Power supply service AI virtual seat assistant

Abstract

According to the power supply service AI virtual seat assistant disclosed by the invention, comprehensive monitoring, intelligent processing and data analysis of the power supply service are realized by integrating a plurality of functional modules, so that the efficiency and the quality of the power supply service are improved; the scheme comprises the following steps: step 1, collecting and recording data; the system comprises an event monitoring module, an intelligent duty module, a unified voice interaction module, an auxiliary command module and an operation and service quality data analysis module. The event monitoring module is used for monitoring and feeding back a monitoring result to the to-do message pool; the intelligent on-duty module adopts a personnel round-robin algorithm model to intelligently distribute the information to a manual agent; the unified voice interaction module comprises an audio acquisition module, a preprocessing module, a voice endpoint detection module, a voice signal feature extraction module and a voice recognition module; the auxiliary command module is used for the AI virtual seat to assist manual command regulation and control business; and the operation and service quality data analysis module.

Description

Technical Field

[0001] The present invention relates to the field of intelligent service technology, and in particular to an AI virtual agent assistant for power supply services. Background Art

[0002] With the rapid development of digital technology, the virtual digital human industry has gradually become an important part of my country's digital economy and an important force in promoting economic transformation and upgrading and innovative development. As a key public service area, the digital transformation of power supply services is crucial to improving service quality and efficiency.

[0003] Traditional power supply services rely on a large amount of human resources. Customer service personnel need to deal with various trivial issues. Service efficiency and quality are difficult to guarantee. They mainly face the following four major problems:

[0004] 1. Low degree of automation in business processing and high labor costs

[0005] In the current power supply service command center, business processing mainly relies on manual operations, such as identifying abnormal fault data, compiling work orders, dispatching work orders, checking work order quality, and reviewing and archiving. These repetitive operations not only take up a lot of the commander's working time, but also consume a lot of manpower costs.

[0006] 2. The system operation is complicated and the learning cost is high

[0007] Since the power supply service business involves multiple systems and modules, commanders need to frequently switch systems and modules when handling different businesses, which is cumbersome and complicated. This not only increases the learning cost of commanders, but also reduces work efficiency. Especially after the business was transferred to district and county companies, the business volume of district and county command centers continued to increase, and the problem of insufficient staffing became increasingly prominent.

[0008] 3. Lack of intelligent event monitoring and auxiliary operation capabilities

[0009] At present, all matters of the power supply service command center at the municipal and county levels rely on manual monitoring, and there are many cases of missed inspections and missed handling. At the same time, the coverage rate of automatic event monitoring is not high, and the supervision overtime reminder mechanism is not intelligent enough. These problems have led to low business processing efficiency and unstable quality.

[0010] 4. Lack of auxiliary data analysis for distribution network operation command services

[0011] In terms of distribution network operation and services, there is currently a lack of intelligent data analysis support. For example, detailed fault statistics, hotspot statistics for fault repair work orders, and analysis and statistics of frequently outage areas all need to be completed manually, with large amounts of data extraction and statistics and complex business analysis logic. This not only increases the workload of the control and regulation personnel in the command center, but may also affect the timely grasp of distribution network operation and service conditions by managers due to omissions and efficiency issues in manual statistics.

[0012] In summary, developing an AI virtual agent assistant for power supply services is an effective way to solve the challenges currently faced by the power supply service command center in terms of business processing, system operation, event monitoring, and data analysis. Make full use of its deep learning capabilities, voice interaction capabilities, voice command execution capabilities, etc. to empower the command center with intelligence. This can not only promote the transformation and upgrading of the command center's operating mode, greatly improve the command and coordination efficiency of the city and county command centers, but also effectively alleviate the problem of staff shortages in the command center. Through human-computer voice interaction, the AI ​​virtual agent assistant can assist manual commanders in completing a large amount of tedious business processing work, monitor the status of equipment and work orders in real time, and intelligently push reminders, thereby greatly improving work efficiency and quality. Summary of the invention

[0013] The present invention aims to address the technical defects of the prior art and provide an AI virtual agent assistant for power supply services. By integrating multiple functional modules, it realizes comprehensive monitoring, intelligent processing and data analysis of power supply services, thereby improving the efficiency and quality of power supply services.

[0014] The power supply service AI virtual agent assistant of the present invention includes an event monitoring module, an intelligent duty module, a unified voice interaction module, an auxiliary command module, and an operation and service quality data analysis module;

[0015] The event monitoring module is used to monitor power outage events and equipment monitoring abnormal events, and the monitoring results are fed back to the pending message pool;

[0016] The intelligent duty module uses a personnel round-robin algorithm model to intelligently dispatch to human agents;

[0017] The unified voice interaction module includes an audio acquisition module, a preprocessing module, a voice endpoint detection module, a voice signal feature extraction module, and a voice recognition module;

[0018] The auxiliary command module is used for AI virtual agents to assist humans in commanding and controlling services, including real-time automatic monitoring of fault work orders, intelligent auxiliary fault handling, automatic dispatching of SMS reminders for overdue work orders, automatic preparation of work orders, automatic dispatching of work orders, automatic verification of work order receipts, auxiliary archiving of work orders, automatic preparation and dispatching of power outage notification SMS, and intelligent auxiliary command of fault events and work order handling methods based on historical operation data and big data analysis technology, assisting the current manual research and command;

[0019] In the operation and service quality data analysis module, the AI ​​virtual agent regularly extracts data and compiles reports according to specific data analysis and statistical methods, including statistics on hot spots and lines for fault repair work orders, intelligent analysis and early warning of areas with frequent power outages, statistics on distribution network heavy overload equipment, statistics on distribution network operation and power supply services, automatic evaluation of proactive work orders, daily reports, weekly reports, and monthly reports.

[0020] Preferably, the audio collection module collects the voice command signal of the human agent through a microphone;

[0021] The pre-processing module includes noise reduction and gain adjustment;

[0022] The speech endpoint detection module is used to determine the start and end time of the user's speech;

[0023] The speech signal feature extraction module uses MFCC to convert the speech signal into a feature representation that can be used by subsequent processing algorithms;

[0024] The speech recognition module adopts a preset large language model, and the preset large language model interacts by inputting text information or the speech corresponding to the text information.

[0025] Preferably, the personnel round-robin algorithm model includes the following steps:

[0026] Manual agent status management: When a manual agent is on duty, he / she needs to log in to the system and click the "On Duty" button, and the system will set his / her status to "Online and Idle". When a manual agent is off duty, he / she needs to click the "Off Duty" button, and the system will update his / her status to "Offline".

[0027] Window processing status monitoring and distribution: The system monitors the window processing status of each online agent in real time, including whether the customer request is being processed, the processing time, and identifies the window processing status;

[0028] Message processing and distribution: The system distributes the event message queue formed by the pending message pool to online and idle human agents in sequence.

[0029] Preferably, the preset large language model performs prompt word optimization training, and the training step includes:

[0030] Step 1, data preparation;

[0031] 1.1, collect industry-related data;

[0032] Text data in the power sector, including power industry documents, technical manuals, regulations, expert documents, and operation manuals;

[0033] Domain-specific data formats, including work order operation records, event processing logs, and data analysis rules, convert data into text format;

[0034] 1.2, data cleaning and preprocessing;

[0035] Data cleaning, removing useless information, standardizing terminology and cleaning up duplicate data;

[0036] Segmentation: split the document into paragraphs suitable for model processing so that the model can learn the relationship between sentences;

[0037] Format into the input format accepted by the model;

[0038] 1.3, Data Annotation: Prepare standard question-answer pairs, dialogue data, common questions and their corresponding professional answers to common questions for fine-tuning tasks;

[0039] Step 2: Fine-tune the large language model;

[0040] 2.1, set fine-tuning parameters; choose a lower learning rate; choose a suitable batch size of 8-32; and train 3-5 rounds;

[0041] 2.2, fine-tuning process; use the Transformer library or open source large model framework to train in a distributed environment, save the model weight file and use checkpoints during the fine-tuning process;

[0042] 2.3. Verify the model effect. Use the validation data set of the power industry to evaluate the accuracy of the model’s answers. Use accuracy, recall, and F1 score to evaluate the model’s question-answering effect and the quality of the generated text.

[0043] Step 3, retrieve the enhanced generation RAG component settings;

[0044] 3.1, Document retrieval settings;

[0045] The vector database FAISS is selected to support vector retrieval;

[0046] Data vectorization, encoding power industry document data into vectors;

[0047] Vector storage and indexing: vectorize documents and store them in the milvus database to build an index structure;

[0048] 3.2, set the search strategy;

[0049] Use Top-k retrieval to retrieve the most relevant k document paragraphs before generating an answer;

[0050] Similarity calculation, using cosine similarity or dot product similarity, calculates the similarity between the input question and each document in the database, and selects documents with high relevance;

[0051] Step 4: fine-tune the large language model and integrate it with RAG;

[0052] 4.1, Retrieval and generation integration;

[0053] The user asks questions and inputs them. The questions asked by the user are vectorized.

[0054] Retrieve the most relevant documents and retrieve k relevant document paragraphs based on vector similarity;

[0055] Generate answers by feeding the question and the retrieved document snippet into a fine-tuned large language model and letting the model generate answers based on relevant information.

[0056] Post-processing: formatting and denoising the generated answers.

[0057] Preferably, the preset large language model adopts the open source ChatGLM3 / Qwen1.5.

[0058] Preferably, the pending message pool is used in combination with Kafka and RabbitMQ;

[0059] Use Kafka to process large amounts of real-time data streams;

[0060] RabbitMQ is used for message transmission and processing.

[0061] Preferably, the event message queue in the pending message pool is implemented by pushing the event task information monitored by the event monitoring module from the background to the foreground through the websocket component.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] The present invention provides an AI virtual agent assistant for power supply services, which is an intelligent system integrating multiple functional modules, aiming to improve the efficiency and quality of power supply services, including:

[0064] 1. Event monitoring module

[0065] This module is mainly responsible for monitoring various events in the power supply system, especially power outage events and equipment monitoring abnormal events. Through real-time monitoring, the system can promptly discover and feedback these problems to ensure the stability and reliability of power supply services. The monitoring results will be fed back to the pending message pool for subsequent processing;

[0066] Power outage event monitoring: The system can monitor power outage events in the power supply network in real time, including key information such as power outage scope and power outage time;

[0067] Equipment monitoring abnormal events: The system also monitors the operating status of power supply equipment (such as transformers, lines, etc.), and immediately provides feedback once an abnormality (such as overheating, overload, etc.) is found;

[0068] 2. Intelligent duty module

[0069] This module uses a personnel round-robin algorithm model to intelligently assign pending tasks to appropriate human agents based on the current workload and personnel status, which helps optimize human resource allocation and improve work efficiency.

[0070] Personnel round-robin algorithm: The system will automatically adjust the staffing configuration according to the preset rules to ensure that each agent can get a reasonable task allocation;

[0071] 3. Unified voice interaction module

[0072] This module is the core part of the voice interaction between AI virtual agents and users, including audio acquisition, preprocessing, voice endpoint detection, voice signal feature extraction and voice recognition.

[0073] 4. Auxiliary command module

[0074] This module is mainly used for AI virtual agents to assist humans in commanding and controlling business, including real-time monitoring of fault work orders, intelligent auxiliary disposal, automatic dispatch of reminder SMS for overdue work orders, etc.

[0075] Real-time automatic monitoring of fault work orders: The system can monitor the status of fault work orders in real time to ensure timely processing;

[0076] Intelligent auxiliary fault handling: Based on historical data and big data analysis technology, the system can provide intelligent handling suggestions for fault events;

[0077] Automatically send reminder SMS for expired work orders: For work orders that are about to expire, the system will automatically send reminder SMS to relevant personnel;

[0078] Automatic preparation, distribution and verification of work orders: The system can automatically prepare new work orders and distribute them to appropriate personnel for processing; at the same time, the system will automatically verify completed work orders to ensure their quality;

[0079] Automatic compilation and distribution of power outage notification SMS: When a power outage occurs, the system will automatically compile and distribute a power outage notification SMS to users;

[0080] 5. Operation and service quality data analysis module

[0081] This module is mainly used to perform data analysis on the operation and service quality of AI virtual agents, including data extraction, report preparation and other functions;

[0082] Data extraction: The system will extract relevant data from the database based on specific data analysis and statistical methods;

[0083] Report preparation: Based on the extracted data, the system will automatically generate various reports, such as statistics on hotspot areas and lines for fault repair work orders, intelligent analysis and early warning of areas with frequent power outages, etc.

[0084] Daily, weekly and monthly reports: The system will also generate daily, weekly and monthly reports regularly so that managers can understand the on-site situation in a timely manner;

[0085] To sum up, the power supply service AI virtual agent assistant is an effective way to solve the challenges currently faced by the power supply service command center in business processing, system operation, event monitoring and data analysis. It makes full use of its deep learning ability, voice interaction ability, voice command execution ability, etc. to empower the command center with intelligence. This can not only promote the transformation and upgrading of the command center's operating mode, greatly improve the command and coordination efficiency of the municipal and county command centers, but also effectively alleviate the problem of staff shortage in the command center. Through human-computer voice interaction, the AI ​​virtual agent assistant can assist manual commanders to complete a large amount of tedious business processing work, monitor equipment and work order status in real time, and intelligently push reminders, thereby greatly improving work efficiency and quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 A schematic diagram of an intelligent duty interface according to a specific embodiment of the present invention;

[0087] Figure 2 Unified interaction diagrams for to-do items;

[0088] Figure 3 This is a schematic diagram to assist command;

[0089] Figure 4 This is a technology selection list. DETAILED DESCRIPTION

[0090] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. All other embodiments obtained by ordinary technicians in the field without making creative work based on the embodiments of the present invention shall fall within the scope of protection of the present invention.

[0091] like Figures 1 to 4 As shown, it shows a specific implementation of the present invention: an AI virtual agent assistant for power supply service, including an event monitoring module, an intelligent duty module, a unified voice interaction module, an auxiliary command module, and an operation and service quality data analysis module;

[0092] The event monitoring module is used to monitor power outage events and equipment monitoring abnormal events, and the monitoring results are fed back to the pending message pool;

[0093] The intelligent duty module uses a personnel round-robin algorithm model to intelligently dispatch to human agents;

[0094] The unified voice interaction module includes an audio acquisition module, a preprocessing module, a voice endpoint detection module, a voice signal feature extraction module, and a voice recognition module;

[0095] The auxiliary command module is used for AI virtual agents to assist humans in commanding and controlling services, including real-time automatic monitoring of fault work orders, intelligent auxiliary fault handling, automatic dispatching of SMS reminders for overdue work orders, automatic preparation of work orders, automatic dispatching of work orders, automatic verification of work order receipts, auxiliary archiving of work orders, automatic preparation and dispatching of power outage notification SMS, and intelligent auxiliary command of fault events and work order handling methods based on historical operation data and big data analysis technology, assisting the current manual research and command;

[0096] In the operation and service quality data analysis module, the AI ​​virtual agent regularly extracts data and compiles reports according to specific data analysis and statistical methods, including statistics on hot spots and lines for fault repair work orders, intelligent analysis and early warning of areas with frequent power outages, statistics on distribution network heavy overload equipment, statistics on distribution network operation and power supply services, automatic evaluation of proactive work orders, daily reports, weekly reports, and monthly reports.

[0097] The system of the present invention achieves rapid response and processing of power failures (such as power failure of distribution transformers) by intelligently distributing message event queues to human agents and combining the conversational interaction capabilities of AI virtual agents. The system not only simplifies the work order distribution process, but also provides comprehensive fault analysis and statistical functions.

[0098] When a power event such as a power failure occurs, Figure 1As shown, the system first automatically prepares the corresponding emergency repair work order. Subsequently, the event and the emergency repair work order are intelligently distributed to the manual agent queue, waiting to be processed. At the same time, the AI ​​virtual agent sends event details and emergency repair work order suggestions to the manual agent through a dialog box text combined with voice broadcast. Interaction between manual agents and AI virtual agents: Manual agents can communicate with AI virtual agents through voice or text. If you agree with the emergency repair work order suggestions put forward by the AI ​​virtual agent, the manual agent can simply answer "yes", and then the AI ​​virtual agent will automatically complete the work order distribution. If you disagree, the manual agent can command the AI ​​virtual agent to adjust the work order through voice or text, and distribute it again. Work order tracking and confirmation: After the work order is distributed, the system continues to track the progress of the emergency repair. Once the emergency repair is completed, the system dispatches the completed emergency repair work order to the manual agent, waiting for confirmation. As Figure 2 As shown in the figure, the human agent confirms whether the work order is passed through voice or text. If it is passed, the process ends; if it is not passed, it returns to the previous level and records the reason for failure. Figure 3 As shown in the figure, the human agent can also send a task briefing to the AI ​​virtual agent through voice or text. The system also provides a wealth of statistical analysis functions, including hotspot area and line statistics for fault repair work orders, intelligent analysis and early warning of frequent power outage areas, statistics on distribution network heavy overload equipment, distribution network operation and power supply service statistics, etc. In addition, the system can also realize the automatic evaluation of active work orders, as well as the automatic generation of daily, weekly, monthly and other periodic reports.

[0099] Through the intelligent collaboration of AI virtual agents and human agents, rapid response and processing of power failures are achieved. Intelligent distribution and automatic dispatching functions greatly reduce the workload of human agents. It provides a wealth of statistical analysis functions to help power companies better understand distribution network operations and power supply services. It supports both voice and text interaction to meet the needs of different users.

[0100] In summary, the power supply service AI virtual agent assistant proposed in the present invention has significant innovation and practicality in the field of power fault handling.

[0101] Preferably, the audio collection module collects the voice command signal of the human agent through a microphone;

[0102] The pre-processing module includes noise reduction and gain adjustment;

[0103] The speech endpoint detection module is used to determine the start and end time of the user's speech;

[0104] The speech signal feature extraction module uses MFCC to convert the speech signal into a feature representation that can be used by subsequent processing algorithms;

[0105] The speech recognition module adopts a preset large language model, and the preset large language model interacts by inputting text information or the speech corresponding to the text information.

[0106] like Figure 1 As shown in the figure, the virtual commander monitors the events of the command center, including power outage events, abnormal equipment monitoring events, and pushes voice and text messages to the online virtual commander when new events occur. It includes audio acquisition and preprocessing, voice endpoint detection, voice signal feature extraction, and speech recognition model. Audio acquisition and preprocessing collects the voice command signal of the artificial commander through a microphone, and performs preprocessing such as noise reduction and gain adjustment. Voice endpoint detection is used to determine the start and end time of the user's speech. MFCC (Mel Frequency Cepstral Coefficients) is used for voice signal feature extraction to convert the voice signal into a feature representation that can be used by subsequent processing algorithms. The speech recognition model is the core part of unified voice interaction. A large language model is used to convert the voice signal into text using (FANASR).

[0107] like Figure 2 As shown in the figure, it provides users with a unified interactive function. Controllers only need to connect to the unified interactive interface to handle all to-do items without clicking on each application menu page. Through voice input commands, the background service uses voice recognition (FANASR), large model semantic analysis and other technologies to realize the call and interaction between operation commands and data interfaces in the business system. The large model is used to analyze the interface interaction results and realize the assembly and generation of feedback messages. Finally, the voice broadcast of the message is realized through speech synthesis technology (KANTTS), and the message is broadcast in the form of text messages; the to-do events are uniformly displayed on the unified voice interaction interface, including event broadcast, work order reminder, processing results, etc.

[0108] Preferably, the personnel round-robin algorithm model includes the following steps:

[0109] Manual agent status management: When a manual agent is on duty, he / she needs to log in to the system and click the "On Duty" button, and the system will set his / her status to "Online and Idle". When a manual agent is off duty, he / she needs to click the "Off Duty" button, and the system will update his / her status to "Offline".

[0110] Window processing status monitoring and distribution: The system monitors the window processing status of each online agent in real time, including whether the customer request is being processed, the processing time, and identifies the window processing status;

[0111] Message processing and distribution: The system distributes the event message queue formed by the pending message pool to online and idle human agents in sequence.

[0112] like Figure 3As shown in the figure, the virtual commander can automatically complete the extraction of relevant reporting data and report preparation according to the instructions, provide command auxiliary services for the command business on the intelligent virtual commander system, and is mainly responsible for the distribution of calls, interaction with the intelligent business layer, and dispatching the MS media service module to complete the control of ASR, TTS, etc. Through the large language model and knowledge graph, message processing, interface service matching, feedback message generation, etc. are constructed. Through the full event monitoring service, event message queue, personnel round-robin algorithm model, etc., intelligent duty, refined message push and other services are realized.

[0113] Personnel polling is monitored in real time through the background program. When there are multiple people on duty, each person's work is marked as busy or idle. If they are all busy at the same time, the message is pushed delayed and put on standby in the task pool. If they are all idle at the same time or some of them are idle, the system distributes the event message queue formed by the to-do message pool to the online and idle manual agents in sequence.

[0114] Preferably, the preset large language model performs prompt word optimization training, and the training step includes:

[0115] Step 1, data preparation;

[0116] 1.1, collect industry-related data;

[0117] Text data in the power sector, including power industry documents, technical manuals, regulations, expert documents, and operation manuals;

[0118] Domain-specific data formats, including work order operation records, event processing logs, and data analysis rules, convert data into text format;

[0119] 1.2, data cleaning and preprocessing;

[0120] Data cleaning, removing useless information, standardizing terminology and cleaning up duplicate data;

[0121] Segmentation: split the document into paragraphs suitable for model processing so that the model can learn the relationship between sentences;

[0122] Format into the input format accepted by the model;

[0123] 1.3, Data Annotation: Prepare standard question-answer pairs, dialogue data, common questions and their corresponding professional answers to common questions for fine-tuning tasks;

[0124] Step 2: Fine-tune the large language model;

[0125] 2.1, set fine-tuning parameters; choose a lower learning rate; choose a suitable batch size of 8-32; and train 3-5 rounds;

[0126] 2.2, fine-tuning process; use the Transformer library or open source large model framework to train in a distributed environment, save the model weight file and use checkpoints during the fine-tuning process;

[0127] 2.3. Verify the model effect. Use the validation data set of the power industry to evaluate the accuracy of the model’s answers. Use accuracy, recall, and F1 score to evaluate the model’s question-answering effect and the quality of the generated text.

[0128] Step 3, retrieve the enhanced generation RAG component settings;

[0129] 3.1, Document retrieval settings;

[0130] The vector database FAISS is selected to support vector retrieval;

[0131] Data vectorization, encoding power industry document data into vectors;

[0132] Vector storage and indexing: vectorize documents and store them in the milvus database to build an index structure;

[0133] 3.2, set the search strategy;

[0134] Use Top-k retrieval to retrieve the most relevant k document paragraphs before generating an answer;

[0135] Similarity calculation, using cosine similarity or dot product similarity, calculates the similarity between the input question and each document in the database, and selects documents with high relevance;

[0136] Step 4: fine-tune the large language model and integrate it with RAG;

[0137] 4.1, Retrieval and generation integration;

[0138] The user asks questions and inputs them. The questions asked by the user are vectorized.

[0139] Retrieve the most relevant documents and retrieve k relevant document paragraphs based on vector similarity;

[0140] Generate answers by feeding the question and the retrieved document snippet into a fine-tuned large language model and letting the model generate answers based on relevant information.

[0141] Post-processing: formatting and denoising the generated answers.

[0142] The present invention aims to successfully train a professional model suitable for power supply services in the power field by conducting professional prompt word training on an open source large language model;

[0143] Through the above training steps, the open source large language model was successfully customized into a professional model suitable for power supply services in the power field. The model can accurately understand the user's questions, quickly retrieve relevant information, and generate professional and accurate answers. This not only improves the efficiency and quality of power supply services, but also provides users with a more convenient and efficient power service experience.

[0144] Preferably, the preset large language model adopts the open source ChatGLM3.

[0145] When the present invention can use the open source ChatGLM3 or Qwen1.5 as the preset large language model, we can further elaborate on its application and training process based on the characteristics and advantages of these two models. Select a suitable model parameter size (such as 13B, 30B, etc.) to ensure that the domain knowledge can be retained after fine-tuning the power data.

[0146] ChatGLM3 is a domestic open source AI model jointly developed by Tsinghua University KEG Laboratory and Zhipu AI. It is a multimodal pre-trained language model with powerful multimodal understanding, code modules, networks and knowledge reasoning capabilities.

[0147] Qwen1.5 is a large language model developed by the Alibaba Cloud team, built on the Transformer architecture. It has basic capabilities, human preference alignment, multi-language processing, and long sequence support, and performs well in multiple benchmarks. Qwen1.5 has been open sourced on platforms such as Hugging Face and ModelScope, and provides multiple deployment and calling methods.

[0148] For the professional model construction of power supply services in the power sector, the training and fine-tuning processes of the above two large models are similar, including data preparation, fine-tuning parameter setting, fine-tuning process, verification and evaluation steps.

[0149] In summary, both ChatGLM3 and Qwen1.5 can be used as a preset large language model to build a professional model for power supply services in the power sector. Through reasonable data preparation, fine-tuning parameter settings, and training and verification processes, we can obtain an efficient, accurate, and professional power supply service model in the power sector.

[0150] like Figure 4 As shown, this is the technology selection used in this patent. At the same time, the present invention can use PaddleSpeech instead of KANTTS to realize speech-to-text, use PaddleSpeech instead of FANASR to realize speech synthesis, and use LLaMA 3 instead of the large language model.

[0151] Preferably, the pending message pool is used in combination with Kafka and RabbitMQ;

[0152] Use Kafka to process large amounts of real-time data streams;

[0153] RabbitMQ is used for message transmission and processing.

[0154] Preferably, the event message queue in the pending message pool is implemented by pushing the event task information monitored by the event monitoring module from the background to the foreground through the websocket component.

[0155] Full event monitoring is achieved by real-time monitoring of data interfaces and kafka message queue data through programs; event message queue is achieved by pushing the monitored event task information from the background to the front end through the websocket component; refined message push is achieved by querying the background data and performing semantic analysis and language and text reorganization through the large model ChatGLM3 artificial model, and generating voice through the TTS open source speech synthesis model in real time and pushing it to the front end for refined broadcasting.

[0156] The present invention develops an AI virtual agent assistant, which is an AI virtual agent assistant that can automatically handle business across systems and modules. It is of great significance to improving the work efficiency of the district and county command center. It also has the ability of intelligent event monitoring and auxiliary operation, can monitor the status of various equipment and work orders in real time, and intelligently push reminders to improve the timeliness and accuracy of business processing. It can automatically perform distribution network operation command service auxiliary data analysis, which is of great significance to improving the accuracy and efficiency of data analysis, improving the efficiency and accuracy of business processing, and reducing labor costs.

[0157] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments. Various changes can be made within the knowledge scope of ordinary technicians in the field without departing from the purpose of the present invention. These changes involve related technologies well known to those skilled in the art, which all fall within the scope of protection of the patent of the present invention.

[0158] Many other changes and modifications may be made without departing from the concept and scope of the present invention.It should be understood that the present invention is not limited to the specific embodiments, and the scope of the present invention is defined by the appended claims.

Claims

1. Power supply service AI virtual agent assistant, characterized by: Including event monitoring module, intelligent duty module, unified voice interaction module, auxiliary command module, operation and service quality data analysis module; The event monitoring module is used to monitor power outage events and equipment monitoring abnormal events, and the monitoring results are fed back to the pending message pool; The intelligent duty module uses a personnel round-robin algorithm model to intelligently dispatch to human agents; The unified voice interaction module includes an audio acquisition module, a preprocessing module, a voice endpoint detection module, a voice signal feature extraction module, and a voice recognition module; The auxiliary command module is used for AI virtual agents to assist humans in commanding and controlling services, including real-time automatic monitoring of fault work orders, intelligent auxiliary fault handling, automatic dispatching of SMS reminders for overdue work orders, automatic preparation of work orders, automatic dispatching of work orders, automatic verification of work order receipts, auxiliary archiving of work orders, automatic preparation and dispatching of power outage notification SMS, and intelligent auxiliary command of fault events and work order handling methods based on historical operation data and big data analysis technology, assisting the current manual research and command; In the operation and service quality data analysis module, the AI ​​virtual agent regularly extracts data and compiles reports according to specific data analysis and statistical methods, including statistics on hot spots and lines for fault repair work orders, intelligent analysis and early warning of areas with frequent power outages, statistics on distribution network heavy overload equipment, statistics on distribution network operation and power supply services, automatic evaluation of proactive work orders, daily reports, weekly reports, and monthly reports.

2. The power supply service AI virtual agent assistant according to claim 1, characterized in that: The audio collection module collects the voice command signal of the human agent through a microphone; The pre-processing module includes noise reduction and gain adjustment; The speech endpoint detection module is used to determine the start and end time of the user's speech; The speech signal feature extraction module uses MFCC to convert the speech signal into a feature representation that can be used by subsequent processing algorithms; The speech recognition module adopts a preset large language model, and the preset large language model interacts by inputting text information or the speech corresponding to the text information.

3. The power supply service AI virtual agent assistant according to claim 2, characterized in that: The personnel round-robin algorithm model includes the following steps: Manual agent status management: when a manual agent is on duty, he / she needs to log in to the system and click the "on duty" button, and the system will set his / her status to "online and idle". When a manual agent is off duty, he / she needs to click the "off duty" button, and the system will update his / her status to "offline". Window processing status monitoring and distribution: The system monitors the window processing status of each online agent in real time, including whether the customer request is being processed, the processing time, and identifies the window processing status; Message processing and distribution: The system distributes the event message queue formed by the pending message pool to online and idle human agents in sequence.

4. The power supply service AI virtual agent assistant according to claim 2 or 3, characterized in that: The preset large language model performs prompt word optimization training, and the training steps include: Step 1, data preparation; 1.1, collect industry-related data; Text data in the power sector, including power industry documents, technical manuals, regulations, expert documents, and operation manuals; Domain-specific data formats, including work order operation records, event processing logs, and data analysis rules, convert data into text format; 1.2, data cleaning and preprocessing; Data cleaning, removing useless information, standardizing terminology and cleaning up duplicate data; Segmentation: split the document into paragraphs suitable for model processing so that the model can learn the relationship between sentences; Format into the input format accepted by the model; 1.3, Data Annotation: Prepare standard question-answer pairs, dialogue data, common questions and their corresponding professional answers to common questions for fine-tuning tasks; Step 2: Fine-tune the large language model; 2.1, set fine-tuning parameters; choose a lower learning rate; choose a suitable batch size of 8-32; and train 3-5 rounds; 2.2, fine-tuning process; use the Transformer library or open source large model framework to train in a distributed environment, save the model weight file and use checkpoints during the fine-tuning process; 2.

3. Verify the model effect. Use the validation data set of the power industry to evaluate the accuracy of the model’s answers. Use accuracy, recall, and F1 score to evaluate the model’s question-answering effect and the quality of the generated text. Step 3, retrieve the enhanced generation RAG component settings; 3.1, Document retrieval settings; The vector database FAISS is selected to support vector retrieval; Data vectorization, encoding power industry document data into vectors; Vector storage and indexing: vectorize documents and store them in the milvus database to build an index structure; 3.2, set the search strategy; Use Top-k retrieval to retrieve the most relevant k document paragraphs before generating an answer; Similarity calculation, using cosine similarity or dot product similarity, calculates the similarity between the input question and each document in the database, and selects documents with high relevance; Step 4: fine-tune the large language model and integrate it with RAG; 4.1, Retrieval and generation integration; The user asks questions and inputs them. The questions asked by the user are vectorized. Retrieve the most relevant documents and retrieve k relevant document paragraphs based on vector similarity; Generate answers by feeding the question and the retrieved document snippet into a fine-tuned large language model and letting the model generate answers based on relevant information. Post-processing: formatting and denoising the generated answers.

5. The power supply service AI virtual agent assistant according to claim 4, characterized in that: The preset large language model adopts the open source ChatGLM3 / Qwen1.

5.

6. The power supply service AI virtual agent assistant according to claim 6, characterized in that: The pending message pool is used in combination with Kafka and RabbitMQ; Use Kafka to process large amounts of real-time data streams; RabbitMQ is used for message transmission and processing.

7. The power supply service AI virtual agent assistant according to claim 6, characterized in that: The event message queue in the to-do message pool is implemented by pushing the event task information monitored by the event monitoring module from the background to the front end through the websocket component.

Images