Power grid user voice consultation processing method based on intelligent robot

Abstract

The invention provides a power grid user voice consultation processing method based on an intelligent robot, and the method comprises the steps: carrying out the recognition and emotion analysis of a user voice, and obtaining a text content and an emotional state; driving intention recognition in combination with emotion information, and determining a business intention and an emergency score; calling a power grid service system to query data, and performing intelligent reasoning completion based on the knowledge graph and the urgency degree weighted graph neural network when a field is missing; according to the structured result, the field confidence coefficient and the multi-round dialogue context, a natural language reply is dynamically generated; after interaction is finished, whether personalized service recommendation is triggered or not is judged in combination with a user portrait, and service quality is evaluated based on dialogue continuity, emotion change and field confidence for strategy optimization. According to the invention, the full-link closed loop of emotion perception, intelligent reasoning, multi-round interaction and self-optimization is realized, and the accuracy, fluency and user experience of the power grid voice consultation service are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of power service technology, and in particular relates to a method for processing voice consultations for power grid users based on intelligent robots. Background Technology

[0002] With the maturity of artificial intelligence technology and the advancement of digital transformation in power services, the service model of power grid companies is gradually shifting from traditional human customer service to intelligent voice interaction systems. Users' expectations for services such as power consultation, payment, and electricity information inquiry are no longer limited to simply receiving answers, but rather focus on the efficiency, accuracy, and experience of the interaction. However, existing voice customer service systems generally suffer from insufficient comprehension and rigid response patterns, making it difficult to handle complex and interconnected multi-turn question-and-answer scenarios in power business. In scenarios such as power outages, electricity price adjustments, and load anomalies, users often experience anxiety or confusion, which current systems cannot identify and respond to, resulting in responses lacking empathy and prioritization. Simultaneously, traditional systems often experience data delays or missing information when interfacing with the business backend, failing to effectively reason and complete data when interface data is incomplete, leading to service interruptions. Furthermore, most existing solutions can only complete a single-turn dialogue, lacking the ability to understand context and state management mechanisms, forcing users to repeatedly express their needs, resulting in a fragmented overall experience. A more prominent problem is the lack of personalization and self-optimization capabilities in current systems, making it impossible to dynamically learn or adjust service strategies based on user history and system interaction results. To address the aforementioned issues, the key direction for the development of intelligent power voice consultation systems lies in how to recognize user voice and emotions, understand their true intentions, integrate real-time business data for intelligent reasoning, and continuously optimize response quality through multi-round interactions. Summary of the Invention

[0003] The purpose of this invention is to design a voice consultation processing method for power grid users based on intelligent robots, which can maintain high interpretability and high reliability in power grid business scenarios with strong emotions, missing information, and high complexity, thereby improving consultation accuracy, multi-turn dialogue fluency, and overall service intelligence.

[0004] To achieve the above objectives, the present invention provides a method for processing voice consultation for power grid users based on an intelligent robot, the method comprising:

[0005] Receive power business consultation requests from users via voice input; perform speech recognition on the power business consultation requests to obtain the corresponding text content, and simultaneously perform sentiment analysis to obtain the user's sentiment category and emotional intensity;

[0006] By combining the text content with the emotion category and emotion intensity, emotion-driven intent recognition is performed to determine the user's business intent and corresponding urgency score.

[0007] According to the business intent, the power grid business system interface is invoked to query data. When there are missing fields in the query results, a graph neural network is used to infer and complete the missing fields based on the pre-built power grid business knowledge graph and the urgency score, generating a structured response result containing the original fields and the completed fields, and assigning confidence to each field.

[0008] Based on the structured response results, field confidence levels, and multi-turn dialogue context states, a response template is dynamically selected and a natural language response is generated.

[0009] After completing the current round of interaction, the system combines user profile, current dialogue status, and historical behavior to determine whether to trigger personalized service recommendations. It also conducts a comprehensive evaluation of the quality of this round of interaction based on dialogue coherence, changes in user emotions, and confidence levels of response fields, generating a quality score for optimization of subsequent service strategies.

[0010] Furthermore, the sentiment analysis adopts a Chinese sentiment classification model based on the BERT architecture. The Chinese sentiment classification model is fine-tuned on the labeled corpus of power customer service scenarios, and outputs sentiment categories including anxiety, doubt, pleasure or neutrality and sentiment intensity values ​​ranging from 0 to 1.

[0011] Furthermore, the intent recognition adopts a Chinese business intent recognition model based on the RoBERTa architecture. The Chinese business intent recognition model integrates emotion modulation embedding vectors in the input embedding stage and enhances the attention weight of emotion-related words according to the emotion intensity in the multi-head attention mechanism, so as to improve the accuracy of intent recognition for ambiguous or emotional expressions.

[0012] Furthermore, the urgency score comprehensively considers semantic sentence vectors, emotional intensity, and semantic certainty, and is used to regulate the priority and strategy of subsequent business queries and inferences.

[0013] Furthermore, the power grid business knowledge graph is constructed with business fields as nodes and field dependencies as edges. During the reasoning process for missing fields, the reasoning intensity of the graph neural network is dynamically adjusted according to the urgency score, so that high-urgency requests can still generate preliminary responses when the data is incomplete.

[0014] Furthermore, each field in the structured response result is accompanied by a confidence level, which reflects whether the field originates directly from the business system or is completed through reasoning, and is used for credibility control in subsequent response generation.

[0015] Furthermore, the process of generating a natural language response based on the structured response result, field confidence, and multi-turn dialogue context state includes: encoding up to six rounds of historical dialogue content through a gated recurrent unit and a self-attention mechanism to generate a current dialogue state vector, and identifying whether the user has shifted the topic based on the current dialogue state vector in order to dynamically adjust the dialogue focus and response strategy.

[0016] Furthermore, the natural language response is generated using a combination of templates and neural generation. Candidate templates are filtered according to intent categories, and the optimal template is selected based on a weighted score of field confidence. Low-confidence fields are handled with fuzzy descriptions or omissions.

[0017] Furthermore, the personalized service recommendation is based on a joint judgment of the user profile vector and the current dialogue state vector. The user profile includes recent consultation frequency, intent distribution, sentiment statistics and service preferences, and the recommendation trigger is suppressed when the user is in a high anxiety state.

[0018] Furthermore, the comprehensive evaluation of interaction quality is achieved by calculating the similarity between the current and previous dialogue states to represent dialogue coherence, combining the changing trend of user emotion intensity between the current and previous rounds, and the weighted average of the confidence scores of each field in the current response to generate a quality score ranging from 0 to 1; when the quality score is lower than the threshold, the current dialogue is marked as an optimization candidate sample for offline iterative updates of the system strategy.

[0019] The beneficial technical effects of the present invention are at least as follows:

[0020] To address the aforementioned issues, this invention provides a method for processing voice consultation for power grid users based on an intelligent robot. It constructs a complete technical system encompassing speech recognition, sentiment analysis, intent understanding, business reasoning, multi-turn interaction, and personalized optimization. The method uses sentiment recognition to drive the intent recognition model, adjusting semantic attention weights based on emotion category and intensity to achieve accurate understanding of ambiguous semantics and emotional expressions. An urgency-weighted graph neural reasoning mechanism connects to the power business system, completing missing field completion reasoning by integrating business knowledge graphs and user urgency parameters. A confidence-weighted template response mechanism manages multi-turn dialogues, enabling the robot to dynamically select language expression methods based on the reliability of business data, thus outputting reasonable responses even with incomplete data. Simultaneously, a personalized recommendation model and quality self-evaluation mechanism based on dialogue state and historical behavior are introduced, automatically determining whether to provide extended services and optimizing strategy parameters after the interaction. Through these designs, a leap from "understanding speech" to "understanding users" is achieved, maintaining high interpretability and reliability in emotion-driven, information-deficient, and highly complex power grid business scenarios, significantly improving consultation accuracy, multi-turn dialogue fluency, and overall service intelligence. Attached Figure Description

[0021] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained from the following drawings without any effort.

[0022] Figure 1 This is a flowchart of a power grid user voice consultation processing method based on an intelligent robot, according to the present invention. Detailed Implementation

[0023] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0024] In one or more embodiments, such as Figure 1 As shown, a method for processing voice consultations for power grid users based on an intelligent robot is disclosed, the method comprising the following:

[0025] S1: Receive a power business consultation request input by the user via voice; perform voice recognition on the power business consultation request to obtain the corresponding text content, and simultaneously perform sentiment analysis to obtain the user's sentiment category and emotional intensity;

[0026] Specifically, in this step, the goal is to convert the voice input of power grid users into structured text that can be processed by the system and to identify their emotional state to support subsequent intent understanding and business processing. User voice input scenarios include initiating voice queries about electricity-related issues via telephone, mobile app voice entry, or home terminal. The audio sampling format is 16kHz, 16bit, and the maximum duration is no more than 30 seconds. The audio signal is first processed by MFCC (MelFrequencyCepstralCoefficient) to extract acoustic features and generate a two-dimensional feature tensor. ,in For frame number, The dimension is 40 for each frame.

[0027] Feature tensor The input is fed into the speech recognition module, which employs an end-to-end speech recognition model based on the Transformer architecture. The model's front end consists of a two-layer convolutional structure for local feature extraction, while the middle section comprises a 12-layer self-attention encoder, each layer containing multi-head attention and a feedforward network, with a hidden unit dimension of 512. This recognition system is pre-trained on a general Mandarin Chinese corpus and fine-tuned using labeled speech data from power customer service scenarios (including scenarios related to electricity bills, power outages, and faults). The model output is a text sequence. For example, a user's voice message "What should I do if my power is out?" is transcribed into... "What should I do if my power goes out?"

[0028] text The data is then fed into an emotion recognition model for sentiment analysis. This model is a Chinese sentiment classification network based on BERT-Base, using a 12-layer Transformer structure with pre-trained parameters frozen and fine-tuned only on industry-specific sentiment corpora. To enable the model to better identify typical emotional states in power grid service scenarios, the training set includes four categories of labels: "anxiety," "pleasure," "doubt," and "neutrality," with a total of approximately 80,000 samples. The model input is the text after speech recognition. The output includes emotion categories. And emotional intensity value The latter is obtained by appending a fully connected regression layer to the [CLS] vector output by BERT, with a numerical range of [value missing]. The closer to 1, the stronger the emotion.

[0029] The overall relationship of the emotion recognition process can be represented as:

[0030] ;

[0031] in, This represents a joint model consisting of BERT as the encoder, followed by an emotion classifier and an emotion intensity regressor. Text data output by the speech recognition module; For the identified emotional labels, such as "anxiety" or "doubt"; The intensity of this emotion is defined as follows, with a value range of [value range missing]. .

[0032] For example, when a user says, "My power suddenly went out, when will it be restored?", the text is obtained through speech recognition. "My power suddenly went out, when will it be restored?" The sentiment model identifies this as... "Anxiety" corresponds to the intensity of emotions. This result will be used directly in subsequent intent identification and prioritization to determine the urgency of the request.

[0033] The final output consists of two variables: the text content. With emotional state .

[0034] S2: Combining the text content with the emotion category and emotion intensity, perform emotion-driven intent recognition to determine the user's business intent and corresponding urgency score;

[0035] Specifically, in this step, the system's task is to process the speech-text output from step one. With emotional state Joint understanding is used to identify the business intent expressed in the user's current voice inquiry. And calculate the urgency score of the intention. The result of this processing will directly determine the business type and query order of the next system query, thus playing the role of "instruction translation" and "task scheduling priority setting".

[0036] The voice inquiries from power grid users are highly specialized (containing terms such as "orderly electricity use" and "tiered electricity pricing"), highly ambiguous (as questions like "Why is there no electricity again?" lack clear instructions), and emotion-driven (for example, users may experience anxiety after a power outage at night but their speech may be unclear). Therefore, it is necessary to construct an emotion-driven intent recognition mechanism to overcome the limitations of traditional static classification models in understanding incomplete expressions.

[0037] The input for this step comes from the output of the previous step, i.e., the text. and emotional state .in, It is the user's query text after speech recognition. It is an emotion category (such as anxiety, doubt). It is its strength rating. First, the text The input is fed into a Chinese business intent recognition network based on the RoBERTa architecture, which consists of a 12-layer Transformer encoder, with each layer containing 768-dimensional hidden units. In the input embedding, in addition to the original token embedding and location embedding, a sentiment modulation embedding vector is also introduced. This is used to enhance the model's ability to perceive emotional signals.

[0038] The embedding vector It is based on emotion category The corresponding vector table is obtained, for example, the preset vector for "anxiety". This is achieved through a lookup table; and it is added to the text embedding at the encoder input, allowing the model to obtain the emotional context during the encoding stage. Unlike the traditional BERT model, which relies entirely on language sequences, this mechanism simulates the "perceive emotion first, then identify the problem" strategy in human customer service processing, improving the model's understanding of non-instructional queries (such as "What's going on?").

[0039] Emotional intensity This involves adjusting the weights of words participating in the multi-head attention mechanism. Traditional Transformer attention weights are obtained through dot product; in this approach, we incorporate a sentiment modifier to strengthen the weight of words with strong emotional expression in attention allocation. Let the attention weights be... , indicating the first The keyword to follow The degree of each word is now adjusted as follows:

[0040] ;

[0041] in, and For the first The first query and the first A key vector, Scaling factor For the first A flag value (0 or 1) indicating whether a word is associated with a sentiment word. This formula introduces additional terms. The function is as follows: if a word is marked as an emotion-related word (such as "annoying" or "constantly out of power"), its attention value is automatically increased when the user's emotional intensity is high. This makes the model pay more attention to these words with subjective emotional connotations, thereby improving the accuracy of intent judgment. This mechanism is particularly suitable for inputs that are dominated by high emotion but have ambiguous semantics, such as "What should I do if there is no power again tonight?", which lacks business keywords but has extremely strong emotional words.

[0042] After encoding, the model uses a fully connected layer and a softmax classifier to identify the intent category. Examples of possible requests include "repair request," "electricity price inquiry," and "balance inquiry." The system also needs to output the urgency level of the request. This information is used for subsequent service decision-making. Unlike methods that rely solely on emotional intensity, this approach combines the importance of language content with emotional intensity to construct an urgency scoring model as follows:

[0043] ;

[0044] in: The sentence vector output by the Transformer; For linear layer parameters; The emotional enhancement coefficient (e.g., 0.3) moderates the emotional impact. The regularization coefficient is used to control the model from being overconfident when the input semantic expression is ambiguous, thus avoiding misjudgment. The term represents a penalty applied to the square of the magnitude of the sentence vector, constraining the model to maintain a conservative output when the expression is ambiguous.

[0045] This combination method ensures It can comprehensively consider the user's expression, emotional intensity, and semantic certainty, making it more reasonable than traditional single sigmoid or regression outputs, and better able to handle user inquiries in real power grid scenarios where they are "unable to explain but very anxious".

[0046] Finally, this step outputs two variables: : Indicates the category of the power service intent corresponding to the current user input; The urgency score of the current intention, ranging from... This will serve as the basis for prioritizing service responses in subsequent steps.

[0047] This step deeply integrates the common problems of "non-command language" and "highly emotion-driven" in power grid voice consultation with natural language modeling, breaking through the limitation of traditional semantic understanding models that can only handle text with clear structure. In particular, it introduces conditioning and confidence control terms designed for emotion-driven scenarios into the multi-attention mechanism and output scoring model, realizing three-dimensional joint modeling of semantics, emotion, and decision.

[0048] S3: According to the business intent, call the power grid business system interface to query data, and when there are missing fields in the query results, use a graph neural network to reason and complete the missing fields based on the pre-built power grid business knowledge graph and the urgency score, generate a structured response result containing the original fields and the completed fields, and assign confidence to each field;

[0049] Specifically, this step aims to address the business intent output in step 2. With urgency score The system queries the power grid business system and, if necessary, uses structured reasoning mechanisms to supplement or infer response information. This step connects the results of natural language understanding with actual business data, ensuring the system not only understands user questions but also provides accurate, timely, and actionable feedback based on the actual state of the power system. In power grid voice consultation scenarios, user questions often have strong business coupling, complex data dependencies, and high real-time requirements. For example, the question "My power suddenly went out, when will it be restored?" involves multiple system interfaces such as outage area identification, load dispatch status, and repair work order progress. When business system data is incomplete or response times are long, the system must possess a certain degree of "intelligent reasoning" capability to supplement potentially missing information and generate a reasonable answer. Furthermore, the output of the previous step... The urgency score not only indicates the urgency of the user's needs, but also influences the scheduling priority and strategy selection of queries and inferences in this step.

[0050] The input for this step is the output of step 2. (Business Intent) and (Urgency score). First, Mapped to a predefined set of power business query interfaces A sub-interface in This mapping is determined by a combination of static rules and dynamic parameters. For example:

[0051] "Power Outage Restoration Time Inquiry" → ;

[0052] "Abnormal electricity bill" → ;

[0053] This API call requires a parameter set. This mainly includes user ID, timestamp, geolocation, device number, etc., which comes from user login status, historical interaction records, or terminal registration data, along with voice and text. This avoids ambiguity regarding the source of information. The call results constitute a preliminary response result vector. This vector structure is defined according to the business interface, and its dimensions are not fixed. For example, for electricity bill anomaly issues, It can include fields such as "This Month's Electricity Bill," "Historical Average," and "Abnormal Alarm Flags." Considering that power grid business data systems often suffer from issues such as missing data, delayed interface responses, and uncertain status fields, especially in emergency scenarios (such as peak power outage periods and system switching times), the system may not be able to immediately obtain complete data. Therefore, we propose an urgency-weighted graph neural completion reasoning mechanism, combined with a knowledge graph structure. With urgency score Perform reasoning to complete missing fields.

[0054] Knowledge Graph The construction is based on a knowledge base of power grid business domains, and each queryable field is modeled as a node in a graph. Dependency modeling is done as edges For example, "Power outage restoration time" depends on fields such as "distribution network dispatch status," "work order dispatch time," and "user's line number." For the initial response... For missing fields in the graph, the inference module extracts their representation in the graph. In Rank-adjacent node graph Node embedding updates are then performed on this subgraph using GCN (Graph Convolutional Network).

[0055] To adjust the strength of the reasoning based on the user's urgency level, reasoning completion is performed using the following formula:

[0056] ;

[0057] in: For the first result in the current query The original value of each field is considered as 0 if it is empty; The completed value after reasoning; For nodes The set of neighbors; Embed the features of neighboring nodes. This represents the representation after graph convolution update; The attention weights between nodes are calculated by the structural attention mechanism; The urgency score output from step 2, range , used as an inference amplification factor, allows the inference mechanism to fill in values ​​more boldly in emergency situations.

[0058] This formula lies in... Controlling the strength of reasoning: for example, when When the user's need is extremely urgent, the system can prioritize providing an estimated value and offering a response even when information is incomplete, thus avoiding a cold reception; while when... In such cases, the system will remain conservative, only outputting inference results under high-confidence conditions. This mechanism ensures an initial response is provided even in situations of "intense emotion but insufficient information," thereby improving system stability and service integrity.

[0059] Furthermore, to avoid bias and overfitting in the inference results, a regularization constraint based on structural complexity is introduced to control the entropy distribution of the attention distribution of the inference nodes:

[0060] ;

[0061] This is used to penalize overly scattered attention distribution during the inference training phase, thereby improving the system's generalization ability on sparse graphs. This is a hyperparameter that controls the regularization strength, typically set between 0.01 and 0.1. By introducing this parameter, the system can suppress "arbitrary reasoning" when the dependency path is ambiguous, enhance its ability to focus on critical paths, and improve its robustness in handling uncertain inputs.

[0062] The final output consists of two results:

[0063] Structured query results include fields returned by the original business interface and fields completed after inference by a graph neural network;

[0064] : Field confidence vector, where each dimension represents whether the field value is a true interface return or inference completion. The closer the value is to 1, the higher the confidence.

[0065] This step addresses the issues of information gaps, complex states, and "response discontinuities" in power grid scenarios by introducing mechanisms such as urgency factors, regular entropy constraints, and structural confidence modeling into structural graph reasoning. This not only improves the response rate and accuracy of robot services but also provides solid structural input support for subsequent speech generation and multi-turn dialogue.

[0066] S4: Based on the structured response results, field confidence levels, and multi-turn dialogue context states, dynamically select a response template and generate a natural language response;

[0067] Specifically, the goal of this step is to output the structured query results in step 3. and field confidence vector Based on this, and combined with the user's historical multi-turn dialogue content, natural language responses are dynamically generated, and the entire dialogue session state is managed. This enables the power grid voice robot to continuously understand the context, follow up on questions in a timely manner, and naturally transition between topics. This is especially crucial when handling continuous, multi-round questions such as "Why is the electricity bill so high?", "Can I pay in installments?", and "Are there any discounts?". Unlike the single-round responses of general customer service systems, the voice consultation scenarios of power grid users are characterized by "strong data-driven approach", "deep logical nesting", and "vague expression". Therefore, in the process of multi-round dialogue, the system must not only be able to access business data in real time, but also be able to identify the boundaries of the context and topics, maintain the focus of the dialogue, avoid repetitive and redundant answers, and guide the user to complete a complete consultation process when necessary.

[0068] This step receives the output from the previous step. and .in, It is a structured business result, in the form of key-value pairs (such as {"Monthly Electricity Bill": 237.5, "Last Month's Average": 180.1, "Exceeded Threshold": True}), each field comes from the API return of the business system or inference from the intelligent reasoning module. It is a with A structured vector representing the confidence level of each field value. ,exist Within the range. For example, when the power outage recovery time is directly returned by the business system, its corresponding... The value is close to 1; if it is inferred from a graph neural network inference model in the case of missing data, then... Lower, such as 0.55.

[0069] First, the system needs to construct a dialogue state vector based on the current dialogue context. We use a contextual representation encoded by user input, system response from the previous round, and records of business fields from the previous round as input. The input sequence passes through a Gated Recurrent Unit (GRU) to generate hidden states, and then a self-attention mechanism assigns weights to information from different rounds to obtain the current dialogue state. It will participate in the generation of the response content in the current round. All dialogue content (user questions, system replies) is embedded through the Transformer encoder. The information in each round is uniformly input as [round identifier + text], and a maximum of 6 round windows are maintained to ensure that the context is not lost but to avoid the distortion caused by compression of long text.

[0070] Next, to generate natural language responses, we employ a strategy combining template generation and neural language modeling. The system predefines multiple response templates, each associated with an intent category. This corresponds to and includes several field slots to be filled. For example, a template for the intent "Electricity Bill Explanation" would be: "Hello, your electricity bill for this month is {curr_fee} yuan, which is {change_type} compared to the average of {avg_fee} yuan last month. The reasons may include {cause}." These fields come from... The value in the system. The system first determines the value based on the current intent. and dialogue status Filter candidate template set Then, the optimal template is selected through a confidence scoring mechanism.

[0071] The scoring function is as follows, used to determine which template response to select in the current round:

[0072] ;

[0073] in, Indicates the first One response template; This is the set of fields that need to be filled in the template; It is a field Content weighting (e.g., "Electricity bill value" has a weight of 0.7, and "causal speculation" has a weight of 0.4). The confidence value for the field comes from... ; This function indexes fields with confidence levels below a set threshold (e.g., 0.6). An innovative feature has been added to this function: The squared penalty, a non-linear control, can rapidly amplify the penalty when the confidence level is low, preventing low-confidence fields from being forcibly filled into the template and generating misleading content for users.

[0074] For example, if a user asks, "Why is my electricity bill higher this month?", the system will... The current intent is determined to be "explanation of abnormal electricity bill," and the query module returns... The data includes "This month's electricity bill = 253", "Last month's average = 180", "Abnormal = True", and "Suggested reason = Frequent air conditioner use", with confidence levels of 0.95, 0.92, 0.9, and 0.55 respectively. In this case, the system will determine that the confidence level of the "Suggested reason" field is too low and is not suitable as a direct answer. Therefore, the template selection function will favor templates that do not contain this field or can express it using vague language, such as "According to system data, your electricity consumption has increased this month. We suggest you pay attention to the use of high-energy-consuming equipment such as air conditioners," rather than directly stating "The reason is excessive air conditioner use."

[0075] After generating a response, to maintain consistency and coherence across multiple rounds of dialogue, the system also needs to control topic shifts. For example, if a user asked "Why is the electricity bill so high?" in the previous round and then asks "Can I pay in installments?" in the current round, the system needs to recognize this as a new intent, "payment method inquiry," and clear the current intent before generating a response. This involves using partial state slots to avoid mis-associating context. This process is implemented through a lightweight intent change detection network that uses bidirectional GRU encoding of the current user input. and If the maximum semantic consistency score is below 0.5, it is judged as a topic shift, and the dialogue strategy is automatically switched.

[0076] S5: After completing the current round of interaction, combine user profile, current dialogue status and historical behavior to determine whether to trigger personalized service recommendations, and comprehensively evaluate the quality of this round of interaction based on dialogue coherence, changes in user emotions and confidence of response fields, and generate a quality score for optimization of subsequent service strategies.

[0077] Specifically, the core task of this step is to combine the user's current dialogue state after completing multiple rounds of dialogue responses. Natural language responses generated by the system This mechanism personalizes and optimizes service strategies for the current service round. On one hand, it dynamically determines whether to recommend extended services by identifying user characteristics, usage preferences, and current behavior. On the other hand, it scores service quality based on quality indicators and contextual consistency of the system's internally generated results, providing a data foundation for subsequent model optimization. This mechanism enables the intelligent voice system to adaptively adjust to individual users, providing not only standard responses but also the ability to expand into an intelligent assistant, automatically determining whether to remind users of payments, recommend energy-saving suggestions, or alert them to electricity price changes after a user inquiry.

[0078] The input variables for this step include the output from the previous round. and , where represents the system response text generated in the current round and the current dialogue state vector represented internally by the system, respectively. It is natural language text generated based on the template selection, slot filling, and confidence assessment mechanism in step 4, representing "what the system said"; while It is a hidden state vector (typically 256-dimensional) generated by comprehensively encoding user input, historical context, and multi-round semantic states, representing "what the system understands." These two variables together constitute the input features for determining whether the current user is suitable for personalized service recommendations.

[0079] Meanwhile, the system dynamically extracts user profile vectors from long-term user behavior logs. It includes the following dimensions:

[0080] Conversation frequency in the last 30 days (e.g., average number of inquiries per day);

[0081] Distribution of business intent in the last 3 conversations (e.g., 70% related to electricity bills, 20% related to power outages).

[0082] Historical emotional state statistics (e.g., anxiety accounted for 0.48%).

[0083] Service response preferences (e.g., whether or not one prefers to accept recommended services);

[0084] Behavioral conversion tags (such as whether you have received a recommended payment reminder and successfully redirected to complete the process).

[0085] This data is collected in real time through the integration logs of the voice interaction system and the power company's business system. The system automatically extracts this data from user interaction records, voice recognition results, intent recognition logs, and business processing records via data integration services, without requiring users to actively provide it. (User profile vector) It is a dense vector with a fixed length of 128, which, after normalization, can be directly compared with... The concatenation is used as input to the recommendation discrimination network.

[0086] To determine whether to trigger personalized recommendation services, such as "whether to remind the user to set a tiered electricity pricing threshold reminder" or "suggest that they enable power outage SMS notifications," we construct a recommendation activation probability model. This model has a two-layer fully connected network structure, with the input being... The recommendation probability is output through the Sigmoid activation function. The expression is as follows:

[0087] ;

[0088] in, and These are the weights and biases of the recommendation discrimination network, respectively. For the Sigmoid function; This is an emotion regulation factor, the value of which depends on the emotional state identified in the current conversation. and its strength (All are from the output of step 1). When "Anxiety" and When, set In other cases, it is 1, which means that the recommendation frequency is automatically reduced when the user is anxious to avoid disturbing the user's emotions and to keep the system interaction gentle and rational.

[0089] For example, if a user completes a round of "electricity bill anomaly inquiry," and the system generates a reply such as, "Hello, your electricity bill this month is 248 yuan, which is significantly higher than the average. We suggest you pay attention to the use of high-frequency appliances," and if the system's historical records show that the user's electricity bill has been increasing month by month for the past three months, and the historical conversations have focused on electricity-related issues, then the user profile is... The "energy saving attention" dimension is relatively high, and the dialogue status is also high. If a slot indicating "Concerned about electricity bills" is activated, the user will be identified as suitable for receiving the "Energy Consumption Monitoring Service" recommendation. The system will then automatically connect to the recommendation module and generate subsequent guiding questions, such as "Would you like to enable the monthly energy-saving analysis report function?"

[0090] After generating the recommendation judgment, we designed a comprehensive scoring function to evaluate the quality of the current dialogue in this round of the system. Taking into account the coherence of the response, emotional changes, and the confidence level of the response:

[0091] ;

[0092] in, The cosine similarity between the current and previous dialogue state vectors represents the dialogue coherence. and These represent the intensity of the user's emotions in the current and previous rounds of the conversation (obtained by the emotion recognition module); This represents the weighted average of the confidence scores of all populated fields in the current response, derived from the field confidence vector in step 3. Three weighting coefficients , and It can be obtained through training with historical data or set manually; common values ​​are... , , .

[0093] The scoring function outputs This is used to monitor the quality of system responses in real time and to help determine whether the current dialogue should be included in the training sample pool or to trigger system policy optimization. For example, when... If the response is not received by the system, it indicates a significant problem, such as drastic emotional fluctuations, low confidence in the response, or failure to recognize contextual transitions. In this case, the system marks the session as an "optimization candidate" for sampling during the model fine-tuning phase. Optimization is not performed online but is handled uniformly by the system during low-load periods (such as early morning) to ensure operational stability.

[0094] Ultimately, this step outputs two variables:

[0095] : Probability of triggering personalized recommendation service;

[0096] : Current round system response quality score.

[0097] Through the aforementioned mechanisms, this step serves a dual purpose: "service extension" and "system self-correction." The personalized recommendation mechanism enables the system not only to respond to user-submitted questions but also to proactively uncover users' potential service needs, thereby expanding the depth of services. Meanwhile, the dialogue quality scoring function constructs a finely controlled system feedback loop, continuously improving system stability and user satisfaction.

[0098] This invention also provides a power grid user voice consultation processing device based on an intelligent robot, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the steps described in the above embodiment of a power grid user voice consultation processing method based on an intelligent robot, for example... Figure 1 The steps S1 to S5 described above; or, when the processor executes the computer program, it implements the functions of each module in the above system embodiments.

[0099] For example, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the intelligent robot-based power grid user voice consultation processing device.

[0100] The aforementioned intelligent robot-based power grid user voice consultation processing device can be a desktop computer, laptop, handheld computer, or cloud server, among other computing devices. This device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the intelligent robot-based power grid user voice consultation processing device may also include input / output devices, network access devices, and buses.

[0101] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASACs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. This processor is the control center of the intelligent robot-based power grid user voice consultation processing device, connecting all parts of the device via various interfaces and lines.

[0102] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the intelligent robot-based power grid user voice consultation processing device by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function, etc.; the data storage area may store data created based on the operation of the air conditioner controller, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0103] The module integrated into the intelligent robot-based power grid user voice consultation processing device, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0104] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0105] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for processing voice consultations for power grid users based on intelligent robots, characterized in that, The method includes: Receive power business consultation requests from users via voice input; perform speech recognition on the power business consultation requests to obtain the corresponding text content, and simultaneously perform sentiment analysis to obtain the user's sentiment category and emotional intensity; By combining the text content with the emotion category and emotion intensity, emotion-driven intent recognition is performed to determine the user's business intent and corresponding urgency score. According to the business intent, the power grid business system interface is invoked to query data. When there are missing fields in the query results, a graph neural network is used to infer and complete the missing fields based on the pre-built power grid business knowledge graph and the urgency score, generating a structured response result containing the original fields and the completed fields, and assigning confidence to each field. Based on the structured response results, field confidence levels, and multi-turn dialogue context states, a response template is dynamically selected and a natural language response is generated. After completing the current round of interaction, the system combines user profile, current dialogue status, and historical behavior to determine whether to trigger personalized service recommendations. It also conducts a comprehensive evaluation of the quality of this round of interaction based on dialogue coherence, changes in user emotions, and confidence levels of response fields, generating a quality score for optimization of subsequent service strategies.

2. The power grid user voice consultation processing method based on intelligent robots according to claim 1, characterized in that, The sentiment analysis employs a Chinese sentiment classification model based on the BERT architecture. This model is fine-tuned on annotated corpus of electricity customer service scenarios and outputs sentiment categories including anxiety, confusion, pleasure, or neutrality, along with sentiment intensity values ​​ranging from 0 to 1.

3. The power grid user voice consultation processing method based on intelligent robots according to claim 1, characterized in that, The intent recognition adopts a Chinese business intent recognition model based on the RoBERTa architecture. The Chinese business intent recognition model integrates emotion regulation embedding vectors in the input embedding stage and enhances the attention weight of emotion-related words according to the emotion intensity in the multi-head attention mechanism, so as to improve the accuracy of intent recognition for ambiguous or emotional expressions.

4. The power grid user voice consultation processing method based on intelligent robots according to claim 1, characterized in that, The urgency score comprehensively considers semantic sentence vectors, emotional intensity, and semantic certainty, and is used to regulate the priority and strategy of subsequent business queries and inferences.

5. The power grid user voice consultation processing method based on intelligent robots according to claim 1, characterized in that, The power grid business knowledge graph is constructed with business fields as nodes and field dependencies as edges. During the reasoning process for missing fields, the reasoning strength of the graph neural network is dynamically adjusted according to the urgency score, so that high-urgency requests can still generate preliminary responses when the data is incomplete.

6. The power grid user voice consultation processing method based on intelligent robots according to claim 1, characterized in that, Each field in the structured response result is accompanied by a confidence level, which reflects whether the field is directly returned by the business system or completed through reasoning, and is used for credibility control in subsequent response generation.

7. The power grid user voice consultation processing method based on intelligent robots according to claim 1, characterized in that, The process of generating a natural language response based on structured response results, field confidence, and multi-turn dialogue context states includes: encoding up to six rounds of historical dialogue content through a gated recurrent unit and a self-attention mechanism to generate a current dialogue state vector, and identifying whether the user has shifted the topic based on the current dialogue state vector in order to dynamically adjust the dialogue focus and response strategy.

8. The power grid user voice consultation processing method based on intelligent robots according to claim 1, characterized in that, The natural language response is generated using a combination of templates and neural network generation. Candidate templates are filtered according to intent category, and the optimal template is selected based on a weighted score of field confidence. Low-confidence fields are handled with fuzzy descriptions or omissions.

9. The power grid user voice consultation processing method based on intelligent robots according to claim 1, characterized in that, The personalized service recommendation is based on a joint judgment of the user profile vector and the current dialogue state vector. The user profile includes recent consultation frequency, intent distribution, sentiment statistics and service preferences, and the recommendation trigger is suppressed when the user is in a high anxiety state.

10. The power grid user voice consultation processing method based on intelligent robots according to claim 1, characterized in that, The comprehensive evaluation of interaction quality is achieved by calculating the similarity between the current and previous dialogue states to represent dialogue coherence, combining the changing trend of user emotional intensity between the current and previous rounds, and the weighted average of the confidence scores of each field in the current response to generate a quality score ranging from 0 to 1; when the quality score is lower than the threshold, the current dialogue is marked as an optimization candidate sample for offline iterative updates of the system strategy.

Images