Gas customer service question and answer method and system based on interactive multi-energy synthesis algorithm

Abstract

The invention relates to the technical field of fuel gas intelligent customer service, and discloses a fuel gas customer service question answering method and system based on an interactive multi-energy synthesis algorithm, and the method comprises the steps: 1, constructing a knowledge base, and obtaining multi-source data information; performing text partitioning on the data, and automatically generating a vectorization index; forming a local question and answer knowledge base; step 2, constructing a scene model, constructing a multi-modal scene mode according to a production scene, a conveying scene and a family scene, and respectively establishing vector indexes of a keyword library and a knowledge library under the corresponding scenes; step 3, input analysis: converting user input content into a text, directly entering a semantic understanding layer, and analyzing and extracting keywords and problem intentions; 4, scene modes are matched according to the keywords, corresponding knowledge bases are called, question and answer service is carried out in the corresponding scene modes, and corresponding results are output. According to the method, the multi-scene application requirements can be met, the service efficiency is improved, the accuracy and the applicability are ensured, and the overall user experience is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of gas intelligent customer service, and in particular to a gas customer service question-answering method and system based on an interactive multi-functional synthesis algorithm. Background Art

[0002] In the current intelligent service practices of the gas industry, the application of intelligent customer service systems is gradually becoming popular, but their actual operation still exposes many limitations. Many systems rely on preset keyword matching or fixed question and answer templates. When users describe gas problems in non-standardized spoken language, the system often returns irrelevant answers or directly refuses to answer due to its inability to understand the semantic associations, significantly reducing the user experience. In particular, it may delay the acquisition of critical information in emergency situations. Moreover, when faced with complex professional problems in the gas field, it is easy to find it difficult to provide accurate answers due to the lack of deep semantic understanding capabilities. In particular, when users use colloquial expressions or involve multimodal information, the response effect is further reduced.

[0003] At the same time, when technicians in industrial environments inquire about equipment failures, the system often provides overly general answers, while ordinary home users who inquire about basic operations may receive complex instructions containing technical terms. This lack of scenario-based capabilities makes it difficult for a single system to simultaneously meet the needs of different user groups.

[0004] Based on this, existing technologies have attempted to introduce generalized large models to improve interactive flexibility, but this has introduced new problems. Due to the lack of strict domain knowledge constraints, the model may freely interpret open-ended questions beyond gas, even generating content that does not conform to industry standards. For example, when a user mistakenly enters a question unrelated to gas, the system fails to effectively identify and intercept it, and instead outputs a misleading answer. Furthermore, most systems' knowledge updates lag behind regulatory changes, making it difficult to keep up with the latest safety standards or policy adjustments, posing compliance risks.

[0005] In addition, in the existing technology, information is usually stored in the form of text summaries, which leads to the loss of visual information and the inability to intuitively display relevant content, thereby reducing the guidance effect of system use and making it difficult to improve the user experience. Summary of the Invention

[0006] The present invention aims to provide a gas customer service question-and-answer method and system based on an interactive multi-functional synthesis algorithm to solve the problems that the existing gas customer service question-and-answer system is difficult to apply to various scenarios, reduces the user experience, and cannot provide accurate question-and-answer services.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions: Solution 1: A gas customer service question-and-answer method based on an interactive multi-functional synthesis algorithm integrates multi-source knowledge bases and multi-modal large model capabilities, combined with a scenario-based question-and-answer model, to achieve safe and accurate intelligent interactive services in the gas field. It includes the following steps: Step 1: Build a knowledge base by acquiring multi-source data from multiple platforms; segment the data into text blocks according to segmentation rules, and automatically generate vectorized indexes through QA segmentation to form a local question-answering knowledge base; Step 2: Build a scenario model. Construct multimodal scenario models based on production scenarios, transportation scenarios, and home scenarios, and establish vector indexes for the keyword library and knowledge base under the corresponding scenarios. Step 3: Input analysis: convert user input into text and directly enter the semantic understanding layer to parse the question intent, analyze and extract keywords and question intent, and combine the scenario keyword library to perform multi-dimensional scenario judgment; Step 4: Match the scenario mode according to the keyword and call the corresponding knowledge base, perform question-answering service in the corresponding scenario mode, and generate differentiated scenario response output results.

[0008] Solution 2: Provide a gas customer service question-and-answer system based on an interactive multi-functional synthesis algorithm, which is applied to the aforementioned gas customer service question-and-answer method based on an interactive multi-functional synthesis algorithm and includes a data fusion module, a scenario construction module, a semantic understanding module, a question-and-answer generation module, and a safety control module. The data fusion module is used to collect and process multi-source data; The scenario construction module is used to construct three-dimensional scenarios including production, transportation, and home, and configure a professional terminology library and spoken expression mapping relationship for each scenario; The semantic understanding module is used to analyze and extract keywords and question intent from input information; The question-answer generation module is used to retrieve and analyze the input questions and output answer information based on the scenario mode; The security control module is used to establish a security monitoring mechanism and retain information.

[0009] The principles and advantages of this solution are: This solution aims to ensure that the Q&A approach can both accurately understand professional questions and answer everyday questions for non-professionals in easy-to-understand terms. This solution divides and matches the knowledge base according to scenario modes, so only one scenario needs to be invoked at the time of use. This significantly reduces the amount of data calculations while ensuring comprehensive and effective problem analysis. It can provide precise guidance including technical parameters, and can also guide everyday problems for ordinary home users, automatically converting them into simple and clear solutions and operational demonstrations.

[0010] Traditional gas customer service question-and-answer systems, in order to ensure professional and understandable problem handling, generally mix professional terminology with everyday language. In actual use, this requires constant switching. While this ensures accurate problem analysis, it significantly increases the amount of data processing, impacting system response efficiency, and can only provide static, standardized answers. At the same time, when dealing with non-professionals, it is difficult to accurately output effective information, and there is no guarantee that users can fully understand the output content, which impacts the user experience. Secondly, when faced with emergencies or complex issues, it is difficult to quickly determine and respond in a timely manner, and the answers given are often fragmented and difficult to accurately apply.

[0011] By integrating multi-channel gas industry knowledge, this solution not only distinguishes between the three major scenarios of factories, pipelines, and homes, but also refines multiple sub-scenarios within each scenario to ensure the accuracy of scenario judgment and understand the differences in expression in different scenarios, thereby providing more appropriate problem analysis and answers. Secondly, this solution breaks through the inherent customer service question-and-answer model of asking first and then answering. By analyzing subtle changes in the conversation process and combining matching scenario models, it can predict the user's real needs in advance, thereby realizing a proactive service model of "service before questions are asked", and then automatically linking into a complete service process, improving service efficiency and ensuring the safety and standardization of gas use, improving the accuracy and fit of docking, meeting the needs of different users, and improving the overall user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 The figure is a flow chart of the gas customer service question-and-answer method based on the interactive multi-energy synthesis algorithm of the present invention.

[0013] Figure 2 The figure is a schematic diagram of the structure of the gas customer service question-and-answer system based on the interactive multi-functional synthesis algorithm of the present invention. DETAILED DESCRIPTION

[0014] The following is further described in detail through specific implementation methods: Example 1 The gas customer service question-and-answer method and system based on the interactive multi-functional synthesis algorithm in this embodiment realizes safe and accurate intelligent interactive services in the gas field by integrating multi-source knowledge bases and multi-modal large model capabilities and combining scenario-based question-and-answer modes.

[0015] In this embodiment, the first solution is a gas customer service question-answering method based on an interactive multi-functional synthesis algorithm, as shown in the attached Figure 1 As shown, the following steps are included: S1, build a local question-answering knowledge base.

[0016] In this embodiment, the local knowledge base is a fusion of multi-source knowledge bases, thereby integrating the knowledge bases to achieve unified management of multi-source knowledge, improve retrieval efficiency and coverage, and avoid knowledge dispersion and update delays.

[0017] The framing process includes the following sub-steps: S1.1, multi-source data acquisition. Obtain data from the internal knowledge base, legal and regulatory base, and industry popular science base to obtain multi-source data. Among them, the internal knowledge base is internal documents, including data such as customer service records, equipment maintenance manuals, and troubleshooting cases. The legal and regulatory base is public regulations or documents, including regulatory documents such as the "Town Gas Management Regulations" and "Gas Engineering Project Specifications." The industry popular science base is data such as relevant texts, safety guidelines, graphic and text popular science, and videos published by associations. At the same time, pre-process the unstructured data, extract key frames or generate text descriptions to ensure that multimodal information is searchable.

[0018] S1.2. Clean and structure the acquired data. In this embodiment, the acquired multi-source data is preprocessed in the following manner, and custom segmentation rules, such as the "###" delimiter, are used to segment the text to ensure that each paragraph contains a complete question-answer pair or knowledge point.

[0019] The internal knowledge base is categorized and stored by business type, such as installation, maintenance, security inspection, and fees, ensuring more accurate structured retrieval. At the same time, typical problem templates, such as "How to apply for a gas meter," are extracted and associated with solutions.

[0020] For the legal and regulatory database, we first verify the timeliness of the legal and regulatory provisions, delete the abolished content, and then mark them in sections according to themes to facilitate accurate retrieval.

[0021] For the industry science popularization library, text summaries are generated for the obtained pictures, videos and other content. For example, the "gas leak emergency treatment video" is summarized as "1. Close the valve; 2. Open the window for ventilation..." and multimedia resources are associated.

[0022] At the same time, professional terms are standardized to avoid semantic ambiguity.

[0023] S1.3, constructing a vectorized index. In this embodiment, the QA segmentation function of the FastGPT platform can be used to convert the segmented text into a vector representation and establish a semantic index to support subsequent similarity retrieval. Associated tags are generated for multimedia data such as images and videos, and a multimedia resource index is established. This index is mapped to the text knowledge base to facilitate multimodal access.

[0024] S2, build a scenario model. In this embodiment, a multimodal scenario model is constructed according to the production scenario, transportation scenario, and home scenario, and a vector index of the keyword library and knowledge base for each corresponding scenario is established. This improves the accuracy of scene recognition, enables the question-and-answer service to better solve practical problems, and improves system user satisfaction. Specifically, it includes the following sub-steps: S2.1, define the scenario classification system. In this embodiment, based on the specific application scenarios of gas, the gas service scenarios are systematically divided into three categories: production scenarios, transportation scenarios, and household scenarios.

[0025] The production scenario involves production operations within factories and production lines, focusing on factory equipment operation and process control, and primarily serves production technicians. The transportation scenario involves pipeline operations between stations, specializing in pipeline network maintenance, and primarily serves maintenance and inspection personnel working on pipelines. The home scenario involves household gas stoves, etc., focusing on daily usage issues, and primarily serves ordinary residential users. For each scenario type, sub-dimensions are further refined, such as breaking down production scenarios into subcategories such as equipment operation, process control, and safety management, to form a hierarchical scenario knowledge system. This structured classification achieves a scenario coverage rate of 98%, which is 40% more accurate than traditional flat classification.

[0026] S2.2. Construct a scenario keyword library. In this embodiment, multi-level keyword extraction is used to extract high-frequency keywords for different scenarios, such as "abnormal pressure, pressure reducing valve, equipment failure" in production scenarios; "pipeline inspection, valve maintenance, leak detection" in transportation scenarios; and "gas stove ignition, fee inquiry, gas leak" in household scenarios. A three-dimensional vocabulary is constructed, encompassing core terms, colloquial expressions, and regional dialects. Through standardization of professional terminology (e.g., "PRV" becomes "pressure reducing valve") and dialect mapping (e.g., "gas hose" becomes "gas hose"), a scenario vocabulary of over 2,000 annotated terms is generated. Furthermore, an expanded synonym library is established, combining industry terminology and user colloquial expressions, and associations are performed to improve intent recognition coverage and ensure accuracy and comprehensiveness. The obtained high-frequency keywords are classified and a scenario vector index is generated to support semantic search.

[0027] S2.3. Associate scenarios with the knowledge base, making the knowledge base scenario-based. In this embodiment, a corresponding knowledge base subset is assigned to each keyword in each scenario, and independent search weights are established to ensure that professional terms are prioritized for production scenarios, while simple answers are prioritized for home scenarios. This optimizes search efficiency and improves the accuracy of the question-answering service.

[0028] In this embodiment, the knowledge base weight is constructed according to the scenario, terminology and knowledge type. In the scenario dimension, the basic weight ratios are set to 0.6 for production, 0.3 for delivery and 0.1 for family; in the terminology dimension, the semantic weight ratios are set to 0.7 for professional terminology and 0.3 for popular expression; in the knowledge type, the content weight ratios are set to 0.5 for SOP, 0.3 for regulations and 0.2 for popular science. In this way, a three-dimensional retrieval weight model is constructed. The corresponding scenario is matched by the final retrieval score. The final retrieval score can be expressed as Score = (scenario matching * α) + (term matching * β) + (knowledge type relevance * δ); Among them, α, β, and δ are corresponding coefficients, and α+β+δ=1; the default settings can be α=0.5, β=0.3, and δ=0.2, and can be dynamically adjusted according to the scenario.

[0029] At the same time, scene features are extracted. In this embodiment, scene keywords can be enhanced based on the improved TF-IDF algorithm. For example, the weight of production scene words is set to original TF-IDF * 1.5; the weight of family scene words is set to original TF-IDF * 0.8. This enhances the weight of keywords in each scene, making it easier to establish the correlation between the scene and the knowledge base, and improving matching accuracy. Based on the constructed weight model and scene features, the scene is associated with the knowledge base, and the knowledge base is annotated according to the weight dimension.

[0030] S3, input analysis, converts user input content into text and directly enters the semantic understanding layer to parse the question intent, analyzes and extracts keywords and question intent, and combines the scene keyword library to perform multi-dimensional scene judgment.

[0031] In this embodiment, the text or voice information input by the user is converted into text using an automated speech recognition (ASR) engine. This engine has been specifically optimized for the gas industry, effectively improving the recognition accuracy of specialized gas terminology. Typos are detected and corrected directly on the text input. Using an edit distance algorithm and domain dictionary matching, the user's input is analyzed to determine whether there are any errors in the logical and semantic meaning. Automatic error correction and standardization are then performed to improve the accuracy of the input and enhance the usability of the original input.

[0032] Use a pre-trained model, such as GasBERT, to analyze the deep semantics of the input information, perform semantic understanding, identify the core intent, obtain its keywords and question intent judgment. In this embodiment, the subject-verb-object structure can be extracted through dependency syntactic analysis, the core verbs and key entities can be identified, and surface analysis can be performed. Then calculate the cosine similarity between the question vector and the intent template, and perform intent judgment based on the calculated similarity, where the cosine similarity I can be expressed as ; Where Q is the question vector, I is the intention template; w is the scene weight coefficient.

[0033] Based on the extracted keywords and the constructed scenario keyword library, the initial scenario probability distribution is calculated to preliminarily determine the scenario to which the problem belongs.

[0034] In this embodiment, scene determination is performed based on multi-feature fusion calculation, and the weighted fusion score of lexical features, semantic features and context features is calculated to calculate the scene probability matching the corresponding scene model. The scene probability calculation method can be expressed as F(a)=0.6*lexical score+0.3*semantic score+0.1*context score; the lexical score calculates word frequency and inverse scene frequency based on the matching scene keyword library; the semantic score calculates the scene semantic density through the scene word vector; and the context score is the probability of extracting scene continuation clues from the conversation history.

[0035] In addition, when the identified keywords involve more than one scenario, that is, when the calculated probability difference of multiple scenarios is less than 0.15, the intelligent questioning mode is triggered. For example, if there is a conflict between the production scenario and the home scenario, and the matching degree difference is small, further questioning will be automatically conducted to improve the accuracy of scene judgment.

[0036] S4, matches the scenario mode according to the keywords and calls the corresponding knowledge base, performs question-answering services in the corresponding scenario mode, and generates differentiated scenario response output results.

[0037] Based on the intent analysis results, the corresponding scenario mode is selected for matching and the corresponding question-and-answer scenario mode is triggered. After invoking the corresponding scenario mode, the relevant knowledge base content for that scenario is optimized and retrieved. Vector similarity calculation is used to match the most relevant answer paragraph from the knowledge base.

[0038] Specifically, the system provides differentiated response strategies based on the characteristics of different scenarios. When a problem is identified as a production scenario, the system accurately accesses the factory's standard operating procedure (SOP) knowledge base, automatically extracts key control points (such as pressure thresholds and torque standards) in the operating steps, and adds regulatory traceability. The output includes professional answers with detailed equipment operating steps. For example, when dealing with abnormal pressure in a pressure reducing valve, the system provides step-by-step instructions for closing the upstream valve and checking the seal, along with relevant regulatory references and emergency contact information.

[0039] For consultations on transportation scenarios, the system will provide technical answers based on the specific terms of the gas engineering specifications. It also integrates GIS spatial analysis. When the problem involves a specific pipe section, it automatically links the maintenance records and responsible person information of the area, ensuring that users can obtain authoritative guidance and quickly contact the responsible personnel when encountering professional problems such as pipeline maintenance or leaks.

[0040] In services for home users, the system proactively translates professional terminology into popular expressions and technical documentation into step-by-step guidance to simplify terminology, such as simplifying "metal bellows" to "gas hose." It also uses multimedia resources such as short videos to more intuitively demonstrate solutions to common problems like troubleshooting gas stove ignition failure, significantly reducing the complexity for average users. This scenario-based intelligent response mechanism not only ensures accurate answers to professional questions, but also enhances the user experience for home users, precisely matching service content with user needs.

[0041] In this embodiment, if the question involves an operation demonstration, etc., while returning the text answer, the corresponding instructional video or schematic diagram will be pushed, and step-by-step answers will be generated for complex questions to improve operability and comprehensibility. This embodiment also includes dynamically adjusting the scene. During the question-answering process, the scene matching degree is determined in real time according to the input content, and the corresponding scene mode is switched according to the scene matching degree.

[0042] When a user enters content into the system, it quickly performs an initial scene prediction, followed by in-depth semantic analysis. During the ongoing conversation, the scene weights are dynamically adjusted. A sliding window is used to analyze the semantic trajectory of the last three rounds of conversation, and the scene transition probability is calculated to determine the current scene's matching degree. If significant scene drift is detected, the system automatically triggers the scene redirection process, dynamically adjusting the scene by recalculating the scene probability and using confirmatory questions to ensure a smooth transition.

[0043] This embodiment also includes security and refusal to answer. The input information is detected and filtered. When unconventional or sensitive words are detected, such as financial, political and other sensitive or security issues, verification prompt information is automatically blocked or output, and it is indicated that it is not a gas issue.

[0044] Option 2 A gas customer service question-answering system based on an interactive multi-energy synthesis algorithm is provided, which is applied to the gas customer service question-answering method based on an interactive multi-energy synthesis algorithm, as shown in the attached figure. Figure 2 As shown, it includes data fusion module, scenario construction module, semantic understanding module, question and answer generation module and security control module.

[0045] In this embodiment, the data fusion module is used to collect and process multi-source data. It includes a data collection unit and a data processing unit. The data collection unit is used to regularly capture the latest regulatory documents via an API interface connected to the government's public data platform. It also uses web crawler technology to obtain popular science materials published by industry associations. It also integrates internal customer service records and equipment maintenance documents. An automated data update mechanism is established to ensure knowledge base synchronization within 24 hours of policy changes.

[0046] The data processing unit is used to filter non-text content using regular expressions, then standardizes the expression using specialized terminology tools, and then verifies the accuracy of key data. Furthermore, an AI-powered automatic tagging function is added to video footage to generate structured metadata descriptions.

[0047] The scenario construction module is used to construct a three-dimensional scenario model covering production, transportation, and home, and configures a professional terminology library and spoken expression mapping relationship for each scenario. It includes a scenario knowledge graph unit and a dynamic scenario judgment unit. Among them, the scenario knowledge graph unit constructs a three-dimensional scenario model based on the experience of industry experts and others. In the production dimension, it includes sub-graphs such as equipment management and process specifications. In the transportation dimension, it covers nodes such as pipeline maintenance and emergency treatment. In the home dimension, it establishes a full life cycle relationship network from installation to troubleshooting. It also uses a graph database to store more than 500 entity nodes and 2000+ relationship edges, and configures a corresponding knowledge base.

[0048] The dynamic scenario determination unit uses fusion rules and machine learning models to configure 200+ keyword rules after primary filtering, and uses a BERT fine-tuned model for intent classification, matching the intent to the corresponding scenario model. Furthermore, this embodiment includes a scenario confusion detection mechanism that initiates a multi-round interactive confirmation process when a user's question contains multiple cross-scenario terms.

[0049] The semantic understanding module analyzes and extracts keywords and question intent from input information. It includes a multimodal input parsing unit and a context awareness unit. The multimodal input parsing unit parses the input information. In this embodiment, a gas-related terminology library is incorporated into the general ASR framework, while coreference resolution technology is also incorporated into text parsing to automatically associate questions with the specific devices or scenarios mentioned above.

[0050] The context-aware unit is used to establish a dialogue state tracking mechanism. It automatically assembles the complete query intent based on the context of the user's input. Furthermore, a regionalized language understanding model is established to cover common expressions in the Northwest region, taking into account the characteristics of local dialects.

[0051] The question-answer generation module performs search and analysis on input questions and outputs answers based on scenario patterns. It includes a search and combination unit and a multimodal presentation unit. The search and combination unit uses a hybrid search approach, using an inverted index to quickly locate relevant documents within the corresponding knowledge base and then using vector similarity for refined ranking. Furthermore, a priority-based approach is used to weight answers related to safety regulations.

[0052] The multimodal presentation unit is used to visualize results based on different scenarios. It automatically generates a step-by-step list for operational questions and presents data query questions in a table format. When a user asks a practical question, it automatically associates the corresponding timestamp of the demonstration video.

[0053] The security control module establishes a security monitoring mechanism and retains information. By implementing bidirectional logging for both input and output, interaction information is retained for 180 days, enabling full-link information tracking. Furthermore, during interactions, the module detects the number of unconventional or sensitive words in real time. If a hit occurs three times in a row, a verification process is automatically triggered to ensure the security and stability of information during interactions and prevent the leakage of critical information.

[0054] In this embodiment, by creating multi-scenario adaptive adjustment, fast and accurate scene matching and switching are achieved, so that the system can provide professional guidance on factory technical issues and return popular answers to daily family consultations like an experienced master. Secondly, a dynamic knowledge graph is established to improve scene adaptability and problem-solving efficiency by integrating meteorological data, equipment status and user habits in real time. At the same time, visual guidance is integrated with spoken answers to shorten the time to understand complex problems and improve overall service efficiency. It has both professional accuracy and civilian affinity, thereby breaking through the constraints of the static knowledge base system and improving service fit.

[0055] Example 2 This embodiment also includes introducing a 3D schematic library based on the scenario type when constructing the scenario model in S2, providing interactive 3D analytical models for the complex equipment structures in each scenario. Specifically, industrial-grade 3D modeling can be used to decompose key gas equipment (such as pressure reducing valves and flow meters) into interactive 3D models. Through parametric modeling, each component is associated with 20-50 technical attributes (material specifications, operating pressure range, etc.), supporting multi-dimensional cross-sectional viewing. When a user queries "pressure reducing valve structure," the system automatically retrieves the 3D model and highlights the relevant components, while also overlaying dynamic data layers (such as pressure transmission animations). This greatly improves the efficiency of understanding complex equipment issues while also reducing training costs.

[0056] At the same time, a text-to-picture comparison system was developed. This system uses a multimodal alignment algorithm based on the Transformer model's attention mechanism to calculate the similarity between text descriptions and visual features. Based on this similarity, a mapping relationship between text terms and visual elements is established. In this embodiment, a similarity greater than 0.85 is considered a match. For example, for documents such as the "Gas Engineering Construction Specifications," computer vision technology is used to automatically extract key elements from illustrations (such as safety distance markings and welding nodes) and establish a bidirectional index with the text terms. When a user queries for "pressure reducing valve structure," the system first performs semantic analysis, identifying the key entity "pressure reducing valve" and querying for the intended meaning, such as "structural description." At this point, the corresponding visual model is simultaneously activated to detect all visual areas in the illustration that may contain "pressure reducing valve," such as the valve body, spring assembly, and pressure gauge interface. Cross-modal alignment and retrieval are then performed. For example, all clauses in the "Gas Engineering Construction Code" that mention "pressure reducing valve structure" are located (such as Article 5.2.3: "The pressure reducing valve shall consist of a valve body, an adjusting spring, a diaphragm..."). The image and text are then mapped. For example, the various components of the pressure reducing valve are framed in the document illustration through target detection and mapped to the descriptions in the text clauses, forming a bidirectional association. Finally, the search results are displayed, such as displaying text on the left side of the display interface and the corresponding image on the right side. When the mouse hovers over any component, the text on the left automatically switches to the detailed content of the corresponding text clause. This establishes a strong association between the safety regulations and the corresponding scene images, greatly improving the efficiency of consulting safety regulations and reducing illegal operations.

[0057] Secondly, in S1.2, the internal knowledge base construction also includes device type identification. This assists with scenario judgment by constructing gas equipment feature terms for each scenario. For example, according to the multiple structured categories set by classification, corresponding device types are added under the corresponding categories, and acquired images, videos, and other information are also associated. In this embodiment, a hierarchical classification approach can be used for device type identification. For example, a quick classification is performed based on the device name. For example, for "pressure reducing valve structure," a lightweight TextCNN (text) or YOLOv8 (image) model is used for a quick initial screening, matching keywords such as "pressure reducing valve, pressure regulator, pressure regulator." Technical parameters are then analyzed for further refined identification, such as extracting technical parameters from associated terms and performing pixel-level dimensional estimation of the detected valve body. Finally, the final application scenario is determined based on the context. This improves the accuracy of scenario judgment through device-related queries, particularly in mixed scenarios, further reducing misjudgments and enhancing the user experience.

[0058] In this embodiment, the user's understanding and operational standardization are enhanced through visual presentation, and the speed of identifying key safety points is increased, thereby improving the user experience of this system, achieving efficient transfer of knowledge in various scenarios, and improving technical service standards in the gas field.

[0059] Example 3 In this embodiment, during the use of the system, a dynamic evolution system is also established: weekly, misjudgment cases are automatically analyzed based on output results and usage feedback to update model parameters; at the same time, professional terminology is manually injected when new equipment is put into production.

[0060] In actual application, we automatically analyze service record data, such as rejected answers and proactive corrections, from user conversation logs stored in internal documents weekly. By extracting semantic and contextual features, we identify potential knowledge gaps and automatically classify similar misjudgment cases using algorithms such as the density-based anomaly detection algorithm (DBSCAN). For frequently misjudged cases, we automatically generate knowledge completion suggestions and push them to the review queue. After confirmation by domain experts, model parameters are automatically updated. This improves the effectiveness and coverage of the knowledge base and reduces the misjudgment rate.

[0061] At the same time, when new equipment is put into production, the equipment model can be constructed through visual annotation to update the internal knowledge base, such as annotating structured data such as standard names, model specifications, and associating applicable scenario rules to update the semantic model, ensure the stability and real-time performance of the model, and ensure the accuracy of new equipment terminology recognition.

[0062] Example 4 This embodiment also includes dynamically adjusting the weight of the Q&A service based on seasonal characteristics in S4. Based on the unique cyclical patterns of the gas industry, the weight of the Q&A service is dynamically adjusted. By analyzing historical service data, environmental parameters, and operational characteristics, a multi-dimensional seasonal impact model is established to achieve automatic optimization and adjustment of knowledge services.

[0063] Among them, the seasonal adjustment coefficient is calculated by weighted fusion, and the question-answering service weight is dynamically adjusted according to the calculated seasonal adjustment coefficient. The seasonal adjustment coefficient can be expressed as Seasonal coefficient = a*historical query distribution + b*meteorological data + c*equipment operating conditions + d*policy events; Among them, a is the basic weight of historical rules, and the seasonal characteristic value can be calculated by counting the frequency of question types in the same period of the past three years; b is the real-time impact weight of temperature / humidity, etc., and the meteorological impact can be calculated by obtaining meteorological indicators such as temperature and precipitation; c is the impact coefficient of equipment operation load factors; d is the impact coefficient of temporary adjustments to policies and regulations. The weight of the Q&A service is dynamically adjusted according to the calculated seasonal coefficient. For example, in winter, the seasonal coefficient is between 1.2-1.5. For home scenarios, the weight of anti-freeze and insulation questions in the Q&A service is increased by 1.8 to improve the applicability of the Q&A service. For production scenarios, standard priority services such as low-temperature operation can be improved; for transportation scenarios, solutions such as pipeline ice blockage can be placed at the top for adjustment.

[0064] For the rainy season, the seasonal coefficient is between 1.0 and 1.3, which can enhance the display of knowledge such as pipeline drainage and anti-corrosion detection, or improve the response speed of underground valve well-related queries.

[0065] At the same time, during peak gas usage periods, when the seasonal coefficient is between 0.8 and 1.0, the interaction process for non-urgent issues can be simplified and multiple high-frequency services can be processed in parallel to expand channel capacity, avoiding service delays or congestion. This improves the efficiency of resolving seasonal issues, prevents risks based on seasonal cycles, and reduces the failure rate caused by extreme weather. Dynamic adjustment can also optimize the allocation of service resources between peak and off-season, reducing resource waste and fluctuations, improving comprehensive management capabilities, and effectively enhancing the user experience to meet actual usage needs.

[0066] The above is only an embodiment of the present invention, and the common knowledge such as the specific technical solutions and / or characteristics in the solution are not described in detail here. It should be pointed out that for those skilled in the art, without departing from the technical solution of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the description can be used to interpret the content of the claims.

Claims

1. A gas customer service question-answering method based on an interactive multi-functional synthesis algorithm, characterized in that: The following steps are involved: Step 1: Build a knowledge base by acquiring multi-source data from multiple platforms; segment the data into text blocks according to segmentation rules, and automatically generate vectorized indexes through QA segmentation to form a local question-answering knowledge base; Step 2: Build a scenario model. Construct multimodal scenario models based on production scenarios, transportation scenarios, and home scenarios, and establish vector indexes for the keyword library and knowledge base under the corresponding scenarios. Step 3: Input analysis: convert user input into text and directly enter the semantic understanding layer to parse the question intent, analyze and extract keywords and question intent, and combine the scenario keyword library to perform multi-dimensional scenario judgment; Step 4: Match the scenario mode according to the keyword and call the corresponding knowledge base, perform question-answering service in the corresponding scenario mode, and generate differentiated scenario response output results.

2. The gas customer service question-answering method based on the interactive multi-functional synthesis algorithm according to claim 1 is characterized by: In the production scenario, when a keyword is triggered, the factory's SOP knowledge base is called, and the equipment operation steps are included in the output results; in the transportation scenario, answers are given in combination with the requirements of gas engineering specifications and associated with the maintenance records and responsible person information of the area; in the home scenario, the terms are simplified when outputting the results, and short video answers are associated.

3. The gas customer service question-answering method based on the interactive multi-functional synthesis algorithm according to claim 1 is characterized in that: The multi-source data includes an internal knowledge base, a legal and regulatory base, and an industry science base; and is pre-processed as follows: The internal knowledge base is structured and classified according to business types; The legal and regulatory database is marked in sections after verification; Industry science database, converted into text summaries and associated with multimedia resources.

4. The gas customer service question-answering method based on the interactive multi-functional synthesis algorithm according to claim 1 is characterized in that: In step 2, the following sub-steps are included: Step 2.1: Define a scenario classification system; systematically divide gas service scenarios into three categories: production scenarios, delivery scenarios, and household scenarios; Step 2.2: Build a scenario keyword library. Extract high-frequency keywords for different scenarios and build a three-dimensional vocabulary library containing core terms, colloquial expressions, and regional dialects. Step 2.3, associate the scenario with the knowledge base; Build retrieval weights for the knowledge base according to scenarios, terms, and knowledge types, calculate retrieval scores, and match corresponding scenarios based on the retrieval scores.

5. The gas customer service question-answering method based on the interactive multi-functional synthesis algorithm according to claim 1 is characterized in that: In step 3, the scene is determined by calculating the weighted fusion score of lexical features, semantic features and context features to match the corresponding scene model, which is expressed as F(a)=0.6*lexical score+0.3*semantic score+0.1*context score.

6. The gas customer service question-answering method based on the interactive multi-functional synthesis algorithm according to claim 1, characterized in that: In step 4, it also includes dynamically adjusting the scene. During the question-answering process, the scene matching degree is determined in real time according to the input content, and the corresponding scene mode is switched according to the scene matching degree.

7. The gas customer service question-answering method based on the interactive multi-functional synthesis algorithm according to claim 1 is characterized by: In step 4, the weight of the question-answering service is dynamically adjusted according to seasonal characteristics. By analyzing historical service data, environmental parameters and operational characteristics, a multi-dimensional seasonal impact model is established. The seasonal adjustment coefficient is calculated by weighted fusion, and the weight of the question-answering service is adjusted according to the seasonal adjustment coefficient. The seasonal adjustment coefficient can be expressed as Seasonal coefficient = a*historical query distribution + b*meteorological data + c*equipment operating conditions + d*policy events; Where a is the basic weight of historical rules; b is the real-time influence weight of temperature / humidity, etc.; c is the influence coefficient of equipment operation load factors; and d is the influence coefficient of temporary adjustments to policies and regulations.

8. A gas customer service question-answering system based on an interactive multi-functional synthesis algorithm, characterized by: A gas customer service question-and-answer method based on an interactive multi-functional synthesis algorithm as described in any one of claims 1 to 7, comprising a data fusion module, a scenario construction module, a semantic understanding module, a question-and-answer generation module, and a safety control module; The data fusion module is used to collect and process multi-source data; The scenario construction module is used to construct three-dimensional scenarios including production, transportation, and home, and configure a professional terminology library and spoken expression mapping relationship for each scenario; The semantic understanding module is used to analyze and extract keywords and question intent from input information; The question-answer generation module is used to retrieve and analyze the input questions and output answer information based on the scenario mode; The security control module is used to establish a security monitoring mechanism and retain information.

9. The gas customer service question-and-answer system based on the interactive multi-functional synthesis algorithm according to claim 8 is characterized by: The scenario construction module includes a scenario knowledge graph unit and a dynamic scenario determination unit; the semantic understanding module includes a multimodal input parsing unit and a context perception unit; and the question and answer generation module includes a retrieval combination unit and a multimodal presentation unit.

10. The gas customer service question-and-answer system based on the interactive multi-functional synthesis algorithm according to claim 8, characterized in that: It also includes the introduction of a three-dimensional schematic library to provide an interactive three-dimensional analytical model for complex equipment structures; and the establishment of a graphic-text comparison system to establish a mapping relationship between text terms and visual elements.

Images