Substation fault interactive reasoning and auxiliary decision method and system based on natural language processing, electronic equipment and storage medium

Abstract

The application provides a power substation fault interactive reasoning and auxiliary decision method and system based on natural language processing, an electronic device and a storage medium, and relates to the technical field of power substation fault diagnosis. The method comprises the following steps: collecting a transformer voiceprint signal and an operation and maintenance personnel voice description text, converting the voiceprint signal into a voiceprint atlas and extracting features, and combining vocabulary information to form a joint feature vector; using a dynamic Bayesian network to perform probability reasoning with a state transition probability matrix as prior knowledge to generate a fault hypothesis set with a confidence rating; when the confidence is insufficient, an inquiry strategy is generated through a cross-modal attention mechanism to guide the operation and maintenance personnel to supplement fault detail descriptions, on the basis of which the state transition probability matrix is updated and reasoning is performed again, and finally a reliable diagnosis decision result is generated. The application improves the accuracy and interactive decision efficiency of power substation fault diagnosis.

Description

Technical Field

[0001] This application relates to the field of substation fault diagnosis technology, and in particular to a substation fault interactive reasoning and auxiliary decision-making method, system, electronic device and storage medium based on natural language processing. Background Technology

[0002] In the field of intelligent operation and maintenance of substation equipment, with the continuous expansion of the power system, higher requirements are placed on the accuracy and real-time performance of fault diagnosis for key equipment such as transformers. However, traditional diagnostic methods that rely on single sensor data are no longer able to cope with the complex and ever-changing fault scenarios on site. Therefore, there is an urgent need for an intelligent decision-making solution that can effectively integrate multi-source heterogeneous data such as voiceprints, video, and gas concentration, and support voice interaction and collaboration among on-site personnel, so as to achieve accurate identification and rapid handling of typical mechanical and discharge faults of transformers.

[0003] Existing solutions mainly employ a multimodal fault diagnosis model based on deep neural networks. This solution simultaneously acquires vibration signals and infrared thermal images of transformers, extracts spatial features from the images using convolutional neural networks, and processes the temporal information of the vibration signals using long short-term memory networks. Finally, it outputs the classification results of the fault categories through a fully connected layer. This solution establishes an end-to-end diagnostic process, thereby achieving automatic identification of some typical fault modes.

[0004] However, this approach still has several limitations in practical applications: First, model training heavily relies on a large amount of well-labeled data, while the scarcity of real-world fault samples results in limited generalization ability of the trained model when facing new scenarios; second, its diagnostic process lacks an effective human-computer interaction mechanism, failing to proactively obtain crucial information from on-site personnel when the model's confidence level is low; third, static model parameters are ill-suited to the dynamic evolution of transformer fault states over time, thus exhibiting insufficient sensitivity to progressive fault diagnosis. These factors collectively affect the practicality and final diagnostic reliability of this method in complex and ever-changing field environments. Summary of the Invention

[0005] This application provides a method, system, electronic device, and storage medium for interactive reasoning and auxiliary decision-making in substation faults based on natural language processing, in order to solve the problems of poor practicality and low reliability of final diagnosis in the complex and ever-changing field environment of existing technologies.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a substation fault interactive reasoning and auxiliary decision-making method based on natural language processing, comprising:

[0007] Collect acoustic signature signals and corresponding fault phenomenon texts of transformers in substations during operation;

[0008] The voiceprint signal is converted into a voiceprint map, and a joint feature vector is formed based on the voiceprint features in the voiceprint map and the vocabulary information in the fault phenomenon text.

[0009] Using a dynamic Bayesian network and a preset state transition probability matrix as prior knowledge, probabilistic reasoning is performed on the joint feature vector to generate a set of fault hypotheses. Each fault hypothesis in the set of fault hypotheses has a corresponding confidence evaluation value.

[0010] When the confidence evaluation value is less than the preset decision threshold, a query strategy is generated through a cross-modal attention mechanism;

[0011] Based on the inquiry strategy, maintenance personnel are guided to supplement the voice description information of the transformer fault. Based on the voice description information, the parameter settings of the state transition probability matrix are updated. Based on the updated state transition probability matrix, the probabilistic reasoning process is re-executed to generate diagnostic decision results.

[0012] Optionally, the generation of the query strategy through the cross-modal attention mechanism includes:

[0013] The correlation strength matrix between the lexical information and the voiceprint features is analyzed using a cross-modal attention mechanism.

[0014] The feature dimension with a first association strength lower than a first preset quantization threshold is identified from the association strength matrix, where the first association strength is the association strength between the vocabulary information and the voiceprint feature;

[0015] When there are multiple feature dimensions, a second association strength is calculated, which is the association strength between every two feature dimensions among the multiple feature dimensions.

[0016] Based on the second association strength, the types of fault information that need to be supplemented and the query priority are determined, and a query strategy is generated according to the fault information types and the query priority.

[0017] Optionally, the step of determining the type of fault information to be supplemented and the query priority based on the second association strength, and generating a query strategy according to the fault information type and the query priority, includes:

[0018] Based on the second association strength, an association matrix is ​​generated using a graph neural network.

[0019] Extract feature dimension pairs from the correlation matrix whose second correlation strength is less than a second preset quantization threshold.

[0020] Based on the second correlation strength of the feature dimension pairs, determine the degree of information loss and the type of fault information that needs to be supplemented;

[0021] Arrange the fault information types in descending order of the degree of information missing, and assign a query priority to each fault information type;

[0022] Based on the inquiry priority, an inquiry sequence is constructed, wherein each inquiry round in the inquiry sequence corresponds to a fault information type of inquiry priority;

[0023] For each round of queries, the content of the query statement is generated using a natural language generation model based on the type of fault information.

[0024] Multiple rounds of queries and their corresponding query statements are integrated into an inquiry strategy.

[0025] Optionally, the step of generating query statement content using a natural language generation model for each query round corresponding to the fault information type includes:

[0026] The fault information type corresponding to each round of inquiry is input into the semantic understanding module of the natural language generation model, and the core element information and contextual information to be inquired are parsed out through the semantic understanding module.

[0027] Based on the core element information and the contextual information, matching professional terms and question templates are retrieved from a preset power industry terminology database;

[0028] By combining and optimizing the specialized vocabulary and question templates using the grammatical constraints in the natural language generation model, multiple candidate questions are generated.

[0029] The evaluation module of the natural language generation model selects the candidate question with the highest sentence score from all the candidate questions as the final question.

[0030] The final question is bound to the corresponding question round to form the question statement content.

[0031] Optionally, the step of converting the voiceprint signal into a voiceprint spectrum, and forming a joint feature vector based on the voiceprint features in the voiceprint spectrum and the lexical information in the fault phenomenon text, includes:

[0032] The voiceprint signal is subjected to sound source separation processing to distinguish the signal components corresponding to the ambient sound source and the device sound source respectively;

[0033] Suppress the signal components corresponding to the ambient sound source, and retain the signal components corresponding to the device sound source;

[0034] Based on the signal components corresponding to the sound source of the device, a sound signature map is constructed;

[0035] Extract the acoustic signature features that reflect the temporal variation of mechanical vibration from the acoustic signature spectrum;

[0036] Identify vocabulary describing the fault phenomenon from the text describing the fault phenomenon;

[0037] The voiceprint features and vocabulary information are fused using multimodal features to form a joint feature vector.

[0038] Optionally, the step of using a dynamic Bayesian network, with a preset state transition probability matrix as prior knowledge, to perform probabilistic reasoning on the joint feature vector to generate a fault hypothesis set, wherein each fault hypothesis in the fault hypothesis set has a corresponding confidence evaluation value, including:

[0039] The joint feature vector is input into the inference model of the dynamic Bayesian network. In the inference model, based on the state transition relationship defined by the state transition probability matrix, the joint feature vector is subjected to probability evolution calculation to obtain the probability distribution result of the fault state.

[0040] Based on the probability distribution results, fault hypotheses are determined, and a reliability evaluation value is assigned to each fault hypothesis based on the numerical value of the probability distribution results.

[0041] All failure hypotheses assigned confidence scores are integrated into a failure hypothesis set.

[0042] Optionally, the step of guiding maintenance personnel to supplement voice description information of transformer faults based on the inquiry strategy, updating the parameter settings of the state transition probability matrix based on the voice description information, and re-executing the probabilistic reasoning process based on the updated state transition probability matrix to generate diagnostic decision results includes:

[0043] In accordance with the order of multiple query rounds in the query strategy, corresponding supplementary requests are sequentially sent to the maintenance personnel;

[0044] Based on the supplementary request, receive the voice description information provided by the maintenance personnel for each round of inquiries;

[0045] The newly added fault features in the voice description information are analyzed, and supplementary feature data is generated based on the newly added fault features;

[0046] Based on the supplementary feature data, adjust the transition probability parameters of the corresponding states in the state transition probability matrix;

[0047] When the confidence evaluation value of all failure hypotheses is less than the preset decision threshold, the probabilistic reasoning process is re-executed based on the adjusted state transition probability matrix, and the confidence evaluation value of each failure hypothesis in the failure hypothesis set is updated by iterative calculation.

[0048] The iteration process terminates when the updated confidence value of at least one fault hypothesis is greater than or equal to a preset decision threshold, and a diagnostic decision result is generated.

[0049] Secondly, this application provides a substation fault interactive reasoning and auxiliary decision-making system based on natural language processing, including:

[0050] The acquisition module is used to collect the acoustic signature signals and corresponding fault phenomenon text of transformers in the substation during operation;

[0051] The conversion module is used to convert the voiceprint signal into a voiceprint spectrum and form a joint feature vector based on the voiceprint features in the voiceprint spectrum and the vocabulary information in the fault phenomenon text.

[0052] The inference module is used to perform probabilistic inference on the joint feature vector using a dynamic Bayesian network and a preset state transition probability matrix as prior knowledge, to generate a set of fault hypotheses, in which each fault hypothesis has a corresponding confidence evaluation value.

[0053] The generation module is used to generate a query strategy through a cross-modal attention mechanism when the confidence evaluation value is less than a preset decision threshold.

[0054] The update module is used to guide maintenance personnel to supplement the voice description information of the transformer fault based on the query strategy, update the parameter settings of the state transition probability matrix based on the voice description information, and re-execute the probabilistic reasoning process based on the updated state transition probability matrix to generate diagnostic decision results.

[0055] Thirdly, this application provides an electronic device, comprising:

[0056] Memory, used to store computer programs;

[0057] A processor, configured to execute the computer program to implement the steps of the natural language processing-based interactive reasoning and decision support method for substation faults as described in the first aspect above.

[0058] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the substation fault interactive reasoning and auxiliary decision-making method based on natural language processing as described in the first aspect above.

[0059] The technical solution provided in this application has the following beneficial effects: First, it realizes the synchronous acquisition of multimodal data, thereby providing a data foundation for subsequent fusion analysis; it completes feature-level fusion of multi-source heterogeneous data to improve the expressive power of fault features; on this basis, it realizes the dynamic inference of fault states by utilizing dynamic Bayesian networks and provides interpretable diagnostic results; at the same time, it establishes an intelligent inquiry mechanism through cross-modal attention mechanism to ensure that key information can be actively obtained when there is uncertainty in diagnosis; in addition, it realizes knowledge update through human-machine collaboration, thereby improving the adaptability of the model; finally, it forms a closed-loop optimization mechanism to improve the reliability of diagnostic results.

[0060] Furthermore, the method employed in this application analyzes the correlation strength matrix between text features and voiceprint features through a cross-modal attention mechanism, identifies feature dimensions with correlation strength below a first preset quantization threshold, and calculates the interaction strength between these weakly correlated dimensions to determine the types of information that need to be supplemented from on-site personnel during the diagnostic process and their inquiry priorities, ultimately generating a clearly targeted interactive inquiry strategy. Therefore, this scheme achieves intelligent inquiry strategy generation based on feature correlation analysis, accurately locates missing information links, and improves human-machine collaboration efficiency and diagnostic completeness by systematically collecting key diagnostic information.

[0061] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 A flowchart illustrating an interactive reasoning and decision support method for substation faults based on natural language processing, provided in this application embodiment;

[0064] Figure 2 This application provides a schematic diagram illustrating a specific implementation of an interactive reasoning and auxiliary decision-making method for substation faults based on natural language processing.

[0065] Figure 3 This is a schematic diagram illustrating another specific implementation of an interactive reasoning and auxiliary decision-making method for substation faults based on natural language processing, provided in this application.

[0066] Figure 4This is a schematic diagram of the structure of a substation fault interactive reasoning and auxiliary decision-making system based on natural language processing, provided in an embodiment of this application. Detailed Implementation

[0067] To address the problems of existing technologies, this application proposes an interactive reasoning and decision support method for substation faults based on natural language processing. The core of this method lies in: firstly, forming a comprehensive feature representation through joint analysis of voiceprint and text features; secondly, simulating the fault evolution process using a dynamic Bayesian network; and thirdly, intelligently generating question-and-answer strategies based on correlation strength analysis, thereby guiding maintenance personnel to supplement key information and dynamically update inference parameters. Therefore, this method overcomes the excessive reliance of traditional models on labeled data, effectively supplements diagnostic information through a human-computer collaborative interaction mechanism, and adapts to fault state evolution using dynamic probabilistic modeling. Ultimately, it improves the accuracy of fault diagnosis and system adaptability in complex scenarios, fundamentally solving the problems of insufficient generalization ability, lack of interaction mechanisms, and limited dynamic adaptability in existing technologies.

[0068] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0069] The core of this application is to provide an interactive reasoning and decision support method for substation faults based on natural language processing. A flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes:

[0070] Step 101: Collect the acoustic signature signal and corresponding fault phenomenon text of the transformer in the substation during operation.

[0071] In step 101, the acoustic signature signal represents the sound signal generated when the transformer is running. The acoustic signature signal may include components such as mechanical vibration sound, discharge sound and environmental noise, and is used to characterize the operating status of the equipment.

[0072] The fault phenomenon text represents a written record of abnormal transformer phenomena described by maintenance personnel through voice, such as "intermittent humming" or "accompanied by popping sounds," which is used to provide semantic information about the fault. The aforementioned maintenance personnel are professional technicians responsible for the daily monitoring, maintenance, and fault handling of substation equipment. Their job responsibilities are directly related to the substation's operating status. In this solution, they serve as the initial providers of transformer fault information and participants in interactive diagnosis.

[0073] For example, in the transformer monitoring scenario of substation A, acoustic sensors first collect acoustic signals containing a fundamental humming sound and a high-frequency crackling sound. At the same time, on-site maintenance personnel use handheld devices to describe the situation in voice, stating that "the transformer has an intermittent humming sound accompanied by a slight crackling sound." This voice description is then converted into corresponding text information by the voice recognition module in the system, thus completing the initial collection of multimodal data.

[0074] Step 102: Convert the voiceprint signal into a voiceprint spectrum, and form a joint feature vector based on the voiceprint features in the voiceprint spectrum and the vocabulary information in the fault phenomenon text.

[0075] In step 102, the voiceprint spectrum represents a two-dimensional spectrum with frequency-time as the coordinate axis and sound intensity as the color depth, which intuitively displays the frequency distribution and temporal changes of the sound signal; the voiceprint features represent the feature parameters extracted from the voiceprint spectrum that reflect the changes of sound over time, such as frequency fluctuations and amplitude changes.

[0076] Lexical information represents key descriptive words extracted from the text of fault phenomena, such as "intermittent" and "explosive sound"; joint feature vector represents a comprehensive feature representation formed by concatenating voiceprint features and text features, which is used for multimodal information fusion.

[0077] In this embodiment, the voiceprint signal undergoes source separation processing to suppress environmental noise and retain the device's sound source components. Then, a voiceprint spectrum is generated based on the sound wave propagation characteristics. Voiceprint features are extracted from this spectrum, and lexical information is parsed from the text. Finally, these two types of features are fused into a joint feature vector through concatenation. The sound wave propagation characteristics refer to the frequency attenuation and time delay characteristics exhibited by sound propagating inside the device.

[0078] For example, the sound source of the acoustic signal from substation A is separated, retaining the 72 dB 100 Hz humming sound and the 68 dB 2000 Hz crackling sound, and generating an acoustic spectrum. The amplitude fluctuation features of the 100 Hz signal and the burst interval features of the 2000 Hz signal are extracted from the spectrum as acoustic features. At the same time, "intermittent" and "crackling sound" are extracted from the text as core words. Finally, the 100-dimensional acoustic features and the 50-dimensional text features are concatenated to form a 150-dimensional joint feature vector.

[0079] Step 103: Using a dynamic Bayesian network, with a preset state transition probability matrix as prior knowledge, perform probabilistic reasoning on the joint feature vector to generate a set of fault hypotheses. Each fault hypothesis in the set of fault hypotheses has a corresponding confidence evaluation value.

[0080] In step 103, the state transition probability matrix represents a probability matrix describing the transition relationship between transformer fault states, where rows and columns represent fault states and element values ​​represent the transition probability between states; the fault hypothesis set contains a set of possible fault types and their confidence evaluation values.

[0081] It should be noted that this embodiment does not impose specific limitations on the structure and parameter design of the dynamic Bayesian network, and can be set accordingly based on the actual situation.

[0082] In this embodiment, the joint feature vector is first input into the dynamic Bayesian network, and a preset state transition probability matrix is ​​loaded as the model prior knowledge; then, the posterior probability distribution of the fault state at the current moment is obtained through probability evolution calculation; finally, a confidence evaluation value is assigned to each potential fault state according to the calculated probability value, thereby forming a structured set of fault hypotheses.

[0083] For example, a joint feature vector of dimension 150 is input into a dynamic Bayesian network, and a preset matrix containing the transition probabilities of various potential fault states such as winding loosening and discharge faults is loaded. Then, the probability distribution of the fault states is calculated by the forward-backward algorithm. Specifically, the probability of discharge fault is 0.55 and the probability of abnormal core vibration is 0.3. Based on this, the fault hypothesis set is formed as {discharge fault: 0.55, abnormal core vibration: 0.3}.

[0084] Step 104: When the confidence evaluation value is less than the preset decision threshold, a query strategy is generated through a cross-modal attention mechanism.

[0085] In step 104, the inquiry strategy refers to the interactive scheme automatically generated by the system based on the current insufficient diagnostic confidence, which includes multiple rounds of inquiries. Each round is aimed at the missing information of a specific type of fault. Through orderly question-and-answer interaction, the system guides the operation and maintenance personnel to gradually supplement key diagnostic information, forming a closed-loop interactive process from system inquiry to user feedback.

[0086] It should be noted that the embodiments of this application do not impose specific limitations on the form and content of the inquiry strategy, and can be set accordingly according to the actual situation.

[0087] In this embodiment, when the highest confidence level in the fault hypothesis set is lower than the decision threshold, a cross-modal attention mechanism is activated to calculate the correlation strength between text features and voiceprint features, and to identify feature dimensions with low correlation strength. Then, based on the correlation strength value, the degree of information loss and the priority of the query are determined, and finally, an ordered query sequence containing specific query content is generated. The decision threshold can be set according to the actual situation.

[0088] For example, when the fault confidence evaluation value of the discharge fault is 0.55, which is lower than the corresponding decision threshold of 0.6, if the correlation strength between the text feature of "pop sound" and the 2000Hz voiceprint feature is calculated to be 0.42, and this correlation strength is lower than the correlation strength threshold of 0.7, then the degree of information loss is calculated based on this correlation strength. For example, if the degree of information loss is 1-0.42=0.58, the first round of query content "Please describe the frequency and duration of the pop sound" is generated accordingly.

[0089] Step 105: Based on the inquiry strategy, guide the maintenance personnel to supplement the voice description information of the transformer fault. Based on the voice description information, update the parameter settings of the state transition probability matrix. Based on the updated state transition probability matrix, re-execute the probabilistic reasoning process to generate diagnostic decision results.

[0090] In step 105, transformer fault refers to an abnormal working state that occurs during the operation of transformer equipment. This abnormal working state specifically includes typical fault types such as loose windings, abnormal core vibration, discharge faults, insulation aging, and poor contact. These fault types are jointly determined by voiceprint feature analysis and text description parsing, and are the specific objects for probabilistic reasoning by dynamic Bayesian networks.

[0091] Voice description information represents the detailed fault description provided by the user based on the inquiry strategy, such as "the popping sound occurs once every 5 minutes and lasts for 2 seconds each time"; parameter update represents the adjustment of the probability values ​​in the state transition probability matrix based on the supplementary information.

[0092] The diagnostic decision result refers to the fault type determination conclusion finally output by the system after iterative reasoning. This conclusion includes the specific fault type name and its corresponding confidence level rating, indicating the reliability of the system's fault determination result. At the same time, the result also includes recommended handling suggestions and risk level assessment, providing maintenance personnel with a complete basis for fault handling solutions.

[0093] In this embodiment, the user is guided to supplement the description in sequence according to the query strategy, and the newly added features in the description are parsed; then the parameters in the state transition probability matrix are adjusted according to the importance of the features, and the probabilistic inference is re-executed using the updated matrix; the confidence is iteratively optimized until the decision threshold is met.

[0094] For example, if a user adds a description that "the popping sound occurs every 5 minutes and lasts for 2 seconds each time", the temporal pattern in the description is analyzed, and the transfer probability of the discharge fault is adjusted from 0.6 to 0.7. After recalculation, the confidence level of the discharge fault increases to 0.65, reaching the decision threshold of 0.6, and the final output diagnosis result is discharge fault.

[0095] This method achieves accurate diagnosis of substation faults through multimodal data fusion and dynamic probabilistic reasoning; at the same time, it actively acquires key information through a human-computer interaction mechanism, thereby solving the problem of traditional methods relying on a large amount of labeled data; and it uses dynamic parameter updates to adapt to the fault evolution process, ultimately improving the diagnostic accuracy and system adaptability in complex scenarios.

[0096] To address the challenge of intelligently generating inquiry strategies when diagnostic confidence is insufficient, some embodiments include step 104: generating an inquiry strategy through a cross-modal attention mechanism, such as... Figure 2 As shown, it includes:

[0097] Step 201: Analyze the correlation strength matrix between the vocabulary information and the voiceprint features through a cross-modal attention mechanism.

[0098] In step 201, the association strength matrix represents the quantification result of the degree of association between text words and voiceprint features. It is calculated through attention weights and is used to identify the matching relationship between the two types of features.

[0099] It should be noted that the specific implementation process of the cross-modal attention mechanism can be found in relevant technologies, and will not be elaborated here.

[0100] In this embodiment of the application, the correlation weight between each text word and the voiceprint feature is calculated through an attention mechanism, and a weight distribution matrix is ​​formed, in which the level of each weight value intuitively reflects the correlation strength between the word and the corresponding voiceprint feature.

[0101] Step 202: Identify feature dimensions from the association strength matrix whose first association strength is lower than a first preset quantization threshold, where the first association strength is the association strength between the vocabulary information and the voiceprint feature.

[0102] In step 202, the feature dimension difference can refer to the part where the correlation strength between text words and voiceprint features is low on a certain dimension. These dimensions represent links where there may be missing information or mismatch in description.

[0103] Step 203: When there are multiple feature dimensions, calculate the second association strength, which is the association strength between every two feature dimensions in the multiple feature dimensions.

[0104] In step 203, the correlation strength is a numerical representation of the degree of correlation between feature dimensions, obtained by calculating the correlation between each pair of feature dimensions, and is used to quantify the degree of information loss.

[0105] In this embodiment, pairwise correlation calculations are performed on feature dimensions with correlation strength lower than a first preset quantization threshold to obtain the correlation strength value between each pair of dimensions. The lower the correlation strength value, the more severe the information loss.

[0106] Step 204: Based on the second association strength, determine the type of fault information that needs to be supplemented and the query priority, and generate a query strategy according to the fault information type and the query priority.

[0107] In step 204, the fault information type is the category of information that needs to be supplemented, determined based on the correlation strength; the query priority is sorted from high to low according to the degree of information missing.

[0108] In this embodiment, information types are sorted according to the correlation strength reflected in the weight distribution matrix to determine the types of information that need to be supplemented to on-site personnel during the diagnosis process and their corresponding priorities. Specific inquiry content is generated for each type of information, and finally, these inquiry contents are combined into a structured interactive inquiry strategy according to priority.

[0109] Here is a specific example:

[0110] In the transformer monitoring scenario of substation A, after detecting a discharge fault with a confidence level of 0.55, which is lower than the decision threshold of 0.6, a cross-modal attention mechanism is initiated to analyze the correlation strength matrix between lexical information and voiceprint features. By calculating the attention weights of text words and voiceprint features, it is found that the correlation strength between the word "explosive sound" and the 2000Hz voiceprint feature is 0.42, while the correlation strength between "intermittent" and the 100Hz voiceprint feature is 0.78. From the correlation strength matrix, it is identified that the correlation strength of "explosive sound" and the 2000Hz feature dimension is 0.42, which is lower than the first preset quantization threshold of 0.7.

[0111] Then, the correlation strength between every two feature dimensions is calculated, where the correlation strength S between the "pop" sound and the 2000Hz feature dimension is calculated using the following formula: ,in This indicates the correlation strength between the textual term "explosive sound" and the 2000Hz voiceprint feature dimension. This represents the word weights in the text; for example, a value of 0.42. This represents the voiceprint feature weight; for example, a value of 0.85 is used to calculate... =1 0.42 0.85|=0.57;

[0112] Next, based on the second correlation strength, the fault information types that need to be supplemented are determined to be pop sound time features and intensity features, where the information missing degree of pop sound time features is 0.43 and the information missing degree of intensity features is 0.3; according to the information missing degree from high to low, the pop sound time feature supplementation request is listed as the highest priority and the intensity feature is listed as the second priority.

[0113] Subsequently, based on the fault information type and query priority, the first round of queries targeting the time characteristics of the popping sound is generated: "Please describe in detail the duration and interval pattern of each popping sound." When the user adds the description "Each popping sound lasts for about 2 seconds and occurs once every 5 minutes," the second round of queries targeting the intensity characteristics is generated: "Please explain whether the loudness of the popping sound changes over time." Finally, a query strategy containing two rounds of queries is formed to guide the user to systematically supplement the key information required for diagnosis.

[0114] In this embodiment of the application, by systematically analyzing the differences in feature associations, the missing information links can be accurately located, thereby generating targeted inquiry strategies, effectively guiding users to supplement key information, and ultimately improving the completeness of diagnosis and the efficiency of human-machine collaboration.

[0115] To further improve the accuracy and systematic nature of query strategy generation, in some embodiments, step 204 involves: based on the second association strength, determining the type of fault information that needs to be supplemented and the query priority; and generating a query strategy according to the fault information type and the query priority, such as... Figure 3 As shown, it includes:

[0116] Step 301: Based on the second association strength, construct the association matrix using a graph neural network.

[0117] In step 301, the association matrix is ​​a two-dimensional matrix constructed by a graph neural network, where rows and columns represent different feature dimensions, and the matrix element values ​​represent the association strength between corresponding feature dimensions, which is used to comprehensively characterize the association relationship between features.

[0118] It should be noted that the embodiments of this application do not limit the specific structure and parameter design of the graph neural network, and can be set accordingly according to the actual situation.

[0119] In this embodiment, each feature dimension is used as a node in a graph neural network, and the correlation strength is used as the edge weight. A correlation matrix is ​​constructed through graph convolution operations to capture complex correlation patterns between feature dimensions.

[0120] Step 302: Extract feature dimension pairs from the correlation matrix whose second correlation strength is less than the second preset quantization threshold.

[0121] In step 302, the preset quantization threshold is a pre-set critical value for association strength, used to filter feature dimension pairs with insufficient association; a feature dimension pair refers to the combination of two feature dimensions corresponding to rows and columns in the association matrix.

[0122] In this embodiment of the application, all elements in the correlation matrix are traversed, and feature dimension pairs with element values ​​less than a preset quantization threshold are extracted. These feature dimension pairs represent feature combinations with missing information.

[0123] Step 303: Determine the degree of information loss and the type of fault information that needs to be supplemented based on the second correlation strength of the feature dimension pair.

[0124] In step 303, the degree of information missing is an index of the severity of missing information calculated based on the correlation strength of the feature dimension pairs.

[0125] In this embodiment of the application, for each selected weakly correlated feature dimension pair, the actual calculated correlation strength is first subtracted from the ideal correlation value of 1.0 to obtain the basic information missing value. At the same time, the specific fault information type that needs to be supplemented is determined according to the actual physical meaning of the feature dimension in fault diagnosis.

[0126] The specific implementation process includes calculating the difference between the correlation strength of each feature dimension pair and the ideal value of 1.0 as its basic missing degree, and then performing a weighted calculation based on the importance weight of the feature dimension in the overall fault diagnosis to finally obtain the comprehensive information missing degree. At the same time, based on the physical meaning of the feature dimension, the specific fault information category that needs to be asked of the on-site personnel is determined.

[0127] Specific example: Assuming the calculated correlation strength between the burst sound feature dimension and the corresponding text description is 0.42, while the ideal correlation value is 1.0, the basic missing degree is 0.58; if the importance weight of the burst sound feature in discharge fault diagnosis is set to 0.8, the comprehensive information missing degree is 0.58 × 0.8 = 0.464; at the same time, based on the physical meaning of the burst sound feature in fault analysis, the type of fault information that needs to be supplemented is determined to be discharge feature information. This information category specifically includes detailed information such as discharge frequency, discharge intensity, and other observable phenomena accompanying the discharge.

[0128] Step 304: Arrange the fault information types in descending order of the degree of information missing, and assign an inquiry priority to each fault information type.

[0129] Step 305: Based on the query priority, construct a query sequence, wherein each query round in the query sequence corresponds to a fault information type of query priority.

[0130] In step 305, the query sequence is an arrangement of query rounds organized according to priority, with each query round corresponding to a fault information type of priority.

[0131] In this embodiment, the query rounds are arranged sequentially according to a priority list to ensure that the fault information type with the highest degree of information loss is processed first in the first round of queries.

[0132] Step 306: For each round of queries, generate query statements using a natural language generation model based on the fault information type.

[0133] In step 306, the content of the query statement is a specific query expression generated by a natural language generation model, generating query sentences that conform to everyday language habits for each fault information type.

[0134] It should be noted that the structure, parameter design, and training process of the natural language generation model are not the focus of this application. You can refer to relevant technologies, and they will not be elaborated here.

[0135] In this embodiment of the application, each fault information type is input into a natural language generation model, which generates easy-to-understand query statements based on the characteristics of the information type, ensuring that users can accurately understand the query intent.

[0136] Step 307: Integrate multiple rounds of inquiries and their corresponding question statements into an inquiry strategy.

[0137] In this embodiment of the application, each round of inquiry and its corresponding inquiry statement content are integrated in order of priority to form a systematic inquiry strategy document.

[0138] Here is a specific example:

[0139] In the transformer monitoring scenario of substation A, based on the obtained feature dimension correlation strength, where the correlation strength r1 of the popping sound time dimension is 0.57 and the correlation strength r2 of the intensity dimension is 0.7, a correlation matrix is ​​constructed through a graph neural network. This matrix contains two nodes in the time dimension and the intensity dimension, with edge weights of 0.57 and 0.7, respectively. A preset quantization threshold θ is set to 0.65, and feature dimension pairs with correlation strength values ​​less than θ are extracted from the matrix.

[0140] Then, based on the correlation strength values ​​of the feature dimension pairs, the information missing degree d is calculated using the formula d = 1 - r, where r is the correlation strength value. Therefore, the information missing degree of the time dimension d1 = 1 - 0.57 = 0.43, and the information missing degree of the intensity dimension d2 = 1 - 0.7 = 0.3. Sorting the information missing degree from high to low, the information missing degree d1 of the time dimension, at 0.43, has the highest priority, while the information missing degree d2 of the intensity dimension, at 0.3, has a lower priority.

[0141] Next, a query sequence is constructed based on this priority order. The first round of queries corresponds to the highest priority in the time dimension, and the second round of queries corresponds to the secondary priority in the intensity dimension. For the first round of queries, the query statement "Please describe in detail the specific duration of each popping sound and the pattern of the interval between occurrences" is generated using a natural language generation model. For the second round of queries, the query statement "Please explain whether the loudness of the popping sound changes significantly in different time periods" is generated.

[0142] Finally, the two rounds of questioning and the corresponding question statements are integrated into a complete questioning strategy. First, the first round of questioning is executed to obtain the popping sound time characteristic information, and then the second round of questioning is initiated based on the feedback results.

[0143] In this embodiment of the application, the efficiency of human-computer interaction and the completeness of diagnostic information are effectively improved through systematic feature association analysis and priority ranking.

[0144] To further improve the accuracy and professionalism of the generated query statements, in some embodiments, step 306: for each query round corresponding to the fault information type, generate query statement content using a natural language generation model, including:

[0145] Step 401: Input the fault information type corresponding to each round of inquiry into the semantic understanding module of the natural language generation model, and use the semantic understanding module to parse out the core element information and contextual information that need to be queried.

[0146] In step 401, core element information refers to the key content elements that need to be queried in the fault information type, and contextual information refers to the background environment and constraints in the current diagnostic scenario.

[0147] It should be noted that the structure, parameter design, and training process of the semantic understanding module are not the focus of this application. You can refer to relevant technologies, and they will not be elaborated here.

[0148] In this embodiment, the semantic understanding module analyzes the deeper meaning of the fault information type, extracts the key points of the specific content to be inquired and the relevant environmental background information, and uses this as the semantic basis for subsequent sentence generation.

[0149] Step 402: Based on the core element information and the context information, retrieve matching professional terms and question templates from the preset power industry terminology database.

[0150] In step 402, the power industry terminology database is a knowledge base containing professional terms related to power equipment faults and common question templates, used to ensure the professional accuracy of the generated statements.

[0151] It should be noted that the technology used for the query can be referenced from relevant technologies, and will not be elaborated here.

[0152] In this embodiment of the application, based on the known core element information and the current diagnostic context, appropriate terms and standard question frames are searched and matched from a preset professional terminology database, thereby providing standardized language materials for the subsequent natural language generation module.

[0153] Step 403: Combine and optimize the professional vocabulary and question templates using the grammatical constraint rules in the natural language generation model to generate multiple candidate questions.

[0154] In step 403, the grammatical constraint rules may include the requirements for the use of professional terminology in the power industry, requiring questions to use standard professional vocabulary. For example, the grammatical constraint rules include sentence structure rules, semantic integrity rules, context adaptability rules, and expression clarity rules.

[0155] Among them, the sentence structure specification requires that the question conforms to the basic subject-verb-object structure and the length is controlled within 20 characters; the semantic integrity specification requires that the question contains key information elements such as time frequency and degree; the context adaptability specification requires that the question be highly relevant to the transformer fault diagnosis scenario; and the expression clarity specification requires that the question avoid using complex clauses and ambiguous expressions; the candidate question is a preliminary combination of multiple question versions that conform to the expression specifications in the power field.

[0156] In this embodiment of the application, the retrieved professional terms and question templates are reasonably combined according to grammatical rules to generate multiple candidate question schemes that conform to the specifications.

[0157] Step 404: Using the evaluation module of the natural language generation model, select the candidate question with the highest sentence score from all the candidate questions as the final question.

[0158] In step 404, the evaluation module is a calculation module used to evaluate the fluency of the question, and the final question is the best version of the question determined after screening. This embodiment does not limit the design of the structural parameters and specific implementation process of the evaluation module, and can be set accordingly according to the actual situation.

[0159] In this embodiment of the application, all candidate questions are scored for fluency, and then the question with the highest score is selected as the final output.

[0160] Step 405: Bind the final question to the corresponding question round to form the question statement content.

[0161] In this embodiment of the application, the best question is associated and encapsulated with the corresponding question round information to form a question statement content that can be used directly.

[0162] In this embodiment, systematic semantic parsing and specialized statement generation can produce accurate, fluent query statements that conform to domain standards, thereby improving the accuracy of human-computer interaction and user experience.

[0163] To further improve the accuracy and effectiveness of multimodal feature fusion, in some embodiments, step 102 involves converting the voiceprint signal into a voiceprint spectrum, and forming a joint feature vector based on the voiceprint features in the voiceprint spectrum and the lexical information in the fault phenomenon text, including:

[0164] Step 501: Perform sound source separation processing on the voiceprint signal to distinguish the signal components corresponding to the ambient sound source and the device sound source respectively.

[0165] In step 501, ambient sound sources refer to non-equipment sound sources such as background noise, while equipment sound sources refer to vibration and discharge sound sources generated by the transformer body.

[0166] In this embodiment, a blind source separation algorithm is used to process the acquired voiceprint signal. By analyzing the statistical characteristics contained in the sound signal, the original mixed signal is decomposed into two independent signal components: the background ambient sound source and the target device sound source. It should be noted that this embodiment does not improve the implementation process of the blind source separation algorithm; its specific implementation process can be referred to related technologies.

[0167] Step 502: Suppress the signal components corresponding to the ambient sound source, and retain the signal components corresponding to the device sound source.

[0168] In this embodiment, the signal components corresponding to the ambient sound source are subjected to noise reduction and filtering, while the separated device sound source signal retains its original integrity and feature clarity.

[0169] Step 503: Construct a sound signature map based on the signal components corresponding to the sound source of the device.

[0170] In this embodiment of the application, based on the frequency distribution and temporal variation characteristics of the device sound source signal, a corresponding acoustic pattern spectrum is generated through time-frequency analysis technology, thereby intuitively displaying the energy distribution and temporal evolution characteristics of the sound signal in different frequency bands.

[0171] Step 504: Extract the acoustic features that reflect the temporal variation law of mechanical vibration from the acoustic pattern.

[0172] Step 505: Identify vocabulary information describing the fault phenomenon from the fault phenomenon text.

[0173] Step 506: Perform multimodal feature fusion of the voiceprint features and the vocabulary information to form a joint feature vector.

[0174] In this embodiment, voiceprint features and text features are fused through feature concatenation to form a joint feature vector that can simultaneously contain the acoustic characteristics of the device and the semantic information of the scene description. The specific implementation process is as follows: first, the voiceprint features are converted into numerical vectors of fixed dimensions, and the lexical information in the text is converted into corresponding semantic vectors through word embedding technology. Then, the voiceprint feature vector and the text semantic vector are directly connected in the feature dimension using feature concatenation. Finally, the feature vector obtained after concatenation is standardized to form a joint feature vector with uniform dimension and consistent scale.

[0175] A specific example is as follows: Suppose a 100-dimensional numerical vector is extracted from the voiceprint signal to represent the vibration frequency and amplitude variation features. At the same time, the text description is converted into a 50-dimensional semantic vector through word embedding technology to represent the semantic information of keywords such as "cracking sound" and "intermittency". These two vectors are directly concatenated to form a 150-dimensional joint feature vector, where the first 100 dimensions are voiceprint features and the last 50 dimensions are text features. Finally, the 150-dimensional vector is subjected to min-max normalization so that all feature values ​​are mapped to the range of 0 to 1, thus forming the final joint feature vector that can be used for subsequent fault diagnosis analysis.

[0176] In this embodiment, multimodal information is effectively fused through systematic voiceprint signal processing and text feature extraction, thereby providing a comprehensive and accurate feature representation for subsequent fault diagnosis.

[0177] To further improve the accuracy and interpretability of fault diagnosis, in some embodiments, step 103 involves using a dynamic Bayesian network with a preset state transition probability matrix as prior knowledge to perform probabilistic reasoning on the joint feature vector, generating a fault hypothesis set. Each fault hypothesis in the fault hypothesis set has a corresponding confidence evaluation value, including:

[0178] Step 601: Input the joint feature vector into the inference model of the dynamic Bayesian network. In the inference model, based on the state transition relationship defined by the state transition probability matrix, perform probability evolution calculation on the joint feature vector to obtain the probability distribution result of the fault state.

[0179] In step 601, the probability distribution result refers to the quantitative representation of the probability of each fault state.

[0180] In this embodiment, after the joint feature vector is input into the dynamic Bayesian network as observation evidence, the forward-backward algorithm is used to infer the state transition relationship defined in the state transition probability matrix, and then the posterior probability distribution of each fault state is calculated.

[0181] It should be noted that this embodiment does not limit the specific implementation process of the forward-backward algorithm, and can be set accordingly according to the actual situation.

[0182] Step 602: Based on the probability distribution results, determine the fault hypothesis, and assign a reliability evaluation value to each fault hypothesis based on the numerical value of the probability distribution results.

[0183] In step 602, the fault hypothesis refers to the fault type judgment, and the confidence level evaluation value refers to the quantitative indicator of the credibility of the judgment.

[0184] In this embodiment of the application, based on the probability values ​​of each fault state in the probability distribution results, the state with the higher probability is determined as the fault hypothesis to be verified, and the calculated probability value is directly used as the confidence evaluation value of the corresponding hypothesis.

[0185] Step 603: Integrate all fault hypotheses assigned confidence scores into a fault hypothesis set.

[0186] To further improve the accuracy and reliability of diagnostic decisions, in some embodiments, step 105 involves: based on the inquiry strategy, guiding maintenance personnel to supplement voice description information of the transformer fault; based on the voice description information, updating the parameter settings of the state transition probability matrix; and based on the updated state transition probability matrix, re-executing the probabilistic inference process to generate diagnostic decision results, including:

[0187] Step 701: In order of the multiple query rounds in the query strategy, send the corresponding supplementary requests to the maintenance personnel in sequence.

[0188] In step 701, the supplementary request is a standardized query instruction generated according to the query strategy, which is used to guide the user to provide specific types of fault information.

[0189] In this embodiment of the application, inquiries for different fault characteristics are sent to the operation and maintenance personnel in sequence according to the order and content defined in the inquiry strategy.

[0190] Step 702: Based on the supplementary request, receive the voice description information provided by the maintenance personnel for each round of inquiries.

[0191] In this embodiment of the application, the user's verbal response is obtained through a voice receiving device and converted into a detailed description in text form.

[0192] Step 703: Analyze the newly added fault features in the voice description information, and generate supplementary feature data based on the newly added fault features.

[0193] In step 703, the newly added fault features are new fault feature information extracted from the user description. The supplementary feature data is a quantitative representation of these features.

[0194] In this embodiment of the application, new feature information is identified from user descriptions using text analysis technology and converted into a data format that can be used for model updates.

[0195] Step 704: Adjust the transition probability parameters of the corresponding states in the state transition probability matrix according to the supplementary feature data.

[0196] In this embodiment of the application, the transition probabilities of relevant states in the state transition probability matrix are appropriately adjusted according to the importance and reliability of the supplementary feature data.

[0197] Step 705: When the confidence evaluation value of all fault hypotheses is less than the preset decision threshold, the probabilistic reasoning process is re-executed based on the adjusted state transition probability matrix, and the confidence evaluation value of each fault hypothesis in the fault hypothesis set is updated by iterative calculation.

[0198] In step 705, the probabilistic reasoning process refers to using a dynamic Bayesian network with a preset state transition probability matrix as prior knowledge to perform probabilistic reasoning on the joint feature vector to generate a set of fault hypotheses; the iterative calculation is a method of gradually optimizing the confidence evaluation value by repeatedly executing the probabilistic reasoning process.

[0199] In this embodiment of the application, the new confidence evaluation value of each fault hypothesis can be calculated by re-performing probabilistic reasoning using the state transition probability matrix updated with interactive information.

[0200] Step 706: When the updated confidence evaluation value of at least one fault hypothesis is greater than or equal to the preset decision threshold, the iteration process is terminated and a diagnostic decision result is generated.

[0201] In step 706, the termination condition is to determine whether the threshold standard required for diagnosis has been met.

[0202] In this embodiment, the practicality and reliability of fault diagnosis are improved through human-computer interaction and dynamic parameter updates.

[0203] Here is a specific example:

[0204] In substation A, located in a city in a certain province, in order to realize intelligent monitoring and diagnosis of the transformer's operating status, a series of specific hardware detection devices were first deployed in key locations around the transformer body. These included a high-precision acoustic sensor array of model ACMxxx, which consisted of six acoustic probes arranged in a ring on the transformer shell. It was used to collect acoustic fingerprint signals of abnormal sounds such as mechanical vibration and partial discharge during transformer operation in real time. Its sampling frequency covered the full frequency band from 20 Hz to 20 kHz.

[0205] Meanwhile, to obtain auxiliary on-site environmental information, a VMSxxx panoramic high-definition video surveillance camera was installed in the substation room to record the appearance of the equipment and the surrounding environment. A GASxxx multi-gas monitoring sensor was also deployed to detect the concentration of fault characteristic gases such as dissolved hydrogen and acetylene in the transformer oil.

[0206] On-site maintenance personnel are equipped with explosion-proof smart safety helmets of model HCSxxx. These helmets integrate noise-canceling microphones, allowing maintenance personnel to directly describe fault phenomena via voice at the inspection site, such as describing it as "hearing a continuous buzzing sound accompanied by intermittent popping sounds." The relevant voice is transmitted in real time to the edge computing server in the station through the wireless communication module of model WLMxxx built into the helmet.

[0207] The server is equipped with an industrial-grade graphics processing unit (model GPCxxx). The received voiceprint signal is first preprocessed and its features are extracted by a dedicated acoustic signal processing chip (model ASPxxx). When the system initially determines that there is a suspected discharge fault through the built-in software algorithm, it will trigger the corresponding emergency response plan. The system will remotely control a PTZxxx pan-tilt infrared thermal imager deployed near the transformer to perform non-contact temperature measurement and scanning of the transformer bushing and connection parts, and at the same time activate a UVSxxx ultraviolet imager to detect possible corona discharge points.

[0208] Real-time data collected by all sensors is aggregated through an industrial IoT gateway (model IOTxxx) to a server cluster (model SCSxxx) in the substation's main control room. The digital twin platform deployed on the cluster performs multi-source data fusion and interactive reasoning analysis. Finally, a complete diagnostic report is generated by a distributed decision system (model DDSxxx) and pushed to an explosion-proof smart terminal (model PADxxx) held by maintenance personnel. This completes the entire hardware-linked closed loop from on-site data acquisition and multimodal interactive diagnosis to final decision output.

[0209] Figure 4This application provides a schematic diagram of the structure of a substation fault interactive reasoning and auxiliary decision-making system based on natural language processing, and the specific implementation section describes the following:

[0210] The acquisition module 41 is used to acquire the acoustic signature signal and corresponding fault phenomenon text of the transformer in the substation during operation.

[0211] The conversion module 42 is used to convert the voiceprint signal into a voiceprint spectrum and form a joint feature vector based on the voiceprint features in the voiceprint spectrum and the vocabulary information in the fault phenomenon text.

[0212] The reasoning module 43 is used to perform probabilistic reasoning on the joint feature vector using a dynamic Bayesian network and a preset state transition probability matrix as prior knowledge, to generate a set of fault hypotheses, in which each fault hypothesis has a corresponding confidence evaluation value.

[0213] The generation module 44 is used to generate a query strategy through a cross-modal attention mechanism when the confidence evaluation value is less than a preset decision threshold.

[0214] The update module 45 is used to guide maintenance personnel to supplement the voice description information of the transformer fault based on the query strategy, update the parameter settings of the state transition probability matrix based on the voice description information, and re-execute the probabilistic reasoning process based on the updated state transition probability matrix to generate diagnostic decision results.

[0215] The substation fault interactive reasoning and auxiliary decision-making system based on natural language processing in this application is used to implement the aforementioned substation fault interactive reasoning and auxiliary decision-making method based on natural language processing. Therefore, the specific implementation of the substation fault interactive reasoning and auxiliary decision-making system based on natural language processing can be found in the embodiment section of the substation fault interactive reasoning and auxiliary decision-making method based on natural language processing mentioned above. The specific implementation can be referred to the description of the corresponding embodiment, which will not be repeated here.

[0216] This application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of any of the above-described natural language processing-based interactive reasoning and auxiliary decision-making methods for substation faults.

[0217] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described interactive reasoning and auxiliary decision-making methods for substation faults based on natural language processing.

[0218] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0219] The embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the substation fault interactive reasoning and auxiliary decision-making method based on natural language processing.

[0220] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. 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.

[0221] The foregoing has provided a detailed description of the interactive reasoning and auxiliary decision-making method, system, electronic device, and storage medium for substation faults based on natural language processing provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A substation fault interactive reasoning and auxiliary decision-making method based on natural language processing, characterized in that, include: Collect acoustic signature signals and corresponding fault phenomenon texts of transformers in substations during operation; The voiceprint signal is converted into a voiceprint map, and a joint feature vector is formed based on the voiceprint features in the voiceprint map and the vocabulary information in the fault phenomenon text. Using a dynamic Bayesian network and a preset state transition probability matrix as prior knowledge, probabilistic reasoning is performed on the joint feature vector to generate a set of fault hypotheses. Each fault hypothesis in the set of fault hypotheses has a corresponding confidence evaluation value. When the confidence evaluation value is less than the preset decision threshold, a query strategy is generated through a cross-modal attention mechanism; Based on the inquiry strategy, the operation and maintenance personnel are guided to supplement the voice description information of the transformer fault. Based on the voice description information, the parameter settings of the state transition probability matrix are updated. Based on the updated state transition probability matrix, the probabilistic reasoning process is re-executed to generate diagnostic decision results. The query strategy generated through the cross-modal attention mechanism includes: The correlation strength matrix between the lexical information and the voiceprint features is analyzed using a cross-modal attention mechanism. The feature dimension with a first association strength lower than a first preset quantization threshold is identified from the association strength matrix, where the first association strength is the association strength between the vocabulary information and the voiceprint feature; When there are multiple feature dimensions, a second association strength is calculated, which is the association strength between every two feature dimensions among the multiple feature dimensions. Based on the second association strength, the types of fault information that need to be supplemented and the query priority are determined, and a query strategy is generated according to the fault information types and the query priority.

2. The method according to claim 1, characterized in that, Based on the second association strength, the type of fault information that needs to be supplemented and the query priority are determined. A query strategy is generated according to the fault information type and the query priority, including: Based on the second association strength, an association matrix is ​​constructed using a graph neural network; Extract feature dimension pairs from the correlation matrix whose second correlation strength is less than a second preset quantization threshold. Based on the second correlation strength of the feature dimension pairs, determine the degree of information loss and the type of fault information that needs to be supplemented; Arrange the fault information types in descending order of the degree of information missing, and assign a query priority to each fault information type; Based on the inquiry priority, an inquiry sequence is constructed, wherein each inquiry round in the inquiry sequence corresponds to a fault information type of inquiry priority; For each round of queries, the content of the query statement is generated using a natural language generation model based on the type of fault information. Multiple rounds of queries and their corresponding query statements are integrated into an inquiry strategy.

3. The method according to claim 2, characterized in that, For each round of queries corresponding to the fault information type, the query statement content is generated using a natural language generation model, including: The fault information type corresponding to each round of inquiry is input into the semantic understanding module of the natural language generation model, and the core element information and contextual information to be inquired are parsed out through the semantic understanding module. Based on the core element information and the contextual information, matching professional terms and question templates are retrieved from a preset power industry terminology database; By combining and optimizing the specialized vocabulary and question templates using the grammatical constraints in the natural language generation model, multiple candidate questions are generated. The evaluation module of the natural language generation model selects the candidate question with the highest sentence score from all the candidate questions as the final question. The final question is bound to the corresponding question round to form the question statement content.

4. The method according to claim 1, characterized in that, The step of converting the voiceprint signal into a voiceprint spectrum, and forming a joint feature vector based on the voiceprint features in the voiceprint spectrum and the lexical information in the fault phenomenon text, includes: The voiceprint signal is subjected to sound source separation processing to distinguish the signal components corresponding to the ambient sound source and the device sound source respectively; Suppress the signal components corresponding to the ambient sound source, and retain the signal components corresponding to the device sound source; Based on the signal components corresponding to the sound source of the device, a sound signature map is constructed; Extract the acoustic signature features that reflect the temporal variation of mechanical vibration from the acoustic signature spectrum; Identify vocabulary describing the fault phenomenon from the text describing the fault phenomenon; The voiceprint features and vocabulary information are fused using multimodal features to form a joint feature vector.

5. The method according to claim 1, characterized in that, The method utilizes a dynamic Bayesian network, with a preset state transition probability matrix as prior knowledge, to perform probabilistic inference on the joint feature vector, generating a set of fault hypotheses. Each fault hypothesis in the set has a corresponding confidence evaluation value, including: The joint feature vector is input into the inference model of the dynamic Bayesian network. In the inference model, based on the state transition relationship defined by the state transition probability matrix, the joint feature vector is subjected to probability evolution calculation to obtain the probability distribution result of the fault state. Based on the probability distribution results, fault hypotheses are determined, and a reliability evaluation value is assigned to each fault hypothesis based on the numerical value of the probability distribution results. All failure hypotheses assigned confidence scores are integrated into a failure hypothesis set.

6. The method according to claim 1, characterized in that, The process involves guiding maintenance personnel to supplement voice description information of the transformer fault based on the inquiry strategy, updating the parameter settings of the state transition probability matrix based on the voice description information, and re-executing the probabilistic inference process based on the updated state transition probability matrix to generate diagnostic decision results, including: In accordance with the order of multiple query rounds in the query strategy, corresponding supplementary requests are sequentially sent to the maintenance personnel; Based on the supplementary request, receive the voice description information provided by the maintenance personnel for each round of inquiries; The newly added fault features in the voice description information are analyzed, and supplementary feature data is generated based on the newly added fault features; Based on the supplementary feature data, adjust the transition probability parameters of the corresponding states in the state transition probability matrix; When the confidence evaluation value of all failure hypotheses is less than the preset decision threshold, the probabilistic reasoning process is re-executed based on the adjusted state transition probability matrix, and the confidence evaluation value of each failure hypothesis in the failure hypothesis set is updated by iterative calculation. The iteration process terminates when the updated confidence value of at least one fault hypothesis is greater than or equal to a preset decision threshold, and a diagnostic decision result is generated.

7. A substation fault interactive reasoning and auxiliary decision-making system based on natural language processing, characterized in that, include: The acquisition module is used to collect the acoustic signature signals and corresponding fault phenomenon text of transformers in the substation during operation; The conversion module is used to convert the voiceprint signal into a voiceprint spectrum and form a joint feature vector based on the voiceprint features in the voiceprint spectrum and the vocabulary information in the fault phenomenon text. The inference module is used to perform probabilistic inference on the joint feature vector using a dynamic Bayesian network and a preset state transition probability matrix as prior knowledge, to generate a set of fault hypotheses, in which each fault hypothesis has a corresponding confidence evaluation value. The generation module is used to generate a query strategy through a cross-modal attention mechanism when the confidence evaluation value is less than a preset decision threshold. The update module is used to guide maintenance personnel to supplement the voice description information of transformer faults based on the query strategy, update the parameter settings of the state transition probability matrix based on the voice description information, and re-execute the probabilistic reasoning process based on the updated state transition probability matrix to generate diagnostic decision results. The query strategy generated through the cross-modal attention mechanism includes: The correlation strength matrix between the lexical information and the voiceprint features is analyzed using a cross-modal attention mechanism. The feature dimension with a first association strength lower than a first preset quantization threshold is identified from the association strength matrix, where the first association strength is the association strength between the vocabulary information and the voiceprint feature; When there are multiple feature dimensions, a second association strength is calculated, which is the association strength between every two feature dimensions among the multiple feature dimensions. Based on the second association strength, the types of fault information that need to be supplemented and the query priority are determined, and a query strategy is generated according to the fault information types and the query priority.

8. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the natural language processing-based interactive reasoning and decision support method for substation faults as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the substation fault interactive reasoning and auxiliary decision-making method based on natural language processing as described in any one of claims 1 to 6.

Images