A large model-based operator comprehensive dispatching system and method

Abstract

This invention discloses a comprehensive operator scheduling system and method based on a large model. The system includes: a data layer for acquiring basic service data, preprocessing the basic service data, and storing it in a database; a model layer for training models based on installation and maintenance scene images and training datasets stored in the database to obtain multiple target models; the target models include a fine-tuned basic large model, an installation and maintenance image recognition small model, and a telecommunications domain installation and maintenance intent recognition small model; and an intelligent agent core layer for acquiring multimodal data related to the task input by the user, processing the multimodal data by calling the database and the target models to obtain multiple sub-tasks of the task, executing them to obtain execution results, and returning a processing solution to the user based on the execution results. Through the cooperation between the data layer, model layer, and intelligent agent core layer, this system can achieve fully automated upgrades of the entire operation and maintenance process.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of artificial intelligence technology, and in particular to a large-scale operator integrated scheduling system and method. Background Technology

[0002] In the existing home broadband installation, maintenance, and integrated dispatch business, the traditional support model has long been constrained by both technical architecture and business processes. The system suffers from isolated data, insufficient automation and intelligence, and difficulty in adapting to the service demands brought about by the expansion of user scale and the increase in business complexity. The current business model relies on manual operation for core processes such as installation and maintenance work order reception, fault diagnosis, and data query. The solution is still centered on manual handling, supplemented only by a single auxiliary tool.

[0003] In terms of data management, data in the Build / Open / Maintain (B / O / M) domains is scattered across different systems, lacking a unified and standardized management mechanism. Some operators have attempted to introduce Artificial Intelligence (AI) technology, but these efforts mostly remain at the level of "single-point assistance," failing to form end-to-end collaborative capabilities. Moreover, general-purpose models have poor adaptability in the vertical domains of operators and have not been specifically optimized for installation and maintenance scenarios. For example, in actual business, simple network outage troubleshooting accounts for over 40% of repetitive work orders, with a large amount of manpower bogged down in repetitive mechanical labor, resulting in response delays of over 30 minutes for complex faults (such as multi-area optical attenuation anomalies); existing solutions only deploy basic dialogue models in intelligent question-and-answer scenarios, unable to be linked to the work order system or automatically generate handling solutions; image recognition technology only achieves simple judgment of optical modem indicator lights, and the recognition results are not linked to the installation and maintenance work order progress; the solution lacks voice interaction functionality, cannot share data with text query data, and lacks multimodal collaborative capabilities. Summary of the Invention

[0004] This invention provides a large-scale operator integrated scheduling system and method to solve the problem of low efficiency caused by the lack of a unified and standardized management and control mechanism in the existing integrated scheduling.

[0005] According to one aspect of the present invention, a large-scale model-based integrated dispatching system for operators is provided, the system comprising: a data layer, a model layer, and an agent core layer; The data layer is used to acquire basic business data, preprocess the basic business data, and store it in the database. The basic business data includes business data from external systems and internal business data of the operator. The internal business data of the operator includes the operator's operation and maintenance cases and manuals, installation and maintenance integrated commissioning records, operator professional knowledge, and installation and maintenance scene pictures. The model layer is used to train models based on the installation and maintenance scene images and the training dataset stored in the database to obtain multiple target models; the target models include a fine-tuned basic large model, an installation and maintenance image recognition small model, and a telecommunications domain installation and maintenance intent recognition small model. The core layer of the intelligent agent is used to acquire multimodal data related to the task input by the user, process the multimodal data by calling the database and the target model, obtain multiple sub-tasks of the task and execute them to obtain execution results, and return a processing solution to the user based on the execution results.

[0006] According to another aspect of the present invention, a large-scale model-based operator integrated scheduling method is provided, which can be executed by any of the large-scale model-based operator integrated scheduling systems described in the embodiments of the present invention. The method includes: Basic business data is obtained through the data layer, preprocessed, and stored in the database. The basic business data includes business data from external systems and internal business data of the operator. The internal business data of the operator includes the operator's operation and maintenance cases and manuals, installation and maintenance integrated commissioning records, operator professional knowledge, and installation and maintenance scene pictures. Multiple target models are obtained by training the model layer based on the installation and maintenance scene images and the training dataset stored in the database. The target models include a fine-tuned basic large model, an installation and maintenance image recognition small model, and a telecommunications domain installation and maintenance intent recognition small model. The intelligent agent core layer acquires multimodal data related to the task from user input, processes the multimodal data by calling the database and the target model, obtains multiple subtasks of the task, executes them to obtain execution results, and returns a processing solution to the user based on the execution results.

[0007] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising: at least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to execute the large-model-based operator integrated scheduling method according to any embodiment of the present invention.

[0008] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the large-model-based operator integrated scheduling method described in any embodiment of the present invention.

[0009] This invention discloses a large-scale operator integrated dispatching system and method. The system includes a data layer, a model layer, and an intelligent agent core layer. The data layer is used to acquire basic business data, preprocess the basic business data, and store it in a database. The basic business data includes business data from external systems and internal business data of the operator. The internal business data of the operator includes operator operation and maintenance cases and manuals, installation and maintenance integrated dispatch records, operator professional knowledge, and installation and maintenance scene images. The model layer is used to train models based on the installation and maintenance scene images and the training dataset stored in the database to obtain target models. The target models include a fine-tuned basic large model, an installation and maintenance image recognition small model, and a telecommunications domain installation and maintenance intent recognition small model. The intelligent agent core layer is used to acquire multimodal data corresponding to the task input by the user, process the multimodal data by calling the database and the target model, obtain multiple sub-tasks of the task, execute them to obtain execution results, and return a processing solution to the user based on the execution results. The system prepares data at the data layer to provide high-precision data support for the model, trains the model at the model layer, and finally processes the user-input tasks through the intelligent agent core layer and returns the processing solution to the user. This enables the full-process automation upgrade of operation and maintenance business, reduces manual operation, and solves the problem of low efficiency caused by the lack of a unified and standardized management and control mechanism in the existing technology for comprehensive scheduling.

[0010] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of the structure of a large-scale operator integrated dispatching system provided in Embodiment 1 of the present invention; Figure 2 This is a first flowchart illustrating the interaction between various intelligent agents in the core layer of an intelligent agent according to Embodiment 2 of the present invention. Figure 3 This is a second flowchart illustrating the interaction between various intelligent agents in the core layer of an intelligent agent according to Embodiment 2 of the present invention. Figure 4This is a flowchart illustrating a large-scale operator integrated scheduling method provided in Embodiment 3 of the present invention. Figure 5 This is a schematic diagram of the electronic device of the operator integrated scheduling method based on a large model according to an embodiment of the present invention. Detailed Implementation

[0013] To enable those skilled in the art to better understand the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention. It should be understood that the various steps described in the method embodiments of the present invention can be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0014] The term "comprising" and its variations as used herein are open-ended inclusions, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the description below.

[0015] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, any variations of the terms "comprising" and "having," etc., are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0016] It should be noted that the terms "a" and "a plurality of" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0017] The names of the messages or information exchanged between the multiple devices in the embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0018] The existing business model has the following problems: 1. High reliance on manual labor and low efficiency. In traditional business processes, 80% of work order processing requires manual intervention, involving repetitive operations and frequent switching between systems. Nearly half of the workload of the dispatch agent is spent on data querying and entry. The self-processing rate of installation and maintenance work orders is less than 30%, and the average number of work orders processed per person per day is only 25, far below the target of 50 orders in intelligent scenarios.

[0019] 2. The data system is chaotic and of questionable quality. Data is scattered and lacks standardized management, resulting in insufficient data completeness (e.g., 12% of work orders lack user location information), low accuracy (device status data deviates from the actual data by more than 10%), and poor timeliness (resource update delays often exceed 2 hours). This makes it impossible to provide high-quality data support for intelligent models and hinders the implementation of core capabilities such as fault diagnosis and work order scheduling.

[0020] 3. Isolated technical architecture and lack of collaboration. Functions such as voice interaction, image recognition, and knowledge-based Q&A are deployed independently, forming "data silos" and "technology silos." For example, user fault descriptions transcribed from voice cannot be automatically synchronized to the work order system, abnormal results of optical modem recognition cannot trigger the fault investigation process, multi-agent collaborative technology is not applied in the operator field, and complex business processes (such as cross-system work order collaborative processing) cannot be closed-loop.

[0021] 4. Poor scenario adaptability and insufficient accuracy. The general-purpose model does not incorporate knowledge of the operator's vertical domain, resulting in low accuracy in core tasks such as intent recognition and fault localization. For example, the accuracy rate for identifying specialized intents such as "Fiber to the Room (FTTR) network failure" and "IP over Ethernet (IPOE) dial-up failure" is below 75%, and it cannot distinguish between "excessive optical attenuation (user-end problem)" and "excessive optical attenuation (line-end problem)." The retrieval enhancement technology does not integrate with a structured database, and question-and-answer responses (such as "query of installation and maintenance work orders in Xiaodian District, XX City in the past 7 days") require secondary manual verification, resulting in an accuracy rate below 80%.

[0022] Based on an analysis of existing business models, this invention addresses the aforementioned problems through the following methods: 1. Achieve full-process automation of business operations and reduce reliance on manual labor. Design a multi-agent collaborative mechanism to break down complex scheduling tasks (such as "regional network outage work order handling") into ordered sub-tasks of "intent recognition - data query - fault diagnosis - solution generation - task execution", realize cross-system automatic scheduling, improve the self-processing rate of installation and maintenance work orders to more than 50%, and shorten the response time of complex faults to within 15 minutes.

[0023] 2. Construct a standardized data management and control system to improve data quality. Establish a "three-level verification + dynamic evaluation" data quality management and control model to solve the problems of data duplication, invalidity, and delay, ensuring data integrity ≥95%, accuracy ≥98%, and timeliness ≤30 minutes, providing high-precision data support for model training and business queries.

[0024] 3. Break through technology silos and achieve multimodal collaboration and tool invocation. Integrate voice, text, and image multimodal data processing capabilities, and open up interfaces between intelligent agents and external systems, such as construction scheduling, operation and maintenance (OM), and resource management systems, to achieve end-to-end collaboration from "speech transcription to intent recognition to work order creation to fault handling". The tool invocation success rate is ≥99%, and the multimodal intent recognition accuracy rate is ≥90%.

[0025] 4. Improve model scenario adaptability and optimize core task accuracy. Through the collaboration of "large model + small model" and retrieval enhancement technology, operator expertise (such as network topology and fault cases) is injected into the model, so that the accuracy of installation and maintenance professional intent recognition is ≥92%, the accuracy of automatic fault delimitation is ≥88%, and the accuracy of question and answer response is ≥90%, meeting the needs of complex scenarios in the operator's vertical field.

[0026] This invention is deeply integrated with operator services, specifically focusing on core business scenarios such as mobile broadband installation and maintenance, and comprehensive dispatching. It integrates large model technology, multi-agent collaboration technology, multimodal data processing technology, and retrieval enhancement technology to construct a comprehensive dispatching intelligent agent platform covering the entire "data-model-intelligent agent-application" link. It aims to solve problems such as high dependence on manual labor, data dispersion, technology isolation, and poor scenario adaptability in traditional dispatching services, and achieve intelligent upgrade of the entire business process to improve dispatching efficiency and service quality.

[0027] The operator integrated scheduling intelligent agent platform proposed in this invention is divided into four layers from bottom to top: data layer, model layer, intelligent agent core layer, and application layer. Each layer functions collaboratively, with a core emphasis on the collaboration between large and small models and the collaboration among multiple intelligent agents, forming a fully intelligent scheduling system. Details are as follows: Example 1 Figure 1This is a schematic diagram of the structure of a large-scale operator integrated dispatching system provided in Embodiment 1 of the present invention. The system is applicable to the full-process automation of business scenarios such as home broadband installation and maintenance, and integrated dispatching.

[0028] like Figure 1 As shown, the first embodiment of the present invention provides an integrated operator scheduling system based on a large model, comprising: a data layer 100, a model layer 200, and an intelligent agent core layer 300; Data layer 100 is used to acquire basic business data, preprocess the basic business data and store it in the database; the basic business data includes business data from external systems and internal business data of operators, the internal business data of operators includes operator operation and maintenance cases and manuals, installation and maintenance integrated commissioning records, operator professional knowledge and installation and maintenance scene pictures; Model layer 200 is used to train models based on the installation and maintenance scene images and the training dataset stored in the database to obtain multiple target models; the target models include a fine-tuned basic large model, an installation and maintenance image recognition small model, and a telecommunications domain installation and maintenance intent recognition small model. The core layer 300 of the intelligent agent is used to acquire multimodal data related to the task input by the user, process the multimodal data by calling the database and the target model, obtain multiple sub-tasks of the task and execute them to obtain execution results, and return a processing solution to the user based on the execution results.

[0029] The data layer 100 is responsible for the collection, cleaning, quality control, and storage of business data, providing high-precision data support for other layers of the system. The database refers to the location where the data layer 100 stores data. Basic business data can include business data from external systems and internal business data of the operator. External systems can refer to business systems such as construction scheduling systems, OM systems, and resource management systems. Business data can be data generated by external systems; this embodiment does not limit this. The operator can refer to the telecommunications operator providing services to users. Internal business data of the operator can refer to data generated or stored by the operator's internal systems, including but not limited to operator operation and maintenance cases and manuals, installation and maintenance integrated commissioning records, operator professional knowledge, and installation and maintenance scene images. Operation and maintenance cases and manuals can refer to cases from the operator's operation and maintenance work process and related manuals, specifically including installation and maintenance operation manuals, fault handling manuals, equipment installation specifications, construction standards, typical fault cases, anomaly handling procedures, experience summary documents, and technical guidance manuals. Installation and maintenance integrated commissioning records can refer to the records made by operators during the installation and maintenance integrated commissioning process; operator professional knowledge can refer to domain knowledge in the operator's field; and installation and maintenance scene pictures can refer to the on-site real-scene image data taken and uploaded by operators during the installation and maintenance process.

[0030] In this embodiment, the data layer 100 can acquire basic business data, such as business data from external systems and internal business data of the operator. The data layer 100 can preprocess this basic business data and store it in the database for use by other layers.

[0031] Model layer 200 can be used to train multiple models, constructing a multi-model collaborative system to improve the accuracy of core tasks. The training dataset can be constructed from internal business data of the operator by data layer 100. The target model may include a fine-tuned basic large model, a small model for installation and maintenance image recognition, and a small model for installation and maintenance intent recognition in the telecommunications field, etc.

[0032] In this embodiment, the model layer 200 can train the model based on the installation and maintenance scene images and the training dataset stored in the database to obtain multiple target models, such as a fine-tuned basic large model, an installation and maintenance image recognition small model, and a telecommunications field installation and maintenance intent recognition small model.

[0033] The core layer 300 of the intelligent agent can be composed of multiple intelligent agents, forming a closed-loop collaboration of "input-planning-execution-feedback". Subsequent intelligent agents all take the output of the preceding intelligent agents as their core input, achieving seamless connection of the entire task process. Tasks can be various business requests initiated by staff or ordinary users in the context of operator integrated dispatching. These tasks can be of various types, such as installation and maintenance, fault handling, knowledge query, data statistics, quality inspection, or business recommendation. For example, tasks could include "processing broadband outage work orders in Xinghualing District, XX City", "querying the self-processing rate of installation and maintenance work orders in the province this month", and "uploading photos of optical modem installation for quality inspection". Multimodal data can refer to multi-source heterogeneous data containing different types and forms. Multimodal data can include one or more combinations of text, voice, images, or other types. Subtasks can refer to the smallest independently executable standardized task unit. Execution results can refer to the response data returned after executing the task, and processing solutions can refer to the handling suggestions or operation instructions formed by summarizing and integrating the execution results.

[0034] In this embodiment, after completing the data preparation of the data layer 100 and the model training of the model layer 200, the agent core layer 300 can obtain the multimodal data related to the task input by the user, process the multimodal data by calling the database and the target model, obtain multiple sub-tasks of the task and execute them to obtain the execution results, and return the processing solution to the user based on the execution results.

[0035] This invention provides a large-scale operator integrated dispatching system, comprising: a data layer, a model layer, and an intelligent agent core layer. The data layer acquires basic business data, preprocesses the basic business data, and stores it in a database. The basic business data includes business data from external systems and internal operator business data, including operator operation and maintenance cases and manuals, installation and maintenance integrated dispatch records, operator professional knowledge, and installation and maintenance scene images. The model layer trains models based on the installation and maintenance scene images and the training dataset stored in the database to obtain target models. The target models include a fine-tuned basic large model, an installation and maintenance image recognition small model, and a telecommunications domain installation and maintenance intent recognition small model. The intelligent agent core layer acquires multimodal data corresponding to a user-input task, processes the multimodal data by calling the database and the target model to obtain multiple sub-tasks of the task, executes them to obtain execution results, and returns a processing solution to the user based on the execution results. The system prepares data at the data layer to provide high-precision data support for the model, trains the model at the model layer, and finally processes the user-input tasks through the intelligent agent core layer and returns the processing solution to the user. This enables the full-process automation upgrade of operation and maintenance business, reduces manual operation, and solves the problem of low efficiency caused by the lack of a unified and standardized management and control mechanism in the existing technology for comprehensive scheduling.

[0036] Based on the above embodiments, modified embodiments of the above embodiments are proposed. It should be noted that, in order to keep the description brief, only the differences from the above embodiments are described in the modified embodiments.

[0037] In one embodiment, the data layer 100 includes a real-time data access module, a file data integration module, a training data construction module, and a knowledge data structuring module; The real-time data access module is used to receive real-time business data from external systems through a standardized channel, preprocess the real-time business data to obtain processed business data, and store the processed business data in a database. The file data integration module is used to parse the operator's unstructured operation and maintenance case and manual files, extract key information from the operation and maintenance case and manual files, convert the key information into a structured format to obtain first structured data, and send the first structured data to the knowledge data structuring module; the operation and maintenance case and manual files include at least installation and maintenance manuals and fault cases; The training data construction module is used to clean the operator's installation and maintenance integrated dispatch records through a three-level verification mechanism to obtain a training dataset, and store the training dataset in the database; the installation and maintenance integrated dispatch records include installation and maintenance support orders and integrated dispatch chat data. The knowledge data structuring module is used to vectorize the text content in the first structured data and the text content in the operator's professional knowledge through an embedding model and store them in a vector database, store the structured data in the professional knowledge in a relational database, and create indexes for the key fields in the relational database.

[0038] The real-time data access module can be used to receive data from external systems in real time, and the database type can be set according to actual needs. The standardized channel refers to the unified data interaction link built between the real-time data access module and the external system for data transmission. The file data integration module can be used to structure unstructured data. Key information in the operation and maintenance case and manual files refers to core business information that supports standardized execution of subsequent installation and maintenance operations, fault location and resolution, model training, and knowledge-based Q&A. The operation and maintenance case and manual files can include at least an installation and maintenance manual and fault cases. The installation and maintenance manual can be a standardized operation guide document developed by the operator for daily maintenance and other installation and maintenance operations, and the fault cases can be a collection of fault cases that occur during the operator's work.

[0039] The training data construction module can be used to build datasets. Installation and maintenance integrated dispatch records can include installation and maintenance support orders and integrated dispatch chat data. Installation and maintenance support orders can be structured business data generated when installation and maintenance personnel request assistance from the backend when encountering problems during operations, within the operator's integrated dispatch system. Integrated dispatch chat data can be unstructured text interaction data between installation and maintenance personnel and backend personnel during the collaborative processing of installation and maintenance support orders. The three-level verification mechanism can include first-level filtering of invalid data, second-level removal of duplicate data, and third-level verification of annotation accuracy. The knowledge data structuring module can be used to store data. The embedding model can refer to a model that transforms text-based business knowledge into computer-understandable vectors. The vector database can be a database that stores vector data. The relational database can be a traditional database that organizes, stores, and manages structured data using a two-dimensional table structure. Key fields can be core data fields in a relational database used for unique identification, fast retrieval, business association, and conditional filtering.

[0040] In this embodiment, the data layer 100 comprises four core modules: a real-time data access module, a file data integration module, a training data construction module, and a knowledge data structuring module. The real-time data access module receives real-time business data from external systems through a standardized channel, preprocesses the real-time business data, and stores the processed business data in a database. The file data integration module parses unstructured operation and maintenance case studies and manuals from operators, extracts key information from these documents, converts the key information into a structured format, and sends the first structured data to the knowledge data structuring module. The training data construction module cleans the operator's installation and maintenance records using a three-level verification mechanism to obtain a training dataset, which is then stored in the database. The knowledge data structuring module vectorizes the text content in the first structured data and the text content in the operator's professional knowledge using an embedding model and stores it in a vector database. It also stores the structured data in the professional knowledge in a relational database and indexes the key fields in the relational database.

[0041] For example, the standardized channel for the real-time data access module can be the Kafka message queue. It can access real-time business data (such as work order progress, equipment status, and network topology) from more than 10 external systems, such as the construction scheduling system, OM system, and resource management system. The data transmission latency of the Kafka message queue is ≤500ms, and it supports both incremental synchronization and full synchronization modes.

[0042] The file data integration module can parse unstructured files (docx, pdf, rtf formats) such as installation and maintenance manuals and fault cases, extract key information (such as fault phenomena and handling steps), and convert them into JSON structured format for easy subsequent vectorized storage.

[0043] The training data construction module can collect 300,000 installation and maintenance support orders and 960,000 comprehensive dispatch chat data. It cleans the data through a "three-level verification" mechanism (first level filtering invalid data, second level removing duplicate data, and third level verifying the accuracy of the annotation) to form a high-quality training dataset with a data validity rate of ≥95%.

[0044] The knowledge data structuring module can embed eight categories of professional knowledge (such as terminal products, home opening process operation, installation common problem handling manual, comprehensive adjustment case library, comprehensive adjustment knowledge sharing, FTTR manual, complaint work order handling, and installation process issues) into a vectorized model (vector dimension d=1024) and store it in the Milvus vector database; structured data (such as user information and device parameters) can be stored in a MySQL relational database and indexed (such as user identifier and device number), with a query response time ≤100ms.

[0045] In one embodiment, the training data construction module is specifically used to: use the equipment maintenance and commissioning records as the original dataset; filter invalid data in the original dataset based on structured query language statements and regular expressions to obtain a first-level valid dataset; remove duplicate data in the first-level valid dataset using a cosine similarity algorithm to obtain a second-level valid dataset; perform annotation verification on the data in the second-level valid dataset through manual sampling and model pre-evaluation to obtain a third-level valid dataset and the corresponding annotation accuracy; when the annotation accuracy is not lower than a preset threshold, use the third-level valid dataset as the training dataset.

[0046] Among them, Structured Query Language (SQL) statements are a set of instruction statements written based on SQL syntax rules, used to manipulate data in a database or manage the database structure. Specific statements can be set according to actual needs. Regular expressions refer to expressions that achieve fast matching, searching, replacing, validating, and filtering of text data through predefined characters, symbols, and combination rules. Cosine similarity algorithms are algorithms used to calculate the similarity of the cosine value of the angle between two high-dimensional vectors in space. Preset thresholds can be set according to actual needs.

[0047] In this embodiment, the training data construction module cleans the operator's installation and maintenance integrated dispatch records through a three-level verification mechanism. First, the installation and maintenance integrated dispatch records are used as the original dataset. A first-level verification is performed based on structured query statements and regular expressions to filter invalid data in the original dataset, resulting in a valid dataset after the first-level verification. Then, a second-level verification is performed using the cosine similarity algorithm to remove duplicate data in the valid dataset after the first-level verification, resulting in a valid dataset after the second-level verification. Finally, a third-level verification is performed through manual sampling and model pre-evaluation to label and verify the data in the valid dataset after the second-level verification, resulting in a valid dataset after the third-level verification and the corresponding labeling accuracy. When the labeling accuracy is not lower than a preset threshold, the valid dataset after the third-level verification can be used as the training dataset.

[0048] For example, Level 1 validation (data extraction stage): Filtering invalid data based on SQL statements and regular expressions. The Level 1 validation formula is as follows: ; in, For the original dataset, This is the valid dataset after the first-level validation. regex is the regular expression for the data format (e.g., mobile phone numbers need to match "1[3-9]\d{9}"). This stage can filter out 80% of invalid data.

[0049] Secondary verification (data cleaning stage): The cosine similarity algorithm is used to remove duplicate data. If the similarity between two pieces of data is ≥95%, it is determined as duplicate and one is retained. The formula of the cosine similarity algorithm is as follows: ; where , are the vectors after data vectorization. In this stage, the data duplication rate can be reduced from 15% to below 3%.

[0050] Tertiary verification (data annotation stage): Manual sampling (sampling ratio 10%) is combined with model pre-evaluation to verify the annotation accuracy. The annotation accuracy threshold T = 95% is set. If the sampling accuracy Acc < T, re-annotation is performed to ensure that the annotation accuracy of the training data is ≥95%.

[0051] Dynamic evaluation: Key Quality Indicator (KQI) metrics (integrity, accuracy, timeliness) for data quality are established, and real-time monitoring is carried out to generate a quality report. For example, integrity = (number of non-empty fields / total number of fields) × 100%. If the integrity < 90%, the data repair process is automatically triggered (such as resending a data request to the business system) to ensure stable data quality.

[0052] In one embodiment, the model layer 200 includes a basic large model selection and fine-tuning module and a special small model training module; The basic large model selection and fine-tuning module is used to select a basic large model from candidate models, and combine part of the data in the training data set to fine-tune the basic large model through a fine-tuning method based on low-rank adaptation to obtain a fine-tuned basic large model; The special small model training module is used to train an image recognition basic model through an optimized loss function and installation and maintenance scenario pictures to obtain a small model for installation and maintenance image recognition; The fine-tuned basic large model is used as a teacher model, and combined with knowledge distillation technology and the training data set to train a text basic model to obtain a small model for recognizing the installation and maintenance intention in the telecommunications field.

[0053] The module for selecting and fine-tuning the basic large model can be used to train basic large models for operator integrated dispatch scenarios. Candidate models can include qwen2.5-72b, qwen2.5-57b, etc. Fine-tuning can be a model optimization method that uses domain datasets to retrain pre-trained models to improve their performance in specific scenarios, including full fine-tuning and lightweight fine-tuning. Low-Rank Adaptation (LoRA) can be an efficient fine-tuning algorithm for large models. The module for training specialized small models can be used to train specialized small models. The image recognition basic model can be a pre-built model architecture. The text basic model can be a basic pre-trained model based on a large-scale general text corpus. The installation and maintenance image recognition small model can be a lightweight visual model for recognizing images of installation and maintenance field equipment, faults, construction specifications, and indicator light status. The telecommunications domain installation and maintenance intent recognition small model can be a lightweight semantic understanding model for recognizing installation and maintenance business intent and extracting business entities from integrated dispatch text.

[0054] In this embodiment, the model layer 200 may include a basic large model selection and fine-tuning module and a specialized small model training module. The basic large model selection and fine-tuning module can select a basic large model from the candidate models, and fine-tune the basic large model by combining some data in the training dataset and using a low-rank adaptation-based fine-tuning method to obtain a fine-tuned basic large model. The specialized small model training module can train the image recognition basic model by optimizing the loss function and installation and maintenance scene images to obtain an installation and maintenance image recognition small model. Alternatively, the fine-tuned basic large model can be used as a teacher model, and combined with knowledge distillation technology and the training dataset to train the text basic model to obtain a telecommunications domain installation and maintenance intent recognition small model.

[0055] For example, the basic large model selection and fine-tuning module: From the candidate models, through intent recognition and fault boundary scenario testing, qwen2.5-72b was selected as the basic large model (initial intent recognition accuracy 72%); a "Fine-tuning + LoRA" hybrid fine-tuning method was adopted, through full fine-tuning of key layers (such as the output layer and key feature layer), and light adjustment of general layers using LoRA, using 100,000 pieces of labeled data in the equipment and maintenance field for optimization, and the fine-tuning loss function is as follows: ; in, Cross-entropy loss is used for intent classification tasks. The mean squared error loss is used to predict continuous parameters in work orders, such as fault location, equipment parameters and other numerical information. α=0.7 and β=0.3 are weighting coefficients. After fine-tuning, the intention recognition accuracy is improved to 94%.

[0056] The specialized small model training module can train two types of specialized small models to meet multimodal processing needs: A small-scale image recognition model for installation and maintenance: Based on PaddleOCR 3.0, a full-process framework of "detection-correction-recognition" is constructed. In the detection stage, the model employs Differentiable Binarization (DB) algorithm. By optimizing the loss function (e.g., adding functions for bounding box regression loss and text region perception loss), the accuracy of locating the text region on the optical modem nameplate can reach 98%. In the correction stage, Thin Plate Spline (TPS) transformation and an orientation classifier are used to standardize text boxes with tilt angles ≤45°, achieving an orientation calibration accuracy of 99%. In the recognition stage, a Residual Network + Connectionist Temporal Classification (ResNet + CTC) architecture is used, combined with 50,000 installation and maintenance scene images (optical modem serial number, speed test screenshots, device tags) for fine-tuning. This achieves an accuracy of ≥95% for optical modem serial number (SN) recognition and ≥90% for speed test results (download / upload speed).

[0057] A small-scale model for installation and maintenance intent recognition in the telecommunications field: Based on a pre-trained model using Enhanced Representation through kNowledge IntEgration (ERNIE), a three-tiered training system of "pre-training-fine-tuning-distillation" is constructed, tailored to the business characteristics of the installation and maintenance domain. The upper layer uses knowledge distillation technology to transfer the professional capabilities of the fine-tuned large model to the small model, while integrating a retrieval enhancement mechanism to achieve real-time knowledge updates. Specifically, this includes: introducing a knowledge base in the installation and maintenance domain (such as covering fault cases, processing procedures, and professional terminology) to build a retrieval index; obtaining relevant knowledge fragments through semantic similarity matching during the inference stage and fusing them with the model output for decision-making; and employing temperature scaling and a knowledge distillation loss function (distillation loss weight 0.3) to use the fine-tuned large model as a teacher model to guide the small model's learning. While maintaining a 92% intent recognition accuracy, the number of model parameters is reduced by 70%, and the inference speed is increased by 3 times, meeting the deployment requirements of edge devices. The bottom layer can incorporate a dictionary of installation and maintenance domains (containing 327 professional terms such as "excessive optical attenuation" and "optical network unit (ONU) offline") to optimize the word segmenter. The middle layer can use multi-label classification loss functions, such as functions that fuse cross-entropy and focal loss, to solve the problem of imbalanced samples with similar intents, such as single-user network outages and regional faults.

[0058] The training data can consist of 300,000 installation and maintenance support orders (labeled with three categories: fault link, phenomenon, and error) and 960,000 integrated dispatch chat data (labeled with user intent type), divided into training / validation / test sets in an 8:1:1 ratio after three levels of validation. For low-frequency scenarios (such as FTTR network failures), a Synthetic Minority Over-sampling Technique (SMOTE) is used to supplement and generate 5,000 synthetic samples, ensuring that the sample size for each scenario is ≥3,000.

[0059] The model's performance metrics and adaptability have both improved: On the test set, the model achieved an accuracy rate of 79.01% in operational scenarios (such as work order dispatch), 87.98% in query scenarios (such as optical power query), and 86.70% in diagnostic scenarios (such as fault localization). The model supports dynamic incremental training; when adding new scenarios (such as 5G home broadband converged faults), only 500 labeled data points are needed, and the model iteration can be completed within 24 hours without full retraining.

[0060] In one embodiment, the core agent layer 300 includes a multimodal input processing agent, a task planning agent, a tool scheduling agent, and a business flow orchestration agent; The multimodal input processing agent is used to acquire multimodal data related to the task input by the user, and based on the multimodal data, calls the installation and maintenance image recognition small model, the telecommunications field installation and maintenance intention recognition small model and the database to generate structured parameters, and sends the structured parameters to the task planning agent. The task planning agent is used to call the fine-tuned basic model to parse the structured parameters to obtain task parameters. Based on the task parameters, the task is divided into multiple sub-tasks in combination with the preset business process graph and atomic capability list. The execution order of each sub-task is determined by the task scoring rules and priority formula to obtain a structured task list. The structured task list is then sent to the business flow orchestration agent. The business flow orchestration agent is used to match standardized process templates for each subtask based on the structured task list, add branch judgment and loop execution logic to the standardized process template, convert it into a standardized execution script, and send the standardized execution script to the tool scheduling agent. The tool scheduling agent is used to call external system interfaces and tools to execute each sub-task, obtain the execution results returned by the external system interfaces and tools, and return the execution results to the task planning agent so that the task planning agent can return a processing plan to the user based on the execution results.

[0061] The structured parameters can include task type T (e.g., fault handling), geographical information L (e.g., XX city XX district), business parameters P (e.g., work order number, user ID), and a parameter list. Task parameters can be parsed work order number, user number, task type, etc., which are not limited in this embodiment. The business process diagram can be a standardized process knowledge set of the entire telecommunications installation and maintenance business process. The atomic capability list can be a capability set composed of the smallest independently executable standardized functional units of the system. The task scoring rules can be standardized criteria for multi-dimensional quantitative scoring of tasks, and the priority formula can be a formula for calculating the priority value of a task. The structured task list can be a set of subtasks of a task, which can include sub-subtask sequences and priority lists. The standardized process template can be a reusable process framework. The standardized execution script can be a set of instructions that clearly defines the execution steps, specifications, and verification rules of subtasks.

[0062] In this embodiment, the core layer 300 may include a multimodal input processing agent, a task planning agent, a tool scheduling agent, and a service flow orchestration agent. The multimodal input processing agent can acquire multimodal data related to the task input by the user. Based on the multimodal data, it calls a small-scale installation and maintenance image recognition model, a small-scale installation and maintenance intent recognition model in the telecommunications field, and the database to generate structured parameters, which are then sent to the task planning agent. The task planning agent can call a fine-tuned basic large model to parse the structured parameters to obtain task parameters. Based on the task parameters, and combined with a preset service flow graph and atomic capability list, the task can be divided into multiple sub-tasks. Then, the execution order of each sub-task is determined through task scoring rules and priority formulas to obtain a structured task list, which is then sent to the service flow orchestration agent. The workflow orchestration agent can match standardized process templates for each subtask based on a structured task list. Branching and loop execution logic can be added to the matched standardized process templates, and the resulting standardized process template is converted into a standardized execution script, which is then sent to the tool scheduling agent. The tool scheduling agent can call external system interfaces and tools to execute each subtask and obtain the execution results returned by these interfaces and tools. The execution results are then returned to the task planning agent, which can then return a processing solution to the user based on these results.

[0063] In one embodiment, the multimodal input processing agent is specifically used to: acquire task-related multimodal data input by the user; call the installation and maintenance image recognition mini-model and the telecommunications domain installation and maintenance intent recognition mini-model to extract key information from the multimodal data; call the database of the data layer based on the key information to match and inherit corresponding core business parameters for the multimodal data; verify the core business parameters, and send structured parameters to the task planning agent when the verification passes; the structured parameters include the key information and the core business parameters.

[0064] Among them, the core service parameters can be various core service parameters such as broadband number, SN, optical line terminal (OLT), and passive optical network (PON).

[0065] In this embodiment, the multimodal input processing agent can acquire multimodal data related to the task input by the user, call the installation and maintenance image recognition mini-model and the telecommunications field installation and maintenance intent recognition mini-model, extract key information from the multimodal data, call the database of the data layer based on the key information, match and inherit the corresponding core business parameters for the multimodal data, verify the core business parameters, and send the structured parameters to the task planning agent when the parameter verification passes.

[0066] For example, Figure 2 This is the first flowchart illustrating the interaction between various intelligent agents in the core layer of an intelligent agent according to Embodiment 2 of the present invention. Figure 3 This is a second flowchart illustrating the interaction between agents in the core layer of an intelligent agent system, as provided in Embodiment 2 of the present invention. Figure 2 and Figure 3 As shown, the multimodal input processing agent provides pre-processing data support for the task planning agent. The multimodal input processing agent supports five input types: voice, text, image, QR code, and local file. Image input supports JPEG and PNG formats. After an image is input, a multi-task image model is automatically triggered to extract key information (such as the optical modem's serial number and fault location). The multi-task image model can be composed of Optical Character Recognition (OCR) and object detection. QR code input can automatically obtain the work order number, user ID, and location information by parsing the work order QR code, with a parameter extraction accuracy of ≥99%.

[0067] The input parameter management system establishes a parameter inheritance and verification mechanism, which automatically inherits 12 core parameters such as broadband number, SN, OLT, and PON from the installation, relocation, and replacement work orders stored in the database, with an inheritance accuracy of ≥99%. During the parameter verification stage, regular expressions (e.g., SN must match "86 [A-Z0-9]{14}") and business rules (e.g., OLT number must exist in the resource management system) can be used for dual verification. When parameters are incorrect, the system will automatically prompt and guide the correction, reducing the failure rate of business processing.

[0068] The task planning agent serves as the entry point for collaboration and the core of task decomposition. It acts as the overall scheduler for multi-agent collaboration, responsible for parsing the original requirements, breaking down them into ordered subtasks, and outputting a priority-based task list, providing clear execution goals for all subsequent agents. Specifically: (1) Task Analysis After receiving the structured parameters corresponding to the user-input task (such as "handling broadband outage work order in Hualing District, XX City"), the data in the structured parameters can be identified through the fine-tuned large model, and the task type T (fault handling), the regional information L (Hualing District, XX City), and the business parameters P (work order number, user ID) can be extracted.

[0069] (2) Task breakdown Based on the business process diagram (including 6 major links: "work order reception - intent recognition - fault diagnosis - solution generation - task execution - result feedback") and the list of 23 atomic capabilities (such as optical attenuation query, account reset, and fault delimitation) compiled by business experts, complex tasks are broken down into ordered sub-tasks. For example, "network outage work order processing" is broken down into four sub-tasks: "user information query → resource status verification → fault delimitation → repair solution generation".

[0070] (3) Task scoring rules: The large model scores based on prompts. For the three metrics of urgency, complexity, and system load of subtasks, explicit prompt rules are designed, and the large model automatically outputs a score of 1-5 (urgency / complexity) or 0-1 (system load): The urgency level (1-5 points) prompt is: "Based on the telecom operator's installation and maintenance / integrated dispatch business rules, score the following sub-tasks (1=lowest urgency, 5=highest urgency): 1 point = non-core business (e.g., historical work order query), no user complaints; 2 points = general consultation tasks (e.g., installation and maintenance process consultation), no service interruption; 3 points = minor single-user fault (e.g., single-household WIFI lag), no complaints; 4 points = serious single-user fault (e.g., single-household network outage), potential complaints if not repaired within 12 hours; 5 points = multi-user fault (e.g., regional network outage / batch device offline), has generated concentrated complaints or affected key customers (818 customers / government and enterprise customers). Current sub-task: {sub-task description}, please output the score." The complexity (1-3 points) prompt is: "Based on the complexity rules of telecom operators' installation and maintenance / integrated dispatching services, score the following sub-tasks (1 = minimum complexity, 3 = maximum complexity): 1 point = single system operation (e.g., querying the status of a single device / work order volume in a single area), no cross-system data required; 2 points = cross-system collaboration (e.g., combining the resource system and OM system to troubleshoot a single user's fault), requiring 1-2 types of data support; 3 points = cross-system collaboration (e.g., regional faults require linkage with the resource / OM / construction scheduling system), requiring multi-type data fusion + topology analysis. Current sub-task: {sub-task description}, please output the score." The system load (0-1 point) prompt is: "Based on the current system resource usage, score the following subtasks (0 = lowest load, 1 = highest load): 0 points = target system (e.g., resource management system) current concurrency < 200; 0.3 points = concurrency 200-500; 0.5 points = concurrency 500-1000; 0.8 points = concurrency 1000-2000; 1 point = concurrency ≥ 2000. The current subtask needs to call the system: {target system name}, current system concurrency: {real-time concurrent data}, please output the score (rounded to 1 decimal place)." (4) Priority calculation and output The execution order of subtasks is calculated using a priority formula, outputting a structured task list containing "subtask identifier, task description, priority score, dependent systems, and core parameters." This provides the execution basis for the tool's scheduling agent and the retrieval-enhanced question-answering agent. The priority formula is: ; in, The priority of the i-th subtask is denoted as , where a higher priority value means that the subtask is scheduled more quickly. The urgency of the i-th subtask is represented by a score ranging from 1 to 5, with higher scores indicating greater urgency. Let be the complexity of the i-th subtask, with a value ranging from 1 to 3. A higher score indicates a more complex task to process. This represents the system load, reflecting the load pressure on the system when processing this subtask. The value ranges from 0 to 1, with the closer the value is to 1, the greater the system load. The weighting coefficient for urgency. The weighting coefficients for complexity. This is the weighting factor for the system load. =0.5、 =0.3、 =0.2, used to balance the impact of these three factors on task priority. Here, urgency has the highest weight, indicating that task urgency is a key factor to consider during resource scheduling. After calculating the priority of subtasks according to this formula, subtasks with higher priority will be scheduled first, and the accuracy of subtask allocation can reach ≥95%, ensuring the effectiveness and accuracy of resource scheduling.

[0071] This embodiment employs a dynamic priority scheduling algorithm combining "rule-based scoring and real-time load awareness" to achieve precise subtask ranking and optimized resource allocation. By quantifying business experience into calculable scoring metrics through Prompt rules, subjective human judgment is avoided, achieving an accuracy rate of ≥95% for urgency and complexity scoring. By introducing dynamic system concurrency data, the priority of corresponding subtasks is automatically reduced when the target system load is too high (e.g., ≥2000 concurrent users), preventing system congestion. Furthermore, by supporting adjustments to the weighting coefficients of urgency, complexity, and load based on business needs (e.g., increasing the urgency weight to 0.6 when prioritizing customer experience), it can adapt to different business scenarios. The priority list output by the algorithm can be directly used as input for tool scheduling agents and business flow orchestration agents without secondary conversion, ensuring high efficiency in multi-agent collaboration and a subtask scheduling accuracy rate of ≥95%.

[0072] The workflow orchestration agent can automatically write scripts using the "task list + priority" output by the task planning agent as its core input, and then output the written scripts to the tool scheduling agent. The input to the workflow orchestration agent is: a sequence of subtasks from the task planning agent (e.g., "user information query → resource status check → fault delimitation → solution generation"); then, process matching is performed: matching the corresponding standardized scenario template based on the subtask type (e.g., "regional network outage handling process"); further, branch judgments are added (e.g., "trigger line maintenance if optical attenuation value > -28dBm") and loop execution is performed (e.g., "retry interface call 3 times") to ensure process closure; finally, a standardized execution process script is output and synchronized to the tool scheduling agent, ensuring that subtasks are executed in an orderly manner according to the logic set by the task planning agent, with a process compliance rate ≥99%.

[0073] The tool scheduling agent serves as the execution platform for the task planning agent. It takes the standardized execution script output by the business flow orchestration agent as its core input and can automatically call external system interfaces and tools to complete sub-task execution, with the execution results fed back to the task planning agent. The input is the standardized execution script from the business flow orchestration agent; then, interface matching is performed: based on the sub-task type (e.g., "resource query" → connecting to the resource management system), an Application Programming Interface (API) is automatically matched; priority scheduling executes calls according to the priority order calculated by the task planning agent, with higher-priority tasks occupying system resources first; if a call fails, an exponential backoff strategy is used for retrying to ensure execution continuity; finally, the output is the sub-task execution result (e.g., device status query result, work order dispatch success receipt), with a call success rate ≥99.2%. The execution results are synchronized to the task planning agent in real time for subsequent task closure.

[0074] In one embodiment, the agent core layer 300 further includes a retrieval-enhanced question-answering agent; The task planning agent is used to send request data to the retrieval enhancement question answering agent when knowledge enhancement support is needed during the generation of subtasks. The retrieval-enhanced question-answering agent is used to invoke the fine-tuned basic model, search for corresponding candidate knowledge from a vector database and / or a relational database based on the requested data, and return the candidate knowledge to the task planning agent so that the task planning agent can continue to execute the process of generating subtasks using the candidate knowledge.

[0075] The enhanced question-answering agent can be an intelligent executor that provides users with intelligent question-answering and fault decision-making services through knowledge base retrieval. Requested data can be relevant data supporting the requested knowledge, such as user identifiers and fault descriptions. Candidate knowledge can include a variety of different types of knowledge.

[0076] In this embodiment, the core layer 300 of the intelligent agent may further include a retrieval-enhanced question-answering intelligent agent. When the task planning intelligent agent needs knowledge enhancement support during the generation of sub-tasks, it can send the request data to the retrieval-enhanced question-answering intelligent agent. The retrieval-enhanced question-answering intelligent agent can call the fine-tuned basic large model, search for one or more corresponding candidate knowledge from the vector database and / or relational database based on the request data, and return the candidate knowledge to the task planning intelligent agent.

[0077] For example, the enhanced question-answering agent provides knowledge and data support for task planning. Knowledge storage includes vectorized unstructured knowledge (installation and maintenance manuals, fault cases) stored in the Milvus vector database, and structured data (work order volume, equipment status) stored in MySQL. The retrieval method can be a hybrid retrieval: after a user asks a question, two types of retrieval will be triggered.

[0078] Knowledge base retrieval involves calculating the query vector Q and the knowledge vector. The cosine similarity is used to select the top 5 most similar knowledge points as candidate knowledge. ; Database retrieval involves converting natural language (e.g., "query the number of installation and maintenance work orders in Xiaodian District, XX City in the past 7 days") into SQL statements using a Natural Language to Structured Query Language (NL2SQL) algorithm. SELECT COUNT (*) FROM work_order WHERE area=' XX City Xiaodian District' AND create_time >= DATE_SUB (NOW(), INTERVAL 7 DAY); When generating the answer, candidate knowledge and database query results are input into the large model, and multi-source information is fused through an attention mechanism. The attention weight formula is as follows: ; in, As candidate knowledge weights, The weights are assigned to the database results. Execution results are synchronized to the task planning agent in real time for subsequent task closure.

[0079] In one embodiment, the core layer 300 of the intelligent agent also includes a memory-learning intelligent agent; The task planning agent is used to generate an execution log based on the execution result after obtaining the execution result and send the execution log to the memory learning agent; The memory learning agent is used to determine the optimization parameters of the task splitting rules and priority formulas based on the execution logs using reinforcement learning, and send the optimization parameters to the task planning agent so that the task planning agent can optimize the task splitting rules and priority formulas.

[0080] Among them, the memory-learning agent can be an agent that continuously improves its decision-making and execution performance through continuous learning. The execution log can be data that records the execution content, time, results, and exception information at each node of the task's entire lifecycle.

[0081] In this embodiment, the core layer 300 of the intelligent agent may further include a memory learning intelligent agent. After obtaining the execution result, the task planning intelligent agent can generate an execution log based on the execution result and send the execution log to the memory learning intelligent agent. Based on the execution log, the memory learning intelligent agent can use reinforcement learning to determine the optimization parameters of the task splitting rules and priority formulas, and send the optimization parameters to the task planning intelligent agent, thereby optimizing the task splitting rules and priority formulas.

[0082] For example, using the execution results of the task planning agent (such as work order processing time and user satisfaction) as input, reinforcement learning is used to optimize task splitting and priority calculation rules, which in turn feeds back into the task planning agent for iteration. The input consists of historical task data of the task planning agent (subtask splitting results, priority scores, execution time, and user satisfaction scores); then, memory storage is performed: task execution data for the past 30 days is cached to establish a task-effect mapping relationship; through reinforcement learning, "user satisfaction" and "processing time" are used as reward signals, and the priority formula weights and subtask splitting rules are optimized using the Deep Q-Network (DQN) algorithm; if the execution accuracy of a certain type of subtask is below 90% for three consecutive days, an anomaly feedback is provided, automatically triggering a data labeling task to supplement training data; finally, the optimized task splitting rules and priority weight parameters (such as adjusting the urgency weight to 0.6) are output and synchronized to the task planning agent, continuously improving its decision-making accuracy, resulting in a 15% improvement in work order processing efficiency after iteration.

[0083] In one embodiment, the system further includes an application layer, which includes a knowledge Q&A assistant, a question and data retrieval assistant, an image quality inspection assistant, an installation and maintenance support assistant, and a sales assistant. The knowledge question-and-answer assistant is used to obtain knowledge question-and-answer tasks sent by staff, send the knowledge question-and-answer tasks to the core layer of the intelligent agent, receive the first processing solution returned by the core layer of the intelligent agent, and display the first processing solution to the staff in a visual manner. The data retrieval assistant is used to obtain data retrieval tasks sent by staff, send the data retrieval tasks to the core layer of the intelligent agent, receive the second processing solution returned by the core layer of the intelligent agent, and display the second processing solution to the staff in a visual manner. The image quality inspection assistant is used to obtain the image quality inspection task sent by the staff, send the image quality inspection task to the core layer of the intelligent agent, receive the third processing solution returned by the core layer of the intelligent agent, and display the third processing solution to the staff in a visual manner. The installation and maintenance support assistant is used to obtain the installation and maintenance support task sent by the staff, send the installation and maintenance support task to the core layer of the intelligent agent, receive the fourth processing solution returned by the core layer of the intelligent agent, and display the fourth processing solution to the staff in a visual manner. The accompanying sales assistant is used to obtain the recommendation task sent by the staff, send the recommendation task to the core layer of the intelligent agent, receive the fifth processing solution returned by the core layer of the intelligent agent, and display the fifth processing solution to the staff in a visual manner.

[0084] The terms "first," "second," "third," "fourth," and "fifth" are used only to distinguish different execution schemes. The knowledge-answering assistant assists users in answering knowledge questions; the data retrieval assistant assists users in processing data; the image quality inspection assistant assists users in processing images; the installation and maintenance support assistant provides installation and maintenance support; and the sales assistant provides product recommendations. The specific content of the knowledge-answering task, data retrieval task, image quality inspection task, installation and maintenance support task, and recommendation task can be set according to actual conditions; this embodiment does not limit this.

[0085] In this embodiment, the system may further include an application layer, which may include a knowledge Q&A assistant, a data retrieval assistant, an image quality inspection assistant, an installation and maintenance support assistant, and a sales assistant. Through these application layer assistants, services can be provided to users in a visual manner. The knowledge Q&A assistant can send knowledge Q&A tasks to the core intelligent agent layer; the data retrieval assistant can send data retrieval tasks to the core intelligent agent layer; the image quality inspection assistant can send image quality inspection tasks to the core intelligent agent layer; the installation and maintenance support assistant can send installation and maintenance support tasks to the core intelligent agent layer; and the sales assistant can send recommendation tasks to the core intelligent agent layer. After receiving the tasks from each assistant, the core intelligent agent layer can process the tasks through its internal intelligent agents and return an execution plan. Each assistant can then visually display the execution plan to the staff.

[0086] For example, based on the core capabilities of intelligent agents, five types of high-frequency applications can be implemented, covering the full range of integrated dispatching services required by operators: (1) Knowledge Q&A Assistant: Integrated into the web page of the integrated dispatch system and the mobile home customer APP, it supports voice and text questions and can answer installation and maintenance business specifications (such as FTTR installation process) and fault handling solutions (such as solutions for excessive optical attenuation). After going online, the usage rate of front-line personnel reached 95%, the accuracy rate of Q&A reached 92%, and the response time was ≤0.8 seconds.

[0087] (2) Data Query Assistant: Provides H5 page and software development kit (SDK) integration, supports natural language data query and automatically generates visual reports (bar chart, line chart, pie chart), such as "query the self-processing rate of installation and maintenance work orders in each city of the province this month", the data query accuracy rate reaches 91%, the cross-table query efficiency is 55% higher than the traditional manual method, and the average daily data retrieval time per person is shortened from 1.5 hours to 0.5 hours.

[0088] (3) Image Quality Inspection Assistant: Embedded in the installation and maintenance work order system, it automatically receives on-site images (optical modem, line, equipment panel) uploaded by installation and maintenance personnel, and detects 10 types of indicators such as "whether the equipment is completely photographed", "whether the indicator light is normal", and "whether the speed test result meets the standard". The quality inspection time is ≤1.5 seconds / image, the accuracy rate reaches 91%, the manual re-inspection rate is reduced from 30% to 5%, and it replaces the traditional manual quality inspection, saving more than 80,000 yuan in labor costs per month.

[0089] (4) Installation and Maintenance Support Assistant: Supports deployment on both the WEB and the Mobile Home Customer APP. The WEB version adds historical session management (filtering records of the last 30 days by time / scenario, displaying the question content, intent tags, and processing results), session query response time ≤ 0.5 seconds, and generates monthly evaluation reports for model and process optimization, improving user satisfaction to over 90%.

[0090] (5) On-demand Sales Assistant: User Profile and Recommendation Algorithm: Constructing a user segmentation model for installation and maintenance scenarios. Users with broadband access within the city are categorized as "818 key customers, ordinary new customers, and purely new customers," and user tags are generated based on consumption data (call charges, data usage, broadband bandwidth), family attributes (number of people, cross-network status), and problem feedback (poor WIFI, low data usage). Users without broadband access outside the city are grouped based on demand tags (data usage demand, voice usage demand, home networking demand). A hybrid recommendation algorithm combining collaborative filtering and rule engine is adopted. High-value customers are prioritized for high-end packages (such as "1000M broadband + 95G data usage"), while price-sensitive customers are recommended cost-effective packages, with a recommendation accuracy rate ≥85%.

[0091] Based on the technical solutions of the above embodiments, this invention provides several specific implementation methods.

[0092] As a specific implementation method, taking the "Handling of Broadband Network Outage Work Order in Xinghualing District, XX City" as an example, we demonstrate the responsiveness and closed-loop collaboration of six types of intelligent agents: Multimodal input processing agent: The user submits the text input "XX City, Xinghualing District, User ID 12345, Home broadband disconnected" through the Palm Home Customer APP. The agent extracts the core parameters (task type: fault handling, region: XX City, Xinghualing District, User ID: 12345) and outputs the structured parameter package to the task planning agent. Task planning agent: After receiving the parameter packet, it automatically scores the subtasks according to the Prompt rule: S1 (user information query) has an urgency of 4 points and a complexity of 1 point, S2 (resource status check) has an urgency of 4 points and a complexity of 2 points, and S3 (fault delimitation) has an urgency of 5 points and a complexity of 3 points; combined with the system load data (all 0.3), it calculates the priority: S3 (3.5) > S2 (2.68) > S1 (2.39), and splits and outputs the priority list of subtasks to the tool scheduling agent and the business flow orchestration agent; Business flow orchestration agent: After receiving the subtask list, it matches the standardized process template of "regional network outage handling" and outputs the execution process script of "fault delimitation → resource verification → user information query → work order generation", which is synchronized to the tool scheduling agent; Tool scheduling agent: Based on priority and process script, first call the OM system interface to execute S3 (fault delimitation), then call the resource management system to execute S2 (resource verification), and finally call the user management system to execute S1 (user information query). The execution results (regional optical attenuation anomaly, 10 users affected, user contact information) are fed back to the task planning agent in real time. The retrieval-enhanced question-answering agent receives the "fault handling solution generation" subtask from the task planning agent, matches "regional optical attenuation anomaly handling steps" from the knowledge base, queries the database for nearby maintenance personnel information, and outputs a standardized repair solution to the task planning agent. Task planning agent: Integrates the results of all sub-tasks to generate a "regional optical attenuation repair work order", which is then assigned to the nearest maintenance personnel by the tool scheduling agent; memory learning agent records the processing time (8 minutes) and user satisfaction (5 points) of the entire process, and subsequently optimizes the priority weight of this type of task through reinforcement learning. The entire process takes ≤10 minutes, which is 70% shorter than traditional manual processing.

[0093] The system in this embodiment constructs a four-layer architecture of "data-model-intelligent agent-application" for a comprehensive operator dispatch intelligent agent platform. Through standardized control at the data layer, collaborative optimization of "large model + small model" at the model layer, multi-dimensional closed-loop collaboration at the intelligent agent layer, and full-scenario implementation at the application layer, it achieves automated upgrades to the entire operation and maintenance process, covering core scenarios such as home broadband installation and maintenance and comprehensive dispatch. By proposing a "three-level verification + dynamic evaluation" data quality control model, it filters invalid data, eliminates duplicate data, and verifies the accuracy of annotations through multi-stage verification, combined with dynamic KQI monitoring, ensuring data integrity ≥95% and accuracy ≥98%. Through the design of a six-level multi-intelligent agent collaboration mechanism, the task planning intelligent agent is based on dynamic... Priority algorithms enable complex task decomposition and resource scheduling; the enhanced question-answering agent integrates knowledge bases and databases to achieve high-precision question answering; and the tool scheduling agent supports automated calls to multimodal tools and external system interfaces, achieving question-answering response times of ≤1 second and image quality inspection times of ≤1.5 seconds per image. A dynamic priority scheduling algorithm combining "rule-based scoring + real-time load awareness" is designed to achieve precise sorting of subtasks and optimized resource allocation. Through the implementation of five core applications—knowledge-based question answering, question data retrieval, image quality inspection, installation and maintenance support, and on-demand sales assistant—precise question answering for installation and maintenance business knowledge, automated data querying, and intelligent image quality inspection are achieved, increasing work order self-processing rates to over 50% and reducing labor costs by 30%.

[0094] By comparing the technical solution of the present invention with the existing solutions, the comparison results are shown in Table 1: Table 1 Comparison of existing solutions with the solution proposed in this application Example 2 Figure 4 This is a flowchart illustrating a large-scale operator integrated scheduling method provided in Embodiment 2 of the present invention. This method is applicable to scenarios involving full-process automation in areas such as home broadband installation and maintenance, and integrated scheduling. For details not covered in this embodiment, please refer to Embodiment 1.

[0095] like Figure 4 As shown in Embodiment 2 of the present invention, a comprehensive operator scheduling method based on a large model includes the following steps: S210. Obtain basic business data through the data layer, preprocess the basic business data and store it in the database.

[0096] The basic business data includes business data from external systems and internal business data from the operator. The internal business data from the operator includes the operator's operation and maintenance cases and manuals, installation and maintenance commissioning records, operator professional knowledge, and installation and maintenance scene pictures.

[0097] In this embodiment, basic business data can be obtained through the data layer, preprocessed, and stored in the database.

[0098] S220. The model is trained by the model layer based on the installation and maintenance scene images and the training dataset stored in the database to obtain multiple target models.

[0099] The target model includes a fine-tuned basic large model, a small model for installation and maintenance image recognition, and a small model for installation and maintenance intent recognition in the telecommunications field.

[0100] In this embodiment, multiple target models can be obtained by training the model layer based on the installation and maintenance scene images and the training dataset stored in the database.

[0101] S230. Obtain multimodal data related to the task input by the user through the core layer of the intelligent agent, process the multimodal data by calling the database and the target model to obtain multiple sub-tasks of the task and execute them to obtain execution results, and return the processing solution to the user based on the execution results.

[0102] In this embodiment, the multimodal data related to the task input by the user can be obtained through the core layer of the intelligent agent. The multimodal data is processed by calling the database and the target model, the task is split into multiple sub-tasks and executed to obtain the execution results. Based on the execution results, the processing solution is returned to the user.

[0103] Embodiment 2 of this invention provides a large-scale operator integrated scheduling method, comprising: acquiring basic service data through a data layer, preprocessing the basic service data and storing it in a database; the basic service data includes service data from external systems and internal operator service data, the internal operator service data including operator operation and maintenance cases and manuals, installation and maintenance integrated scheduling records, operator professional knowledge and installation and maintenance scene images; training models based on the installation and maintenance scene images and the training dataset stored in the database through a model layer to obtain multiple target models; the target models include a fine-tuned basic large model, an installation and maintenance image recognition small model, and a telecommunications domain installation and maintenance intent recognition small model; acquiring multimodal data related to the task input by the user through an intelligent agent core layer, processing the multimodal data by calling the database and the target models to obtain multiple sub-tasks of the task and executing them to obtain execution results, and returning a processing solution to the user based on the execution results. This method prepares data at the data layer to provide high-precision data support for the model, trains the model at the model layer, and finally processes the user-inputted task through the intelligent agent core layer and returns the processing solution to the user. This enables the full-process automation upgrade of operation and maintenance business, reduces manual operation, and solves the problem of low efficiency caused by the lack of a unified and standardized management and control mechanism in the existing technology for comprehensive scheduling.

[0104] Example 3 Figure 5 A schematic diagram of an electronic device 10, which can be used to implement embodiments of the present invention, is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0105] like Figure 5As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0106] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0107] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as a large-model-based operator integrated scheduling method.

[0108] In some embodiments, the large-model-based operator integrated scheduling method can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the large-model-based operator integrated scheduling method described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to execute the large-model-based operator integrated scheduling method by any other suitable means (e.g., by means of firmware).

[0109] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0110] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0111] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0112] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0113] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0114] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0115] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0116] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A large-scale operator integrated dispatching system, characterized in that, The system includes: a data layer, a model layer, and an agent core layer; The data layer is used to acquire basic business data, preprocess the basic business data, and store it in the database. The basic business data includes business data from external systems and internal business data of the operator. The internal business data of the operator includes the operator's operation and maintenance cases and manuals, installation and maintenance integrated commissioning records, operator professional knowledge, and installation and maintenance scene pictures. The model layer is used to train models based on the installation and maintenance scene images and the training dataset stored in the database to obtain multiple target models; the target models include a fine-tuned basic large model, an installation and maintenance image recognition small model, and a telecommunications domain installation and maintenance intent recognition small model. The core layer of the intelligent agent is used to acquire multimodal data related to the task input by the user, process the multimodal data by calling the database and the target model, obtain multiple sub-tasks of the task and execute them to obtain execution results, and return a processing solution to the user based on the execution results.

2. The system according to claim 1, characterized in that, The data layer includes a real-time data access module, a file data integration module, a training data construction module, and a knowledge data structuring module; The real-time data access module is used to receive real-time business data from external systems through a standardized channel, preprocess the real-time business data to obtain processed business data, and store the processed business data in a database. The file data integration module is used to parse the operator's unstructured operation and maintenance case and manual files, extract key information from the operation and maintenance case and manual files, convert the key information into a structured format to obtain first structured data, and send the first structured data to the knowledge data structuring module; the operation and maintenance case and manual files include at least installation and maintenance manuals and fault cases; The training data construction module is used to clean the operator's installation and maintenance integrated dispatch records through a three-level verification mechanism to obtain a training dataset, and store the training dataset in the database; the installation and maintenance integrated dispatch records include installation and maintenance support orders and integrated dispatch chat data. The knowledge data structuring module is used to vectorize the text content in the first structured data and the text content in the operator's professional knowledge through an embedding model and store them in a vector database, store the structured data in the professional knowledge in a relational database, and create indexes for the key fields in the relational database.

3. The system according to claim 2, characterized in that, The training data construction module is specifically used for: Using the installation and maintenance integrated survey records as the original dataset, and based on structured query language statements and regular expressions, invalid data in the original dataset is filtered out to obtain a valid dataset after first-level verification; The duplicate data in the valid dataset after the first-level verification is removed by the cosine similarity algorithm to obtain the valid dataset after the second-level verification. Through manual sampling and model pre-evaluation, the data in the effective dataset after the second-level verification are labeled and verified to obtain the effective dataset after the third-level verification and the corresponding labeling accuracy. When the annotation accuracy is not lower than a preset threshold, the valid dataset after three-level verification is used as the training dataset.

4. The system according to claim 1, characterized in that, The model layer includes a basic large model selection and fine-tuning module and a specialized small model training module; The basic large model selection and fine-tuning module is used to select a basic large model from the candidate models, and combine it with some data in the training dataset to fine-tune the basic large model using a low-rank adaptation-based fine-tuning method to obtain the fine-tuned basic large model. The specialized small model training module is used to train the basic image recognition model by optimizing the loss function and installation and maintenance scene images to obtain a small installation and maintenance image recognition model; the fine-tuned basic large model is used as the teacher model, and combined with knowledge distillation technology and the training dataset, the text basic model is trained to obtain a small model for installation and maintenance intent recognition in the telecommunications field.

5. The system according to claim 1, characterized in that, The core layer of the intelligent agent includes a multimodal input processing intelligent agent, a task planning intelligent agent, a tool scheduling intelligent agent, and a business flow orchestration intelligent agent; The multimodal input processing agent is used to acquire multimodal data related to the task input by the user, and based on the multimodal data, calls the installation and maintenance image recognition small model, the telecommunications field installation and maintenance intention recognition small model and the database to generate structured parameters, and sends the structured parameters to the task planning agent. The task planning agent is used to call the fine-tuned basic model to parse the structured parameters to obtain task parameters, and based on the task parameters, combined with the preset business process graph and atomic capability list, the task is split into multiple sub-tasks. The execution order of each subtask is determined by the task scoring rules and priority formulas to obtain a structured task list, which is then sent to the business flow orchestration agent. The business flow orchestration agent is used to match standardized process templates for each subtask based on the structured task list, add branch judgment and loop execution logic to the standardized process template, convert it into a standardized execution script, and send the standardized execution script to the tool scheduling agent. The tool scheduling agent is used to call external system interfaces and tools to execute each sub-task, obtain the execution results returned by the external system interfaces and tools, and return the execution results to the task planning agent so that the task planning agent can return a processing plan to the user based on the execution results.

6. The system according to claim 5, characterized in that, The multimodal input processing agent is specifically used for: Obtain multimodal data related to the task input by the user, call the installation and maintenance image recognition small model and the telecommunications field installation and maintenance intention recognition small model, and extract key information from the multimodal data; Based on the key information, the database of the data layer is invoked to match and inherit the corresponding core business parameters for the multimodal data; The core business parameters are verified, and the structured parameters are sent to the task planning agent when the verification passes. The structured parameters include the key information and core business parameters.

7. The system according to claim 5, characterized in that, The core layer of the intelligent agent also includes a retrieval-enhanced question-answering intelligent agent; The task planning agent is used to send request data to the retrieval enhancement question answering agent when knowledge enhancement support is needed during the generation of subtasks. The retrieval-enhanced question-answering agent is used to invoke the fine-tuned basic model, search for corresponding candidate knowledge from a vector database and / or a relational database based on the requested data, and return the candidate knowledge to the task planning agent so that the task planning agent can continue to execute the process of generating subtasks using the candidate knowledge.

8. The system according to claim 5, characterized in that, The core layer of the intelligent agent also includes a memory-learning intelligent agent; The task planning agent is used to generate an execution log based on the execution result after obtaining the execution result and send the execution log to the memory learning agent; The memory learning agent is used to determine the optimization parameters of the task splitting rules and priority formulas based on the execution logs using reinforcement learning, and send the optimization parameters to the task planning agent so that the task planning agent can optimize the task splitting rules and priority formulas.

9. The system according to claim 1, characterized in that, The system also includes an application layer, which includes a knowledge Q&A assistant, a question and data retrieval assistant, an image quality inspection assistant, an installation and maintenance support assistant, and a sales assistant. The knowledge question-and-answer assistant is used to obtain knowledge question-and-answer tasks sent by staff, send the knowledge question-and-answer tasks to the core layer of the intelligent agent, receive the first processing solution returned by the core layer of the intelligent agent, and display the first processing solution to the staff in a visual manner. The data retrieval assistant is used to obtain data retrieval tasks sent by staff, send the data retrieval tasks to the core layer of the intelligent agent, receive the second processing solution returned by the core layer of the intelligent agent, and display the second processing solution to the staff in a visual manner. The image quality inspection assistant is used to obtain the image quality inspection task sent by the staff, send the image quality inspection task to the core layer of the intelligent agent, receive the third processing solution returned by the core layer of the intelligent agent, and display the third processing solution to the staff in a visual manner. The installation and maintenance support assistant is used to obtain the installation and maintenance support task sent by the staff, send the installation and maintenance support task to the core layer of the intelligent agent, receive the fourth processing solution returned by the core layer of the intelligent agent, and display the fourth processing solution to the staff in a visual manner. The accompanying sales assistant is used to obtain the recommendation task sent by the staff, send the recommendation task to the core layer of the intelligent agent, receive the fifth processing solution returned by the core layer of the intelligent agent, and display the fifth processing solution to the staff in a visual manner.

10. A comprehensive operator scheduling method based on a large model, characterized in that, This method can be executed by the large-model-based operator integrated scheduling system as described in claim 1, and the method includes: Basic business data is obtained through the data layer, preprocessed, and stored in the database. The basic business data includes business data from external systems and internal business data of the operator. The internal business data of the operator includes the operator's operation and maintenance cases and manuals, installation and maintenance integrated commissioning records, operator professional knowledge, and installation and maintenance scene pictures. Multiple target models are obtained by training the model layer based on the installation and maintenance scene images and the training dataset stored in the database. The target models include a fine-tuned basic large model, an installation and maintenance image recognition small model, and a telecommunications domain installation and maintenance intent recognition small model. The intelligent agent core layer acquires multimodal data related to the task from user input, processes the multimodal data by calling the database and the target model, obtains multiple subtasks of the task, executes them to obtain execution results, and returns a processing solution to the user based on the execution results.

Images