Voice interaction methods, devices, equipment and readable storage media

By parsing user intent through voice interaction methods and a substation operation and maintenance language model, and generating execution strategies, the problem of rapid decision-making caused by the diversity of substation operation and maintenance equipment types is solved, improving operation and maintenance efficiency and accuracy, and adapting to the operation and maintenance needs of both complex and simple scenarios.

CN123119902APending Publication Date: 2026-11-17GUANGDONG DIANWANG GONGSI YUNFU POWER SUPPLY BUREAU
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611556716.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-09
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

There are many types and models of substation maintenance equipment, making it difficult for maintenance personnel to quickly handle uncommon faults. Existing document-based methods are inefficient, lack information correlation, and cannot meet the needs for accurate and real-time queries.

Method used

By employing a voice interaction method, the system analyzes user intent through a substation operation and maintenance language model and knowledge graph, generates execution strategies, integrates multiple types of information, and provides structured operation and maintenance solutions.

Benefits of technology

It improves the efficiency and accuracy of operation and maintenance decision-making, lowers the operational threshold for operation and maintenance personnel, adapts to complex and simple scenarios, shortens the knowledge acquisition time, and ensures the safe and stable operation of power equipment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123119902A_ABST
    Figure CN123119902A_ABST
Patent Text Reader

Abstract

This application discloses a voice interaction method, apparatus, device, and readable storage medium. The method includes converting user speech into initial text; acquiring a substation operation and maintenance language model and a substation operation and maintenance knowledge graph; wherein the substation operation and maintenance language model includes an intent recognition level and a solution generation level; the substation operation and maintenance knowledge graph includes multiple triples, each triple consisting of a corresponding substation operation and maintenance equipment identifier, equipment attribute parameters, and multiple operation and control information; generating user intent using the substation operation and maintenance language model; determining whether the user intent matches two or more operation and control information; if so, generating an execution strategy that satisfies the user intent based on each target control information using the substation operation and maintenance language model; if not, using the operation and control information matched by the user intent as the execution strategy. Therefore, this application can accommodate both simple and complex operation and maintenance scenarios, lower the operational threshold for operation and maintenance personnel, and alleviate the knowledge reserve pressure on operation and maintenance personnel.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of power grid technology, and more specifically, to a voice interaction method, apparatus, device, and readable storage medium. Background Technology

[0002] In the current power industry's substation operation and maintenance field, the diverse types, models, and technical parameters of substation equipment, along with significant differences in daily operating procedures, inspection processes, and fault handling solutions for different types of equipment, place high demands on the knowledge and practical skills of maintenance personnel. In practice, when encountering uncommon or complex faults, frontline maintenance personnel often struggle to handle them quickly and need to consult experienced personnel. However, such consultations are time-consuming, potentially delaying fault resolution and even causing power supply interruptions.

[0003] To address these issues, the industry has compiled fault cases and operational procedures into booklets for maintenance personnel to consult. However, this documented approach has significant drawbacks, especially in complex scenarios with multiple faults. The information is stored linearly and lacks correlation, requiring maintenance personnel to manually search and match segment by segment, which is inefficient, error-prone, and cannot meet the need for accurate and real-time queries. Summary of the Invention

[0004] In view of this, this application provides a voice interaction method, apparatus, device, and readable storage medium to meet the information query needs of substation operation and maintenance personnel.

[0005] To achieve the above objectives, the following solution is proposed: A voice interaction method, comprising: Convert user speech into initial text; Obtain a substation operation and maintenance language model and a substation operation and maintenance knowledge graph; wherein, the substation operation and maintenance language model is obtained through training and includes an intent recognition level and a solution generation level; the substation operation and maintenance knowledge graph contains multiple triples, each triple consisting of the corresponding substation operation and maintenance equipment's equipment identifier, equipment attribute parameters, and multiple operation and control information; The initial text is parsed using the intent recognition hierarchy of the substation operation and maintenance language model to generate user intent; Determine whether the user intent matches two or more operation control information; If so, all operation and control information matching the user intent will be used as the target control information; using the scheme generation hierarchy of the substation operation and maintenance language model, an execution strategy that satisfies the user intent will be generated based on the target control information. If not, the operation control information matching the user intent will be used as the execution strategy.

[0006] Optionally, the step of parsing the initial text and generating user intent using the intent recognition hierarchy of the substation operation and maintenance language model includes: Using the intent recognition hierarchy, entity recognition and context awareness are performed on the initial text to generate user intents that include target device entities, target event entities, and / or information request types.

[0007] Optionally, the scheme generation hierarchy utilizing the substation operation and maintenance language model, based on various target control information, generates an execution strategy that satisfies the user's intent, including: The scheme generation hierarchy of the substation operation and maintenance language model is used to prioritize the execution of various target control information, resulting in a ranking. The target control information is then integrated and processed according to the ranking to form an initial execution strategy. The initial execution strategy is then validated based on the equipment attribute parameters matched to each target control information in the substation operation and maintenance knowledge graph to determine its feasibility. If feasible, the initial execution strategy is adopted as the final execution strategy. If not feasible, the initial execution strategy is updated based on the equipment attribute parameters matched to each target control information, and the process returns to the step of validating the initial execution strategy based on the equipment attribute parameters matched to each target control information in the substation operation and maintenance knowledge graph, until the initial execution strategy is deemed feasible.

[0008] Optionally, obtain the substation operation and maintenance language model, including: Construct an LLaMA large language model that includes an initial intent recognition layer and an initial scheme generation layer; Obtain substation operation and maintenance standards and procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans; Based on the aforementioned substation operation and maintenance standards and procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans, the LLaMA large language model is trained using LoRA fine-tuning technology. The LLaMA large language model obtained after training is the substation operation and maintenance language model. The initial intent recognition level obtained after training is the intent recognition level of the substation operation and maintenance language model, and the initial solution generation level obtained after training is the solution generation level of the substation operation and maintenance language model.

[0009] Optionally, the step of training the LLaMA large language model based on the substation operation and maintenance standard procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans, using LoRA fine-tuning technology, includes: The substation operation and maintenance standard procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans are processed to form multiple types of training samples. One type of training sample is extracted from the substation operation and maintenance standard procedures, another type is extracted from the substation operation and maintenance defect data, yet another type is extracted from the substation operation and maintenance fault handling plans, and a third type is extracted from both the substation operation and maintenance defect data and the substation operation and maintenance fault handling plans. The LLaMA large language model was fine-tuned using LoRA training with various training samples.

[0010] Optionally, the data processing of the substation operation and maintenance standard procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans forms multiple types of training samples, including: The standard processing procedure for each substation operation and maintenance equipment is extracted from the substation operation and maintenance standard procedures, and the standard processing procedure for each substation operation and maintenance equipment is integrated to form one type of training sample; Event extraction is performed on the substation operation and maintenance defect data to extract the defect performance characteristics, defect types and defect levels under different typical faults of each substation operation and maintenance equipment; based on the defect performance characteristics, defect types and defect levels under each typical fault, another type of training sample is constructed. Key information is extracted from the substation operation and maintenance fault handling plan to generate handling schemes under different typical faults, and the handling schemes under each typical fault are integrated to form another type of training sample. The defect characteristics and solutions of at least two typical faults are integrated to form another type of training sample.

[0011] Optionally, obtain the substation operation and maintenance knowledge graph, including: Obtain the equipment identification, rated parameters, normal operating parameter range, safety constraints, equipment inspection procedures, equipment fault characteristics, equipment fault handling plans and / or equipment operation procedures for different substation operation and maintenance equipment; The equipment identification, rated parameters, normal operating parameter range, safety constraints, equipment inspection procedures, equipment fault characteristics, equipment fault handling plan and / or equipment operation process of the same substation operation and maintenance equipment are processed to form a corresponding triplet. The various ternary sets are integrated to form a substation operation and maintenance knowledge graph.

[0012] A voice interaction device, comprising: The conversion module is used to convert user speech into initial text; The acquisition module is used to acquire the substation operation and maintenance language model and the substation operation and maintenance knowledge graph; wherein, the substation operation and maintenance language model is obtained through training and includes an intent recognition level and a solution generation level; the substation operation and maintenance knowledge graph contains multiple triples, each triple consisting of the corresponding substation operation and maintenance equipment's equipment identifier, equipment attribute parameters, and multiple operation and control information. The parsing module is used to parse the initial text and generate user intent by utilizing the intent recognition hierarchy of the substation operation and maintenance language model; The judgment module is used to determine whether the user intent matches two or more operation control information; if yes, the generation module is called; if no, the determination module is called. The generation module is used to take all the operation and control information that match the user intent as the target control information; and to generate an execution strategy that meets the user intent based on the target control information using the scheme generation hierarchy of the substation operation and maintenance language model. The determination module is used to use the operation control information that matches the user intent as the execution strategy.

[0013] A voice interaction device, including a memory and a processor; The memory is used to store programs; The processor is used to execute the program to implement the various steps of the above-described voice interaction method.

[0014] A readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the various steps of the above-described voice interaction method.

[0015] As can be seen from the above technical solutions, the voice interaction method provided in this application can convert user voice into initial text. Based on this, this application can be triggered by voice, which is simple and adaptable to the needs of different operation and maintenance scenarios, improving the applicability of the solution. This application can obtain a substation operation and maintenance language model and a substation operation and maintenance knowledge graph. Using the intent recognition level of the substation operation and maintenance language model, the initial text is parsed to generate user intent. Since the substation operation and maintenance language model is trained and includes an intent recognition level and a solution generation level, this application can use a language model adapted to the substation operation and maintenance field to convert the initial text into user requirements, remove redundant information from the initial text, and clarify the user's operation and maintenance needs. This application determines whether the user intent matches two or more operation and control information in the substation operation and maintenance knowledge graph; wherein, the substation operation and maintenance knowledge graph includes... Multiple triples are used, each consisting of the corresponding substation maintenance equipment's equipment identifier, equipment attribute parameters, and multiple operation and control information. Based on this, this application can use a substation maintenance knowledge graph with structured storage of substation maintenance information to differentiate the complexity of user needs and determine whether directly using the existing operation and control information is sufficient to meet user needs. If it cannot meet the needs, all operation and control information matching the user intent is used as the target control information. Using the scheme generation hierarchy of the substation maintenance language model, an execution strategy that meets the user intent is generated based on the target control information. Thus, for complex maintenance needs such as multiple fault superposition and multi-device collaborative management, this application can integrate all related operation and control information to generate executable and implementable execution strategies, avoiding the tedious operation of maintenance personnel manually integrating multiple types of information, reducing human omissions and decision-making biases, and improving the efficiency and accuracy of maintenance decisions in complex scenarios. If a single operation and control information can meet the user's needs, this application directly uses the operation and control information that matches the user's intent as the execution strategy, which simplifies the process, shortens the response time, avoids unnecessary strategy generation steps, improves query and response efficiency in simple scenarios, and meets the needs of operation and maintenance personnel to quickly obtain basic information.As can be seen, this application can provide matching execution solutions for the ambiguous voice-based needs of maintenance personnel, solving the pain point of inconvenience for maintenance personnel to manually input queries during on-site operations, improving the convenience of interaction, adapting to the high-efficiency needs of maintenance personnel in on-site operations, breaking the drawbacks of traditional consultation delays and inefficient document retrieval, and significantly shortening the time for maintenance personnel to acquire knowledge and execute strategies. At the same time, this application can adapt to single operation and control information needs as well as multiple operation and control information fusion needs, taking into account both simple and complex operation and maintenance scenarios, adapting to the characteristics of diverse equipment and complex faults in the substation operation and maintenance field, improving the accuracy of operation and maintenance decisions in different scenarios; lowering the operation threshold for maintenance personnel, eliminating the need for manual retrieval and integration of information, and enabling them to quickly obtain suitable execution strategies through voice interaction, alleviating the knowledge reserve pressure on maintenance personnel; after equipment updates or new faults, data maintenance can be achieved by directly modifying the substation operation and maintenance knowledge graph, ultimately improving the overall efficiency and quality of substation operation and maintenance work, and ensuring the safe and stable operation of substation equipment. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0017] Figure 1 This is a flowchart of a voice interaction method disclosed in an embodiment of this application; Figure 2 This is a structural block diagram of a voice interaction device disclosed in an embodiment of this application; Figure 3 This is a hardware structure block diagram of a voice interaction device disclosed in an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] This application provides a voice interaction method that can be applied to various substation operation and maintenance systems or substation operation and maintenance equipment repair systems, as well as various computer terminals or smart terminals. The executing entity can be the processor or server of the computer terminal or smart terminal.

[0020] Next, combine Figure 1 The voice interaction method of this application is described in detail, including the following steps: Step S1: Convert the user's voice into initial text.

[0021] Specifically, speech recognition technology can be used to convert the voice information input by maintenance personnel into text form, which can then be used as the initial text for subsequent processing.

[0022] Step S2: Obtain the substation operation and maintenance language model and substation operation and maintenance knowledge graph.

[0023] Specifically, the substation operation and maintenance language model is a finely tuned and trained LLaMA large language model, which may include an intent recognition level and a scheme generation level.

[0024] The substation operation and maintenance knowledge graph is used to store information in the field of substation operation and maintenance in a structured manner. It contains multiple triples, each of which consists of the equipment identifier, equipment attribute parameters, and multiple operation and control information of the corresponding substation operation and maintenance equipment.

[0025] For example, equipment identification could be 220kV main transformer, 101 switch, etc.

[0026] Equipment attribute parameters can include equipment rated parameters, normal operating parameter range, and safety constraints.

[0027] The rated parameters of the equipment can include rated capacity, rated voltage, etc. Normal operating parameter ranges can include oil level standard ranges, etc.

[0028] Operation and control information can include equipment topology relationships, equipment inspection procedures, equipment fault characteristics, equipment fault handling solutions, and equipment operation procedures.

[0029] Step S3: Using the intent recognition level of the substation operation and maintenance language model, parse the initial text to generate user intent.

[0030] Specifically, entity recognition technology in the intent recognition layer can be used to extract all entities from the initial text, and combined with context-aware technology, the relationships between the various entities in the initial text can be understood, thereby generating user intent.

[0031] Step S4: Determine whether the user intent matches two or more operation control information; if yes, proceed to step S5; if no, proceed to step S6.

[0032] Specifically, it can be determined whether the user's intent requires the integration of multiple operational control information. If yes, then step S5 is executed; otherwise, step S6 is executed.

[0033] Specifically, this can be achieved by analyzing keywords in user intent and matching them with triples in the substation operation and maintenance knowledge graph.

[0034] If the keywords in the user intent are associated with operation and control information in multiple triples, it is determined that multiple operation and control information needs to be integrated. For example, if the user intent simultaneously mentions the defect manifestation characteristics corresponding to multiple typical faults or multiple keywords correspond to different triples in the power supply operation and maintenance knowledge graph, it indicates that multiple operation and control information needs to be integrated.

[0035] Step S5: Take all the operation and control information that match the user intent as each target control information; use the scheme generation hierarchy of the substation operation and maintenance language model to generate an execution strategy that satisfies the user intent based on each target control information.

[0036] Specifically, the scheme generation hierarchy can be used to integrate and process the control information of various objectives to initially form an initial execution strategy; The initial execution strategy is validated based on the equipment attribute parameters matched with the target control information in the substation operation and maintenance knowledge graph.

[0037] By comparing the device attribute parameters, it is determined whether the initial execution strategy is compliant and feasible. If so, the initial execution strategy can be directly used as the execution strategy. If not, the initial execution strategy needs to be adjusted to ensure compliance.

[0038] Once the execution is complete, if the feedback indicates that the execution strategy is effective and error-free, the knowledge graph can be updated according to the user's intent and the corresponding execution strategy to achieve continuous evolution and optimization of the knowledge system, thereby continuously improving the system's intelligence and adaptability.

[0039] Step S6: Use the operation control information that matches the user intent as the execution strategy.

[0040] Specifically, a single operational control information matching the user's intent can be directly used as the execution strategy.

[0041] As can be seen from the above technical solutions, the voice interaction method provided in this application can convert user voice into initial text. Based on this, this application can be triggered by voice, which is simple and adaptable to the needs of different operation and maintenance scenarios, improving the applicability of the solution. This application can obtain a substation operation and maintenance language model and a substation operation and maintenance knowledge graph. Using the intent recognition level of the substation operation and maintenance language model, the initial text can be parsed to generate user intent. Since the substation operation and maintenance language model is trained and includes an intent recognition level and a solution generation level, this application can use a language model adapted to the substation operation and maintenance field to convert the initial text into user requirements, remove redundant information from the initial text, and clarify the user's operation and maintenance needs. This application determines whether the user intent matches multiple operation and control information in the substation operation and maintenance knowledge graph; wherein, the substation operation and maintenance knowledge graph contains multiple... Each triple consists of a corresponding substation maintenance equipment identifier, equipment attribute parameters, and multiple operation and control information. Based on this, this application can use a substation maintenance knowledge graph with structured storage of substation maintenance information to differentiate the complexity of user needs and determine whether directly using the existing operation and control information is sufficient to meet user needs. If it cannot meet the needs, all operation and control information matching the user intent will be used as the target control information. Using the scheme generation hierarchy of the substation maintenance language model, an execution strategy that meets the user intent will be generated based on the target control information. Thus, for complex maintenance needs such as multiple fault superposition and multi-device collaborative management, this application can integrate all related operation and control information to generate executable and implementable execution strategies, avoiding the tedious operation of manually integrating multiple types of information by maintenance personnel, reducing human omissions and decision-making biases, and improving the efficiency and accuracy of maintenance decisions in complex scenarios. If a single operation and control information can meet the user's needs, this application directly uses the operation and control information that matches the user's intent as the execution strategy, which simplifies the process, shortens the response time, avoids unnecessary strategy generation steps, improves query and response efficiency in simple scenarios, and meets the needs of operation and maintenance personnel to quickly obtain basic information.As can be seen, this application can provide matching execution solutions for the ambiguous voice-based needs of maintenance personnel, solving the pain point of inconvenience for maintenance personnel to manually input queries during on-site operations, improving the convenience of interaction, adapting to the high-efficiency needs of maintenance personnel in on-site operations, breaking the drawbacks of traditional consultation delays and inefficient document retrieval, and significantly shortening the time for maintenance personnel to acquire knowledge and execute strategies. At the same time, this application can adapt to both single operation and control information needs and multiple operation and control information fusion needs, taking into account both simple and complex operation and maintenance scenarios, adapting to the characteristics of diverse equipment and complex faults in the substation operation and maintenance field, improving the accuracy of operation and maintenance decisions in different scenarios; lowering the operational threshold for maintenance personnel, eliminating the need for manual retrieval and integration of information, and enabling them to quickly obtain suitable execution strategies through voice interaction, alleviating the knowledge reserve pressure on maintenance personnel, and ultimately improving the overall efficiency and quality of substation operation and maintenance work, ensuring the safe and stable operation of substation equipment.

[0042] Furthermore, this application can receive user voice data collected from various wearable terminals such as smart helmets, smart glasses, and wristband devices. The collected user voice data is transmitted to the processing unit via a dedicated encrypted channel, effectively ensuring the security and privacy of data transmission and preventing information leakage.

[0043] This application can also be executed via a wearable terminal.

[0044] When the wearable terminal is in a low-power state, this application can automatically activate a resource-aware intelligent scheduling strategy to ensure the normal operation of core functions. At this time, this application can reduce the frequency of calling the substation operation and maintenance language model. For frequently occurring standard questions such as common equipment parameter queries and basic operation procedure inquiries, the substation operation and maintenance knowledge graph is prioritized for rapid response, ensuring that users' routine needs are met in a timely manner. For non-urgent requests, they are temporarily cached and processed after the device's power is restored.

[0045] This low-power adaptive mechanism maximizes power saving and extends the usage time of wearable terminals while ensuring that the core services of this application are not interrupted, thus meeting the needs of maintenance personnel for long-term operations.

[0046] It can convert execution strategies into speech and broadcast them, adapting to a variety of flexible output methods. It can quickly play standard templates cached locally to meet the need for rapid response to simple commands, or synthesize natural, fluent, and emotional speech through cloud-based emotional TTS text-to-speech technology to enhance the interactive experience.

[0047] Voice output can be achieved using bone conduction or directional speakers, which not only ensures that maintenance personnel can clearly hear the response in noisy substation maintenance sites, but also effectively protects the privacy of voice information and prevents it from being accessed by unauthorized personnel.

[0048] This application is adaptable to the resource constraints of wearable terminals, fully meets the strict safety regulations such as "two tickets and three systems" in power field, and provides comprehensive and high-quality intelligent support for substation operation and maintenance.

[0049] As can be seen from the above technical solutions, the voice interaction method provided in this application can convert user voice into initial text. Based on this, this application can be triggered by voice, which is simple and adaptable to the needs of different operation and maintenance scenarios, improving the applicability of the solution. This application can obtain a substation operation and maintenance language model and a substation operation and maintenance knowledge graph. Using the intent recognition level of the substation operation and maintenance language model, the initial text can be parsed to generate user intent. Since the substation operation and maintenance language model is trained and includes an intent recognition level and a solution generation level, this application can use a language model adapted to the substation operation and maintenance field to convert the initial text into user requirements, remove redundant information from the initial text, and clarify the user's operation and maintenance needs. This application determines whether the user intent matches multiple operation and control information in the substation operation and maintenance knowledge graph; wherein, the substation operation and maintenance knowledge graph contains multiple... Each triple consists of a corresponding substation maintenance equipment identifier, equipment attribute parameters, and multiple operation and control information. Based on this, this application can use a substation maintenance knowledge graph with structured storage of substation maintenance information to differentiate the complexity of user needs and determine whether directly using the existing operation and control information is sufficient to meet user needs. If it cannot meet the needs, all operation and control information matching the user intent will be used as the target control information. Using the scheme generation hierarchy of the substation maintenance language model, an execution strategy that meets the user intent will be generated based on the target control information. Thus, for complex maintenance needs such as multiple fault superposition and multi-device collaborative management, this application can integrate all related operation and control information to generate executable and implementable execution strategies, avoiding the tedious operation of manually integrating multiple types of information by maintenance personnel, reducing human omissions and decision-making biases, and improving the efficiency and accuracy of maintenance decisions in complex scenarios. If a single operation and control information can meet the user's needs, this application directly uses the operation and control information that matches the user's intent as the execution strategy, which simplifies the process, shortens the response time, avoids unnecessary strategy generation steps, improves query and response efficiency in simple scenarios, and meets the needs of operation and maintenance personnel to quickly obtain basic information.As can be seen, this application can provide matching execution solutions for the ambiguous voice-based needs of maintenance personnel, solving the pain point of inconvenience for maintenance personnel to manually input queries during on-site operations, improving the convenience of interaction, adapting to the high-efficiency needs of maintenance personnel in on-site operations, breaking the drawbacks of traditional consultation delays and inefficient document retrieval, and significantly shortening the time for maintenance personnel to acquire knowledge and execute strategies. At the same time, this application can adapt to both single operation and control information needs and multiple operation and control information fusion needs, taking into account both simple and complex operation and maintenance scenarios, adapting to the characteristics of diverse equipment and complex faults in the substation operation and maintenance field, improving the accuracy of operation and maintenance decisions in different scenarios; lowering the operational threshold for maintenance personnel, eliminating the need for manual retrieval and integration of information, and enabling them to quickly obtain suitable execution strategies through voice interaction, alleviating the knowledge reserve pressure on maintenance personnel, and ultimately improving the overall efficiency and quality of substation operation and maintenance work, ensuring the safe and stable operation of substation equipment.

[0050] The following will provide a detailed explanation of this application through specific examples.

[0051] When the initial text obtained from the conversion is "The oil level of switch 101 is abnormally low, accompanied by slight leakage. How to handle it?", the user's intent can be determined through the substation operation and maintenance language model as "to resolve the abnormally low oil level and slight leakage of switch 101". By combining the substation operation and maintenance knowledge graph, it is possible to match the operation and control information corresponding to "low oil level" and "minor leakage". "Abnormally low oil level and minor leakage of switch 101" is a composite scenario. The solution generation hierarchy can be utilized by the substation operation and maintenance language model. It integrates the equipment parameters of switch 101, the basic handling rules for abnormal oil levels, the relevant safety constraints for leakage faults in the substation operation and maintenance knowledge graph, and the extensive experience in handling complex faults learned through training. The generated execution strategy is as follows: "Immediately stop the relevant operation and report to the duty officer", "Request equipment power outage for maintenance", "Set up a warning zone in the fault area to prevent unauthorized personnel from entering", "Arrange professional personnel with tools to rush to the site to investigate and repair the leakage point", and "Recheck the oil level after repair to ensure that it meets the standard before restoring equipment operation". It can automatically record detailed information about the interaction process; After industry experts reviewed and confirmed that the execution strategy was correct, it was injected into the substation operation and maintenance knowledge graph as a new knowledge entry, enriching the coverage of complex fault handling scenarios.

[0052] In some embodiments of this application, the process of obtaining the substation operation and maintenance knowledge graph in step S2 is described in detail, and the steps are as follows: S20. Obtain the equipment identification, rated parameters, normal operating parameter range, safety constraints, equipment inspection procedures, equipment fault characteristics, equipment fault handling plans and / or equipment operation procedures for different substation operation and maintenance equipment.

[0053] Specifically, it can obtain the equipment identification, rated parameters, normal operating parameter range, safety constraints, equipment inspection procedures, equipment fault characteristics, equipment fault handling plans and / or equipment operation procedures for each substation operation and maintenance equipment.

[0054] The same substation maintenance equipment can correspond to one or more equipment identifiers.

[0055] The equipment identification of substation operation and maintenance equipment can be used to identify its corresponding equipment model and version, as well as its corresponding deployment area.

[0056] S21. Process the equipment identification, rated parameters, normal operating parameter range, safety constraints, equipment inspection procedures, equipment fault characteristics, equipment fault handling plan and / or equipment operation process of the same substation operation and maintenance equipment to form a corresponding triplet.

[0057] Specifically, the rated parameters, normal operating parameter range, and safety constraints of the same substation operation and maintenance equipment can be used as the equipment attribute parameters of that substation operation and maintenance equipment; The equipment inspection procedures, equipment fault characteristics, equipment fault handling plans and equipment operation procedures of the same substation maintenance equipment can be used as the operation and control information of that substation maintenance equipment. The equipment identifier, equipment attribute parameters, and operation and control information of the same substation operation and maintenance equipment are combined to form the corresponding triplet for that substation operation and maintenance equipment.

[0058] S22. Integrate the various ternary sets to form a substation operation and maintenance knowledge graph.

[0059] Specifically, based on the topological relationships of various substation operation and maintenance equipment, the various triples can be integrated to form a substation operation and maintenance knowledge graph.

[0060] As can be seen from the above technical solution, this embodiment provides an optional method for obtaining a substation operation and maintenance knowledge graph. Through this method, a substation operation and maintenance knowledge graph can be constructed efficiently and accurately, providing a solid data foundation for subsequent voice interaction.

[0061] In some embodiments of this application, the process of obtaining the substation operation and maintenance language model in step S2 is described in detail, and the steps are as follows: S20. Construct an LLaMA large language model that includes an initial intent recognition level and an initial scheme generation level.

[0062] Specifically, a basic framework for building the LLaMA large language model can be constructed, which has powerful language processing capabilities and scalability.

[0063] The basic framework of the LLaMA large language model can include core components such as the input layer, hidden layer, and output layer.

[0064] The input layer can be used to receive initial text data, the hidden layer can perform deep processing and feature extraction on the input data through a complex neural network structure, and the output layer can generate results related to user intent and execution strategies that satisfy user intent.

[0065] An initial intent recognition layer and an initial scheme generation layer can be built on the basic framework respectively.

[0066] Specifically, in the hidden layer of the basic framework, an initial intent recognition layer can be constructed by introducing an attention mechanism and a neural network module to capture key information in the initial text and identify the core intent expressed by the user.

[0067] Key information may include querying equipment parameters, requesting troubleshooting solutions, or understanding operating procedures.

[0068] The initial scheme generation hierarchy is also designed in the hidden layer. Deep learning algorithms and a large amount of substation operation and maintenance data can be used to train the LLaMA large language model, enabling this application to achieve accurate parsing, safe reasoning, and intelligent generation of substation operation and maintenance instructions, supporting various operation and maintenance tasks such as consultation and Q&A, operation guidance, and fault diagnosis.

[0069] S21. Obtain substation operation and maintenance standards and procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans.

[0070] Specifically, substation operation and maintenance standards and procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans can be obtained through multi-source data acquisition methods.

[0071] These data come from a wide range of sources, including but not limited to national and enterprise-level substation operation and maintenance regulations, core industry standards such as the "General Management Regulations for Substation Operation and Maintenance of State Grid", a typical equipment failure case library containing descriptions and handling records of 2,000+ real events, providing rich practical experience references for the model; standardized operating procedures (SOPs), defect reporting process templates, historical fault handling methods, historical operation and maintenance records, and technical manuals provided by equipment manufacturers.

[0072] Substation operation and maintenance standards and procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans contain a large number of professional terms and expressions in the field of substation operation and maintenance.

[0073] The collected data needs to undergo preprocessing steps such as cleaning, deduplication, and format standardization to ensure data quality and consistency, resulting in substation operation and maintenance standard procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans.

[0074] S22. Based on the substation operation and maintenance standard procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plan, LoRA fine-tuning technology is used to train the LLaMA large language model. The LLaMA large language model obtained after training is the substation operation and maintenance language model. The initial intent recognition level obtained after training is the intent recognition level of the substation operation and maintenance language model. The initial scheme generation level obtained after training is the scheme generation level of the substation operation and maintenance language model.

[0075] Specifically, a training dataset can be constructed based on the preprocessed substation operation and maintenance standard procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans. LoRA fine-tuning technology was used to perform targeted training and parameter fine-tuning on the initial intent recognition layer and the initial scheme generation layer of the LLaMA large language model.

[0076] LoRA fine-tuning technology introduces a low-rank adaptation matrix into the original parameters of the LLaMA large language model, which can quickly adapt the LLaMA large language model into a model specifically for substation operation and maintenance without significantly increasing the amount of computation.

[0077] During training, the LLaMA language model can gradually learn the professional terminology system, fault feature association rules, and standard handling procedures in the substation operation and maintenance field. For example, it can accurately identify equipment status descriptions such as "abnormal oil level" and "leakage" and associate them with standardized operations such as "apply for power outage maintenance" and "set up a warning zone." Through multiple rounds of iterative optimization, the final substation operation and maintenance language model has reached industry application standards in terms of intent recognition accuracy and solution generation compliance. The intent recognition level can achieve an accuracy rate of over 95% in understanding complex scenarios, and the execution strategies generated at the solution generation level have a compliance rate of over 98% with industry procedures, effectively supporting the intelligent implementation of subsequent voice interaction processes.

[0078] After training, the final LLaMA language model can be used as the substation operation and maintenance language model, the final initial intent recognition level can be used as the intent recognition level of the substation operation and maintenance language model, and the final initial scheme generation level can be used as the scheme generation level of the substation operation and maintenance language model.

[0079] As can be seen from the above technical solution, this embodiment provides an optional method for obtaining a substation operation and maintenance language model. Through this method, specialized optimization training can be performed on professional terminology in the substation operation and maintenance field, enabling accurate recognition of various professional expressions, ensuring the accuracy of speech-to-text conversion, and avoiding operational errors caused by terminology recognition mistakes. This gives the substation operation and maintenance language model a strong ability to understand professional terminology, accurately grasping the complex professional vocabulary and expressions in the substation operation and maintenance field; excellent multi-turn dialogue context modeling capabilities, clearly outlining dialogue logic and accurately understanding user intent; rigorous operational compliance judgment capabilities, enabling compliance verification of various operational instructions according to industry standards; and excellent causal reasoning capabilities, deriving reasonable conclusions based on known information, ultimately generating a natural language response that fully complies with power safety regulations.

[0080] In some embodiments of this application, the process of training the LLaMA large language model using LoRA fine-tuning technology based on the substation operation and maintenance standard procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plan is described in detail. The steps are as follows: S220. Perform data processing on the substation operation and maintenance standard procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans to form multiple types of training samples; one type of training sample is extracted from the substation operation and maintenance standard procedures, another type of training sample is extracted from the substation operation and maintenance defect data, yet another type of training sample is extracted from the substation operation and maintenance fault handling plans, and yet another type of training sample is extracted from both the substation operation and maintenance defect data and the substation operation and maintenance fault handling plans.

[0081] Specifically, key data can be extracted from substation operation and maintenance standards and procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans to form three types of training samples. Key data extraction and fusion are performed on substation operation and maintenance defect data and substation operation and maintenance fault handling plans to form another type of training sample.

[0082] S221. Using various training samples, perform LoRA fine-tuning training on the LLaMA large language model.

[0083] Specifically, the resulting training samples can be divided into training set, validation set and test set according to a certain ratio.

[0084] The training set is used for model parameter updates, the validation set is used for hyperparameter tuning, and the test set is used for final performance evaluation.

[0085] During the LoRA fine-tuning training process, the basic parameters in the LLaMA large language model are first initialized, and a low-rank adaptation matrix is ​​introduced to expand the parameters of the initial intent recognition layer and the initial scheme generation layer.

[0086] During training, gradient descent algorithm can be used, and cross-entropy loss function can be used to measure the difference between model output and true label. Model parameters can be updated through backpropagation.

[0087] To ensure the adaptability of the final trained substation operation and maintenance language model to the substation operation and maintenance field, the cross-entropy loss function includes terminology consistency constraints and operational compliance regularization terms, ensuring that the generated execution strategy conforms to both language logic and power safety regulations.

[0088] After each training round, the model performance can be evaluated on the validation set. Training can be terminated early if the validation loss does not decrease for three consecutive rounds to prevent overfitting.

[0089] Finally, the model's generalization ability was verified through a test set to ensure that it met industry application standards in core indicators such as intent recognition accuracy and solution generation compliance.

[0090] As can be seen from the above technical solution, this embodiment provides an optional method for training the LLaMA large language model based on the aforementioned substation operation and maintenance standard procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans, using LoRA fine-tuning technology. This method fully utilizes multi-source heterogeneous data in the substation operation and maintenance field to construct a comprehensive and clearly structured training sample set, providing high-quality data support for model training. LoRA fine-tuning technology, by introducing a low-rank adaptation matrix, achieves precise embedding of substation operation and maintenance expertise while maintaining the original model's powerful language processing capabilities. This lightweight parameter adjustment method significantly reduces computational resource consumption, enabling the model to quickly adapt to the specific needs of the substation operation and maintenance field, demonstrating superior performance in intent recognition and solution generation.

[0091] In some embodiments of this application, the process of processing the substation operation and maintenance standard procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans to form multiple types of training samples is described in detail, and the steps are as follows: S2200. Extract the standard processing procedure for each substation operation and maintenance equipment from the substation operation and maintenance standard procedures, and integrate the standard processing procedure for each substation operation and maintenance equipment to form one type of training sample.

[0092] Specifically, when processing data from substation operation and maintenance standards and procedures, the standard processing procedures related to equipment operation specifications, safety standards, and inspection requirements for different substation operation and maintenance equipment can be extracted.

[0093] One type of standardized training sample can be formed according to the structure of "operation object-operation type-specification requirements".

[0094] For example, one of the training samples in this class could be "Transformer - Oil level check - Values ​​need to be read and deviations recorded when the ambient temperature is stable".

[0095] S2201. Extract events from the substation operation and maintenance defect data, and extract the defect performance characteristics, defect types and defect levels of each substation operation and maintenance equipment under different typical faults; construct another type of training sample based on the defect performance characteristics, defect types and defect levels of each typical fault.

[0096] Specifically, when processing substation operation and maintenance defect data, the description of real events can be structured and analyzed to extract core fields such as defect performance characteristics, defect type and defect level under different typical faults of each substation operation and maintenance equipment. Another type of standardized training sample can be formed according to the structure of "substation operation and maintenance equipment - defect performance characteristics - defect type - defect level".

[0097] For example, one of the training samples in this category could be "110kV circuit breaker - SF6 low pressure alarm - defect type is high voltage equipment defect - defect level is severe".

[0098] S2202. Extract key information from the substation operation and maintenance fault handling plan, generate handling schemes for different typical faults, and integrate the handling schemes for each typical fault to form another type of training sample.

[0099] Specifically, for the substation operation and maintenance fault handling plan, the handling priority and operation steps under different typical faults of each substation operation and maintenance equipment can be extracted, and the multi-step operation can be broken down into a hierarchical training sample of "defect manifestation characteristics - associated equipment - handling priority - operation steps".

[0100] For example, one of the training samples in this category could be "110kV circuit breaker - SF6 low pressure alarm - associated equipment is protection device and monitoring system - priority of handling is emergency - operation steps are to immediately check the SF6 pressure gauge reading and record it, turn on the ventilation equipment, isolate the faulty circuit breaker, and contact maintenance personnel to replenish gas".

[0101] S2203. Integrate the defect characteristics and handling solutions of at least two typical faults to form another type of training sample.

[0102] Specifically, the defect manifestation characteristics and handling solutions of different typical faults can be integrated to form a complete logical chain sample of "defect manifestation characteristics - associated equipment - handling priority - operation steps", making this application adaptable to complex scenarios where multiple typical faults coexist.

[0103] Furthermore, all samples must be reviewed by industry experts to ensure that the terminology used complies with national standards such as the "Guidelines for Supervision of Electric Power Technology," ultimately forming a million-level structured training dataset covering scenarios such as single-device operation, multi-device collaboration, and complex fault handling.

[0104] As can be seen from the above technical solution, this embodiment provides an optional method for processing the substation operation and maintenance standard procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans to form multiple types of training samples. Through this method, professional knowledge in the substation operation and maintenance field can be transformed into structured training samples, providing standardized and high-quality learning materials for the model. By extracting core elements such as equipment standard procedures, defect features, and fault handling plans at different levels, a complete training system covering single-equipment operation to multi-fault collaborative handling is constructed, further improving the applicability of the substation operation and maintenance language model of this application.

[0105] In some embodiments of this application, the process of parsing the initial text and generating user intent using the intent recognition level of the substation operation and maintenance language model in step S3 is described in detail as follows: S30. Using the intent recognition level, perform entity recognition and context awareness on the initial text to generate user intents that include target device entities, target event entities, and / or information demand types.

[0106] Specifically, after receiving the initial text, named entity recognition technology at the intent recognition level can be used to locate all target fields contained in the initial text; The target field can be any one or more of the target device entity and the target event entity.

[0107] The target equipment entity can be a specific equipment name such as "transformer" or "circuit breaker"; The target event entity can be a fault description or operation request such as "abnormal oil level" or "low SF6 pressure alarm".

[0108] The context-aware module at the intent recognition level can be used to analyze the related information in the initial text by combining all target fields to determine the type of information needs of the user.

[0109] Information requests can include "querying equipment parameters", "requesting troubleshooting solutions", or "understanding operating procedures".

[0110] For example, for the initial text "110kV circuit breaker SF6 low pressure alarm, please provide processing steps", the target equipment entity can be identified as "110kV circuit breaker", the target event entity is "SF6 low pressure alarm", and the information request type is determined to be "request for fault handling solution".

[0111] As can be seen from the above technical solution, this embodiment provides an optional method for parsing the initial text and generating user intent using the intent recognition level of the substation operation and maintenance language model. This method accurately captures key equipment and event information in the user's voice input, and accurately determines the type of user need by combining contextual logic, providing a reliable basis for generating targeted responses. Because the intent recognition level of the substation operation and maintenance language model, through the substation operation and maintenance domain knowledge accumulated during the pre-training phase, can effectively distinguish similar terms, avoiding intent recognition deviations caused by differences in expression. When dealing with complex sentences, the intent recognition level of the substation operation and maintenance language model can trace the previous dialogue history through context awareness, and eliminate ambiguity by combining background information such as equipment operating status and historical fault records. For example, when a user continuously asks "What should be done after the transformer oil level is abnormal?" and "Which departments should be contacted?", the model can identify that the second question is an extension of the previous fault handling request, rather than an independent query request. By introducing an attention mechanism to strengthen the weight of key fields, this level can still maintain an accuracy rate of over 92% in long text parsing, significantly outperforming the intent recognition effect of general language models.

[0112] In some embodiments of this application, the process of generating an execution strategy that satisfies the user intent based on various target management and control information using the scheme generation hierarchy of the substation operation and maintenance language model in step S5 is described in detail, and the steps are as follows: S50. Utilizing the scheme generation hierarchy of the substation operation and maintenance language model, prioritize the execution of each target control information to obtain a ranking result. Integrate and process the target control information according to the ranking result to form an initial execution strategy. Verify the initial execution strategy based on the equipment attribute parameters matched by each target control information in the substation operation and maintenance knowledge graph to determine its feasibility. If feasible, adopt the initial execution strategy as the final execution strategy. If not feasible, update the initial execution strategy based on the equipment attribute parameters matched by each target control information, and return to the step of verifying the initial execution strategy based on the equipment attribute parameters matched by each target control information in the substation operation and maintenance knowledge graph until the initial execution strategy is feasible.

[0113] Specifically, considering that different target control information has varying degrees of urgency in substation operation and maintenance scenarios, and that there are certain operational correlations between different target control information, the priority evaluation module in the scheme generation layer of the substation operation and maintenance language model can be used to sort the execution priorities of each target control information, combined with the equipment relationships, fault propagation paths, and operational safety specifications in the substation operation and maintenance knowledge graph, to obtain a ranking result representing the execution priority of each target control information.

[0114] The step fusion module in the scheme generation hierarchy of the substation operation and maintenance language model can be used to fuse the control information of each target according to the sorting result, delete redundant steps, and form an initial execution strategy.

[0115] The Monte Carlo simulation algorithm in the scheme generation level of the substation operation and maintenance language model can be used to perform timing conflict detection on each processing operation in the initial execution strategy. Combined with the compliance check engine, the initial execution strategy is matched and verified with the security constraint rules in the substation operation and maintenance knowledge graph to determine whether the initial execution strategy is feasible.

[0116] Safety constraints may include key parameters such as equipment state transition conditions and operation interval time limits.

[0117] If feasible, the initial execution strategy will be used as the final execution strategy. If it is not feasible, the constraint solver in the scheme generation hierarchy of the substation operation and maintenance language model can be used to update the initial execution strategy by adjusting the operation sequence, replacing associated equipment, or inserting intermediate state transition steps.

[0118] An iterative optimization mechanism can be used to update the initial execution strategy, and after each correction, the strategy is re-verified until all constraints are met.

[0119] The solution generation hierarchy can also include an emergency plan library. When the initial execution strategy is detected to involve high-risk operations, it can compare and analyze the handling plans of similar historical cases, evaluate the expected effects of different strategies through a deep reinforcement learning model, and select the optimal execution path.

[0120] All generated execution strategies can be accompanied by detailed operational guidelines, including referenced procedure clause numbers, real-time equipment status parameters, and risk assessment levels, providing traceable decision support for operation and maintenance personnel. Through this closed-loop optimization mechanism, the solution generation hierarchy can ensure that the final output execution strategy not only complies with power safety regulations but also possesses practical operability.

[0121] As can be seen from the above technical solution, this embodiment provides a solution generation hierarchy utilizing the substation operation and maintenance language model to generate optional execution strategies that meet the user's intent based on various target management and control information. Through this method, dynamic optimization and security verification of the execution strategy can be achieved by combining the substation operation and maintenance knowledge graph. The solution generation hierarchy prioritizes the urgency of multi-target management and control information through a priority evaluation module, ensuring that critical operations are executed first; the step fusion module generates a concise and efficient initial execution strategy by deleting redundant operations and merging related steps. A Monte Carlo simulation algorithm and a compliance check engine constitute a dual verification mechanism; the former detects operation timing conflicts, and the latter matches security constraint rules to ensure that the strategy meets key parameter requirements such as equipment state transition conditions and operation interval limits. When an infeasible operation is detected, the constraint solver iteratively optimizes by adjusting the operation order, replacing related equipment, or inserting intermediate state transition steps until all constraints are met. This closed-loop optimization mechanism enables the solution generation hierarchy to dynamically adapt to complex operation and maintenance scenarios, significantly improving fault handling efficiency while ensuring power safety.

[0122] Next, we will combine Figure 2 The voice interaction device provided in this application will be described in detail. The voice interaction device described below can be compared with the voice interaction method described above.

[0123] See Figure 2 It can be observed that voice interaction devices may include: Conversion module 10 is used to convert user speech into initial text; Specifically, the conversion module 10 may include a voice input unit, an encrypted communication unit, a voice recognition submodule optimized with substation terminology, and a voice output unit that supports local template playback or cloud-based emotional TTS synthesis.

[0124] The acquisition module 20 is used to acquire the substation operation and maintenance language model and the substation operation and maintenance knowledge graph; wherein, the substation operation and maintenance language model is obtained through training and includes an intent recognition level and a solution generation level; the substation operation and maintenance knowledge graph contains multiple triples, each triple consisting of the corresponding substation operation and maintenance equipment's equipment identifier, equipment attribute parameters, and multiple operation and control information. The parsing module 30 is used to parse the initial text and generate user intent by utilizing the intent recognition level of the substation operation and maintenance language model; The judgment module 40 is used to determine whether the user intent matches two or more operation control information; if yes, the generation module 50 is called; if no, the determination module 60 is called. The generation module 50 is used to take all the operation and control information that match the user intent as each target control information; and to generate an execution strategy that satisfies the user intent based on each target control information using the scheme generation hierarchy of the substation operation and maintenance language model. The determination module 60 is used to use the operation control information that matches the user intent as the execution strategy.

[0125] Specifically, to fully ensure the safety and compliance of substation operation and maintenance, the voice interaction device may include a safety verification module.

[0126] The security verification module is used to perform multimodal identity authentication mechanisms, combining voice wake-up word voiceprint recognition with various biometric technologies such as fingerprint verification or iris recognition.

[0127] Specifically, the security verification module can be combined with a multimodal identity authentication mechanism for identity authentication. After successful authentication, the conversion module and its subsequent modules can be called to effectively prevent unauthorized personnel from making mistakes or malicious operations.

[0128] In addition, the voice interaction device is also equipped with an operation log recording mechanism, which can be used to automatically record all interactive behaviors.

[0129] The recorded content may include key data such as the specific content of the instruction, the operator's unique ID, the operation timestamp, and centimeter-level positioning information.

[0130] This mechanism forms a complete, auditable, and traceable operational traceability chain, fully meeting the stringent requirements of "full-process traceability" in the power industry, and providing solid and powerful support for the safety supervision and responsibility identification of operation and maintenance work.

[0131] Furthermore, the parsing module 30 may include: The user intent generation unit is used to perform entity recognition and context awareness on the initial text using the intent recognition level, and generate user intents that include target device entities, target event entities and / or information demand types.

[0132] Furthermore, the generation module 50 may include: The target management and control information sorting unit is used to prioritize the execution of each target management and control information according to the scheme generation hierarchy of the substation operation and maintenance language model, obtain the sorting result, and integrate and process the target management and control information according to the sorting result to form an initial execution strategy. Based on the equipment attribute parameters matched to each target management and control information in the substation operation and maintenance knowledge graph, the initial execution strategy is verified to determine its feasibility. If feasible, the initial execution strategy is adopted as the final execution strategy; if not feasible, the initial execution strategy is updated based on the equipment attribute parameters matched to each target management and control information, and the process returns to the step of verifying the initial execution strategy based on the equipment attribute parameters matched to each target management and control information in the substation operation and maintenance knowledge graph, until the initial execution strategy is feasible.

[0133] Furthermore, the acquisition module 20 may include: The model building unit is used to build an LLaMA large language model that includes an initial intent recognition level and an initial scheme generation level. The substation operation and maintenance defect data acquisition unit is used to acquire substation operation and maintenance standard procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans. The model training unit is used to train the LLaMA large language model based on the substation operation and maintenance standard procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans, using LoRA fine-tuning technology. The LLaMA large language model obtained after training is the substation operation and maintenance language model, the initial intent recognition level obtained after training is the intent recognition level of the substation operation and maintenance language model, and the initial solution generation level obtained after training is the solution generation level of the substation operation and maintenance language model.

[0134] Furthermore, the model training unit may include: The training sample acquisition subunit is used to process the substation operation and maintenance standard procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans to form multiple types of training samples. One type of training sample is extracted from the substation operation and maintenance standard procedures, another type is extracted from the substation operation and maintenance defect data, yet another type is extracted from the substation operation and maintenance fault handling plans, and a third type is extracted from both the substation operation and maintenance defect data and the substation operation and maintenance fault handling plans. The training sample utilizes sub-units to perform LoRA fine-tuning training on the LLaMA large language model using various training samples.

[0135] Furthermore, the training sample acquisition subunit may include: The first training sample acquisition subunit is used to extract the standard processing procedure for each substation operation and maintenance equipment from the substation operation and maintenance standard procedures, and integrate the standard processing procedure for each substation operation and maintenance equipment to form one type of training sample. The second training sample acquisition subunit is used to extract events from the substation operation and maintenance defect data, extract the defect performance characteristics, defect types and defect levels under different typical faults of each substation operation and maintenance equipment; and construct another type of training sample based on the defect performance characteristics, defect types and defect levels under each typical fault. The third training sample acquisition subunit is used to extract key information from the substation operation and maintenance fault handling plan, generate handling schemes under different typical faults, and integrate the handling schemes under each typical fault to form another type of training sample. The fourth training sample acquisition subunit is used to integrate the defect characteristics and handling solutions of at least two typical faults to form another type of training sample.

[0136] Furthermore, the acquisition module 20 may also include: The equipment identification acquisition unit is used to acquire the equipment identification, rated parameters, normal operating parameter range, safety constraints, equipment inspection procedures, equipment fault characteristics, equipment fault handling plans and / or equipment operation procedures of different substation operation and maintenance equipment. The ternary unit is used to process the equipment identifier, equipment rated parameters, normal operating parameter range, safety constraints, equipment inspection procedures, equipment fault characteristics, equipment fault handling plan and / or equipment operation process of the same substation operation and maintenance equipment to form a corresponding ternary unit. The ternary integration unit is used to integrate various ternary sets to form a substation operation and maintenance knowledge graph.

[0137] The voice interaction device provided in this application embodiment can be applied to voice interaction devices, such as PC terminals, cloud platforms, servers, and server clusters. Optionally, Figure 3 The hardware structure block diagram of the voice interaction device is shown below. Figure 3 The hardware structure of a voice interaction device may include: at least one processor 1, at least one communication interface 2, at least one memory 3, and at least one communication bus 4. In this embodiment, the number of processor 1, communication interface 2, memory 3, and communication bus 4 is at least one, and processor 1, communication interface 2, and memory 3 communicate with each other through communication bus 4. Processor 1 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. Memory 3 may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage device; The memory stores a program, which the processor can call. The program is used for: Convert user speech into initial text; Obtain a substation operation and maintenance language model and a substation operation and maintenance knowledge graph; wherein, the substation operation and maintenance language model is obtained through training and includes an intent recognition level and a solution generation level; the substation operation and maintenance knowledge graph contains multiple triples, each triple consisting of the corresponding substation operation and maintenance equipment's equipment identifier, equipment attribute parameters, and multiple operation and control information; The initial text is parsed using the intent recognition hierarchy of the substation operation and maintenance language model to generate user intent; Determine whether the user intent matches two or more operation control information; If so, all operation and control information matching the user intent will be used as the target control information; using the scheme generation hierarchy of the substation operation and maintenance language model, an execution strategy that satisfies the user intent will be generated based on the target control information. If not, the operation control information matching the user intent will be used as the execution strategy.

[0138] Optionally, the refined and extended functions of the program can be referred to the above description.

[0139] This application embodiment also provides a readable storage medium that can store a program suitable for execution by a processor, the program being used for: Convert user speech into initial text; Obtain a substation operation and maintenance language model and a substation operation and maintenance knowledge graph; wherein, the substation operation and maintenance language model is obtained through training and includes an intent recognition level and a solution generation level; the substation operation and maintenance knowledge graph contains multiple triples, each triple consisting of the corresponding substation operation and maintenance equipment's equipment identifier, equipment attribute parameters, and multiple operation and control information; The initial text is parsed using the intent recognition hierarchy of the substation operation and maintenance language model to generate user intent; Determine whether the user intent matches two or more operation control information; If so, all operation and control information matching the user intent will be used as the target control information; using the scheme generation hierarchy of the substation operation and maintenance language model, an execution strategy that satisfies the user intent will be generated based on the target control information. If not, the operation control information matching the user intent will be used as the execution strategy.

[0140] Optionally, the refined and extended functions of the program can be referred to the above description.

[0141] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0142] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0143] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. The various embodiments of this application can be combined with each other. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A voice interaction method, characterized in that, include: Convert user speech into initial text; Obtain a substation operation and maintenance language model and a substation operation and maintenance knowledge graph; wherein, the substation operation and maintenance language model is obtained through training and includes an intent recognition level and a solution generation level; the substation operation and maintenance knowledge graph contains multiple triples, each triple consisting of the corresponding substation operation and maintenance equipment's equipment identifier, equipment attribute parameters, and multiple operation and control information; The initial text is parsed using the intent recognition hierarchy of the substation operation and maintenance language model to generate user intent; Determine whether the user intent matches two or more operation control information; If so, all operation and control information matching the user intent will be used as the target control information; using the scheme generation hierarchy of the substation operation and maintenance language model, an execution strategy that satisfies the user intent will be generated based on the target control information. If not, the operation control information matching the user intent will be used as the execution strategy.

2. The voice interaction method according to claim 1, characterized in that, The step of parsing the initial text and generating user intent using the intent recognition hierarchy of the substation operation and maintenance language model includes: Using the intent recognition hierarchy, entity recognition and context awareness are performed on the initial text to generate user intents that include target device entities, target event entities, and / or information request types.

3. The voice interaction method according to claim 1, characterized in that, The scheme generation hierarchy utilizing the substation operation and maintenance language model generates execution strategies that satisfy the user's intent based on various target control information, including: The scheme generation hierarchy of the substation operation and maintenance language model is used to prioritize the execution of various target control information, resulting in a ranking. The target control information is then integrated and processed according to the ranking to form an initial execution strategy. The initial execution strategy is then validated based on the equipment attribute parameters matched to each target control information in the substation operation and maintenance knowledge graph to determine its feasibility. If feasible, the initial execution strategy is adopted as the final execution strategy. If not feasible, the initial execution strategy is updated based on the equipment attribute parameters matched to each target control information, and the process returns to the step of validating the initial execution strategy based on the equipment attribute parameters matched to each target control information in the substation operation and maintenance knowledge graph, until the initial execution strategy is deemed feasible.

4. The voice interaction method according to claim 1, characterized in that, Obtain the substation operation and maintenance language model, including: Construct an LLaMA large language model that includes an initial intent recognition layer and an initial scheme generation layer; Obtain substation operation and maintenance standards and procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans; Based on the aforementioned substation operation and maintenance standards and procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans, the LLaMA large language model is trained using LoRA fine-tuning technology. The LLaMA large language model obtained after training is the substation operation and maintenance language model. The initial intent recognition level obtained after training is the intent recognition level of the substation operation and maintenance language model, and the initial solution generation level obtained after training is the solution generation level of the substation operation and maintenance language model.

5. The voice interaction method according to claim 4, characterized in that, The process involves training the LLaMA large language model using LoRA fine-tuning technology, based on the substation operation and maintenance standard procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans. The substation operation and maintenance standard procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans are processed to form multiple types of training samples. One type of training sample is extracted from the substation operation and maintenance standard procedures, another type is extracted from the substation operation and maintenance defect data, yet another type is extracted from the substation operation and maintenance fault handling plans, and a third type is extracted from both the substation operation and maintenance defect data and the substation operation and maintenance fault handling plans. The LLaMA large language model was fine-tuned using LoRA training with various training samples.

6. The voice interaction method according to claim 5, characterized in that, The process involves data processing of the substation operation and maintenance standards and procedures, substation operation and maintenance defect data, and substation operation and maintenance fault handling plans to generate multiple types of training samples, including: The standard processing procedure for each substation operation and maintenance equipment is extracted from the substation operation and maintenance standard procedures, and the standard processing procedure for each substation operation and maintenance equipment is integrated to form one type of training sample; Event extraction is performed on the substation operation and maintenance defect data to extract the defect performance characteristics, defect types and defect levels under different typical faults of each substation operation and maintenance equipment; based on the defect performance characteristics, defect types and defect levels under each typical fault, another type of training sample is constructed. Key information is extracted from the substation operation and maintenance fault handling plan to generate handling schemes under different typical faults, and the handling schemes under each typical fault are integrated to form another type of training sample. The defect characteristics and solutions of at least two typical faults are integrated to form another type of training sample.

7. The voice interaction method according to claim 1, characterized in that, Obtain the substation operation and maintenance knowledge graph, including: Obtain the equipment identification, rated parameters, normal operating parameter range, safety constraints, equipment inspection procedures, equipment fault characteristics, equipment fault handling plans and / or equipment operation procedures for different substation operation and maintenance equipment; The equipment identification, rated parameters, normal operating parameter range, safety constraints, equipment inspection procedures, equipment fault characteristics, equipment fault handling plan and / or equipment operation process of the same substation operation and maintenance equipment are processed to form a corresponding triplet. The various ternary sets are integrated to form a substation operation and maintenance knowledge graph.

8. A voice interaction device, characterized in that, include: The conversion module is used to convert user speech into initial text; The acquisition module is used to acquire the substation operation and maintenance language model and the substation operation and maintenance knowledge graph; wherein, the substation operation and maintenance language model is obtained through training and includes an intent recognition level and a solution generation level; the substation operation and maintenance knowledge graph contains multiple triples, each triple consisting of the corresponding substation operation and maintenance equipment's equipment identifier, equipment attribute parameters, and multiple operation and control information. The parsing module is used to parse the initial text and generate user intent by utilizing the intent recognition level of the substation operation and maintenance language model; The judgment module is used to determine whether the user intent matches two or more operation control information; if yes, the generation module is called; if no, the determination module is called. The generation module is used to take all the operation and control information that match the user intent as the target control information; and to generate an execution strategy that meets the user intent based on the target control information using the scheme generation hierarchy of the substation operation and maintenance language model. The determination module is used to use the operation control information that matches the user intent as the execution strategy.

9. A voice interaction device, characterized in that, Including memory and processor; The memory is used to store programs; The processor is configured to execute the program to implement the various steps of the voice interaction method as described in any one of claims 1-7.

10. A readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the various steps of the voice interaction method as described in any one of claims 1-7.