Intelligent power distribution cabinet AI voice interaction method

Abstract

The invention discloses an AI voice interaction method for an intelligent power distribution cabinet, and the method comprises the steps: receiving a voice instruction of an operator, and analyzing the expression intention and experience score of the operator; matching a task framework according to the expression intention, and calculating personalized interaction strategy parameters through a preset strategy decision model in combination with experience scores; and generating voice interaction output content according to the personalized interaction strategy parameters and the expression intention. According to the invention, the interaction strategy can be adaptively adjusted according to the ability level of the operator, and the operation and maintenance efficiency and operation safety guarantee of the full-spectrum operator from a green hand to an expert are realized.

Description

Technical Field

[0001] This invention relates to the field of power distribution cabinet technology, and in particular to an AI voice interaction method for intelligent power distribution cabinets. Background Technology

[0002] Distribution cabinets (including indoor and outdoor high and low voltage switchgear, power cabinets, etc.) are key nodes in the power transmission and distribution system. Regular inspection, preventive maintenance and fault repair are important links to ensure the stable operation of the power grid.

[0003] Currently, the operation and maintenance of power distribution cabinets mainly relies on manual labor. The general process is as follows: maintenance personnel go to the site according to the plan, and check the status of switches, lines and other components in the cabinet by observing with their senses and measuring with handheld instruments (such as thermometers and current measuring clamps). They then manually record the data, and finally judge the health status of the equipment based on their personal experience and perform simple maintenance or report the fault.

[0004] With the upgrading to intelligent systems, some advanced power distribution cabinets have begun to integrate sensors (such as temperature and current sensors), human-machine interfaces (display screens, speakers, microphones), and communication modules, and have introduced basic voice interaction functions. For example, operators can call out the switch number by voice, and the system will display the real-time and historical parameters of that switch on the screen and provide voice announcements.

[0005] However, the voice interaction mechanisms currently used in power distribution cabinets only offer a "question-broadcast" mode. The system's response content and form lack substantial differentiation, failing to distinguish between operators with different maintenance types and experience levels, thus hindering the provision of specialized proactive guidance or warnings. Therefore, it remains merely a simple tool. Consequently, for experienced experts, the system's redundant broadcasts and cumbersome confirmation steps can interfere with their efficient operation.

[0006] For novices, the system lacks necessary step-by-step guidance and safety prompts. Operational errors could lead to serious safety incidents, causing them to heavily rely on experienced experts when facing complex faults, resulting in high training costs and low efficiency. Furthermore, due to the significant knowledge gap between novices and experts in the field of power distribution cabinet maintenance, short-term training with a voice interaction mechanism that densely outputs technical jargon necessitates novices relying on other tools to aid their understanding, greatly impacting maintenance efficiency.

[0007] Furthermore, when implementing intelligent and deep interaction, the current hardware configuration of power distribution cabinets primarily focuses on fulfilling telemetry, remote control, and remote adjustment functions. Standard configurations typically include a display screen, an integrated microphone, a communication module, and necessary status sensors. If additional non-standard equipment such as dedicated high-precision cameras are configured to perform behavioral analysis and assess operator capabilities, it will significantly increase costs and weaken the market competitiveness of current power distribution cabinets.

[0008] Therefore, in order to enhance the depth of human-machine collaboration, this application proposes an intelligent power distribution cabinet AI voice interaction method based on the existing standard hardware configuration of the power distribution cabinet. This method can dynamically sense and evaluate the experience level of operators and then adaptively adjust the strategy and content of voice interaction. This reduces the dependence on the experience of novice operators and reduces meaningless content output to experienced operators, thereby achieving operational efficiency and safety assurance for operators of all levels, from novices to experts. Summary of the Invention

[0009] The purpose of this invention is to address the shortcomings of existing technologies by proposing an AI voice interaction method for intelligent power distribution cabinets. This method can adaptively adjust the interaction strategy according to the operator's skill level, thereby improving the operational efficiency and safety of operators from novices to experts across the entire spectrum.

[0010] Firstly, in order to achieve the above objectives, the present invention provides the following technical solution:

[0011] AI voice interaction methods for intelligent power distribution cabinets include:

[0012] Receive voice commands from operators and analyze the operators' expressive intentions and experience ratings;

[0013] Based on the expressed intent, a task framework is matched, and personalized interaction strategy parameters are calculated through a pre-set strategy decision model in conjunction with experience scoring.

[0014] Based on the personalized interaction strategy parameters and the expressive intent, voice interaction output content is generated.

[0015] By adopting this implementation method, the voice interaction system can adaptively adjust the interaction strategy according to the operator's skill level, thereby achieving operational efficiency and safety assurance for operators of all skill levels, from novices to experts.

[0016] In one implementation, the process of analyzing the operator's expressed intent and experience score further includes:

[0017] An endpoint detection algorithm based on energy and zero-crossing rate and a spectral subtraction noise reduction algorithm are used to preprocess the speech command, generate speech command text, and perform word segmentation and part-of-speech tagging on the speech command text.

[0018] In one implementation, parsing the operator's expressed intent includes:

[0019] An intent classifier, implemented by fine-tuning a pre-trained language model based on the Transformer architecture, parses the pre-processed speech instruction text and uses a sequence labeling model to fill semantic slots to extract the operation object and parameters, thereby obtaining the expressed intent.

[0020] In one implementation, the analysis of the operator's experience score includes:

[0021] Semantic features, including terminology expression features and linguistic logic features, are extracted from the voice command text; and acoustic prosodic features are extracted from the voice command.

[0022] Based on the semantic features, an experience score for the operator is generated.

[0023] This implementation method constructs an experience scoring calculation method that is more suitable for the operation and maintenance of power distribution cabinets. It can calculate the experience score based on the voice commands issued by the operator during each interaction, solving the problem of providing corresponding interaction strategies when operators have different levels of familiarity with different operation and maintenance operations, and further improving the degree of personalization.

[0024] In one implementation, the generation of the operator's experience score includes:

[0025] Extract the voiceprint features from the voice commands;

[0026] Based on the voiceprint feature matching database, the historical score of the operator is output;

[0027] An experience score for the operator is generated based on the historical scores, terminological expression features, linguistic logic features, and acoustic prosodic features.

[0028] This implementation constructs a more stable and comprehensive experience score calculation architecture that not only responds sensitively to the operator's current performance but also calculates based on an understanding of their past abilities. This ensures that a novice's occasional fluent command will not result in an inflated score, and an experienced operator's occasional slip of the tongue will not be directly misjudged as a novice, thus ensuring the generation of more reasonable and personalized voice interaction output.

[0029] In one implementation, the method further includes updating the historical experience score, including:

[0030] After analyzing the operator's experience score, the experience score is overwritten with the historical score stored in the operator's voiceprint feature database.

[0031] In one embodiment, the terminology expression features include at least one of terminology accuracy and terminology usage density;

[0032] The language logic features include at least the integrity of the instruction logic structure;

[0033] The acoustic prosodic features include at least one of speech rate, intonation, and pause density.

[0034] In one implementation, the operator's experience score is generated. include:

[0035] Calculate this score ;

[0036] Calculate experience score ;

[0037] in, This indicates the score for this test; This indicates the preset historical scoring weights; This indicates the predefined terminology expression feature weights; This indicates the pre-defined weights of the language logic features; This represents the preset acoustic prosodic feature weights, and ∈[0,1], + + =1;

[0038] The original components represent the expressive features of the terminology. Representing the original components of linguistic logical features, Represents the primitive components of acoustic prosodic features. , and All ∈ [0, 100]; Indicates the benchmark score for the expressive characteristics of the terminology; Indicates the benchmark score for the logical features of the language; The baseline score represents the acoustic prosodic features. , and All are preset values.

[0039] This implementation method constructs a specific calculation method for experience scores, which can simultaneously cover historical scores and the current score of the current interaction. This ensures that no matter what kind of accidental event occurs, the fluctuation of experience scores will not exceed expectations, and is more in line with the actual expectations of any operator in any maintenance operation.

[0040] In one implementation, the historical rating weights in the calculation of the experience rating are... Determined by the following method:

[0041] Extract query keywords from the voice command text, wherein the query keywords include at least one of interrogative words and quantifiers;

[0042] Calculate the query depth score based on the query keywords;

[0043] The query depth score is mapped to historical score weights according to predefined mapping rules. The mapping rule satisfies the condition that the higher the query depth score, the lower the historical score weight α.

[0044] By employing this implementation method, the "sensitivity" of experience scoring is intelligently adjusted: for shallow queries, the score changes gradually, relying more on the operator's past abilities; for deep exploration, the score changes more sensitively, paying more attention to the operator's performance in the current interaction. This makes experience scoring calculation more intelligent and better able to distinguish between "smoothness due to simple tasks" and "precision resulting from superior abilities."

[0045] In one embodiment, extracting query keywords from the voice command text includes:

[0046] Based on a predefined or pre-trained query keyword model library, the system matches query keywords in the voice command text and outputs the keyword weight corresponding to each query keyword.

[0047] The calculation of the query depth score based on the query keywords includes:

[0048] The query depth score is calculated based on the query keywords and the first weight.

[0049] In one implementation, the step of overwriting the historical scores stored in the operator's voiceprint feature database with the experience score includes:

[0050] ;

[0051] Where β represents the preset historical rating update weight, and β∈(0,1). This indicates the updated rating and will be used for the next interaction. , This indicates the current historical rating, i.e., the rating before this interaction. , This represents the experience score calculated during this interaction.

[0052] By adopting this implementation method, the "capability profile" maintained by the voice interaction system for each operator possesses both anti-disruption and memory capabilities. It can both slowly reflect the operator's true growth or decline trend over time and filter out accidental noise in a single interaction, ensuring the stability and reliability of the "historical score," which is one of the foundations for strategic decision-making.

[0053] In one embodiment, the original component of the terminology expression feature = ;

[0054] in The raw score representing the accuracy of technical terms, The raw score representing the density of technical terms used; This indicates the preset accuracy weight of technical terms. This indicates that the default technical terms use density weights.

[0055] In one embodiment, the extraction of terminological features from speech instructions includes:

[0056] Based on a pre-built standard library of terms in the field of power distribution cabinets, the word segments are matched and identified, and the identified words are divided into completely correct words, fuzzy terminology words and non-terminology words.

[0057] The term "completely correct" means that it is completely consistent with the terminology in the standard library of terminology in the field of power distribution cabinets or is its standard abbreviation;

[0058] The fuzzy term representation is consistent with the core word root of a completely correct term or has a semantic similarity exceeding a preset threshold.

[0059] The non-terminal terms mentioned do not match any of the terms in the standard terminology library for the field of distribution cabinets.

[0060] In one embodiment, the original score of the accuracy of technical terms in the terminology expression features. The calculation includes: calculating the ratio of the number of completely correct words to the total number of word segments, and mapping it to a score range of [0, 100] as the raw score for the accuracy of the terminology.

[0061] In one embodiment, the original distribution of the density of specialized terminology usage in the terminology expression features. The calculation includes: calculating the sum of the number of completely correct words and the number of ambiguous term words, then calculating the ratio with the total number of word segments, and mapping it to a score range of [0, 100] as the raw score of the density of term usage.

[0062] In one embodiment, the original language logic features are... = ;

[0063] in, The original component representing the integrity of the logical structure. This represents the preset logical structure integrity weight.

[0064] In one implementation, the original components of the logical structure integrity The calculations include:

[0065] Based on the knowledge base of power distribution cabinet operation and maintenance, the core semantic roles that a complete operation instruction should contain are defined, and the matching rules of the core semantic roles in the syntax tree are defined.

[0066] The speech command text is parsed by a dependency parser to extract the grammatical dependency relations between each word segment, including at least one of subject-predicate relations, verb-object relations, and attributive-head relations, and to obtain a syntax tree.

[0067] According to the matching rules and the syntax tree, when the dependency relationship of the core semantic role is missing, a score is deducted according to a preset value, and the score of logical structure integrity is output.

[0068] The scores are mapped to a score range of [0, 100] and used as the raw scores for the accuracy of the terminology.

[0069] In one embodiment, the original acoustic prosodic features are... = ( ) ;

[0070] in, This indicates the preset speech rate weight. This represents the preset pause density weight, and + =1; The original score represents the speech rate. Indicates the original density of pauses. Indicates the original part of intonation.

[0071] In one embodiment, the original speech rate score The calculations include:

[0072] The total duration of voice commands was detected using a VAD model.

[0073] The number of syllables can be detected using a syllable boundary detection model or a short-time energy envelope method.

[0074] The speech rate is obtained by calculating the ratio of the number of syllables to the total duration of the voice command;

[0075] The difference between the stated speech rate and the predefined standard speech rate for instructions in the field of power distribution cabinet operation and maintenance is calculated, and the difference is mapped to a predefined speech rate scoring rule to obtain a speech rate score.

[0076] The speech rate score is mapped to the score range of [0, 100] to obtain the raw speech rate score;

[0077] The speech rate scoring rule states that the smaller the difference, the higher the speech rate score.

[0078] In one embodiment, the original pause density is... The calculations include:

[0079] The non-speech segments and the duration of each non-speech segment are detected using the syllable boundary detection model or the short-time energy envelope method.

[0080] Calculate the number of non-speech segments whose duration exceeds a preset duration to obtain the number of pauses;

[0081] Calculate the ratio of the number of pauses to the total duration of the voice command, and inversely map the ratio to the score range of [0, 100] to obtain the original score of pause density.

[0082] In one embodiment, the original intonation segment The calculations include:

[0083] By extracting the fundamental frequency sequence of the voice command;

[0084] The fundamental frequency sequence is smoothed and fitted to obtain the trend function of the voice command;

[0085] Calculate the residual sequence between the fundamental frequency sequence and the trend function, and further calculate the standard deviation of the residual sequence;

[0086] The standard deviation is mapped to a predefined intonation scoring rule to obtain an intonation score;

[0087] The intonation score is mapped to the score range of [0, 100] to obtain the raw intonation score;

[0088] The tone scoring rule states that the smaller the standard deviation, the higher the tone score.

[0089] Using this implementation method, although the overall change in the raw scores of acoustic prosody features has little impact on the calculation of experience scores, ensuring that normal physiological fatigue will not have a disruptive impact on personnel ability assessment, the multiplicative model endows the raw scores of acoustic prosody features with high internal sensitivity during calculation. This allows any psychological signal of "disfluency" or "lack of confidence" to be amplified into a significantly low score. As a result, the final interaction strategy always tends to increase the degree of guidance and safety prompts. This makes the voice interaction system always lean more towards reliability than efficiency, which is in line with the characteristics of the power distribution cabinet operation and maintenance field, which prioritizes absolute safety. This ensures that any potential risks caused by the operator's condition can receive auditable and traceable additional attention.

[0090] In one implementation, the task framework for matching based on expressed intent includes:

[0091] The expressed intent is matched with a pre-defined operation and maintenance task map to output the corresponding task framework;

[0092] The task framework is a structured data object, which includes at least one of the following: task type identifier, task object identifier, task risk level, task default security confirmation level, and task advisory information output dimension set.

[0093] In one implementation, the task type identifier includes at least one of status query, parameter setting, fault diagnosis, status control, and contingency plan execution; the task risk level and the task default safety confirmation level are predefined based on power safety regulations.

[0094] By adopting this implementation method, the objects to be affected by experience scoring are constructed, which proves that all subsequent personalized interaction strategies of the system will be intelligently adjusted within the framework allowed by security procedures and within the reasonable scope of business logic, reserving a flexible, computable and traceable decision space for the strategy decision-making model based on experience scoring.

[0095] In one implementation, the input to the strategy decision model is a feature vector formed by concatenating the experience score with the task framework;

[0096] The model output is a set of personalized interaction strategy parameters, including information detail levels. , guidance intervention density level Security Confirmation Level and broadcast style identifier .

[0097] In one implementation, the information detail level Used to control the granularity of information retrieved from the database and the length of additional explanations;

[0098] The guided intervention density level Used to determine the number of nodes that proactively offer suggestions for the next action in a single conversation;

[0099] The security confirmation level Used to determine the complexity of the independent verification steps that must be passed before performing operations involving state changes in the task framework;

[0100] The broadcast style identifier Used to index a set of pre-configured speech synthesis parameters, including speech rate, intonation, and sentence templates.

[0101] By adopting this implementation method, the strategic decision-making process is made transparent, explainable, and optimizable, ensuring that the AI ​​model has reliable auditability when applied to the field of power distribution cabinet operation and maintenance with stringent safety requirements, thus providing a more trustworthy basis for personalized interaction in this field.

[0102] In one embodiment, the method for constructing the input feature vector of the policy decision model includes:

[0103] The experience score is used as the first continuous feature;

[0104] The task type identifier, criticality level, default security confirmation level, and information output dimension set in the task framework are encoded and converted into a second feature group.

[0105] The first continuous feature and the second feature group are concatenated in a predetermined order to form the feature vector.

[0106] In one implementation, the strategy decision model is a classification model trained using historical interaction logs;

[0107] Each training sample in the historical interaction log contains a feature vector as input and annotated, and a set of personalized interaction strategy parameters as output and annotated.

[0108] The feature vector is a first array formed by concatenating vectors obtained by encoding the experience score and the task framework;

[0109] The personalized interaction strategy parameter set includes information detail levels. , guidance intervention density level Security Confirmation Level and broadcast style identifier The second array is formed.

[0110] In one embodiment, the annotation method for the personalized interaction strategy parameter set includes: automatically or manually determining the personalized strategy parameter set of the interaction as a valid label based on the operator's positive confirmation feedback, task completion efficiency, and safety records in a historical interaction, and storing it in the training sample library after verification.

[0111] In one implementation, the strategy decision model is a classification model based on gradient boosting trees.

[0112] In one implementation, when historical interaction logs are lacking in the early stages of the deployment of the strategy decision model, a rule-based strategy decision engine is used to directly output personalized interaction strategy parameters by looking up tables based on experience scores and task frameworks.

[0113] Once the accumulated interaction logs of the strategy decision model reach a preset number, the training of the machine learning-based strategy decision model is initiated, and it gradually replaces the strategy decision engine.

[0114] This implementation method greatly enhances the engineering practicality and resilience of the technical solution, ensuring that it can be smoothly integrated into the existing operation and maintenance environment and continuously optimize itself.

[0115] Secondly, in order to achieve the above objectives, the present invention also provides the following technical solution:

[0116] An intelligent power distribution cabinet AI voice interaction system is used to implement the above method. The system includes:

[0117] The voice processing module is used to receive voice commands from the operator and parse the operator's intended meaning.

[0118] An experience score generation module is used to extract semantic features and voiceprint features from the voice commands and generate an experience score for the operator.

[0119] The strategy decision module is used to match the task framework according to the expressed intent, and calculate personalized interaction strategy parameters through a preset strategy decision model in combination with the dynamic experience score.

[0120] The content generation and output module is used to generate and output voice interaction content based on the personalized interaction strategy parameters and the expression intent.

[0121] Thirdly, in order to achieve the above objectives, the present invention also provides the following technical solution:

[0122] A computer program product, including a computer program, which, when executed by a processor, implements the method described above.

[0123] Fourthly, in order to achieve the above objectives, the present invention also provides the following technical solutions:

[0124] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0125] Fifthly, in order to achieve the above objectives, the present invention also provides the following technical solutions:

[0126] AI intelligent power distribution cabinet includes a cabinet, power components installed inside the cabinet, and the system described above;

[0127] The system is connected to at least the status sensors of the power components to acquire information about each power component; the system also includes a human-machine interface panel, a voice acquisition module and a broadcasting unit, with the voice acquisition module and the broadcasting unit both located on the human-machine interface panel of the cabinet.

[0128] Compared with the prior art, the beneficial effects of this invention are: it can adaptively adjust the interaction strategy according to the operator's ability level, so as to achieve operation and maintenance efficiency and work safety assurance for operators of all levels, from novices to experts. Attached Figure Description

[0129] Figure 1 This is a flowchart of the AI ​​voice interaction method for intelligent power distribution cabinets proposed in this invention;

[0130] Figure 2 This is a topology diagram of the AI ​​voice interaction system for intelligent power distribution cabinets proposed in this invention. Detailed Implementation

[0131] In the operation and maintenance of power distribution cabinets (including switch cabinets, electrical cabinets, etc.), operators (such as inspection and maintenance personnel) can use the integrated voice interaction system and display panel to assist their work. For example, when it is necessary to query the historical records of a switch, the operator can directly issue a voice command such as "historical records of switch XX", and the system will automatically retrieve the historical data of the corresponding power component and present it on the display panel and announce it via voice. This method reduces the workload of manual inspection, facilitates quick judgment of equipment conditions, and significantly improves operation and maintenance efficiency.

[0132] However, due to the relatively basic development of current voice interaction mechanisms in the field of power distribution cabinets, their response patterns are fixed and undifferentiated. Operators of all experience levels complete their tasks through the same voice interaction mechanism, which reveals the following obvious limitations in practical applications:

[0133] For experienced personnel, the system output often does not match their expectations: in simple or familiar work scenarios, the broadcast content is often too long, interfering with their efficient work; while in critical or complex work situations, the output may be too brief, forcing operators to repeatedly issue instructions to obtain sufficient information, which in turn reduces efficiency.

[0134] Furthermore, for newcomers, the standardized output content presents another problem: the broadcast information is dense with technical jargon, exceeding their current comprehension. This forces them to either repeatedly ask for clarification on the meaning of the terms or seek help from on-site experts. This not only slows down the work progress but also makes it difficult for them to effectively accumulate experience through practical work, resulting in slow growth.

[0135] The above situation illustrates that the voice interaction mechanism currently configured in the field of power distribution cabinet operation and maintenance is still a relatively simple tool. It cannot provide appropriate output content based on the different experience levels and abilities of operation and maintenance personnel, and therefore cannot provide "just right" assistance based on the actual capabilities of different operation and maintenance personnel, thus restricting the further improvement of the overall efficiency and safety level of operation and maintenance work.

[0136] It should be understood that the following embodiments are all intended to provide a voice interaction method that can dynamically sense the operator's experience level and provide personalized interactive content accordingly, so as to meet the actual needs of different maintenance personnel and balance the operational efficiency and rational guarantee of various maintenance tasks.

[0137] In the following embodiments, voice commands should be considered as all commands issued by the operator during the voice interaction process, that is, the voice content received by the voice interaction system and output by the operator during the voice interaction process; output content refers to the content replied by the voice interaction system during the voice interaction process; all calculation processes in the interaction process, including the parsing and extraction of voice commands, the calculation of experience scores, the calculation of personalized interaction strategy parameters, and the generation and output of the final voice interaction content, are implemented through the existing voice interaction system of the power distribution cabinet, such as through existing cloud computing, edge computing, etc., or through other methods; the preprocessing, parsing, and extraction steps after the operator issues voice commands can all be implemented through existing related voice processing technology solutions, and the following embodiments are all based on this prior art.

[0138] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0139] Example 1:

[0140] This invention provides the following technical solution: an AI voice interaction method for intelligent power distribution cabinets, comprising:

[0141] Receive voice commands from operators and analyze their intended meaning and experience rating;

[0142] The task framework is matched based on the expressed intent, and personalized interaction strategy parameters are calculated through a pre-set strategy decision model by combining experience scores.

[0143] Based on personalized interaction strategy parameters and expressive intent, generate voice interaction output content.

[0144] Figure 1 The core logic of this voice interaction mechanism is illustrated. As an optional implementation of this invention, in a complete voice interaction round: after the operator issues a voice command, the voice receiving module (e.g., a microphone) configured in the power distribution cabinet receives the command and automatically parses it, extracting the operator's expressive intent and experience score. The expressive intent represents the task the operator needs the voice interaction system to perform, and the experience score represents the operator's skill level. By simultaneously acquiring the operator's expressive intent and experience score, the voice interaction system can understand "what needs to be done" and "what kind of operator is being served." Subsequently, the system matches the corresponding task framework based on the expressive intent and calculates personalized interaction strategy parameters based on the experience score, thereby dynamically adjusting the content of the current voice command response (i.e., the voice interaction output content) and its presentation.

[0145] In this embodiment, the system analyzes voice commands to assess the operator's experience level and uses the interaction strategy parameters to determine which interaction strategy to adopt (e.g., different strategies adapted to novices, qualified personnel, or senior personnel), and finally generates appropriate voice interaction content for the operator.

[0146] In some possible implementations, if the operator's experience score is high, indicating they are a senior personnel (i.e., an experienced veteran) in the field of power distribution cabinet maintenance, the voice interaction system can adopt a concise and professional response method. This means the voice interaction output is more refined, uses more technical terms, and reduces the level of detail and length, making the system an efficient auxiliary tool and avoiding lengthy information from interfering with efficient work. The speaking speed should be steady or slightly faster. Conversely, if the experience score is low, indicating a novice, the system should adopt a detailed and guided response method. This means the voice interaction output is more detailed, minimizing or explaining technical terms, using a more conversational and visual style, increasing the depth and length of the content, and speaking at a slightly slower pace. This allows the voice interaction system to act as a patient and meticulous "expert mentor," ensuring the novice can fully understand and safely and correctly complete various maintenance tasks.

[0147] In some possible implementations, since the voice interaction system primarily provides query and analysis functions in current actual operation and maintenance work, operators can independently judge and perform subsequent operations such as maintenance and reporting based on the query or analysis results provided by the system, combined with their own experience or system prompts. Typical operation and maintenance scenarios include: querying the status or historical data of specific power components when the distribution cabinet is normal; or, when a problem occurs, the operator can ask the voice interaction system to analyze the cause.

[0148] Therefore, whether experienced or novice, all personnel will inevitably issue instructions containing intents such as queries and commands during actual operations. Providing corresponding services based on these instructions is the fundamental role of current power distribution cabinet voice interaction systems. It should be noted that existing technologies commonly use the intent expressed in input instructions to match a task framework and thus determine the scope of the output content. This explanation aims to illustrate the specific meanings of "expressed intent" and "task framework" within this field; the matching method itself is not the focus or limitation of this embodiment.

[0149] To enable those skilled in the art to better understand the implementation process of this embodiment, the following experimental examples are based on voice commands that may or have occurred in real-world operation and maintenance processes, specifically illustrating the role of experience scoring in operation and maintenance work in this embodiment:

[0150] Experimental Example a:

[0151] The operator gave the voice command: "Check the tripping status of switch 301";

[0152] The voice interaction system interprets the following intent as: query the tripping status of switch 301, then retrieve the historical logs of the switch in the log database, and extract the logs tagged "tripping";

[0153] Assuming the experience score calculated by the voice interaction system indicates that the operator is a novice with little experience, the specific interaction strategy could include: 1. The voice interaction output should include the reason for the switch tripping (if the switch has tripped), the corresponding maintenance procedure in the maintenance procedure library, and if there is a safety risk in the maintenance procedure, prioritize broadcasting the safety tips from the safety operation procedure library; 2. When broadcasting, the voice interaction output should be detailed and conversational, the speech should be slow, and the volume should be increased when broadcasting safety tips for emphasis.

[0154] Experimental Example b: The operator gives the voice command: "Check the tripping status of switch 301";

[0155] The voice interaction system interprets the intended message as follows: query the tripping status of switch 301, and then retrieve the historical logs of the switch in the log database and extract the logs tagged "tripping".

[0156] Assuming the experience score calculated by the voice interaction system indicates that the operator is a senior expert, the specific interaction strategy could include: 1. The voice interaction output should include the reason why the switch tripped (if the switch has tripped); 2. The voice interaction output should be concise and professional, and the speaking speed should be slightly faster than usual.

[0157] In both experimental cases, we can clearly see that when the experience score is high, this voice interaction system serves as a simple auxiliary tool for experienced operators, making the overall operation and maintenance work more efficient. When the experience score is low, this voice interaction system acts as a professional senior expert and mentor, providing detailed guidance to less experienced novices to safely perform various operation and maintenance tasks, ensuring that the operation and maintenance tasks are completed safely and correctly, without requiring novices to use other tools or seek help from other senior experts. Therefore, it also indirectly improves the efficiency of novices in operation and maintenance work and helps them accumulate experience in the field of power distribution cabinet operation and maintenance.

[0158] As an optional implementation of this invention, in a complete voice interaction round: when the operator first issues a voice command, the voice receiving module (e.g., a microphone) configured in the power distribution cabinet receives the voice command, and then automatically parses the voice command, extracting the operator's expressive intent and experience score contained therein. The expressive intent indicates what the operator wants to do, and the experience score indicates the operator's ability level. Therefore, through the expressive intent and experience score, the voice interaction system can understand what the operator wants to do and know the operator's ability level. Subsequently, the task framework is matched according to the expressive intent, and personalized interaction strategy parameters are calculated in combination with the experience score, thereby adjusting the interaction strategy of the voice interaction system, that is, adjusting the content of the voice interaction system's response to the voice command (i.e., the voice interaction output content).

[0159] In some possible implementations, experience scoring can be calculated using existing technologies. For example, in the recruitment field, there are technologies that assess an applicant's ability level based on their expression (e.g., "How NLP Helps AI Recruitment Gain Insights into Candidates' Deeper Traits" (CSDN-2025-03-18). This technology analyzes the candidate-interviewer dialogue in real time during the interview process using NLP technology. It converts speech to text through speech recognition and then uses semantic analysis to understand the dialogue content. It can not only assess a candidate's expression ability and communication skills but also gain insights into their logical thinking, problem-solving abilities, and professional qualities from their answers. For instance, analyzing a candidate's answers to complex questions can determine whether their logical thinking is clear and whether they can articulate their views in an organized manner. In particular, NLP technology can capture subtle differences in a candidate's language. By analyzing word choice, grammatical structure, and expression, it can determine a candidate's level of expertise, language habits, and personality traits. For example, a candidate who frequently uses industry terminology and expresses it accurately may have deep professional knowledge in that field; while a candidate whose language is creative and flexible may possess strong innovative abilities).

[0160] It is understood that in this embodiment, the technology can be directly used to construct the operator's experience and ability profile, thereby obtaining the operator's experience score (i.e., the professional level of the applicant in the prior art). Alternatively, it can be based on other existing technologies that can judge the experience and ability of the personnel through the content of their conversations. It can also be based on the experience score calculation method defined in the following content. It should be understood that this embodiment does not limit which experience score calculation method to be adopted in specific implementation.

[0161] In some possible implementations, the pre-set strategy decision model is essentially a machine learning model that takes numerical experience scores and structured task framework features as input, and outputs a set of parameters (i.e., personalized interaction strategy parameters) used to adjust the final output content. Therefore, it is essentially a machine learning classification / regression model, which can be implemented, for example, using the widely used Gradient Boosting Decision Tree algorithm. This model learns the mapping relationship between input features (i.e., the operator's real-time experience scores and structured task framework) and output labels (i.e., the set of personalized interaction strategy parameters) through training. The specific training method of the model is also well-known in the art and will not be elaborated here. Of course, this embodiment can also be implemented based on the specific limitations of the strategy decision model described later. It should be understood that any machine learning model that can use numerical values ​​representing a certain ability and features representing certain tasks as input to calculate interaction strategies falls within the scope of the strategy decision model in this embodiment.

[0162] By adopting this implementation method, the voice interaction system can adaptively adjust the interaction strategy according to the operator's real-time ability level, thereby achieving operational efficiency and safety assurance for operators of all skill levels, from novices to experts.

[0163] Example 2:

[0164] Please see Figure 1 The present invention provides the following technical solution: an AI voice interaction method for intelligent power distribution cabinets, comprising:

[0165] S101: Receives voice commands from the operator;

[0166] S102: An endpoint detection algorithm based on energy and zero-crossing rate and a spectral subtraction noise reduction algorithm are used to preprocess the speech commands, generate the speech command text, and perform word segmentation and part-of-speech tagging on the speech command text.

[0167] S103: Analyze the operator's expressed intentions and experience score, including:

[0168] S1031: An intent classifier, implemented by fine-tuning a pre-trained language model based on the Transformer architecture, parses the pre-processed speech command text and uses a sequence labeling model to fill semantic slots to extract the operation object and parameters, thereby obtaining the expressed intent.

[0169] S1032: Extract semantic features from the voice command text, including terminology expression features and language logic features, and extract acoustic prosodic features from the voice commands; generate an operator experience score based on the semantic features;

[0170] Among them, the terminology expression characteristics include at least one of the accuracy of professional terminology and the density of professional terminology use;

[0171] The logical features of a language include at least the completeness of its instruction logical structure;

[0172] Acoustic prosodic features include at least one of speech rate, intonation, and pause density;

[0173] S20: Match the task framework according to the expressed intent, and calculate personalized interaction strategy parameters through a pre-set strategy decision model in combination with experience scoring;

[0174] S30: Generates voice interaction output content based on personalized interaction strategy parameters and expressive intent.

[0175] As an optional implementation of the present invention, compared with Embodiment 1, this embodiment further specifies the specific method for parsing and obtaining the expressive intent and experience score. It is understood that the preprocessing of voice commands and the parsing of expressive intent can both be implemented based on mature existing technologies, and this embodiment will not elaborate on these aspects.

[0176] Regarding the analysis method for experience scoring, although Example 1 mentions that it can be based on existing technologies in the recruitment field, such technologies typically require analyzing multiple rounds of rich dialogue to construct a complete profile of an employee's capabilities. The resulting "professionalism" score (i.e., the experience score in this example), while reflecting the applicant's comprehensive abilities, is essentially a global assessment based on a long-term, multi-topic dialogue. Directly deploying this computational architecture in the field of power distribution cabinet maintenance will make it difficult to obtain accurate experience scores. This is because, in the power distribution cabinet maintenance scenario, each voice interaction corresponds to an independent service request, and regardless of the operator's skill level, their voice commands are usually extremely concise (the language complexity may be an order of magnitude lower than in recruitment interviews), with a high probability of highly simplified commands. Therefore, directly applying existing technologies often leads to two problems: first, experience scores calculated based on a single, brief command are generally inflated, failing to accurately distinguish skill levels; second, multiple rounds of interaction are needed for the score to slowly converge to the operator's true skill level. This is not well-suited for the operation and maintenance of power distribution cabinets, where high operational efficiency is required, so targeted improvements are necessary.

[0177] This embodiment calculates real-time and accurate experience scores for single, brief maintenance voice commands. Each time an operator issues a command, the system executes the entire process from feature extraction to score generation, achieving "one interaction, one evaluation." Specifically, by extracting multi-dimensional features (terminology expression features, linguistic logic features, and acoustic prosody features) from a single voice command and calculating an experience score, the voice interaction system can perceive the instantaneous ability level demonstrated by the operator when performing the current specific task in real time.

[0178] Compared to existing technologies, although all involve the analysis of terminological expression features, linguistic logic features, and acoustic prosody features at the technical principle level, this embodiment achieves accurate evaluation based on a single command in ultra-short text and real-time scenarios, rather than relying on a global dialogue-based assessment of ability levels. This effectively avoids evaluation biases caused by varying degrees of familiarity with different maintenance operations by operators (i.e., the "ability profile" changes with the task). For example, an experienced operator might have a high experience score for the "routine inspection" command, but in the next interaction, they might issue a less frequently encountered command like "fault waveform analysis," and their single expression might show uncertainty. The system can accurately capture this difference in proficiency for a specific task through this evaluation, rather than giving a general, inaccurate high score. In this way, the interaction strategy invoked by the voice interaction system in each interaction will always match the operator's actual ability to handle the current task, thus laying a reliable foundation for personalized interaction.

[0179] This implementation method constructs an experience scoring calculation method that is more suitable for the operation and maintenance of power distribution cabinets. It can calculate the experience score based on the voice commands issued by the operator during each interaction, solving the problem of providing corresponding interaction strategies when operators have different levels of familiarity with different operation and maintenance operations, and further improving the degree of personalization.

[0180] In some possible embodiments, generating an operator's experience score includes:

[0181] Extracting voiceprint features from voice commands;

[0182] Based on the voiceprint feature matching database, output the operator's historical score;

[0183] Based on historical scores, terminological expression features, linguistic logic features, and acoustic prosodic features, an experience score for the operator is generated.

[0184] After analyzing the operator's experience score, the experience score is overwritten with the historical score stored in the operator's voiceprint feature database.

[0185] While the aforementioned embodiments derive the operator's experience score from voice commands during a single interaction, making the final voice interaction output more aligned with the operator's actual expectations during that interaction, this method is highly susceptible to interference from accidental factors. For example, a seasoned user might, due to verbal tics, fatigue, or a momentary slip of the tongue, exhibit unusual ambiguity or logical inconsistencies when issuing commands; conversely, a novice might, when given a highly procedural and simple task, occasionally produce a very fluent and standard expression. If the voice interaction system relies solely on the experience score from this single interaction, it might misjudge a seasoned user as "incompetent" and provide redundant guidance, or misjudge a novice as an "expert" and provide a concise reply with insufficient information. Both of these scenarios would damage the interactive experience and contradict the operator's actual expectations.

[0186] It is understandable that even if there are situations where a novice is very familiar with a certain operation and maintenance task or an experienced person is unfamiliar with a certain operation and maintenance task, the voice interaction system in the aforementioned embodiment has an unnecessary "cliff-like" error when calculating experience scores. In this case, the actual significance of experience scores is rather lacking.

[0187] As an optional implementation of the present invention, this embodiment further specifies the specific method of integrating personal historical records into the experience score generation. First, the voice interaction system extracts the voiceprint feature, a unique biometric identifier, from the current voice command. Next, the voice interaction system matches this voiceprint feature with a pre-established voiceprint feature database to uniquely identify the current operator and retrieve the operator's previously accumulated "historical scores." Finally, the voice interaction system does not view the current command in isolation, but rather integrates the features obtained from the real-time analysis—terminological expression, linguistic logic, and acoustic prosody—with the identified "historical scores" to generate the current "experience score." This binds the ability level demonstrated in the current interaction to the operator's historical ability level, preserving the real-time assessment capability based on a single command while incorporating an understanding of the operator's long-term ability level. This avoids situations where a momentary slip of the tongue or a perfectly executed expression results in a significantly different experience score from the operator's actual ability level—that is, a large error between the calculated experience score and the expected value.

[0188] It is important to note that this embodiment, by introducing historical scores, does not directly discard mistakes in the current performance. In fact, this embodiment not only allows for occasional perfect expressions by novices and occasional mistakes by experienced users, but also allows for novices who are indeed capable of expressing themselves perfectly in a certain situation and experienced users who become rusty in a certain situation. Therefore, this embodiment aims to reduce the degree of change in the interaction strategy when such occasional mistakes or perfect performances occur, that is, to reduce the impact of such occasional events on the overall interaction strategy, and to avoid directly generating voice interaction content "unchanged" based on the current score, which would greatly damage the interaction experience.

[0189] In some embodiments, after the voice interaction system extracts the voiceprint features of the current voice command, if an existing record can be matched from the voiceprint feature database, the system immediately outputs the operator's historical score. If no match is found, it indicates that the operator is a new person (but not necessarily a novice), and the system can output a historical score that represents a medium level of ability.

[0190] This implementation constructs a more stable and comprehensive experience score calculation architecture that not only responds sensitively to the operator's current performance but also calculates based on an understanding of their past abilities. This ensures that a novice's occasional fluent command will not result in an inflated score, and an experienced operator's occasional slip of the tongue will not be directly misjudged as a novice, thus ensuring the generation of more reasonable and personalized voice interaction output.

[0191] In some possible embodiments, an operator experience score is generated. include:

[0192] Calculate this score ;

[0193] Calculate experience score ;

[0194] in, This indicates the score for this test; This indicates the preset historical scoring weights; This indicates the predefined terminology expression feature weights; This indicates the pre-defined weights of the language logic features; This represents the preset acoustic prosodic feature weights, and ∈[0,1], + + =1;

[0195] The original components represent the expressive features of the terminology. Representing the original components of linguistic logical features, Represents the primitive components of acoustic prosodic features. , and All ∈ [0, 100]; Indicates the benchmark score for the expressive characteristics of the terminology; Indicates the benchmark score for the logical features of the language; The baseline score represents the acoustic prosodic features. , and All are preset values.

[0196] As an optional implementation of the present invention, this embodiment further defines the overall framework for calculating the experience score, which is a summary of the experience score calculation methods in the foregoing embodiments.

[0197] First, when calculating this score, the raw score for each feature ( , and Each of these will subtract a preset baseline score. , and This benchmark score is derived from a statistical analysis of a large sample of qualified operators, representing a "neutral score," "passing score," or "default score" in the field of distribution cabinet operation and maintenance. By subtracting the corresponding benchmark score from the raw score of each feature, benchmark discrepancies can be eliminated, ensuring that this score represents the deviation of the performance of each semantic feature contained in this voice command from the average level. Then, each deviation value is multiplied by its corresponding weighting coefficient and summed to obtain the current score. .

[0198] This ensures that different features can contribute differently to the total score, and is also configurable and interpretable. For example, in a voice command issued by an operator, the logical structure completeness and terminology accuracy often have a greater impact on judging whether the operator is professional than the speech rate and pause density. Of course, the weight design of each semantic feature can be customized with reference to existing standards.

[0199] This implementation method constructs a specific calculation method for experience scores, which can simultaneously cover historical scores and the current score of the current interaction. This ensures that no matter what kind of accidental event occurs, the fluctuation of experience scores will not exceed expectations, and is more in line with the actual expectations of any operator in any maintenance operation.

[0200] In some other possible embodiments, the historical rating weights in the experience rating are calculated. Determined by the following method:

[0201] Based on a predefined or pre-trained query keyword model library, it matches query keywords in voice command text and outputs the keyword weight corresponding to each query keyword; query keywords include at least one of interrogative words and quantifiers;

[0202] Calculate the query depth score based on the query keywords and the first weight;

[0203] Map the query depth score to the historical score weights according to predefined mapping rules. The mapping rule satisfies the condition that the higher the query depth score, the lower the historical score weight α.

[0204] As understood in the previous embodiment, in order to make the voice interaction output generated by the voice interaction system more in line with the actual expectations of any operator in any maintenance operation, and to solve the problem when accidental events occur, it is necessary to include both historical scores and the current score of the current interaction in the calculation of the experience score.

[0205] However, it was also mentioned that if an experienced user is genuinely unfamiliar with a certain operation and maintenance task, and it is not due to forgetting or a slip of the tongue, or if a novice happens to be very familiar with a certain operation and maintenance task and it is not by chance that their expression is accurate, this means that experience scoring should focus more on the current score rather than historical scores.

[0206] Conversely, if the operation and maintenance work is indeed a case of an experienced person making a slip of the tongue or a novice expressing themselves perfectly by chance, it means that experience scoring should focus more on historical scores rather than the current score.

[0207] Therefore, how to design the weighting of historical scores and current scores will affect the actual experience of operators.

[0208] In this embodiment, the historical scoring weight is dynamically determined based on the query depth of the voice command. In real-world operations and maintenance scenarios, when operators raise complex, in-depth, and exploratory questions, it better reflects their true depth of thought and professional curiosity. The quality of their responses also better reflects their actual skill level; therefore, this score should be awarded accordingly. Higher weight ( This means reducing the weight of historical ratings. .

[0209] Compared to the previous embodiment where the historical scoring weights were set to preset values, the dynamic weight design allows the voice interaction system to better distinguish between situations where "an experienced person makes an accidental mistake / a novice expresses themselves perfectly" and "a novice is indeed familiar / an experienced person is indeed unfamiliar," which is also more in line with common sense in the field of power distribution cabinet operation and maintenance.

[0210] Furthermore, we have limited the association between depth scoring and query keywords. It should be understood that query keywords and quantifiers are words used in traditional semantic understanding. For example, query keywords include interrogative words such as "why," "what," and "where," reflecting the different levels of concern the operators have for the maintenance work (e.g., "When did it burn out?", "Why didn't the alarm go off after tripping for so long? When did it start burning out?"). Quantifiers include words indicating quantity such as "times" and "numbers," reflecting the scope of the operators' attention to the maintenance work (e.g., "Did the three switches on the left trip simultaneously? Check the tripping status of all switches. Compare the monthly line burnout situation over the past six months.").

[0211] Understandably, interrogative words in the query keywords reflect the vertical depth of the operator's query in this voice command, while quantifiers reflect the breadth of the query. In various practical fields, existing knowledge in psychology and education, especially Bloom's Taxonomy of Cognitive Objectives, reveals the phenomenon that "experienced individuals are often able to ask questions at higher cognitive levels." Therefore, in this embodiment, the different depths and breadths of the questions asked by the operator naturally reflect their ability level. It is important to note that any operator can ask very shallow or very deep questions, depending on their level of thought, attention, importance, and interest. However, only relatively more "professional" individuals can ask more sophisticated questions. Therefore, if the depth score is directly used as one component of the semantic features, it becomes difficult to discern the operator's true experience level.

[0212] In this embodiment, the depth score is used as the weight of the historical score. The adjustment factor allows the voice interaction system to determine whether to focus more on the operator's current performance or their historical performance during the experience score calculation. For example, if the operator only issues simple commands during the interaction, the experience score calculation will focus more on the operator's historical scores, and the voice interaction service can be completed directly according to the operator's historical habits, which is more in line with the operator's current skill level.

[0213] By employing this implementation method, the "sensitivity" of experience scoring is intelligently adjusted: for shallow queries, the score changes gradually, relying more on the operator's past abilities; for deep exploration, the score changes more sensitively, paying more attention to the operator's performance in the current interaction. This makes experience scoring calculation more intelligent and better able to distinguish between "smoothness due to simple tasks" and "precision resulting from superior abilities."

[0214] In some possible embodiments, overriding historical scores stored in the operator's voiceprint feature database with experience scores includes:

[0215] ;

[0216] Where β represents the preset historical rating update weight, and β∈(0,1). This indicates the updated rating and will be used for the next interaction. , This indicates the current historical rating, i.e., the rating before this interaction. , This represents the experience score calculated during this interaction.

[0217] As an optional implementation of the present invention, since the aforementioned embodiments mentioned the introduction of historical scores, but historical scores are not necessarily static, in fact, such a situation also exists in the field of power distribution cabinet operation and maintenance: experienced personnel become rusty after not having been in operation and maintenance positions for a long time, while novices gradually accumulate experience after being in operation and maintenance positions for a long time. When they are rusty, the "mentor" role of the voice interaction system naturally needs to be emphasized, and when their abilities improve, the "tool" attribute of the voice interaction system naturally needs to be emphasized.

[0218] Therefore, the experience score obtained after each interaction in this embodiment can be used as the historical score for the next interaction. Preferably, β∈(0,0.01) results in a slower trend in the historical score, which is more in line with the actual growth process of operators.

[0219] By adopting this implementation method, the "capability profile" maintained by the voice interaction system for each operator possesses both anti-disruption and memory capabilities. It can both slowly reflect the operator's true growth or decline trend over time and filter out accidental noise in a single interaction, ensuring the stability and reliability of the "historical score," which is one of the foundations for strategic decision-making.

[0220] In some possible embodiments, the terminology expresses the original characteristics. = ;

[0221] in The raw score representing the accuracy of technical terms, The raw score representing the density of technical terms used; This indicates the preset accuracy weight of technical terms. The default terminology uses density weights;

[0222] The original score of the accuracy of professional terms in the characteristics of terminology expression The calculation includes: calculating the ratio of the number of completely correct words to the total number of words segmented, and mapping it to a score range of [0, 100] as the raw score for the accuracy of technical terms;

[0223] Raw score of the density of professional terminology use in terminology expression characteristics The calculation includes: calculating the sum of the number of completely correct words and the number of fuzzy term words, then calculating the ratio with the total number of word segments, and mapping it to the score range of [0, 100] as the raw score of the density of use of technical terms.

[0224] As an optional implementation of this invention, the calculation method for the raw scores of term expression features is further defined, and the raw scores of term expression features are output through a relatively simple weighted calculation method. It should be understood that the configuration of each weight can be directly based on existing linguistic knowledge, or can be directly customized, and this embodiment does not limit it in this way.

[0225] In some possible embodiments, extracting terminological features from voice commands includes:

[0226] Based on a pre-built standard library of terms in the field of power distribution cabinets, the word segments are matched and identified, and the identified words are divided into completely correct words, fuzzy terminology words and non-terminology words.

[0227] A completely correct term indicates a term that is completely consistent with or is a standard abbreviation of a term in the terminology standard library for the field of electrical distribution cabinets.

[0228] The fuzzy terminology is consistent with the core word root of a completely correct term or has a semantic similarity exceeding a preset threshold.

[0229] The non-terminal terminology does not match any of the terms in the standard terminology library for the distribution cabinet field.

[0230] As an optional implementation of the present invention, the method for extracting terminology expression features is further defined. Of course, this embodiment belongs to the existing mature technical solution, so it will not be described in detail.

[0231] In some possible embodiments, the original language logic features are divided = ;

[0232] in, The original component representing the integrity of the logical structure. Represents the pre-defined logical structure integrity weight;

[0233] The original division of logical structure integrity The calculations include:

[0234] Based on the knowledge base of power distribution cabinet operation and maintenance, we define the core semantic roles that a complete operation instruction should contain, and define the matching rules of the core semantic roles in the syntax tree.

[0235] The speech command text is parsed by a dependency parser to extract the grammatical dependency relations between each word segment, including at least one of subject-predicate relations, verb-object relations, and attributive-head relations, and to obtain a syntax tree.

[0236] Based on the matching rules and syntax tree matching, when the dependency relationship of the core semantic role is missing, a score is deducted according to the preset value, and the score of logical structure integrity is output.

[0237] The scores are mapped to the score range of [0, 100] and used as the raw scores for the accuracy of the technical terms.

[0238] As an optional implementation of the present invention, the extraction method of language logic features and the calculation method of the original scores of language logic features are further defined. Of course, this embodiment belongs to the existing mature technical solution, so it will not be described in detail.

[0239] In some possible embodiments, the original acoustic prosodic features are... = ( ) ;

[0240] in, This indicates the preset speech rate weight. This represents the preset pause density weight, and + =1; The original score represents the speech rate. Indicates the original density of pauses. Indicates the original part of intonation;

[0241] Among them, speech rate original score The calculations include:

[0242] The total duration of voice commands was detected using a VAD model.

[0243] The number of syllables can be detected using a syllable boundary detection model or a short-time energy envelope method.

[0244] The speech rate is obtained by calculating the ratio of the number of syllables to the total duration of the voice command;

[0245] The difference between the speaking speed and the predefined standard speaking speed for instructions in the field of power distribution cabinet operation and maintenance is calculated, and the difference is mapped to the predefined speaking speed scoring rules to obtain the speaking speed score.

[0246] Map the speech rate score to the score range of [0, 100] to obtain the raw speech rate score. ;

[0247] Among them, the speech rate scoring rule means that the smaller the difference, the higher the speech rate score;

[0248] Voice interaction systems do not believe that "the faster the better." They use a VAD model to obtain the effective speech duration, syllable count through syllable detection, and calculate the raw speech rate (syllables per second). Then, this speech rate is compared to a predefined "expert standard speech rate range." This range is derived through statistical analysis of a large number of speech samples from experienced, calm operators. Speech rates falling within this range receive the highest score. Speech rates that are too fast (potentially due to nervousness) or too slow (potentially due to hesitation) will be penalized based on the degree of deviation, ultimately mapping to... ;

[0249] Pause density original score The calculations include:

[0250] Non-speech segments and the duration of each non-speech segment are detected using a syllable boundary detection model or a short-time energy envelope method.

[0251] Calculate the number of non-speech segments whose duration exceeds a preset duration to obtain the number of pauses;

[0252] Calculate the ratio of the number of pauses to the total duration of the voice command, and then inversely map this ratio to a score range of [0, 100] to obtain the raw pause density score. ;

[0253] The voice interaction system detects silent segments with voice energy below a threshold and classifies silent segments lasting longer than a preset duration (e.g., 250ms) as "effective pauses" (brief breaths are not counted). It calculates the number of effective pauses per unit time; fewer pauses result in a higher score (inverse proportionality). ;

[0254] Original intonation The calculations include:

[0255] By extracting the fundamental frequency sequence of the voice commands;

[0256] The trend function of the voice command is obtained by smoothly fitting the fundamental frequency sequence.

[0257] Calculate the residual sequence between the fundamental frequency sequence and the trend function, and further calculate the standard deviation of the residual sequence;

[0258] The standard deviation is mapped to a predefined intonation scoring rule to obtain an intonation score;

[0259] Map the intonation score to the score range of [0, 100] to obtain the raw intonation score. ;

[0260] Among them, the tone scoring rule means that the smaller the standard deviation, the higher the tone score;

[0261] The voice interaction system extracts the fundamental frequency (F0) trajectory of speech, but does not directly evaluate the level of the fundamental frequency. Instead, it first obtains a smooth trend function by fitting, and then calculates the residual standard deviation between the original fundamental frequency and the trend. This standard deviation measures the unnecessary jitter that deviates from the natural intonation trend. A steady and confident intonation has a small residual standard deviation and a high score; a trembling and fluctuating intonation has a large residual standard deviation and a low score.

[0262] As an optional implementation of this invention, although extracting acoustic prosodic features such as speech rate, intonation, and pause density from speech commands is a mature existing technology, the original analysis of acoustic prosodic features... However, the calculation is not a conventional and simple weighted calculation.

[0263] We know that tone of voice, speaking speed, and pause density reflect a person's fluency and confidence, thus indirectly reflecting their expressive ability and professionalism. However, in the high-intensity work environment of power distribution cabinet maintenance, operators are prone to physiological fatigue due to physical exertion, manifesting as a natural slowing of speaking speed, a lower tone, or involuntary pauses due to panting. If these fatigue-affected acoustic characteristics are simply included in the scoring, it will mislead the system, causing it to deviate from its judgment of the operator's true professional competence. Therefore, in the overall experience scoring, the pre-set weighting of acoustic prosodic features is actually... The values ​​are usually relatively low to prevent operator fatigue from excessively interfering with the overall experience score.

[0264] While this ensures that experience scoring doesn't overemphasize the operator's vocal rhythm, it introduces another problem: when operators experience unfamiliarity with the task, nervousness, or hesitation leading to "psychological insecurity," their voices may exhibit trembling (unsteady tone) or hesitation (abnormal pauses). If these signals are ignored due to insufficient weighting, the system will fail to capture this potential risk state, resulting in inadequate guidance or safety prompts that cannot meet the operator's real-time needs.

[0265] To balance the overall experience-based scoring without overemphasizing the operator's acoustic prosody, while still capturing the operator's psychological state when issuing voice commands, this embodiment uses speech rate and pause density as the basic fluency of voice commands, and intonation as the operator's confidence level. The raw score of acoustic prosody features is obtained by weighting speech rate and pause density and then multiplying it by intonation. .

[0266] When operators fail to deliver instructions fluently or confidently, any deficiency will be amplified. This creates a delicate balance in the voice interaction system: on the one hand, because acoustic features have a low weight in the overall experience score, normal physiological fatigue will not have a drastic impact on the assessment of personnel capabilities; on the other hand, the multiplicative model endows the system with extremely high internal sensitivity, capable of amplifying any psychological signal of "disfluency" or "lack of confidence" into a significant one. Low score. This aligns with the reality of operations and maintenance. Only operators who consistently demonstrate confidence and fluent command delivery are truly seasoned experts who require no further instruction.

[0267] Of course, if the operator's voice commands have even the slightest acoustic imperfections, the voice interaction system will tend to provide more cautious guidance or safety confirmation, ensuring that in the power distribution cabinet, an area where safety is of paramount importance, any potential risks caused by the operator's condition receive systematic, traceable, and auditable additional attention.

[0268] Using this implementation method, although the overall change in the raw scores of acoustic prosody features has little impact on the calculation of experience scores, ensuring that normal physiological fatigue will not have a disruptive impact on personnel ability assessment, the multiplicative model endows the raw scores of acoustic prosody features with high internal sensitivity during calculation. This allows any psychological signal of "disfluency" or "lack of confidence" to be amplified into a significantly low score. As a result, the final interaction strategy always tends to increase the degree of guidance and safety prompts. This makes the voice interaction system always lean more towards reliability than efficiency, which is in line with the characteristics of the power distribution cabinet operation and maintenance field, which prioritizes absolute safety. This ensures that any potential risks caused by the operator's condition can receive auditable and traceable additional attention.

[0269] In some other possible embodiments, the original acoustic prosodic features are... = Multiplying the raw scores for intonation, speech rate, and pause density directly amplifies the "weakest link effect," meaning that even a minor flaw in any feature can have a more significant impact on the raw scores for acoustic prosodic features. This more stringent evaluation model is suitable for benchmark distribution cabinet maintenance companies or power units with a high proportion of senior experts and extremely high requirements for the psychological stability and standardized expression of operators. It helps to identify top-tier operators who can maintain comprehensive stability even under high voltage conditions.

[0270] In some possible embodiments, the task matching framework based on expressed intent includes:

[0271] The intended message is matched with a pre-defined operation and maintenance task map, and the corresponding task framework is output.

[0272] The task framework is a structured data object, which includes at least one of the following: task type identifier, task object identifier, task risk level, task default security confirmation level, and task advisory information output dimension set;

[0273] The task type identifier must include at least one of the following: status query, parameter setting, fault diagnosis, status control, and contingency plan execution; the task risk level and the task default safety confirmation level are based on the predefined power safety regulations.

[0274] As an optional implementation of this invention, since the expressed intent represents the operator's expectation and is the primary task performed by the voice interaction system, and considering the strict safety regulations and operational procedures of power distribution cabinet maintenance, a simple intent label (such as "query") is far from sufficient to support the subsequent generation of a safe and efficient personalized interaction strategy. Without a unified and structured task context definition, the strategy decision-making model will struggle to make accurate and procedural decisions. In this embodiment, the specific task framework further defines, in addition to the essential task type identifier (what kind of task, such as status query) and task object identifier (what kind of power component, such as switch 301, this identifier is directly derived from the extraction of semantic slots in the expressed intent), which are indispensable in the voice interaction output and represent the basic tasks contained in the voice commands, other data objects are also defined.

[0275] It should be understood that although the task risk level, the task default safety confirmation level, and the set of task suggestion information output dimensions already exist in the existing voice interaction mechanism in the field of power distribution cabinet operation and maintenance, as mentioned above, the existing voice interaction mechanism is too lengthy for some highly skilled senior operators. In other words, the existing lengthy voice interaction output content often already includes the task risk level, the task default safety confirmation level, and the set of task suggestion information output dimensions related to the task.

[0276] In this embodiment: the task risk level is based on the predefined power safety regulations and is used to distinguish the risks of different task type identifiers and task object identifiers. For example, high-risk power components may include main switch, transformer, fuse, main line, etc., high-risk task types include status control, parameter setting, etc., low-risk power components include branch switches, indicator lights, ambient temperature sensors, etc., low-risk task type identifiers include status query, etc. The default security confirmation level for a task can be mapped based on different task risk levels and existing national power safety regulations. This indicates that some intended expressions require security verification. For example, task type identifiers representing queries generally do not require security verification, while task type identifiers representing status control may require stronger security verification. If high-risk task objects are involved in high-risk task types, extremely strong security verification processes are required (for example, before an operator performs maintenance operations, they will issue the instruction "disconnect the xx main switch." Before the voice interaction system automatically disconnects the xx switch, it will first remind the operator to confirm that they have put on protective equipment, or verify the operator's identity. After disconnecting the xx switch, it will automatically broadcast the relevant operating procedures that comply with national power safety regulations for the operator to confirm and verify). The suggested information output dimension set indicates what information the task typically requires (such as real-time values, historical curves, alarm lists).

[0277] It should be noted that the above-mentioned data objects can be directly matched with the operation and maintenance task map based on the threshold in this embodiment. This is already quite mature in the existing technology. Of course, the task map can also be customized by those skilled in the art based on power knowledge. This embodiment does not limit this.

[0278] By adopting this implementation method, the objects to be affected by experience scoring are constructed, which proves that all subsequent personalized interaction strategies of the system will be intelligently adjusted within the framework allowed by security procedures and within the reasonable scope of business logic, reserving a flexible, computable and traceable decision space for the strategy decision-making model based on experience scoring.

[0279] In some possible embodiments, in step S20: the input to the strategy decision model is a feature vector formed by concatenating experience scores and task frameworks; the model output is a personalized interaction strategy parameter set, including information detail levels. , guidance intervention density level Security Confirmation Level and broadcast style identifier ;

[0280] Among them, information detail level Used to control the granularity of information retrieved from the database and the length of additional explanations; guides the level of intervention density. Used to determine the number of nodes that proactively offer suggestions for the next step in a single conversation; security confirmation level. Used to determine the complexity of the independent verification steps that must be passed before performing operations involving state changes within the task framework; broadcast style identifier. Used to index a set of pre-configured speech synthesis parameters, including speech rate, intonation, and sentence templates;

[0281] Methods for constructing the input feature vector of a strategy decision-making model include:

[0282] Experience ratings were used as the first continuous feature.

[0283] The task type identifier, criticality level, default security confirmation level, and information output dimension set in the task framework are encoded and transformed into the second feature group.

[0284] The first continuous feature and the second feature group are concatenated in a predetermined order to form a feature vector.

[0285] As an optional implementation of the present invention, this embodiment further defines the working mechanism of the strategy decision-making model, namely, as a key conversion mechanism from understanding "people" and "things" to generating specific "action instructions". After the voice interaction system has learned the operator's experience score and expression intention, it needs to transform the abstract understanding into executable interaction parameters. As mentioned in the previous embodiment, the task framework represents what aspects of voice interaction output content the voice interaction system can output, while the experience score determines what aspects of content to output, as well as the guidance and prompting level contained in each aspect. That is, the core task of the strategy decision-making model is to decide: what details to tell the operator, whether to actively prompt the next step, how many layers of safety confirmation to be performed, and what tone to use for the announcement, in order to truly achieve operational efficiency and safety assurance for operators of all skill levels, from novices to experienced users.

[0286] This embodiment transforms the complex art of interaction into a computable and optimizable engineering problem by defining a standardized input interface and a quantifiable set of output strategies for the model.

[0287] Specifically, firstly, experience scores representing "human" abilities are used as the core continuous numerical features. Then, the task framework representing "tasks" is decomposed and encoded: task types (such as "status queries") are converted into one-hot encodings, criticality levels and default security levels are used as ordinal features, and suggested information dimension sets (such as "historical curves, real-time values") are converted into multi-hot encodings. Finally, these encoded features are concatenated in a predetermined order to form a unified feature vector. This process "translates" the business scenario into a "language" that the machine learning model can understand.

[0288] Subsequently, the model's task is not to generate the final natural language, but to output a set of "control knobs" that govern the final generation, namely a set of personalized interaction strategy parameters, including: information detail level. Used to determine the coarseness of data retrieved from the database and the length of the interpreted text; guiding the intervention density level. Used to control the frequency with which a voice interaction system actively interrupts conversations and provides suggestions; security confirmation level. : Used to specify the complexity of the verification steps that must be passed before performing hazardous operations; broadcast style identifier : Used to directly call a set of preset speech synthesis parameters to achieve a style switch from "calm and professional" to "patient guidance".

[0289] Finally, the model makes a decision: the strategy decision model receives the feature vector, calculates and outputs the optimal set of strategy parameters through internal complex nonlinear mapping, and the subsequent content generation module will strictly follow this parameter set to arrange the final voice interaction output content.

[0290] By adopting this implementation method, the strategic decision-making process is made transparent, explainable, and optimizable, ensuring that the AI ​​model has reliable auditability when applied to the field of power distribution cabinet operation and maintenance with stringent safety requirements, thus providing a more trustworthy basis for personalized interaction in this field.

[0291] In some possible embodiments, the policy decision model is a gradient boosting tree-based classification model trained using historical interaction logs;

[0292] Each training sample in the historical interaction log contains a feature vector that is labeled as input and a set of personalized interaction strategy parameters that is labeled as output.

[0293] The feature vector is the first array formed by concatenating the vectors obtained by encoding the experience score and the task framework;

[0294] The personalized interaction strategy parameter set includes information detail levels , guidance intervention density level Security Confirmation Level and broadcast style identifier The second array is formed.

[0295] As an optional implementation of the present invention, this embodiment further defines the training method of the aforementioned strategy decision-making model. The strategy decision-making model of this technical solution originates from learning from a large amount of historical successful interaction experience. Compared with manually writing fixed rules for massive combinations of scenarios such as "novice + critical task," this data-driven approach is easier to build, maintain, and extend. This embodiment uses a supervised learning classification model based on gradient boosting trees to achieve this process. This method is a mature technology in the field of machine learning, capable of effectively learning complex mapping relationships from data. Its specific training process is well known to those skilled in the art and will not be elaborated here.

[0296] In some possible embodiments, the annotation method for personalized interaction strategy parameter sets includes: automatically or manually determining the personalized strategy parameter set of an interaction as a valid label based on the operator's positive confirmation feedback, task completion efficiency, and safety records in a historical interaction, and storing it in the training sample library after verification.

[0297] As an optional implementation of this invention, this embodiment provides an efficient solution to the problem of obtaining high-quality "policy labels" required for supervised learning model training. In the field of power distribution cabinet operation and maintenance, relying entirely on manual labeling of the optimal policy for massive historical interaction data, while providing stronger subsequent auditability, is inefficient and cannot meet the needs of rapid model iteration.

[0298] Therefore, this embodiment introduces an automatic annotation mechanism based on interaction result feedback. This mechanism can automatically determine the effectiveness of the strategy used in a single interaction based on multi-dimensional results such as positive confirmation feedback from the operator, task completion efficiency, and safety records, and annotate it as a training sample, thereby significantly improving the speed and scale of training sample accumulation. Constructing a model training sample library using automated methods falls within the scope of existing mature technologies, and its specific implementation will not be elaborated upon in this embodiment.

[0299] In some possible embodiments, when historical interaction logs are lacking in the early stages of the deployment of the strategy decision model, a rule-based strategy decision engine is used to directly output personalized interaction strategy parameters by looking up tables based on experience scores and task frameworks.

[0300] Once the strategy decision model has accumulated a preset number of interaction logs, the training of the machine learning-based strategy decision model will be initiated and gradually replace the strategy decision engine.

[0301] As an optional implementation of the present invention, this embodiment considers the "cold start" problem of the strategy decision-making model in the early stages of a project. Given that basic voice interaction functionality is already common in the current field of power distribution cabinet operation and maintenance, this system can adopt a gradual deployment strategy.

[0302] In the initial deployment phase, when historical interaction logs are lacking, the system can initially activate a rule-based policy decision engine. This engine, based on experience scoring and the task framework, directly outputs a set of conservative and safe default policy parameters using a lookup table method, ensuring basic system usability. Subsequently, the system automatically accumulates interaction logs during operation and generates labeled samples through the aforementioned mechanism. Once the sample size reaches a preset threshold, training of a machine learning-based policy decision model can begin, and the trained intelligent model will gradually replace the initial rule engine.

[0303] This deployment approach of "first ensuring deployment with rules, then driving evolution with data" is a common and mature engineering practice in many AI application fields, and will not be elaborated on in detail in this embodiment.

[0304] This implementation method greatly enhances the engineering practicality and resilience of the technical solution, ensuring that it can be smoothly integrated into the existing operation and maintenance environment and continuously optimize itself.

[0305] Example 3:

[0306] Please see Figure 2 The present invention also provides the following technical solutions:

[0307] The intelligent power distribution cabinet AI voice interaction system is used to implement the methods of Examples 1-2. The system includes:

[0308] The voice processing module is used to receive voice commands from operators and parse the operators' intended meaning.

[0309] The experience score generation module is used to extract semantic and voiceprint features from voice commands and generate an experience score for the operator.

[0310] The strategy decision module is used to match the task framework according to the expressed intent and combine dynamic experience scoring to calculate personalized interaction strategy parameters through a pre-set strategy decision model.

[0311] The content generation and output module is used to generate and output voice interaction content based on personalized interaction strategy parameters and expressive intent.

[0312] Example 4:

[0313] The present invention also provides the following technical solutions:

[0314] Computer program products, including computer programs, which, when executed by a processor, implement the methods described above.

[0315] Example 5:

[0316] The present invention also provides the following technical solutions:

[0317] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0318] Example 6:

[0319] The present invention also provides the following technical solutions:

[0320] AI intelligent power distribution cabinet, including cabinet, power components installed inside the cabinet and the above-mentioned system;

[0321] The system is connected to at least the status sensors of the power components to acquire information from each power component. The system also includes a human-machine interface panel, a voice acquisition module, and a broadcasting unit. The voice acquisition module and the broadcasting unit are both located on the human-machine interface panel of the cabinet.

[0322] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. AI voice interaction method for intelligent power distribution cabinets, wherein, include: Receive voice commands from operators and analyze the operators' expressive intentions and experience ratings; Based on the expressed intent, a task framework is matched, and personalized interaction strategy parameters are calculated through a pre-set strategy decision model in conjunction with experience scoring. Based on the personalized interaction strategy parameters and the expressive intent, voice interaction output content is generated.

2. The method according to claim 1, wherein, The analysis of the operator's experience score includes: Extract semantic features from the voice commands, including terminology features, linguistic logic features, and acoustic prosodic features; Based on the semantic features, an experience score for the operator is generated.

3. The method according to claim 2, wherein, The experience score for the operator involved in the generation process includes: Extract the voiceprint features from the voice commands; Based on the voiceprint feature matching database, the historical score of the operator is output; An experience score for the operator is generated based on the historical scores, terminological expression features, linguistic logic features, and acoustic prosodic features.

4. The method according to claim 2 or 3, wherein, The terminology expression characteristics include at least one of the following: the accuracy of the terminology and the density of terminology usage. The language logic features include at least the integrity of the instruction logic structure; The acoustic prosodic features include at least one of speech rate, intonation, and pause density.

5. The method according to claim 4, wherein, The experience score of the operator is generated. include: Calculate this score ; Calculate experience score ; in, This indicates the score for this test; This indicates the preset historical scoring weights; This indicates the predefined terminology expression feature weights; This indicates the pre-defined weights of the language logic features; This represents the preset acoustic prosodic feature weights, and + + =1; The original components represent the expressive features of the terminology. Representing the original components of linguistic logical features, Represents the original components of acoustic prosodic features; Indicates the benchmark score for the expressive characteristics of the terminology; Indicates the benchmark score for the logical features of the language; The baseline score represents the acoustic prosodic features.

6. The method according to claim 1, wherein, The framework for matching tasks based on expressed intent includes: The expressed intent is matched with a pre-defined operation and maintenance task map to output the corresponding task framework; The task framework is a structured data object, which includes at least one of the following: task type identifier, task object identifier, task risk level, task default security confirmation level, and task advisory information output dimension set.

7. The method according to claim 6, wherein, The task type identifiers include: status query, parameter setting, fault diagnosis, status control, and contingency plan execution; the task risk level and the default task safety confirmation level are predefined based on power safety regulations.

8. The method according to claim 6, wherein, The input to the strategy decision model is a feature vector formed by concatenating the experience score with the task framework; The model output is a set of personalized interaction strategy parameters, which includes information detail levels. , guidance intervention density level Security Confirmation Level and broadcast style identifier .

9. The method according to claim 8, wherein, The level of information detail Used to control the granularity of information retrieved from the database and the length of additional explanations; The guided intervention density level Used to determine the number of nodes that proactively offer suggestions for the next action in a single conversation; The security confirmation level Used to determine the complexity of the independent verification steps that must be passed before performing operations involving state changes in the task framework; The broadcast style identifier Used to index a set of pre-configured speech synthesis parameters, including speech rate, intonation, and sentence templates.

10. The method according to claim 8, wherein, The strategy decision-making model is a classification model trained using historical interaction logs; Each training sample in the historical interaction log contains a feature vector input and a personalized interaction strategy parameter set label; the feature vector input is a feature vector concatenated from experience scoring and task framework, and the personalized interaction strategy parameter set label is a feature vector composed of information detail level. , guidance intervention density level Security Confirmation Level and broadcast style identifier A collection pieced together.

Images