic features, knowledge graph features, and dialogue features are analyzed based on an attention mechanism. A feature fusion network dynamically concatenates the time-aligned and logically related features, i.e., dynamically concatenates and fuses them according to time and logical order (e.g., user raises a question - customer service A initially replies - user provides supplementary information - customer service B escalates the process - final resolution), generating a work order story flow text containing a complete event timeline as the global context for subsequent analysis. Core semantic information is extracted from the lengthy work order story flow text to provide supervision signals for training and generate queries for retrieval. A Large Language Model (LLM) is used as an information extractor to analyze the work order story flow. Through design, the LLM instruction extracts core structured fields from the work order story flow text. These core structured fields include issue summary tags: multi-label classification of the work order's core demands (e.g., [complaint, refund]), and urgency scores: continuous numerical values ​​based on text descriptions (e.g., 2.0 / 4.5). High-quality structured labeled data is output. A classification loss function is constructed, and the issue summary tags and urgency scores are used as supervisory signals for training on the work order training data. The issue summary tags are converted into query vectors, which are then retrieved from a policy knowledge relational database and a triplet vector database to obtain the Top K most relevant issue summary anchor points and Top M historical solutions for the current work order. Specifically, the vector database calculates the cosine similarity score [-1, 1] between the query vector and all historical vectors in the database. The database returns the Top K historical work orders with similarity scores higher than a set threshold, and the returned results include both the similarity score and the associated original work order information. The retrieval results are used as external memory and concatenated with the original work order text to form enhanced input. This enhanced input is then fed into a hot-swappable subnetwork, where it undergoes dual-task training with external memory assistance. The training is supervised by the question summary labels for multi-label classification and by the urgency score for regression. Task 1: Multi-label classification, supervised by the question summary labels using binary cross-entropy loss; Task 2: Regression, supervised by the urgency score using mean squared error loss or smoothed L1 loss. Through this process, the hot-swappable subnetwork learns how to utilize external memory (policies and historical cases) to assist in classification and scoring tasks, making its decisions more evidence-based. Unused training data is used as a reserved set for evaluation. A weighted scoring criterion combining classification accuracy, F1 score, and regression error is established. Training is complete when the model's overall score on the validation set exceeds a preset threshold, resulting in the final deployable trained subnetwork.

[0078] Then, the title and content of the customer request ticket are input into the trained sub-network, outputting formatted data containing the request category and urgency score: the title and content of the customer request ticket, after the same preprocessing as during training, are input into the trained hot-swappable sub-network, outputting the original score for the request category: a vector of length N (N = total number of categories), where each element is a probability value between 0 and 1 processed by the Sigmoid function. These represent the original confidence scores for each category. Using calibration methods such as orthographic regression or PlattScaling, on a preserved calibration dataset, the model obtains the original urgency score and its corresponding true urgency level for each sample, fitting a monotonic mapping function f: original score > calibrated score. If the original score A > B, then the calibrated score f(A) > f(B). At the same time, the original scores are calibrated to a scale that better matches the actual business level distribution. Finally, a calibrated urgency score that is more in line with business intuition is output. For each probability value in the original score vector of the request category, a threshold is applied, and the category with a probability value greater than the threshold is determined as the final label of the work order. The final result is encapsulated into a structured data object, usually in JSON format, which is convenient for network transmission and parsing by other systems.

[0079] In this embodiment, a condition-triggered retrieval enhancement generation process is also included. When the trained subnetwork identifies the request category of the work order, the system automatically triggers the subsequent RAG process if and only if the output primary request category or category combination contains "policy consultation". The problem summary or complete work order text of the work order is converted into a query vector. In the policy and regulation vector library, the query vector is used to perform high-precision semantic similarity retrieval. The system recalls the Top K policy articles most relevant to the current work order problem. The retrieved policy articles are concatenated with the original work order content according to a preset template. The concatenated enhanced context (i.e., the above-mentioned prompt words) is input into the subnetwork, and an automatically generated policy response text is output. This text directly addresses the user's question. The system directly sends the generated response text to the user and marks the work order as replied, realizing second-level automatic response to policy consultation work orders. This frees customer service from the tedious work of finding and interpreting policy articles and greatly reduces processing time.

[0080] When vectorizing the current work order based on formatted data and recalling historical cases with similarity exceeding a threshold from the historical work order database as decision reference data to generate a qualified proposed response, the specific steps include:

[0081] The current work order is vectorized based on formatted data. Customer request work orders are vectorized and vector similarity is searched in the historical work order database. Historical work orders with similarity to the current work order that are higher than a set threshold and their processing results are recalled as decision reference data.

[0082] Customer service agents read the decision-making reference data provided by the system and, combined with their own experience, manually draft responses. Simultaneously, the system can automatically concatenate the decision-making reference data with the current work order content, input it into a text generation model, and automatically generate a candidate response draft. The draft response generated in the previous step, along with the context of the current work order, is then input into the trained sub-network, which outputs a quantified quality score. If the score is unsatisfactory, a rewrite is prompted until a satisfactory response is output.

[0083] After the aforementioned cycle of "generation-scoring-judgment-rewriting," the final output is a proposed response that meets the quality score standard. This approach combines historical success (retrieval) with model quality assurance (scoring), ensuring both reliability and effectiveness. All scored and modified response data can be fed back for incremental training of the quality scoring model, creating a positive feedback loop that improves with use. This seamlessly integrates intelligent retrieval, content generation, and quality assessment, constructing a robust and adaptive content production pipeline. This is a key technological guarantee for ensuring the reliability, credibility, and usability of the intelligent customer service system's output, achieving a fundamental shift from simply "having a response" to "having a high-quality response."

[0084] As one possible implementation, this embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements a work order intelligent analysis method based on closed-loop self-evolution.

[0085] As one possible implementation, this embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a work order intelligent analysis method based on closed-loop self-evolution.

[0086] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A work order intelligent analysis method based on closed-loop self-evolution, characterized in that, The specific steps include the following: Historical work order texts, policy and regulatory documents, and customer service dialogue records are obtained, preprocessed, and used to form a structured database; A text encoder based on a pre-trained language model is constructed. Text content from a structured database is concatenated into a unified text input to the text encoder to obtain a semantic vector representation of the work order text. Based on the semantic vector representation, a hot-swappable subnetwork with the ability to identify work order request categories and score urgency is constructed. Training data is extracted from the structured database, including text semantic features obtained from the work order vector database, knowledge graph features obtained from the policy and regulation database, and temporal dialogue features obtained from the customer service dialogue database. The hot-swappable subnetwork is then jointly trained to obtain the trained subnetwork. Input the title and content of the customer request work order into the trained subnetwork, and output formatted data containing the request category and urgency score; Based on the formatted data, the current work order is vectorized, and historical cases with similarity higher than the threshold are recalled from the historical work order database as decision reference data to generate a proposed response content that meets the quality standards. The proposed response content is fed back to the structured database for periodic retraining of the subnetwork and updating of the graph to obtain the work order processing model.

2. The intelligent work order analysis method based on closed-loop self-evolution as described in claim 1, characterized in that, The preprocessing of historical work order texts, policy and regulatory documents, and customer service dialogue records includes: Historical work orders are divided into triples consisting of sentence + preceding sentence + following sentence, and the triple text is converted into vectors. Natural language processing is used to identify the structure of policy and regulatory documents, resulting in a four-level heading group of articles, clauses, items, and subheadings. An anchor ID is created based on the subheading, and text is associated through the anchor ID to obtain a policy knowledge relational database that can be quickly queried and traversed. The real-time voice dialogue stream in the customer service chat log is segmented and marked, and the segmented and marked voice segments are converted into corresponding text records.

3. The intelligent work order analysis method based on closed-loop self-evolution as described in claim 1, characterized in that, The hot-swappable subnetwork constructed based on semantic vector representation, capable of identifying work order request categories and scoring urgency, specifically includes: Build the initial model; The title and content of the work order are concatenated into a unified text, which is then input into the initial model for encoding to obtain a multi-dimensional sentence vector. Two feedforward branches, multi-label classification and linear regression, are connected in parallel to the multi-dimensional sentence vector. The two feedforward branches are used to output the multi-dimensional request category recognition of the work order and the urgency score, respectively. The initial model is combined with the two branches to form a dual-task initial subnetwork that simultaneously has the ability to identify work order request categories and score urgency.

4. The intelligent work order analysis method based on closed-loop self-evolution as described in claim 1, characterized in that, Training data is extracted from the structured database, including textual semantic features from the work order vector database, knowledge graph features from the policy and regulation database, and temporal dialogue features from the customer service dialogue database. The hot-swappable subnetwork is then jointly trained to obtain the trained subnetwork, specifically including: Align text semantic features, knowledge graph features, and temporal dialogue features along the timeline to determine the temporal relationships between features; Based on the attention mechanism, we analyze the logical relationships between semantic features, graph features, and dialogue features; By using a feature fusion network to dynamically stitch together time-aligned and logically related features, a work order story flow containing a complete event timeline is generated. The work order story flow is analyzed using a large language model to extract core structured fields, including issue summary tags and urgency scores. A classification loss function is constructed, and the problem summary label and urgency score are used as supervision signals for training on the work order training data; If the subnetwork's overall score on the reserved set exceeds the set threshold during verification, the training is considered complete, and the trained subnetwork is obtained.

5. The intelligent work order analysis method based on closed-loop self-evolution as described in claim 4, characterized in that, When training on work order training data using problem summary labels and urgency scores as supervisory signals, the following steps are included: The problem summary tags are converted into query vectors, which are then searched in the policy knowledge relational database and the triple vector library to obtain the top K problem summary anchor points and top M historical solutions that are most relevant to the current work order. The search results are used as external memory and vector-concatenated with the original work order text to form an enhanced input. The enhanced input is fed into the hot-swappable subnetwork, and multi-label classification supervision is performed using problem summary labels, and regression supervision is performed using urgency scores, to conduct dual-task training with the assistance of external memory.

6. The work order intelligent analysis method based on closed-loop self-evolution as described in claim 1, characterized in that, The process of inputting the title and content of a customer request work order into the trained subnetwork and outputting formatted data containing the request category and urgency score specifically includes: The title and content of the customer request work order are input into the trained subnetwork, which has the ability to identify multi-label request categories and urgency regression. The trained subnetwork outputs the raw score for the appeal category and the raw score for urgency. The original urgency score is causally calibrated to obtain a calibrated urgency score; Encapsulate the appeal category and the calibrated urgency score into formatted data and return it.

7. The intelligent work order analysis method based on closed-loop self-evolution as described in claim 6, characterized in that, When the request category output by the trained subnetwork is policy consultation, the retrieval enhancement generation mechanism is triggered first before customer service processing. Relevant policy provisions are retrieved through vector similarity retrieval, and the relevant policy provisions are concatenated with the work order content and input into the subnetwork to automatically generate policy response text.

8. The intelligent work order analysis method based on closed-loop self-evolution as described in claim 1, characterized in that, The proposed response content that meets the quality standards specifically includes: The customer request work order is vectorized and the vector similarity is searched in the historical work order database. Historical work orders with a similarity higher than a set threshold and their processing results are recalled as decision reference data. The trained subnetwork is invoked to score the quality of the customer service response content generated using decision reference data. If the score is not up to standard, a rewrite is prompted until a satisfactory response content is output.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the work order intelligent analysis method based on closed-loop self-evolution as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the work order intelligent analysis method based on closed-loop self-evolution as described in any one of claims 1 to 8.

Images