Control Method and Equipment for Power Question Answering Robot Based on Multi-Source Knowledge Decision Making

Abstract

The application discloses a power question and answer robot control method and equipment based on multi-source knowledge decision, and relates to the technical field of intelligent question and answer robots. The method comprises the following steps: determining the business scene and risk level label of a user question through an intention recognition model; obtaining a plurality of candidate knowledge segments from a multi-source knowledge base according to the business scene and risk level label; extracting decision features and obtaining a comprehensive score through model processing; combining power grid real-time operation state data, performing logic-state alignment verification on high-score candidate segments to generate an executability confidence; generating an output strategy according to the comprehensive score and the confidence, and outputting a state conflict prompt when there is a logic conflict. The application is used to solve the problem that a traditional power question and answer system only relies on static text matching and is easily to output incorrect operation guidance by being separated from dynamic physical working conditions of a power grid.

Description

Technical Field

[0001] This invention relates to the field of intelligent question-answering robot technology, and more specifically, to a control method and device for an electric question-answering robot based on multi-source knowledge decision-making. Background Technology

[0002] The operation and maintenance of power systems is a multi-disciplinary, high-risk, and complex business. In scenarios such as customer service inquiries, on-site maintenance guidance, and dispatch assistance, staff and users frequently need accurate operating procedures, regulations, and electricity price information. To improve business processing efficiency, power-related Q&A robots have been widely adopted to automatically analyze questions and provide consultation or guidance services. This requires the Q&A system to not only possess accurate semantic understanding capabilities but also ensure the absolute safety, compliance, and high reliability of the output instructions.

[0003] Currently, most mainstream power-related question-answering robots employ one-way matching techniques based on retrieval-enhanced generation or large language models. Their typical workflow is as follows: after receiving a user's question, they calculate the semantic vector similarity between texts, retrieve the highest-scoring knowledge fragment from a pre-built static procedure document library, and then directly output this fragment, or have it summarized by a generative model and returned to the user.

[0004] However, the aforementioned existing technologies have significant limitations when dealing with complex power operations. First, existing solutions do not perform two-dimensional anchoring of specific business scenarios and potential risks at the front end, resulting in a lack of differentiated anti-misoperation interventions under high-risk operations. Second, existing matching mechanisms sever the connection between "static knowledge" and "dynamic power grid," making it easy for high-scoring static text retrieved to deviate from the real-time topology of the power grid; for example, if standard switching operation knowledge is blindly output without considering the current equipment interlock status, it is very easy to violate the five-prevention interlocking logic and cause major safety accidents. In addition, existing single-location reliability output lacks multi-source consistency verification and dual-threshold risk control verification, making it difficult to achieve anti-misoperation blocking and rigorous step-by-step instruction guidance when safety operations are involved. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a power question-answering robot control method based on multi-source knowledge decision-making. This method anchors business scenarios and risks through intent recognition, performs initial screening through multi-source knowledge comprehensive scoring, and, importantly, introduces a logic-state alignment verification mechanism to fuse and verify candidate knowledge fragments with the real-time operating status of the power grid to generate the final output strategy. This addresses the problem that traditional power question-answering systems rely solely on static text similarity matching, which is detached from the dynamic physical conditions of the power grid and is prone to outputting incorrect operation guidance.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A power question-answering robot control method based on multi-source knowledge decision-making includes the following steps: determining the business scenario label and risk level label of the user's question through a power business intent recognition model; obtaining multiple candidate knowledge fragments from a multi-source knowledge base containing a real-time state knowledge base based on the business scenario label; extracting the decision feature vector of each candidate knowledge fragment and obtaining a comprehensive score through the multi-source knowledge decision-making model; combining the real-time operating status data of the real-time state knowledge base, performing logic-state alignment verification to generate executability confidence for candidate knowledge fragments with a comprehensive score higher than the baseline; generating an output strategy based on the comprehensive score and executability confidence; if the executability confidence is lower than a preset safety threshold, the output strategy includes outputting a state conflict prompt.

[0007] In a preferred embodiment, the execution logic-state alignment verification to generate executability confidence includes: performing semantic dependency analysis on candidate knowledge fragments to extract operation logic triples containing preconditions and action instructions; obtaining corresponding real-time running state data from a real-time state knowledge base based on the preconditions; and substituting the real-time running state data into the operation logic triples for logical operations to obtain executability confidence.

[0008] In a preferred embodiment, the step of substituting real-time operating status data into the operation logic triplet for logical operation includes: identifying the device interlocking constraints and energized state constraints in the operation logic triplet; obtaining the current switch position signal, telemetry value, and associated timestamp of the target device corresponding to the action command in the real-time operating status data; constructing a local power grid topology sub-graph of the area to which the target device belongs at the current moment; simulating the execution of the action command in the local power grid topology sub-graph; if the simulation result triggers the preset five-prevention interlocking logic, then determining that the executability confidence is lower than the preset safety threshold.

[0009] In a preferred embodiment, generating an output strategy based on a comprehensive score and executability confidence level includes: when the executability confidence level reaches a preset safety threshold, calculating the difference between the highest and second-highest comprehensive scores among multiple candidate knowledge fragments; when the difference is less than a first threshold and the risk level of the user's question is a preset high-risk level, generating an output strategy based on a preset template script for answering or transferring to human service; when the difference is between the first and second thresholds, adjusting the comprehensive score of the candidate knowledge fragments based on the consistency of values ​​of key business fields in different knowledge sources; otherwise, taking the candidate knowledge fragment with the highest comprehensive score as the target knowledge fragment and generating an output strategy based on the target knowledge fragment.

[0010] In a preferred embodiment, the decision feature vector includes at least the following features: semantic similarity features between the user question and the candidate knowledge fragment; matching degree features between the business scenario label of the user question and the preset applicable scenario label of the candidate knowledge fragment; and adaptation degree features between the risk level label and the preset risk level label of the candidate knowledge fragment.

[0011] In a preferred embodiment, the power business intent recognition model is a deep learning text classification model based on a self-attention mechanism, and the power business intent recognition model is trained using historical power customer service dialogue data containing intent labels and risk level labels.

[0012] In a preferred embodiment, the real-time status knowledge base is structured data that changes dynamically over time; the structured data includes at least power grid operation status data, planned maintenance data, smart meter data, time-of-use electricity price data, and environmental status data; each piece of structured data has a corresponding timestamp and geographic location identifier.

[0013] In a preferred embodiment, adjusting the comprehensive score of the corresponding candidate knowledge fragment based on the consistency of values ​​of key business fields in different knowledge sources includes: extracting key business fields from the candidate knowledge fragment, wherein the key business fields include electricity price values, time ranges, or equipment identifiers; calculating the consistency ratio of the key business fields in their respective knowledge sources; and using the consistency ratio as a weighting coefficient to adjust the comprehensive score of the candidate knowledge fragment.

[0014] This invention provides a power question-answering robot control device based on multi-source knowledge decision-making, comprising: a question acquisition module, used to acquire user questions and determine the business scenario label and risk level label of the user questions based on a power business intent recognition model; a knowledge extraction module, used to retrieve multiple candidate knowledge fragments for the user questions from a multi-source knowledge base according to the business scenario labels; a comprehensive scoring module, used to construct a decision feature vector for each candidate knowledge fragment, input it into a preset multi-source knowledge decision-making model, and obtain a comprehensive score for each candidate knowledge fragment; and a strategy output module, used to control the question-answering robot to output a strategy based on the comprehensive score and the business scenario label of the user questions.

[0015] A power question-answering robot control device based on multi-source knowledge decision-making includes a memory and a processor: the memory is used to store programs; the processor is used to execute the programs to implement the various steps of the power question-answering robot control method based on multi-source knowledge decision-making.

[0016] The technical effects and advantages of the power question-answering robot control method based on multi-source knowledge decision-making in this invention are as follows: This invention uses a power business intent recognition model to determine the scenario and risk label of user questions, establishing a preliminary classification basis for knowledge matching. Based on this label, candidate fragments are obtained from a multi-source knowledge base containing real-time status data. A multi-source knowledge decision model calculates a comprehensive score, achieving quantitative initial screening of candidate knowledge. A logic-state alignment verification mechanism is then introduced to combine high-scoring knowledge fragments with real-time operating status data for verification, generating executability confidence. Finally, the output strategy is determined based on the comprehensive score and confidence, and a state conflict warning is output when the confidence falls below a preset safety threshold. This solution changes the traditional question-answering system's output model, which relies solely on static text similarity. By establishing a correlation verification between static knowledge text and the dynamic real-time operating conditions of the power grid, it blocks erroneous guidance outputs that violate the current underlying operating state of the power grid, improving the safety and accuracy of the power question-answering robot in business decision-making. Attached Figure Description

[0017] Figure 1 A schematic diagram of the control method for a power question-answering robot based on multi-source knowledge decision-making provided in an embodiment of the present invention; Figure 2 The accuracy / loss variation curve during the training process of the power business intent recognition model provided in this embodiment of the invention; Figure 3 A heatmap of the business scenario tag classification confusion matrix provided in this embodiment of the invention; Figure 4 A heatmap of the risk level label classification confusion matrix provided in this embodiment of the invention; Figure 5 A block diagram of the power question-answering robot control device based on multi-source knowledge decision-making provided in an embodiment of the present invention; Figure 6 This is a structural block diagram of an exemplary electronic device provided for implementing embodiments of the present disclosure. Detailed Implementation

[0018] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1, Figure 1 The present invention provides a control method for an electric question-answering robot based on multi-source knowledge decision-making, comprising the following steps: S1: Obtain user issues and determine the business scenario label and risk level label of the user issues through the power business intent recognition model.

[0020] It should be noted that, before determining the business scenario label and risk level label, this embodiment of the invention first needs to obtain the user's question. The system receives business consultation requests initiated by users through accessed multi-channel human-computer interaction interfaces, including voice entry points from web clients, mobile terminal applications, and smart wearable devices. In specific implementation, when a user asks a question via voice, the system calls a preset voice acquisition module to obtain the raw voice signal containing the power business consultation content, and inputs the voice signal into a voice recognition engine for endpoint detection, acoustic feature extraction, and decoding to obtain the corresponding Chinese text sequence; when a user inputs a question via text through the customer service system or APP, the user's text input is directly used as the text sequence of the user's question.

[0021] Furthermore, to form standardized model input, the system preprocesses the aforementioned text sequence, including standardizing full-width and half-width symbols, filtering abnormal characters, and using regular expressions to identify and replace special identifiers such as electricity meter numbers and work order numbers with unified placeholders. Through this standardization process, synonyms and abnormal codes in the user input are mapped to unified word forms, thereby effectively improving the consistency and robustness of intent recognition.

[0022] In this embodiment, the power business intent recognition model is a deep learning text classification model based on a self-attention mechanism. The power business intent recognition model is trained using historical power customer service dialogue data containing intent tags and risk level tags.

[0023] Specifically, the training and inference principles of the deep learning text classification model (e.g., using a Transformer or BERT-based architecture) are as follows: During the model training phase, dialogue samples containing complete question-and-answer rounds are first extracted from historical customer service recordings and online customer service chat logs. User questions in each sample are then manually or semi-automatically labeled to form corresponding business scenario tag sets and risk level tag sets. The business scenario tags are used to distinguish different electricity business scenarios such as installation and repair requests, electricity bill inquiries, fault reporting, planned power outage notifications, and electricity safety reminders, and are denoted as a predefined set of business scenario categories. M represents the total number of business scenario categories; risk level labels are used to distinguish different service risk levels such as high risk, medium risk, and low risk, and are denoted as a predefined set of risk level categories. N represents the total number of risk level categories.

[0024] Taking common power business scenarios as an example, the system can divide common power business scenarios into several preset categories and map them to business scenario tags. The example configuration is shown in Table 1. The table lists typical business scenario tags and their corresponding descriptions, where scenarios 1 to 5 correspond to... arrive .

[0025] Table 1

[0026] When building the model, the labeled user question text mentioned above is used as input, and the corresponding business scenario labels and risk level labels are used as supervision signals for joint training: First, a standardized user question text sequence of length L is processed. Perform word segmentation or sub-word splitting, and map it to a sequence of word vectors. ,in Let E be the embedding vector of the Lth word or subword; input E into a Transformer or BERT encoder to obtain the corresponding contextual semantic representation sequence. ,in This is the context vector representation of the Lth position. When using the BERT structure, the vector corresponding to the special marker [CLS] at the beginning of the sequence is taken. As a semantic representation of the whole sentence, Input the business scenario classification subnetwork and the risk level classification subnetwork, respectively, and pass them through a fully connected layer and a nonlinear transformation to obtain the unnormalized score vector for business scenario classification. and the unnormalized score vector of risk level classification ,in Indicates the corresponding business scenario category The score, Indicates the corresponding risk level category The scores are then calculated; the score vectors are then subjected to softmax normalization to obtain the predicted probability distribution of the labels. (Business scenario label prediction probability) The calculation formula is as follows: , Risk level label prediction probability The calculation formula is as follows: , in, User issues fall under the category of business scenarios. The probability, User issues fall under the risk level category. The probability. During training, manually labeled real-world business scenario tags will be used. and true risk level label One-hot encoding is used to represent the data, which are denoted as one-dimensional vectors of length M and N respectively. The cross-entropy between the predicted probability distribution and the true label is used as the loss function.

[0027] Business scenario classification loss The calculation formula is as follows: , Risk level classification loss The calculation formula is as follows: , in, The tag indicates the actual business scenario. The value is 1 if it is true, and 0 otherwise. The label indicates the true risk level. The value is 1 if it is true, and 0 otherwise. Let represent the natural logarithm function. To achieve joint optimization of business scenario identification and risk level identification, and are weighted and summed according to preset weight coefficients to obtain the total loss L. The formula for calculating the total loss L is as follows: , in and These are non-negative weighting coefficients used to balance the importance of business scenario classification tasks and risk level classification tasks. The values ​​of α and β are determined by performance tuning on the validation set. The model parameters are updated using the backpropagation algorithm and gradient descent optimizer until the total loss L converges or the preset number of training epochs is reached.

[0028] In one specific implementation, during model training, the accuracy and total loss of business scenario identification and risk level identification on the validation set are recorded at different training epochs, and the change curves of the training process are plotted, such as... Figure 2 As shown in the figure, the horizontal axis represents the number of training rounds, the left vertical axis represents the accuracy of business scenario recognition and risk level recognition on the validation set, and the right vertical axis represents the corresponding total loss value. Figure 2 It can be seen that as the number of training rounds increases, the total loss gradually decreases and tends to stabilize, while the classification accuracy gradually increases and remains at a high level after convergence. This indicates that the constructed power business intent recognition model can effectively learn the semantic features in historical power customer service dialogue data.

[0029] During the online inference phase, after obtaining the standardized user question text, it is input into the trained power business intent recognition model, and the predicted probability distribution of the business scenario category is obtained by following the same encoding and classification process as described above. and the predicted probability distribution of risk level categories ,according to The business scenario tag corresponding to the category with the highest probability is used as the business scenario tag for this user's problem. ,Right now ,according to The risk level label corresponding to the category with the highest probability is used as the risk level label for this user's problem. ,Right now Meanwhile, the maximum probability value is used as the confidence index of the corresponding label, and samples with confidence scores below a preset threshold can be marked as awaiting manual review or downgraded as "unknown scenarios" to reduce the interference of misclassification on subsequent multi-source knowledge decision-making.

[0030] Furthermore, to enhance performance in specialized domains, a power industry-specific dictionary, including electricity pricing terms, power grid equipment names, and work order business codes, is introduced into the model training data. Custom encoding of these specialized terms during word segmentation and embedding stages further improves the accuracy and sensitivity to domain-specific terminology in the intent recognition model within power business scenarios. To visually demonstrate the classification performance of the power business intent recognition model across different business scenario labels and risk level labels, the true and predicted labels on the validation set or online samples can be statistically analyzed, and a heatmap of the classification confusion matrix can be generated using Python. Figure 3 and 4 As shown in the figure, the horizontal axis represents the true label and the vertical axis represents the predicted label. The color intensity represents the sample size or proportion. Figure 3 For business scenario tag obfuscation matrix, Figure 4 The risk level label confusion matrix shows that most samples are concentrated near the main diagonal, indicating that the model has high classification accuracy in major business scenarios and risk levels, while some easily confused scenario combinations can be the focus of subsequent optimization.

[0031] This step combines speech recognition and text standardization technologies during the user question acquisition phase to unify multi-channel inputs into a structured text representation. It also utilizes a deep learning power business intent recognition model based on Transformer or BERT to conduct supervised training on historical power customer service dialogues. This enables the automatic and accurate output of business scenario labels and risk level labels after a user initiates an inquiry.

[0032] S2, based on business scenario tags, retrieves multiple candidate knowledge fragments from a multi-source knowledge base containing a real-time status knowledge base.

[0033] It should be noted that, before performing multi-source knowledge retrieval, this embodiment of the invention first extracts and organizes data such as internal regulations, business process documents, structured business data, historical dialogue records, and power grid operation status of the power company during the system deployment phase. These data are then divided into multiple knowledge sub-bases to form a multi-source knowledge base. The multi-source knowledge base includes at least a regulations document knowledge base, a business process and operation guidance knowledge base, a structured business data knowledge base, a historical dialogue and typical case knowledge base, and a real-time status knowledge base. Each knowledge record is segmented into appropriately granular knowledge fragments when it is entered into the database, and each knowledge fragment is assigned a unique identifier and associated metadata such as business scenario tags, risk level tags, knowledge source type identifiers, applicable time intervals, and applicable geographical location / transformer area identifiers. In this embodiment, the real-time status knowledge base is structured data that changes dynamically over time; the structured data includes at least power grid operation status data, planned maintenance data, smart meter data, time-of-use electricity price data, and environmental status data; each piece of structured data has a corresponding timestamp and geographic location identifier.

[0034] Specifically, to meet the need for accurate filtering and matching based on time and space dimensions during online retrieval, this real-time status knowledge base is specifically used to store the aforementioned dynamic structured status data related to power business.

[0035] During the real-time online retrieval phase, when the business scenario tag determined in step S1 is... Risk level label After obtaining standardized user question text, the system first tags the questions according to the business scenario. Scenario filtering is performed on multi-source knowledge bases, prioritizing those containing business scenario tags. or with Similar procedural document knowledge bases, business process and operational guidance knowledge bases, structured business data knowledge bases, and historical dialogue and typical case knowledge bases were used as the search targets, while also being tagged with risk level labels. It can be restricted that certain high-risk scenarios can only be retrieved from the procedural document knowledge base and the process guidance knowledge base, in order to avoid low-reliability knowledge sources from participating in the answering of questions in high-risk business.

[0036] For the text-type knowledge fragments to be retrieved, the system calls the same text encoding network as in step S1 to encode the user's question text into a semantic vector u, and reads the pre-stored semantic vectors of each knowledge fragment. Calculate the semantic similarity between the semantic vector of the user question and the semantic vector of the knowledge fragment, wherein the semantic similarity The calculation formula is as follows: , Where u is the semantic vector of the user's question. Let k be the semantic vector of the k-th knowledge fragment. Represents the vector dot product. and These are the L2 norms of the corresponding vectors. The higher the value, the higher the semantic relevance of the knowledge fragment to the user's question. The system can weight and fuse the semantic similarity with the text relevance score obtained based on word segmentation and keyword matching according to a preset ratio to obtain a comprehensive retrieval score, and select several text-type knowledge fragments as candidate knowledge fragments according to the comprehensive score from high to low.

[0037] For the real-time status knowledge base, the system determines the target transformer area identifier and query time interval corresponding to the user's question based on the user's identity information, electricity address, electricity meter number, line name and other entity information explicitly or implicitly contained in the user's question, as well as the current time or the time range mentioned in the question. Then, it filters status data records in the real-time status knowledge base that have the same geographical location / transformer area identifier as the target transformer area identifier and whose status timestamp falls within the query time interval, and converts these status data into concise text descriptions as status knowledge fragments.

[0038] At the same time, state-related knowledge fragments are sorted according to their time freshness. Preferably, this can be based on the time difference between the state record time and the current time. Calculate the time freshness score The time freshness score The calculation formula is as follows: , in The time decay coefficient, It is the difference between the current time and the time of the m-th status data record. The smaller the value, the closer the state record is to the current moment, and the higher the corresponding time freshness score. Therefore, the time freshness score is used as one of the sorting criteria to select several state-related knowledge fragments as real-time state candidate knowledge fragments.

[0039] Through the above text-based knowledge retrieval and real-time operational status data retrieval, the system finally obtains a set of multi-source candidate knowledge fragments for the current user's problem. Each candidate knowledge fragment in this set is accompanied by metadata information such as its source sub-library type, business scenario tag, risk level tag, timestamp, and geographical location / station area identifier, which serve as the basis for subsequent steps to construct decision feature vectors and input them into the multi-source knowledge decision model to calculate the comprehensive score.

[0040] This step utilizes the business scenario tags identified in step S1 to perform scenario-driven limited retrieval of the multi-source knowledge base, and combines semantic similarity calculation and spatiotemporal constraint filtering based on real-time running status data. This significantly improves the semantic relevance and timeliness of candidate knowledge fragments to user questions while ensuring retrieval recall.

[0041] S3: Extract the decision feature vector of each candidate knowledge fragment, and obtain a comprehensive score through the multi-source knowledge decision model.

[0042] It should be noted that, in performing multi-source knowledge decision-making in this embodiment of the invention, the total set of multi-source candidate knowledge fragments obtained in the previous step S2 is preferably used. As input, each candidate knowledge fragment can be uniquely identified. Identify and label a given user question and its business scenario. and risk level labels The system first constructs multi-dimensional numerical features for each candidate knowledge fragment based on the semantic similarity and scene and risk-related information calculated in the aforementioned retrieval stage, and then splices these multi-dimensional numerical features to form a decision feature vector.

[0043] In this embodiment, the decision feature vector includes at least the following features: semantic similarity features between the user question and the candidate knowledge fragment; matching degree features between the business scenario label of the user question and the preset applicable scenario label of the candidate knowledge fragment; and adaptation degree features between the risk level label and the preset risk level label of the candidate knowledge fragment.

[0044] Specifically, in order to accurately characterize the multidimensional attributes of each candidate knowledge fragment, the system performs the following quantitative calculation rules for the above three key features: (1) Semantic similarity features between user questions and candidate knowledge fragments: For the corresponding candidate knowledge fragment k, the semantic similarity calculated in step S2 is denoted as ,Will Semantic similarity features are obtained by mapping to the [0,1] interval through linear normalization or standardization transformation. semantic similarity features The calculation formula is as follows: , in and In the current candidate set The minimum and maximum semantic similarity of all candidate knowledge fragments within the range. season To avoid the denominator being zero, A larger value indicates that the candidate knowledge fragment is semantically closer to the user's question.

[0045] (2) Matching characteristics between business scenario tags for user problems and preset applicable scenario tags for candidate knowledge fragments: Considering that each candidate knowledge fragment k is associated with one or more preset applicable scenario tag sets during the knowledge entry stage. The system predefines a business scenario similarity matrix in the label space. It is used to measure the similarity between different scene labels, when compared with When they are exactly the same The value is 1, when the two are completely unrelated. A value of 0 represents a business scenario tag for user issues. The set of applicable scenario labels for candidate knowledge fragment k It can be calculated With sets The maximum similarity between each label is used as the scene matching feature. Scene matching features The calculation formula is as follows: , Where c is a label for a specific applicable scenario of candidate knowledge fragment k. This is a set of applicable scenario tags for this segment.

[0046] (3) Regarding the compatibility characteristics between the risk level tags and the preset risk level tags of the candidate knowledge fragments: considering that the candidate knowledge fragment k is associated with preset risk level tags when it is entered into the database. (For example, high risk, medium risk, low risk), the system also pre-constructs a risk level compatibility matrix. This is used to indicate the degree of suitability of a certain knowledge fragment for answering the current question across different risk levels. And belonging to the same risk level The value is 1, when For low risk A value of 0 is used for high risk to indicate incompatibility; based on this, the risk level fit feature is defined. for Risk level suitability characteristics The calculation formula is as follows: , in The user problem risk level label determined in step S1, Preset risk level labels for candidate knowledge fragments.

[0047] After acquiring the above features, the system concatenates the three features in a predetermined order to form a decision feature vector for input to the multi-source knowledge decision-making model. If the total number of decision feature dimensions is D=3, then It can be represented as a three-dimensional real column vector. ,in , , In other alternative embodiments, in addition to the three basic features mentioned above, additional numerical features may be added as needed according to specific business requirements, and incorporated into the decision feature vector in the same manner as in this embodiment. However, this embodiment focuses on the case where the decision feature vector consists only of the above three features.

[0048] The multi-source knowledge decision-making model is a machine learning model obtained through supervised training on the optimal answer annotation data from historical question-and-answer samples. Specifically: A large number of user questions and their corresponding, actually adopted or confirmed correct answers were extracted from historical electricity customer service dialogue records. These fragments were then mapped in the current multi-source knowledge base to obtain a result in the form of ( , , , , The training samples, of which This represents the user question text of the nth training sample. and These respectively represent its business scenario label and risk level label. This represents the set of candidate knowledge fragments that can be retrieved given the current state of the knowledge base for this question. This represents the identifier of the candidate segment designated as the "optimal answer" for each training sample n and its candidate segments. Construct decision feature vectors in the same way as during online inference. and assign them supervisory labels. When k = season = 1, indicating that the candidate fragment is the optimal answer; otherwise, let = 0 indicates that the candidate fragment is a non-optimal answer; Multi-source knowledge decision-making models can employ machine learning models such as logistic regression, gradient boosting tree, or feedforward neural networks. In this embodiment, a feedforward neural network model with one hidden layer is preferred for nonlinear feature fusion, which integrates the decision feature vector. As input to the model, the corresponding candidate segments are output as the comprehensive score of the final answer. Specifically, let the input weight matrix of the decision model be denoted as . The bias vector is The hidden layers use ReLU or other non-linear activation functions, and the output layer weight vector is... The bias scalar is Then, for a candidate segment k in the training sample n, its hidden layer representation is... The calculation formula is as follows: , in Represents the product of a matrix and a vector. The bias vector has the same dimension as the hidden layer. Each component undergoes an element-wise nonlinear mapping; the output layer performs a linear transformation on the hidden layer representation and applies it through the Sigmoid function. Normalization was performed to obtain the comprehensive score. ), overall score The calculation formula is as follows: , in This represents the transpose of the output layer weight vector. This is the hidden layer vector representation after omitting the sample index. The value can be interpreted as the probability or confidence level of a candidate knowledge fragment k being selected as the final answer, given a user question, its scenario, and risk labels. During the model training phase, the comprehensive score corresponding to each candidate segment k for each sample n is calculated. With supervision label To compare the results, binary cross-entropy is used as the loss function to measure the deviation between the predicted results and the true labels. The loss is applied to sample n. The calculation formula is as follows: , The summation is performed on all candidate knowledge fragments k of the current sample n. Represents the natural logarithm function. Total training loss. The total training loss is the sum or average of the losses for all training samples. The calculation formula is as follows: , in This represents the number of training samples. The number is calculated using the backpropagation algorithm. The gradients of each parameter of the model are calculated, and stochastic gradient descent or its variants are used to optimize the model with a preset learning rate. The parameters are iteratively updated, and the general form of parameter update is: , Where θ represents any model parameter to be updated. and These are the parameter values ​​before and after the update, respectively. This represents the partial derivative of the total loss with respect to this parameter. After multiple rounds of iterative training, when the total loss converges or the performance metric on the validation set reaches a preset threshold, a well-trained multi-source knowledge decision-making model is obtained and can be used for online inference.

[0049] During the online operation phase, for the set of candidate knowledge fragments corresponding to the current user question obtained in step S2, for each candidate fragment, the system calculates the decision feature vector in the same way and inputs it into the trained multi-source knowledge decision model to obtain a comprehensive score. The comprehensive score is used as the comprehensive score value for double threshold comparison in subsequent steps. The higher the comprehensive score value, the better the candidate knowledge fragment is in terms of semantic relevance, scene matching, risk adaptability and other multi-source features, and the more suitable it is as the final answer.

[0050] To enable the multi-source knowledge decision-making model to adapt to business changes and knowledge base updates over time, the multi-source knowledge decision-making model also includes online updates of the weights of each feature dimension based on user feedback and manual review results, specifically: During actual system operation, each question-and-answer interaction records the user's explicit feedback on the answer (such as satisfaction / dissatisfaction, whether to continue asking questions) and the manual review conclusions of maintenance personnel or quality inspectors on some question-and-answer samples. These feedbacks and review results are mapped to reward or penalty tags for the output answer candidate segments. When an answer segment is actually output in a question-and-answer session... When a sample is confirmed to be correct and receives positive feedback, it is considered a positive sample and its label is recorded. When an answer is deemed incorrect or causes a user complaint, it is considered a negative sample and its label is recorded. And based on the feature vectors of candidate segments in the session and corresponding comprehensive score Constructing incremental loss The calculation formula is as follows: , The system uses a small learning rate without interrupting online services. The model parameters are fine-tuned online, and the parameter update formula is as follows: , in A learning rate significantly lower than that used in offline training ensures that model parameters adjust slowly upon receiving online feedback, preventing drastic model oscillations caused by isolated abnormal feedback. This can be combined with sliding window or exponential decay strategies to assign higher weights to newer feedback, allowing the model to adapt more quickly to recent business changes. Through this long-term, cumulative online update mechanism, the multi-source knowledge decision-making model can automatically adjust the implicit weights of different feature dimensions in the overall score, thereby weakening feature combinations that frequently lead to incorrect answers and strengthening feature patterns that have historically performed well.

[0051] This step constructs a multi-dimensional decision feature vector for each candidate knowledge fragment, including semantic similarity features, business scenario matching features, and risk level suitability features. It then uses a multi-source knowledge decision model trained under supervision with the best answer annotation data from historical question-and-answer samples to perform nonlinear fusion and comprehensive scoring of the above features. At the same time, it updates the model parameters online by combining user feedback and manual review results. This allows the question-and-answer robot to no longer rely solely on a single similarity index when selecting candidate knowledge fragments, but to make global optimization decisions under multiple constraints such as power business scenarios, risk control, and knowledge source credibility.

[0052] S4. Combining the real-time running status data of the real-time status knowledge base, perform logic-state alignment verification on candidate knowledge fragments with a comprehensive score higher than the baseline to generate executability confidence.

[0053] It should be noted that, in order to avoid the situation of "high-scoring misanswers" where the semantic score is high but violates the hard constraints of the current power grid operation, after obtaining the comprehensive score of each candidate knowledge segment, the system will perform a strict logic-state alignment verification mechanism on the candidate knowledge segments with scores higher than the preset baseline (e.g., 0.6), and perform secondary calculation and fusion of static text knowledge and dynamic physical state of the power grid.

[0054] In this embodiment, the execution logic-state alignment verification to generate executability confidence includes: performing semantic dependency analysis on candidate knowledge fragments to extract operation logic triples containing preconditions and action instructions; obtaining corresponding real-time running state data from the real-time state knowledge base based on the preconditions; and substituting the real-time running state data into the operation logic triples for logical operations to obtain the executability confidence.

[0055] Specifically, the system first uses natural language processing techniques (such as semantic dependency analysis) to parse the text of candidate knowledge fragments, extracting the implicit operational constraints in the text into structured "operational logic triples," denoted as... ,in For prerequisite entities (such as "10kV busbar sectionalizing switch"), For logical relationships (such as "in position", "locked"), This is an action command (such as "Do not close" or "Switching operation"). Subsequently, based on the transformer area identifier and device ID identified in step S1, the system indexes the real-time operating status data of the associated device from the real-time status knowledge base. Simulation calculations are then performed to determine the executability confidence of the action instruction.

[0056] In terms of the implementation details of the logical operation, the step of substituting real-time operating status data into the operation logic triplet for logical operation includes: identifying the device interlocking constraints and energized state constraints in the operation logic triplet; obtaining the current switch position signal, telemetry value and associated timestamp of the target device corresponding to the action command in the real-time operating status data; constructing a local power grid topology sub-graph of the area to which the target device belongs at the current moment; simulating the execution of the action command in the local power grid topology sub-graph, and if the simulation result triggers the preset five-prevention interlocking logic, then determining that the executability confidence is lower than the preset safety threshold.

[0057] In practical implementation, the system identifies the aforementioned logical triples and extracts the underlying equipment interlocking constraints (such as the five-prevention logic of "locking the isolating switch when the bus tie switch is closed") and energized state constraints (such as the safety logic of "prohibiting the closing of the grounding switch when the line is energized"). Based on this, the system... The system accurately acquires the current switching position signal (i.e., remote signaling data) of the target device corresponding to the action command, as well as the voltage and current values ​​(i.e., telemetry data) representing the energized state and the associated timestamps. Next, the system constructs a local power grid topology sub-graph of the area to which the target device belongs, and simulates the execution of the action command within this sub-graph.

[0058] Specifically, the system maps and activates the device interlocking constraints and energized state constraints identified from the knowledge fragments as the "preset five-prevention blocking logic" in the topology subgraph engine. For example, if the device interlocking constraint implicit in the candidate knowledge fragment is "if the bus tie switch is in the closed position, then the operation of the isolating switch is prohibited," and the real-time operating status data obtained by the system shows that the bus tie switch in the current physical power grid is indeed "closed," then when the system tentatively simulates the action command of "operating the isolating switch" in the local power grid topology subgraph, since the real-time status node of the bus tie switch is already closed, this simulated action will directly trigger the aforementioned preset five-prevention blocking logic (i.e., the topology link verification fails). The system thus determines that there is a logical conflict, that is, it determines that the executability confidence of the operation command is lower than the preset safety threshold (e.g., the confidence level is lowered). If the five-prevention interlocking logic is met or there are no logic constraints, then set it to the threshold of safety (e.g., 0). Set to 1). This executability confidence level. This will be passed as a decisive mask signal to subsequent steps.

[0059] Step S5: Generate an output strategy based on the comprehensive score and executability confidence level; if the executability confidence level is lower than a preset safety threshold, the output strategy includes an output state conflict prompt.

[0060] It should be noted that traditional question-answering robot output control typically employs a strategy of "maximum single score" or "maximum single model confidence." Specifically, for each candidate knowledge fragment, the highest-scoring fragment is selected based solely on semantic similarity or retrieval score and fed into a generative language model or template engine to generate the final answer. This approach fails to differentiate between the risk levels of different power business scenarios and does not perform consistency checks on key business fields from different knowledge sources. Consequently, in high-risk scenarios involving electricity price explanations, power outage scope descriptions, and safe operation guidance, even slight differences in the overall score may lead to the selection of fragments from unreliable sources or with outdated key fields, resulting in erroneous or non-compliant answers. This poses a risk of misleading users and violating power safety or regulatory requirements. This embodiment introduces dual threshold comparison and key field consistency checks on top of the overall score, and adopts different output strategies based on business scenario tags. This enables the question-answering robot to automatically limit freely generated and uncertain answers when facing high-risk power businesses, significantly reducing the risk of misanswers, while maintaining answer efficiency and naturalness in low-risk scenarios.

[0061] In this embodiment, the output strategy is generated based on the comprehensive score and the executability confidence level, including: when the executability confidence level reaches a preset safety threshold, calculating the difference between the highest and second-highest comprehensive scores among multiple candidate knowledge fragments; when the difference is less than a first threshold and the risk level of the user's question is a preset high-risk level, generating an output strategy based on a preset template script for answering or transferring to human service; when the difference is between the first and second thresholds, adjusting the comprehensive score of the candidate knowledge fragments based on the consistency of values ​​of key business fields in different knowledge sources; otherwise, taking the candidate knowledge fragment with the highest comprehensive score as the target knowledge fragment, and generating an output strategy based on the target knowledge fragment.

[0062] Specifically, the system processes the candidate knowledge fragment set obtained in step S2. The overall score of each candidate fragment k Sort the scores to obtain a descending order. ,in The candidate segment with the highest overall score is identified. The candidate segment with the second-highest overall score is identified, and the highest overall score is recorded as . The second highest overall score was The difference between the two Defined as The system pre-sets a first threshold. Second threshold And satisfy The first threshold is used to determine whether the distinction between the best and second-best candidates in the overall score is sufficient to support direct selection. The second threshold is used to determine whether key field consistency adjustments are needed to improve decision reliability. At that time, it was believed that the confidence levels among the current candidates were relatively close and there was significant uncertainty. At that time, it was believed that there was a certain degree of differentiation, but multi-source consistency verification was still needed. At that time, it was believed that the optimal candidate already had a clear advantage.

[0063] It should be noted that if the logic-state alignment check in step S4 shows that all high-scoring candidate segments have logical conflicts (i.e., This indicates that the user's query is prohibited or extremely dangerous under the current power grid conditions. In this case, regardless of the risk level, the system will trigger a "state blocking strategy," directly controlling the question-and-answer robot to output a state conflict warning message.

[0064] For example, the output could be "Based on the current real-time operating mode, XX equipment is under maintenance. The operation (fragment content) you requested is currently not executable. Please verify." Instead of directly answering the question about the content of the knowledge fragment, the output could proactively display the difference between the real-time operating status data and the operating conditions in the knowledge fragment.

[0065] When performing scenario-based output control, the system first determines the user problem risk level label based on step S1. Determine whether the current scene belongs to the preset high-risk level set. For example, scenarios involving on-site operational guidance, interpretation of power safety regulations, explanation of electricity pricing policies, and confirmation of power outage scope and time periods are marked as high-risk scenarios. and When the difference between the highest and second-highest comprehensive scores is less than the first threshold and the risk level of the user's question is a preset high-risk level, the system considers that the distinguishability between the current candidate knowledge fragments is insufficient to support the direct selection of any candidate as the final answer in a high-risk scenario. To avoid bias caused by the generative language model's free generation, this embodiment controls the question-answering robot to answer only based on preset templates or directly triggers a transfer to human service. Specifically, in the case of high risk and a small ΔS, the system no longer calls the generative language model to freely generate candidate knowledge fragments, but instead selects those from a preset high-risk template library that match the business scenario tags. and risk labels The matching standard notification template is filled with verified structured business data or original procedure content. For key information that cannot be automatically confirmed, users are prompted to contact human or on-site personnel with conservative wording. If necessary, the conversation is directly escalated and transferred to a human agent. The chatbot outputs a uniform message such as "To ensure safety / policy accuracy, a human customer service representative will provide further answers" to achieve a safety net for high-risk scenarios.

[0066] when Between and Between, that is At this time, the system determines that there is a certain degree of differentiation between the current best candidate and the second best candidate, but there may still be conflicts in knowledge sources or inconsistencies in key fields. Therefore, before selecting the final target knowledge segment, this embodiment adjusts the comprehensive score of the corresponding candidate knowledge segments based on the consistency of the values ​​of key business fields in different knowledge sources. This includes: extracting key business fields from the candidate knowledge segments, where the key business fields include electricity price values, time ranges, or equipment identifiers; calculating the consistency ratio of the key business fields in their respective knowledge sources; and using the consistency ratio as a weighting coefficient to adjust the comprehensive score of the candidate knowledge segments.

[0067] Specifically, the system predefines a set of key fields. This describes fields that have a decisive impact on security, compliance, and billing accuracy in electricity business responses, such as electricity price, price execution method, effective and expiration dates, power outage start and end times, transformer area or line identification, operation ticket number, and important equipment names. For each candidate knowledge fragment k and its corresponding knowledge source type... The key fields involved in the fragment The value of is denoted as At the same time, candidate fragments are divided into different source category sets based on the knowledge source type. For example, source documents, business system sources, historical dialogue sources, real-time status sources, etc.

[0068] To measure the consistency of values ​​for the same key field across different knowledge sources, the system uses the current candidate set. Within the scope, for each key field and each possible value Count the number of candidate fragments with this value in each source category and record them in the key field under knowledge source type s. Values The number of candidate segments is The total number across all knowledge source types is , field The total number of candidates with all different values ​​is Then the value is taken as follows In the field Global consistency ratio Defined as: , Consistency ratio The larger the value, the more candidate knowledge fragments give the same value in that field. For a given candidate fragment k, its value in the set of key fields... The set of field values ​​involved is This allows you to define a comprehensive key field consistency index for the segment. This is a weighted average of the consistency ratios of each field. It considers the overall consistency of key fields. The calculation formula is as follows: , in This represents the subset of key fields actually involved in candidate fragment k. For fields The preset weighting coefficients are used to reflect the importance of different key fields to business security, and to meet the requirements. and Furthermore, the credibility of the knowledge source type to which the candidate fragment belongs can be used to further... Make corrections, such as assigning higher basic credibility to fragments from procedural documents and business systems, and lower basic credibility to fragments from historical dialogues, so that fragments with the same field consistency can still reflect differences in credibility between different knowledge sources.

[0069] In obtaining Subsequently, this embodiment uses a consistency coefficient to evaluate the overall score. After weighted adjustment, the adjusted comprehensive score is obtained. The adjusted comprehensive score calculation formula can be expressed as: , in This is a consistency amplification factor, used to control the degree of influence of consistency on the final score. To establish a consistency baseline, a value of 0.5 can be used, or it can be set to the average level of key field consistency based on historical statistical data. When... When the value exceeds the interval [0,1], it can be restricted to a predetermined range by truncation or normalization.

[0070] Based on the results of the logic-state alignment check above, the system first checks the executability confidence of each candidate segment before performing the final sorting. .for Even the candidate segments, even those with a comprehensive semantic score At worst, the system will forcibly mark it as a "logical conflict" state and remove it from the preferred queue, or keep it only as "negative warning" material.

[0071] Subsequently, the system is based on The candidate segments were re-sorted, and the segment with the highest overall score after the adjustment was recorded as... Its corresponding adjusted overall score is This is the target segment that is given priority consideration in subsequent output control. When the difference between the highest and second-highest comprehensive scores is not less than the second threshold, the system considers the current optimal candidate segment to have a sufficient advantage over other candidates and can directly select it. As the target knowledge fragment, no consistency adjustment is required; in this case, the target fragment is denoted as... The overall score is .

[0072] After selecting the final target knowledge fragment or triggering a conflict blocking mechanism, the output generation strategy based on the target knowledge fragment includes: when the business scenario label indicates an on-site maintenance scenario, parsing the target knowledge fragment into step-by-step operation instructions and outputting them step by step via voice; when the business scenario label indicates an electricity customer service scenario, filling the target knowledge fragment into a preset script template and then outputting it; when an output status conflict prompt is received, actively displaying the comparison difference between the real-time operating status data and the preconditions in the candidate knowledge fragment.

[0073] Specifically: in determining the target knowledge fragment or Then, the system uses the user problem business scenario tags determined in step S1. The question-answering robot is controlled using different output strategies, specifically: First, when When the target knowledge fragment belongs to a predefined set of on-site maintenance scenarios, the system parses the text content of the target knowledge fragment according to the operation steps, identifies the operation actions, sequential relationships, and safety prompts, and breaks it down into a sequence of step-by-step operation instructions ordered by execution order. The system controls the Q&A robot's output by step-by-step voice announcements during output generation. After each operation instruction is output, the system listens to the voice feedback from the personnel on site through the voice recognition module of a smart wearable device or mobile terminal, and identifies confirmation instructions or status feedback instructions. Only when a voice instruction indicating that the current step has been completed or the on-site status meets the conditions for proceeding to the next step is the Q&A robot allowed to announce the next operation content. Otherwise, the current step is repeated or the user is prompted for confirmation, preventing on-site personnel from skipping operation steps without completing the necessary preparations or confirming safety measures.

[0074] Second, when When the scenario falls within a predefined set of electricity customer service scenarios, the system uses a template-filling method to output the response. Key information extracted from the target knowledge fragment is mapped to parameter positions in a pre-defined dialogue template. For example, fields such as electricity price name, effective time, electricity address, outage area, and outage time period are filled into the corresponding dialogue slots, generating a standardized and consistent customer service response text. Subsequently, the text generation or polishing module can be called as needed to moderately optimize some natural language expressions, but the values ​​and semantics of key fields are not modified. The response text generated by template filling is denoted as... Depending on the channel, the information can be presented as text on the customer service interface for human relay, or converted into speech by a speech synthesis module and output to the user. To ensure the complete output of key information in high-risk scenarios, important terms and restrictions are explicitly marked in the template, and a more prominent tone or repetition strategy is used during voice broadcast to ensure that users are fully aware of important information. For non-on-site maintenance and other scenarios outside of power customer service, such as marketing activity recommendations and general inquiries, generative language models can be appropriately used to extend the answers based on the template when the risk level is not high and the overall scoring advantage is obvious. However, it is still required that fields such as electricity price, electricity volume, and time in the generated content must be based on structured data, and the model is not allowed to make arbitrary assumptions.

[0075] Third, when a status conflict is detected (i.e., when the blockage is triggered in step S4 due to a violation of the five-prevention interlocking logic), the system does not directly respond with the main text of the knowledge fragment. Instead, it proactively displays a comparison between the referenced real-time operating status data and the preconditions in the candidate knowledge fragment, either on the customer service interface or through voice broadcast. For example, the system will explicitly output comparative information such as "The knowledge fragment requires the operation precondition to be 'the bus tie switch is in the open position,' but the current real-time topology shows 'the bus tie switch is in the closed position.'" This intuitive display of differences allows users (especially on-site operators) to clearly understand the fundamental physical reason for the operation being blocked, avoiding blind retries or business-related doubts.

[0076] This step introduces a dual-threshold comparison mechanism based on the comprehensive score. On the one hand, when the difference between the highest and second-highest comprehensive scores is small and the risk level is high, free generation is prohibited, and a preset template or manual service is mandatory. This avoids misanswers due to model uncertainty in security and compliance-sensitive scenarios. On the other hand, in the medium confidence interval, the global consistency ratio of key fields in different proportions of knowledge sources is calculated. The consistency index is used to weight and adjust the comprehensive score of candidate segments, and the cross-verification results of key business fields among multiple knowledge sources are explicitly incorporated into the decision-making process. This prioritizes candidate segments with high consistency of key fields across multiple knowledge sources as the basis for the final answer. In addition, this embodiment distinguishes different application scenarios such as on-site maintenance and power customer service based on business scenario tags. In the on-site maintenance scenario, a step-by-step operation command and voice confirmation linkage output strategy is adopted. In the customer service scenario, a template filling and key information highlighting output strategy is adopted, enabling the question-answering robot to adopt the most suitable interaction method and safety control method in different scenarios.

[0077] Example 2, Figure 5 A control device for a power question-answering robot based on multi-source knowledge decision-making is presented, including: The problem acquisition module is used to acquire user problems and determine the business scenario label and risk level label of the user problems based on the power business intent recognition model; The knowledge extraction module is used to retrieve multiple candidate knowledge fragments for the user's question from a multi-source knowledge base based on the business scenario tags. The comprehensive scoring module is used to construct a decision feature vector for each candidate knowledge fragment, input it into a preset multi-source knowledge decision model, and obtain a comprehensive score for each candidate knowledge fragment. The strategy output module is used to control the question-answering robot to output a strategy based on the comprehensive score and the business scenario tags of the user's question.

[0078] Example 3, Power question-answering robot control equipment based on multi-source knowledge decision-making, such as Figure 6 As shown, it includes a memory and a processor: the memory is used to store a program; the processor is used to execute the program to implement any of the embodiments in Example 1.

[0079] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0080] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0081] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0082] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0083] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0084] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A control method for a power question-answering robot based on multi-source knowledge decision-making, characterized in that, Includes the following steps: The business scenario label and risk level label of user issues are determined by the power business intent recognition model; Based on business scenario tags, multiple candidate knowledge fragments are obtained from a multi-source knowledge base containing a real-time status knowledge base; The decision feature vectors of each candidate knowledge fragment are extracted and a comprehensive score is obtained through a multi-source knowledge decision model. Based on the real-time running status data of the real-time status knowledge base, logic-state alignment verification is performed on candidate knowledge fragments with a comprehensive score higher than the baseline to generate executability confidence. The output strategy is generated based on the comprehensive score and the feasibility confidence level. If the executability confidence level is lower than a preset security threshold, the output strategy includes outputting a state conflict warning.

2. The power question-answering robot control method based on multi-source knowledge decision-making according to claim 1, characterized in that, The execution logic-state alignment check generates executability confidence, including: Semantic dependency analysis is performed on candidate knowledge fragments to extract operation logic triples containing preconditions and action instructions; Based on the preconditions, obtain the corresponding real-time running status data from the real-time status knowledge base; The real-time running status data is substituted into the operation logic triplet for logical operation to obtain the executability confidence score.

3. The power question-answering robot control method based on multi-source knowledge decision-making according to claim 2, characterized in that, The step of substituting real-time operating status data into the operation logic triplet for logical operations includes: Identify the device interlock constraints and energized state constraints in the operation logic triplet; Acquire the current switch position signal, telemetry value, and associated timestamp of the target device corresponding to the action command from the real-time operating status data; Construct a local power grid topology subgraph of the area where the target device belongs at the current moment; In the local power grid topology subgraph, if the simulation result triggers the preset five-prevention blocking logic, the executability confidence level is determined to be lower than the preset safety threshold.

4. The power question-answering robot control method based on multi-source knowledge decision-making according to claim 1, characterized in that, The output strategy generated based on the comprehensive score and executability confidence includes: When the executability confidence level reaches the preset safety threshold, calculate the difference between the highest and second-highest comprehensive scores among multiple candidate knowledge fragments; When the difference is less than the first threshold and the risk level of the user's question is a preset high-risk level, an output strategy is generated based on a preset template script to answer or transfer to human service. When the difference is between the first threshold and the second threshold, the comprehensive score of the candidate knowledge fragment is adjusted based on the consistency of the values ​​of key business fields in different knowledge sources. Otherwise, the candidate knowledge fragment with the highest comprehensive score is selected as the target knowledge fragment, and an output strategy is generated based on the target knowledge fragment.

5. The power question-answering robot control method based on multi-source knowledge decision-making according to claim 4, characterized in that, The output generation strategy based on target knowledge fragments includes: When the business scenario tag indicates an on-site maintenance scenario, the target knowledge fragment is parsed into step-by-step operation instructions and output step by step via voice. When the business scenario tag indicates an electricity customer service scenario, the target knowledge fragment is filled into the preset script template and then output; When a status conflict is detected, the system actively displays a comparison between the real-time running status data and the preconditions in the candidate knowledge fragments.

6. The power question-answering robot control method based on multi-source knowledge decision-making according to claim 1, characterized in that, The decision feature vector includes at least the following features: Semantic similarity features between user questions and candidate knowledge fragments; Matching characteristics between the business scenario tags of user questions and the preset applicable scenario tags of candidate knowledge fragments; The compatibility feature between the risk level label and the preset risk level label of the candidate knowledge fragment.

7. The power question-answering robot control method based on multi-source knowledge decision-making according to claim 1, characterized in that, The power business intent recognition model is a deep learning text classification model based on a self-attention mechanism. The power business intent recognition model is trained using historical power customer service dialogue data that includes intent labels and risk level labels.

8. The power question-answering robot control method based on multi-source knowledge decision-making according to claim 1, characterized in that, The real-time status knowledge base consists of structured data that changes dynamically over time. The structured data includes at least power grid operation status data, planned maintenance data, smart meter data, time-of-use electricity price data, and environmental status data. Each piece of structured data has a corresponding timestamp and geographic location identifier.

9. The power question-answering robot control method based on multi-source knowledge decision-making according to claim 4, characterized in that, The adjustment of the comprehensive score of the corresponding candidate knowledge fragments based on the consistency of values ​​of key business fields in different knowledge sources includes: Extract key business fields from candidate knowledge fragments, including electricity price values, time ranges, or equipment identifiers; The consistency ratio of key business fields across their respective knowledge sources is statistically analyzed. The overall score of the candidate knowledge fragments is adjusted by using the consistency ratio as a weighting coefficient.

10. A power question-answering robot control device based on multi-source knowledge decision-making, characterized in that, Including memory and processor: The memory is used to store programs; The processor is used to execute the program to implement the power question-answering robot control method based on multi-source knowledge decision-making as described in any one of claims 1-9.

Images