Power operation question and answer method and system based on offline knowledge base matching, medium and product

Abstract

This invention discloses a method, system, medium, and product for power operation question-and-answering based on offline knowledge base matching, relating to the field of artificial intelligence technology. The method includes: background processing: generating question-and-answer pairs based on power operation instructions using a large language model, the question-and-answer pairs including preset questions and corresponding standard answers; converting preset questions into question vectors, and packaging question-and-answer pairs and question vectors into an offline knowledge data package; terminal deployment: pre-storing the offline knowledge data package in the local storage module of a work recorder; offline question answering: in the absence of a network, the work recorder converts the collected voice questions into question text, then into question vectors, retrieves the question vector with the highest similarity to the question vector and greater than or equal to a preset threshold in the local storage module, extracts the corresponding standard answer, and broadcasts it via voice; this method forwards the intelligent reasoning task of the large language model to the background server, with the terminal only performing lightweight vector matching and voice interaction.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, system, medium, and product for power operation question answering based on offline knowledge base matching. Background Technology

[0002] Electrical work is characterized by high risk and high standardization. To ensure safety, workers often need to consult work instructions or safety regulations before and during operations. With the development of IoT and AI technologies, work recorders with intelligent question-and-answer capabilities have emerged. These recorders typically upload questions to a backend server via voice recognition, where a Large Language Model (LLM) retrieves the answer from the knowledge base.

[0003] However, the above solution has the following drawbacks: 1. Power operation sites are often located in areas without 4G / 5G signal coverage, such as the basement of substations, remote mountainous areas, or underground pipe corridors, where real-time communication cannot be guaranteed, causing the intelligent question and answer function to completely fail.

[0004] 2. Some existing solutions attempt to directly port large models to the recorder terminal, but the computing power of embedded devices is limited, and there are problems such as slow response speed (tens of seconds of delay), high power consumption, and poor understanding accuracy when running large models, which makes it difficult to meet the needs of real-time interaction on site.

[0005] Therefore, there is an urgent need for a technical solution that can quickly and accurately respond to power operation inquiries in a network-free environment. Summary of the Invention

[0006] Based on the technical problems existing in the background technology, this invention proposes a power operation question-answering method, system, medium and product based on offline knowledge base matching. The intelligent reasoning task of the large language model is moved to the back-end server to be completed, and the terminal only performs lightweight vector matching and voice interaction.

[0007] The power operation question-answering method proposed in this invention based on offline knowledge base matching includes: Backend processing: Using a large language model, question-answer pairs are generated based on the power operation instructions. Each question-answer pair includes a preset question and a corresponding standard answer. The preset questions are converted into question vectors, and the question-answer pairs and question vectors are packaged into an offline knowledge data package. Terminal deployment: The offline knowledge data package is pre-stored in the local storage module of the work recorder; Offline Q&A: In the absence of a network, the work recorder converts the collected voice questions into question text, then into question vectors. It then retrieves the question vector with the highest similarity to the question vector and a value greater than or equal to a preset threshold from the local storage module, extracts the corresponding standard answer, and broadcasts it on voice.

[0008] Furthermore, in the background processing, the large language model generates corresponding preset questions and standard answers based on preset prompt words for key operation nodes, safety specifications, and equipment parameters in the work instructions.

[0009] Furthermore, the step of retrieving the question vector in the local storage module that has the highest similarity to the question vector and is greater than or equal to a preset threshold specifically uses cosine similarity to calculate the similarity between the question vector and the question vector. The preset threshold is between 0.85 and 0.95.

[0010] Furthermore, in the terminal deployment, the work recorder selectively downloads the corresponding offline knowledge data package from the background server according to the job type. The job type includes at least one of the following: switching operation, transmission line inspection, cable head fabrication, and tower climbing operation.

[0011] Furthermore, the backend server uses a text vectorization model to convert the preset question into a question vector.

[0012] An electrical work recorder, comprising: The local storage module is configured to store offline knowledge data packages downloaded in advance from the backend server. The offline knowledge data packages are generated in advance by the backend server in the following ways: using a large language model to generate question-answer pairs based on the power operation instructions, the question-answer pairs include preset questions and corresponding standard answers; converting the preset questions into question vectors; and packaging the question-answer pairs and question vectors into offline knowledge data packages. The voice acquisition module is configured to collect voice questions from staff. The speech recognition module is used to convert voice questions into text questions. The vector calculation module is configured to convert the question text into a question vector and calculate the similarity between the question vector and each question vector in local storage; The retrieval and matching module is configured to select the question vector with the highest similarity and greater than or equal to a preset threshold, and extract the corresponding standard answer. The speech synthesis module is configured to read out the standard response. The main control module is electrically connected to the local storage module, voice acquisition module, voice recognition module, vector calculation module, retrieval and matching module, and voice synthesis module to control their coordinated operation.

[0013] Furthermore, it also includes: The communication module is configured to communicate with the backend server when a network is available to download or update the offline knowledge data package; The interface module is configured to import the offline knowledge data package to the work recorder via USB or a local wireless network.

[0014] Furthermore, the offline knowledge data package stored in the local storage module does not contain large language model parameters, but only question-answer pairs and question vectors.

[0015] A computer-readable storage medium, characterized in that the computer-readable storage medium stores a plurality of computer programs, the computer programs being invoked by a processor and used to execute the method described thereon.

[0016] A computer program product includes a computer program, characterized in that the computer program is executed by a processor to implement the steps of the method.

[0017] The advantages of the power operation question-answering method, system, medium, and product based on offline knowledge base matching provided by this invention are as follows: It adopts a cloud-side preprocessing and edge-side retrieval architecture, moving the intelligent reasoning task of the large language model to the backend server for completion, while the terminal only performs lightweight vector matching and voice interaction; it eliminates dependence on 4G / 5G networks, making it suitable for all signal blind spots in power operations; the terminal only performs ASR, vector calculation (millisecond level), and TTS, avoiding the generation delay of running a large model and meeting real-time interaction requirements; the question-answer pairs are generated by the backend large model based on a professional knowledge base, ensuring the authority and accuracy of the content; the vector matching method avoids comprehension biases of local small models; the recorder does not require a high-performance AI chip and can run smoothly using a conventional embedded processor, facilitating large-scale deployment. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the architecture of the present invention; Figure 2 This is a flowchart of the offline question-answering method in Example 2; Figure 3 This is a structural block diagram of the power operation recorder in Example 3. Detailed Implementation

[0019] 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.

[0020] like Figures 1 to 3 As shown, the power operation question-answering method based on offline knowledge base matching proposed in this invention includes: Backend processing: In the backend server, a large language model is used to generate question-answer pairs based on the power operation instructions. The question-answer pairs include preset questions and corresponding standard answers. The preset questions are converted into question vectors, and the question-answer pairs and question vectors are packaged into an offline knowledge data package. Terminal deployment: The offline knowledge data package is pre-stored in the local storage module of the work recorder; Offline Q&A: In the absence of a network, the work recorder converts the collected voice questions into question text, then into question vectors. It then retrieves the question vector with the highest similarity to the question vector and a value greater than or equal to a preset threshold from the local storage module, extracts the corresponding standard answer, and broadcasts it on voice.

[0021] This embodiment adopts a cloud-side preprocessing and edge-side retrieval architecture, which moves the intelligent reasoning task of the large language model to the backend server for completion, while the terminal only performs lightweight vector matching and voice interaction.

[0022] Specifically, the backend server first uses a high-performance large language model to automatically analyze the work instructions, generating pre-set questions that staff might ask and their corresponding standard answers, forming question-answer pairs. Simultaneously, a text vectorization model (Embedding model) is used to convert each pre-set question into a high-dimensional question vector. The question-answer pairs and question vectors are then packaged into an offline knowledge data package.

[0023] Understandably, the text vectorization model uses the open-source sentence-transformers / paraphrase-multilingual-MiniLM-L12-v2 model. This model is based on the Transformer architecture, supports Chinese and multiple languages, and outputs a 384-dimensional vector. The model has approximately 120 million parameters and can complete inference in an offline environment on embedded devices with low computing power (approximately 50ms for a single text-to-vector conversion @ ARM Cortex-A55).

[0024] The backend server uses a text vectorization model to convert each preset question text into a 384-dimensional floating-point vector and stores it in binary format. The work logger has a built-in identical text vectorization module (either directly reusing the same model or using a quantized 8-bit integer version) to convert the question text into a vector of the same dimension.

[0025] It should be noted that in the terminal deployment, the work recorder selectively downloads the corresponding offline knowledge data package from the backend server according to the job type. The job type includes at least one of the following: switching operation, transmission line inspection, cable head fabrication, and tower climbing operation.

[0026] This embodiment provides a power work recorder, including: The local storage module is configured to store offline knowledge data packages downloaded in advance from the backend server. The offline knowledge data package generation process is as follows: using a large language model to generate question-answer pairs based on the power operation instructions, the question-answer pairs include preset questions and corresponding standard answers; converting the preset questions into question vectors, and packaging the question-answer pairs and question vectors into an offline knowledge data package; The voice acquisition module is configured to collect voice questions from staff. The speech recognition module is used to convert voice questions into text questions. The vector calculation module is configured to convert the question text into a question vector and calculate the similarity between the question vector and each question vector in local storage; specifically, cosine similarity is used to calculate the similarity between the question vector and the question vector.

[0027] The retrieval and matching module is configured to select the question vector with the highest similarity and greater than or equal to a preset threshold (for example, the preset threshold is set to 0.85 to 0.95) and extract the corresponding standard answers; The speech synthesis module is configured to read out the standard response. The main control module is electrically connected to the local storage module, voice acquisition module, voice recognition module, vector calculation module, retrieval and matching module, and voice synthesis module to control their coordinated operation.

[0028] The local storage module can use an eMMC chip, such as the Kingston EMMC08G-M325 or other compatible chips with equivalent or larger storage capacity, and is soldered onto the recorder's main circuit board via a 153-ball BGA package. This eMMC chip integrates NAND flash memory and an MMC controller, providing 8GB of storage space, sufficient to hold offline knowledge data packages containing thousands of question-answer pairs.

[0029] The speech synthesis module uses the SYN6288 speech synthesis chip from Beijing Yuyin Tianxia Technology Co., Ltd. This chip connects to the main control module via an asynchronous serial port (UART) to receive the text data to be synthesized and realize the text-to-speech conversion. The SYN6288 operates at 3.3V~5V, has a built-in audio amplifier, can directly drive speakers, and features a simple hardware interface, fast response speed, and low cost, making it suitable for applications in embedded systems.

[0030] Before setting off, staff download the offline knowledge data package to the local storage module of the work logger via wired or wireless means, depending on the type of task for the day. This offline knowledge data package contains only text and vector data (question-answer pairs and question vectors), is small in size, and does not contain large language model parameters.

[0031] During on-site operations, the recorder captures staff's voice questions via microphone, which are then converted into question text by a speech recognition module (e.g., using a lightweight offline ASR chip). The recorder's built-in lightweight vector computing engine converts this text into question vectors and performs similarity calculations (e.g., cosine similarity) with all locally stored preset question vectors, retrieving the target question with the highest similarity and a value greater than or equal to a preset threshold. Finally, the corresponding standard answer is extracted and broadcast to the staff via a speech synthesis module.

[0032] Understandably, in the backend server, the large language model generates corresponding preset questions and standard answers based on preset prompts for key operation nodes, safety specifications, and equipment parameters in the work instructions.

[0033] Regarding preset prompts, here's an example: Generate question-and-answer pairs (JSON format, each element containing a question and an answer) based on the following power operation instructions. Requirements: Questions should focus on key operations, safety regulations, and equipment parameters; answers should be strictly based on the original text. Instruction manual content: {Insert text}; Output: [{"question":"...","answer":"..."}, ...].

[0034] Example 1: System Architecture; like Figure 1 As shown in the figure, this embodiment provides an intelligent question-and-answer system for power operations, including a backend server and at least one work recorder.

[0035] The backend server includes: a large language model module, used to read and understand the homework instructions and generate preset questions and standard answers; a text vectorization module, used to convert preset questions into vectors; and a data packaging module, used to package question-answer pairs and vectors into offline knowledge data packages.

[0036] The work recorder connects to the backend server via a transmission interface (wired USB or Wi-Fi) to download offline knowledge data packages before the job begins.

[0037] Example 2: Offline Question-and-Answer Process; like Figure 2 As shown, the offline question-answering method in this embodiment includes the following steps: S1: Before starting work, the operator selects the "switching operation" work type through the recorder interface. The recorder then downloads the corresponding offline knowledge data package from the backend server to the local storage module of the work recorder.

[0038] S2: The voice acquisition module captures questions from staff in the substation basement (where there is no 4G signal): "What conditions need to be confirmed before the circuit breaker tripping operation?" S3: Use the speech recognition module to perform speech recognition and obtain the question text.

[0039] S4: The recorder's built-in vector calculation module converts the question text into a 512-dimensional question vector and calculates the cosine similarity between this vector and 500 question vectors in the local storage module.

[0040] S5: The search and matching module found a preset question with a similarity of 0.94 (greater than the preset threshold of 0.9): "What states should be confirmed before performing a circuit breaker tripping operation?".

[0041] S6: Extract the standard answer corresponding to the preset question: "1. Confirm that the circuit breaker control power supply is normal; 2. Confirm that the protection device has no abnormal alarms; 3. Confirm that the operating handle is in the correct position...".

[0042] S7: Uses a speech synthesis module to synthesize and broadcast voice messages to staff.

[0043] It should be noted that if the retrieval and matching module finds a match with a similarity score less than the threshold of 0.9 (i.e., no matching answer), the system will execute the following fallback processing procedure: (a1) Voice prompt: The speech synthesis module announces: No standard answer matching your question was found. Please try to describe it again or refer to the paper instruction manual. (a2) Local records: Write the text of the unmatched question, the question time, and the job type to the local log file (stored in a specific partition of the local storage module) for subsequent analysis; (a3) Mark for synchronization: Set a pending upload flag in the log. When the work recorder connects to a network environment (such as Wi-Fi or 4G) again, it will automatically upload the unmatched question text to the backend server for administrator review and may supplement new question-and-answer pairs; (a4) Handling of repeated questions: If the same user submits two consecutive questions with similarity scores below the threshold during the same operation, the system will announce: Your question cannot be answered temporarily. Please suspend the operation according to the procedure and contact the technical manager. This measure is to prevent blindly continuing operations on critical security issues.

[0044] Example 3: Recorder hardware structure; like Figure 3 As shown, the recorder includes: The main control module uses an ARM Cortex-M series processor.

[0045] The voice acquisition module includes a microphone array.

[0046] The voice recognition module uses a lightweight offline ASR chip.

[0047] The local storage module is used to store offline knowledge data packages.

[0048] The vector calculation module is used to perform text vector conversion and similarity calculation.

[0049] The retrieval and matching module is used to perform vector matching and result extraction.

[0050] A speech synthesis module is used to convert text into speech and read out the standard response.

[0051] The communication module is configured to communicate with the backend server when a network is available to download or update the offline knowledge data package.

[0052] The interface module is configured to import the offline knowledge data package to the work recorder via USB or a local wireless network.

[0053] Example 4: Data update method; When the recorder returns to a network environment (such as the work site), the communication module automatically detects whether the backend server has a new version of the offline knowledge data package (e.g., due to an update of the work instruction version). It only downloads the incremental part of the question-and-answer pairs and merges them into the local storage module, avoiding the time consumption caused by downloading the full amount.

[0054] This embodiment has the following advantages: No network available: Eliminates reliance on 4G / 5G networks and is suitable for all signal blind spots in power operations.

[0055] High real-time performance: The terminal only performs ASR, vector calculation (millisecond level) and TTS, avoiding the generation delay of running large models and meeting the requirements of real-time interaction.

[0056] High accuracy: Question-answer pairs are generated by a large backend model based on a professional knowledge base, ensuring the authority and accuracy of the content; vector matching avoids the comprehension bias of local small models.

[0057] Low hardware cost: The recorder does not require a high-performance AI chip and can run smoothly using a conventional embedded processor, making it easy to promote on a large scale.

[0058] Based on the above description of the embodiments, those skilled in the art will understand that the power operation question-and-answer method, system, medium, and product based on offline knowledge base matching described in this embodiment can be implemented in pure software or deployed and run on a general-purpose or dedicated computing hardware platform. Based on this essence, the technical solution of this embodiment can be specifically implemented in the form of a software product containing program instructions. This software product can be stored on various non-volatile storage media or directly deployed as a local or cloud service. The program instructions are used to cause computer devices with processing capabilities—including but not limited to personal computers, server clusters, mobile terminals, or other network devices—to execute the steps described in this embodiment.

[0059] 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. A power operation intelligent question and answer method based on offline knowledge base matching, characterized in that, include: Backend processing: Using a large language model, question-and-answer pairs are generated based on the power operation instructions. The question-and-answer pairs include preset questions and corresponding standard answers. Convert the preset questions into question vectors, and package the question-answer pairs and question vectors into an offline knowledge data package; Terminal deployment: The offline knowledge data package is pre-stored in the local storage module of the work recorder; Offline Q&A: In the absence of a network, the work recorder converts the collected voice questions into question text, then into question vectors. It then retrieves the question vector with the highest similarity to the question vector and a value greater than or equal to a preset threshold from the local storage module, extracts the corresponding standard answer, and broadcasts it on voice.

2. The method of claim 1, wherein, In the background processing, the large language model generates corresponding preset questions and standard answers based on preset prompt words for key operation nodes, safety specifications, and equipment parameters in the work instructions.

3. The method of claim 1, wherein, The step of retrieving the question vector with the highest similarity to the question vector and greater than or equal to a preset threshold in the local storage module specifically uses cosine similarity to calculate the similarity between the question vector and the question vector. The preset threshold is between 0.85 and 0.

95.

4. The method of claim 1, wherein, In the terminal deployment, the work recorder selectively downloads the corresponding offline knowledge data package from the background server according to the job type. The job type includes at least one of the following: switching operation, transmission line inspection, cable head fabrication, and tower climbing operation.

5. The method according to claim 1, characterized in that, The backend server uses a text vectorization model to convert the preset question into a question vector.

6. A power operation recorder, characterized in that, include: The local storage module is configured to store offline knowledge data packages downloaded in advance from the backend server. The offline knowledge data packages are generated in advance by the backend server in the following ways: using a large language model to generate question-answer pairs based on the power operation instructions, the question-answer pairs include preset questions and corresponding standard answers; converting the preset questions into question vectors; and packaging the question-answer pairs and question vectors into offline knowledge data packages. The voice acquisition module is configured to collect voice questions from staff. The speech recognition module is used to convert voice questions into text questions. The vector calculation module is configured to convert the question text into a question vector and calculate the similarity between the question vector and each question vector in local storage; The retrieval and matching module is configured to select the question vector with the highest similarity and greater than or equal to a preset threshold, and extract the corresponding standard answer. The speech synthesis module is configured to read out the standard response. The main control module is electrically connected to the local storage module, voice acquisition module, voice recognition module, vector calculation module, retrieval and matching module, and voice synthesis module, respectively.

7. The power operation recorder according to claim 6, characterized in that, Also includes: The communication module is configured to communicate with the backend server when a network is available to download or update the offline knowledge data package; The interface module is configured to import the offline knowledge data package to the work recorder via USB or a local wireless network.

8. The power operation recorder according to claim 6, characterized in that, The offline knowledge data package stored in the local storage module does not contain large language model parameters, but only question-answer pairs and question vectors.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of computer programs, which are used to be invoked by a processor and to execute the method as described in any one of claims 1-5.

10. A computer program product, comprising a computer program, characterized in that, The computer program is executed by a processor to implement the steps of the method according to any one of claims 1 to 5.

Images