Intelligent voice interaction method and system for anti-electricity-stealing field inspection
By combining a large language model system with front-end devices and back-end servers, real-time and standardized operation guidance for anti-electricity theft inspections has been achieved, solving the problem of unreliable output in existing technologies and improving the standardization and efficiency of inspections.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-22
- Publication Date
- 2026-04-03
AI Technical Summary
In existing technologies, intelligent voice assistants for anti-electricity theft inspections cannot provide deep understanding, their output is unreliable, and they are difficult to provide efficient and standardized on-site operation guidance in complex environments.
A large language model system combining front-end devices and back-end servers enables real-time and standardized job guidance through wake word activation, offline control command recognition, speech-to-text conversion, knowledge injection, and prompt word rule constraints.
It has improved the standardization and efficiency of anti-electricity theft inspections, ensured the accuracy and reliability of output, adapted to complex field environments, reduced network dependence and data leakage risks, and improved the uniformity and accuracy of inspections.
Smart Images
Figure CN121789666A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent power operation and artificial intelligence application technology, and in particular to an intelligent voice interaction method and system for on-site inspection of anti-electricity theft. Background Technology
[0002] Power supply service recorders are standard equipment for power field workers. Currently, their main functions are limited to recording and playing back audio and video, with a low level of intelligence. In complex anti-electricity theft inspection scenarios, inspectors need to be proficient in numerous regulations, procedures, and criteria for determining electricity theft. During on-site operations, personnel often rely on memory or manually referencing paper documents, resulting in inefficiency, inconsistent standard implementation, and the potential for overlooking key inspection points.
[0003] In the existing technology, there are some general-purpose voice assistants or mobile office apps, such as: 1. Mobile intelligent voice assistants used in electricity marketing operations focus on routine office processes such as work order inquiries and information retrieval, and cannot provide step-by-step real-time guidance for highly specialized on-site anti-electricity theft inspections.
[0004] 2. Some systems attempt to use voice for standardized information broadcasting or violation alarms, but their interaction is preset and one-way, and cannot cope with the complex and ever-changing free-response needs on site.
[0005] Furthermore, although large language models demonstrate strong conversational capabilities in general domains, their direct application in professional power field inspections carries risks such as uncontrollable output content, potential generation of speculative content inconsistent with knowledge documents, or inconsistencies with standard operating procedures, making it difficult to meet high reliability requirements.
[0006] Therefore, there is an urgent need in this field for a field operation guidance solution that can deeply understand specific business scenarios, output accurate and reliable content, and enable efficient human-machine collaboration through natural speech. Summary of the Invention
[0007] Based on the technical problems existing in the background technology, the present invention proposes an intelligent voice interaction method and system for on-site inspection of anti-electricity theft, which improves the standardization, efficiency and accuracy of inspection operations.
[0008] This invention proposes an intelligent voice interaction method for on-site inspection of electricity theft, comprising: The front-end device is activated by recognizing the set wake word; then, if the user's voice is recognized as a preset offline control command, the corresponding control operation is executed; if it is not recognized as a control command, the user's voice is converted into user question text and sent to the back-end server. The backend server sends the user's question text into a large language model for reasoning. The large language model generates a response text and returns it to the front-end device to provide real-time and standardized operational guidance to on-site personnel.
[0009] Furthermore, the backend server feeds the user's question text into a large language model for inference, specifically as follows: After receiving the user's question text from the front-end device, the back-end server determines whether the injection of knowledge documents and prompt word rules has been performed in this session; If this has not been performed before, first inject the knowledge document and prompt word rules into the large language model, and then send in the user question text; If this has already been done, then the user's question text will be sent directly.
[0010] Furthermore, the large language model generates a response text and returns it to the front-end device, specifically as follows: Return the response text generated by the large language model to the front-end device; The front-end device converts the text response into speech and plays it to the user.
[0011] Furthermore, if the user's voice is identified as a preset offline control command, the corresponding control operation is executed, specifically: the preset offline control command corresponding to the user's voice is identified and executed through the offline command word recognition module.
[0012] Furthermore, the knowledge document includes a pocket guide to on-site investigations of electricity theft.
[0013] Furthermore, the prompt word rules are used to constrain the behavior of the large language model. The prompt word rules include simulating the role of an on-site instructor, strictly answering questions based on document content, limiting the length of answers, and a unified response strategy for questions that are beyond the scope of the subject.
[0014] An intelligent voice interaction system for on-site inspection of anti-electricity theft includes a front-end device and a back-end server. The front-end device includes a voice wake-up module, an offline command word recognition module, and an offline voice dictation module. The back-end server includes a large language model. The voice wake-up module is used to recognize the set wake-up word; The offline command word recognition module is used to identify and execute preset offline control commands; The offline speech dictation module is used to convert user questions into user questions in speech and into text. The large language model performs inference based on the user question text sent by the front-end device, generates a response text, and returns it to the front-end device to provide real-time and standardized operation guidance for on-site personnel.
[0015] Furthermore, the front-end device also includes a speech synthesis module, used to convert the response text generated by the large language model into speech and play it to the user.
[0016] Furthermore, the backend server also includes a session management and knowledge injection module, which is configured to inject knowledge documents of a specified domain and predefined prompt word rules into the context of the large language model in one go during the first session that requires processing by the large language model.
[0017] Furthermore, the prompt word rules are used to constrain the behavior of the large language model. The prompt word rules include simulating the role of an on-site instructor, strictly answering questions based on document content, limiting the length of answers, and a unified response strategy for questions that are beyond the scope of the subject.
[0018] The advantages of the intelligent voice interaction method and system for on-site anti-electricity theft inspection provided by this invention are as follows: It provides real-time, standardized operational guidance to on-site personnel through voice question-and-answer; by deeply customizing the large language model into an "anti-electricity theft inspection domain expert" through "knowledge injection" and "prompt word engineering," it solves the problem of unreliable output from general AI in professional scenarios. It achieves a "hands-free" interaction mode, allowing inspectors to quickly obtain guidance in complex on-site environments without manually searching for information, significantly improving inspection efficiency. The hybrid architecture of "simple offline commands at the front end + complex online reasoning at the back end" ensures both the real-time performance of control operations and the intelligence of complex question-and-answer, with low network dependence and suitability for on-site environments. Through knowledge injection and rule constraints, it ensures that the output of the large language model is strictly based on domain knowledge documents, avoiding the generation of output content not based on injected knowledge. Attached Figure Description
[0019] Figure 1 This is a schematic diagram of the overall architecture of the present invention; Figure 2 This is a complete workflow diagram of the interaction method. Detailed Implementation
[0020] The technical solution of the present invention will now be described in detail through specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.
[0021] like Figure 1 and 2 As shown, the present invention proposes an intelligent voice interaction method for on-site inspection of electricity theft, comprising: The front-end device is activated by recognizing the set wake word; then, if the user's voice is recognized as a preset offline control command, the corresponding control operation is executed; if it is not recognized as a control command, the user's voice is converted into user question text and sent to the back-end server. The backend server sends the user's question text into a large language model for reasoning. The large language model generates a response text and returns it to the front-end device to provide real-time and standardized operational guidance to on-site personnel.
[0022] This embodiment can deeply integrate general large language model capabilities with professional knowledge of anti-electricity theft inspection, and provide on-site inspectors with a real-time, intelligent and reliable "AI instructor" through voice interaction, thereby improving the standardization, efficiency and accuracy of inspection operations.
[0023] That is, the key points of this embodiment are: (1) solidification of domain knowledge: through the "knowledge injection" of the first conversation, a general large language model is instantly transformed into a professional domain model, which is the core of achieving reliability. (2) prompt words as technical means: the prompt words in this embodiment are not simple questions, but a kind of "technical constraint means" used to precisely control the output behavior of the large language model and ensure that it conforms to industry standards, which is an important technical feature in itself.
[0024] In this embodiment, the above-mentioned interaction method is implemented based on an interaction system configured as follows, such as... Figure 1 As shown, specifically: It includes a front-end device and a back-end server. The front-end device is based on the existing power service recorder hardware and integrates multiple voice algorithm modules; the back-end server communicates with the front-end device.
[0025] The front-end equipment includes a voice wake-up module, an offline command word recognition module, an offline speech dictation module, and a speech synthesis module. The back-end server includes a large language model and a session management and knowledge injection module.
[0026] The voice wake-up module is used to recognize the set wake-up word (such as "Xiao e, Xiao e"). The offline command word recognition module is used to identify and execute preset offline control commands, such as "start recording", "stop recording", and "take a picture". The offline voice dictation module is used to convert user questions into spoken text. The speech synthesis module is used to convert the response text generated by the large language model into speech and play it to the user.
[0027] The large language model provides natural language understanding and generation capabilities, performs inference based on the user question text sent by the front-end device, generates response text, and returns it to the front-end device; The session management and knowledge injection module is configured to inject knowledge documents (such as the "Pocket Book on On-site Investigation of Electricity Theft") and predefined prompt word rules into the context of the large language model in the first session that requires processing by the large language model. The prompt word rules are used to constrain the behavior of the large language model. The prompt word rules include simulating the role of an on-site instructor, strictly answering according to the content of the document, limiting the length of the answer, and a unified response strategy for questions that are beyond the scope of the subject.
[0028] The purpose of setting up an offline command word recognition module in this embodiment is to ensure the absolute reliability and real-time performance of core interactive functions in environments without a network or with a weak network, while also strengthening system security and privacy protection. Its specific function is reflected in (a1) to (a4): (a1) Ensuring network-independent and instantaneous response for basic control: In power field settings (such as basements, remote substations, and signal-shielded areas), network connections may be unstable or completely interrupted. The offline command word recognition module can directly recognize and execute key voice commands on the front-end device in environments without or with weak network coverage. This local processing achieves millisecond-level response, avoiding basic operational failures due to network latency or interruptions, and ensuring the continuous execution of workflows and the acquisition of core data.
[0029] (a2) Reduce communication load and cloud computing pressure: Separate high-frequency, fixed, low-complexity control commands from free question and answer that requires complex semantic understanding.
[0030] (a3) Enhanced privacy and data security: All voice commands related to equipment control are recognized and executed locally, and their voice data does not need to be uploaded to the backend server. This avoids exposing control voices that may contain ambient sounds and fragments of personnel conversations to network transmission and cloud processing, meeting the high requirements of power operations for the confidentiality of on-site information and reducing the risk of data leakage.
[0031] (a4) Achieving deterministic, standardized process-driven operations: For standardized operating procedures such as anti-electricity theft, key inspection nodes (such as "verify meter information," "inspect junction box," and "measure current and voltage") can be set as command words. This guides operators to follow the standardized procedures. This enforces adherence to standardized procedures, reduces arbitrariness and omissions, and ensures uniformity in the inspection process.
[0032] The speech synthesis module in this embodiment is not only for text-to-speech conversion, but also a crucial component for improving human-computer collaboration efficiency and on-site adaptability, as detailed in (b1) to (b3): (b1) Achieving natural voice broadcasting of information: At anti-electricity theft inspection sites, the visual attention (observing meters and lines) and hands (operating tools and recording) of the operators are often heavily occupied. The speech synthesis module converts the text responses, process prompts, and key parameter announcements generated by the large language model into clear and natural speech in real time, which is then output through headphones or speakers. This allows operators to receive system guidance information through hearing without interrupting visual observation and manual operation.
[0033] (b2) Supports offline broadcasting: Even in the event of network interruption or loss of connection with the backend server, received text responses or pre-set standard operation prompts can still be synthesized into voice for broadcasting. This ensures that basic guidance information can still be reliably delivered during the most critical operational phases or emergency situations, and will not be lost due to network problems.
[0034] (b3) Complete the interactive loop and provide a smooth dialogue experience: As the system response (voice output) link for user voice input (asking questions), the speech synthesis module, together with the voice wake-up and voice dictation modules, constitutes a complete human-computer interaction loop based on natural speech. This makes the interaction process closer to human-to-human dialogue, lowers the threshold for use and learning costs, and makes the technology tools easier for frontline personnel to accept and use.
[0035] In this embodiment, during the initial dialogue, knowledge documents and prompt word rules are injected into the large language model. This ensures that the generation process of the large language model is not arbitrary, but rather a targeted, precise retrieval and organization within the "rule cage" and "knowledge repository" constructed by the conversation management and knowledge injection module. Its core is the combination of "prompt word engineering" and "context injection," as detailed in (c1) to (c4): (c1) Session initialization and strong constraint injection: By injecting role and behavior constraints and domain knowledge, the generation process of the large language model is restricted to a specific scope: Role and Behavioral Constraints: Includes predefined prompt rules, such as: "You are a strict, meticulous, and experienced anti-electricity theft field inspection supervisor. You must and can only answer questions based on the provided 'Anti-Electricity Theft Field Inspection Pocket Book.' Answers should be concise and direct, focusing on operational steps, judgment criteria, and safety points, and should not exceed three sentences per answer. If a user's question is outside the scope of this pocket book, you should uniformly answer: 'According to current operating procedures, this issue is not currently covered; it is recommended to record it and submit it for review.'" Domain knowledge injection: All or key content of standardized documents such as the "Pocket Book for On-site Investigation of Electricity Theft" is fully injected into the context in the form of structured text, serving as the sole source of knowledge for the large language model to answer questions.
[0036] (c2) User question input and controlled reasoning: When the user inputs the question text through the offline speech dictation module, the question text is attached to the above-initialized context and submitted to the large language model.
[0037] (c3) Ensure that the large language model answers based on the provided knowledge documents and strictly adheres to the preset rule constraints.
[0038] (c4) Context Management During Session: Throughout the entire inspection session, the session management module maintains this dialogue context. Each question and answer is recorded in history to ensure that the model remembers the previous dialogue content and the initial constraints it always follows when answering subsequent questions, thus maintaining logical coherence and the persistence of constraints.
[0039] This embodiment sets constraints in the large language model to solve (d1) to (d3): (d1) It solves the problem of "unprofessionalism and unreliability" of general voice assistants / large models in professional scenarios: through "knowledge injection + strong rule constraints", the powerful language capabilities of large language models are accurately "trained" to specific vertical fields, making their output content accurate, standardized and usable, and meeting the high reliability requirements of power field operations.
[0040] (d2) It solves the problems of inefficiency and inconsistent standards caused by on-site personnel relying on memory or manual review: It provides a real-time, intelligent, and conversational "personal expert" that allows inspectors to conduct natural language Q&A anytime and anywhere to obtain accurate guidance, ensuring the uniformity and integrity of complex standard execution and reducing human negligence and differences.
[0041] (d3) An "interactive" intelligent work guidance system was constructed, surpassing existing one-way broadcasting or simple retrieval systems: it realizes human-computer collaboration based on natural language, bidirectional, deep, and context-aware. The system can not only answer preset questions, but also understand the free, varied, and follow-up questions raised by inspectors in complex on-site situations, and can provide coherent guidance based on conversation history, thus forming a collaborative tool that can improve the level of intelligence in on-site operations.
[0042] Example 1 like Figure 2 As shown, the steps of the intelligent voice interaction method are as follows: S1: Voice Wake-up and Command Recognition Steps: The front-end device is activated via the voice wake-up module, and then attempts to match control commands via the offline command word recognition module. If the match is successful, the corresponding operation is executed and the process ends.
[0043] S2: Free speech processing steps: If no control command is matched, the user's speech is converted into text through the offline speech dictation module and sent to the backend server.
[0044] S3: Intelligent Reasoning and Constraint Output Steps: After receiving the user's question text from the front-end device, the back-end server determines whether the knowledge document and prompt word rules have been injected in this session. If yes, proceed to step S4; otherwise, proceed to step S5.
[0045] S4: Inject knowledge documents and prompt word rules into the large language model, and send the user question text into the large language model.
[0046] S5: The user's question text is directly fed into the large language model. After inference, the large language model generates the response text under the constraints of prompt word rules.
[0047] S6: Voice feedback steps: The backend server returns the response text to the frontend device, which then converts it into speech through the speech synthesis module and plays it to the user.
[0048] Example 2 Take a case study of an on-site inspection to prevent electricity theft.
[0049] (a1) The inspector says the wake word "Xiao e, Xiao e" and the recorder is awakened.
[0050] (a2) When the person says "start recording", the offline command word recognition module recognizes the instruction, the recorder starts recording, and the process ends.
[0051] (a3) The staff member asks, "How many steps are there in total?" This question is not a control command; the process then proceeds to a free-response question-and-answer session. The speech is transcribed into text and sent to the backend server.
[0052] (a4) The backend server determines that this is the first session requiring processing by the large language model during this inspection, and therefore injects the full text of the "Pocket Book for On-site Investigation of Electricity Theft" and the following prompts as system prompts into the large language model: "You are an anti-electricity theft field inspection supervisor. Please answer strictly according to the provided 'Pocket Guide to Anti-Electricity Theft Field Inspection', using a friendly tone and keeping your answers concise. If the question is outside the scope of the document, please answer: 'Sorry, we could not find the question you requested.'"
[0053] (a5) After receiving the question "How many steps are there in total?", the large language model generates a standardized six-step inspection process text based on the digested document content and returns it to the background.
[0054] (a6) The background sends the text back to the recorder, and the recorder announces in a clear voice: "Hello! The on-site inspection is divided into 6 standard steps: 1. Inspection of the outside of the metering box and the box body...".
[0055] (a7) Subsequently, inspectors can ask questions step by step according to the process, such as "What exactly do we do in the first step?" The system can provide accurate guidance based on the documentation within the same session context.
[0056] This embodiment has the following advantages (f1) to (f4): (f1) Deep integration of scenarios: For the first time, a large language model is deeply customized into an "expert in the field of anti-electricity theft inspection" through "knowledge injection" and "prompt word engineering", which solves the industry problem of unreliable output of general AI in professional scenarios.
[0057] (f2) Significantly improved efficiency: It realizes the "talk-only" interactive mode, allowing inspectors to quickly obtain guidance in complex on-site environments without having to manually look up information, thus greatly improving inspection efficiency.
[0058] (f3) Standardization of inspection process: AI ensures the uniformity of inspection process, standards and criteria, reduces missed inspections or misjudgments caused by differences in personnel experience, and improves the quality of operation.
[0059] (f4) System architecture optimization: The system adopts a hybrid architecture of "simple offline instructions at the front end and complex online reasoning at the back end", which not only ensures the real-time performance of control operations, but also realizes the intelligence of complex question and answer, with low network dependence and suitable for field environment.
[0060] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. An intelligent voice interaction method for on-site inspection of electricity theft, characterized in that, include: The front-end device is activated by recognizing the set wake word; Subsequently, if the user's voice is recognized as a preset offline control command, the corresponding control operation is executed. If the user's voice is not recognized as a control command, the user's voice will be converted into text and sent to the backend server. The backend server sends the user's question text into a large language model for reasoning. The large language model generates a response text and returns it to the front-end device to provide real-time and standardized operational guidance to on-site personnel.
2. The interaction method according to claim 1, characterized in that, The backend server sends the user's question text into a large language model for inference, specifically: After receiving the user's question text from the front-end device, the back-end server determines whether the injection of knowledge documents and prompt word rules has been performed in this session; If this has not been performed before, first inject the knowledge document and prompt word rules into the large language model, and then send in the user question text; If this has already been done, then the user's question text will be sent directly.
3. The interaction method according to claim 1, characterized in that, The large language model generates the response text and returns it to the front-end device, specifically as follows: Return the response text generated by the large language model to the front-end device; The front-end device converts the text response into speech and plays it to the user.
4. The interaction method according to claim 1, characterized in that, If the user's voice is identified as a preset offline control command, the corresponding control operation is executed. Specifically, the preset offline control command corresponding to the user's voice is identified and executed through the offline command word recognition module.
5. The interaction method according to claim 2, characterized in that, The knowledge document includes a pocket guide to on-site investigations and handling of electricity theft.
6. The interaction method according to claim 2, characterized in that, The prompt word rules are used to constrain the behavior of the large language model. The prompt word rules include simulating the role of an on-site instructor, strictly answering questions based on document content, limiting the length of answers, and a unified response strategy for questions that are beyond the scope of the subject.
7. An intelligent voice interaction system for on-site inspection of electricity theft, characterized in that, It includes a front-end device and a back-end server. The front-end device includes a voice wake-up module, an offline command word recognition module, and an offline voice dictation module. The back-end server includes a large language model. The voice wake-up module is used to recognize the set wake-up word; The offline command word recognition module is used to identify and execute preset offline control commands; The offline speech dictation module is used to convert user questions into user questions in speech and into text. The large language model performs inference based on the user question text sent by the front-end device, generates a response text, and returns it to the front-end device to provide real-time and standardized operation guidance for on-site personnel.
8. The interactive system according to claim 7, characterized in that, The front-end device also includes a speech synthesis module, which is used to convert the response text generated by the large language model into speech and play it to the user.
9. The interactive system according to claim 7, characterized in that, The backend server also includes a session management and knowledge injection module, which is configured to inject knowledge documents of a specified domain and predefined prompt word rules into the context of the large language model in one go during the first session that requires processing by the large language model.
10. The interactive system according to claim 9, characterized in that, The prompt word rules are used to constrain the behavior of the large language model. The prompt word rules include simulating the role of an on-site instructor, strictly answering questions based on document content, limiting the length of answers, and a unified response strategy for questions that are beyond the scope of the subject.