Power grid equipment fault diagnosis method based on time series modeling and knowledge enhanced reasoning

Abstract

The application discloses a power grid equipment fault diagnosis method based on time sequence modeling and knowledge enhanced reasoning, comprising the following steps: extracting multi-modal features of equipment multi-source data, and fusing to obtain a fusion feature vector; based on the fusion feature vector, combining equipment static structured parameters, jointly modeling to obtain a state evolution encoding vector; based on the fusion feature vector and the state evolution encoding vector, adopting knowledge enhanced multi-path reasoning to obtain fault confidence of each reasoning path; and performing explainability fusion and adaptive optimization on each fault confidence to obtain an optimal fault diagnosis strategy. Through the fusion of multi-modal features, the structured and time sequence joint modeling, the whole process of the operation state evolution is accurately described; through the knowledge enhanced multi-path reasoning, the effective integration of data driving and expert experience knowledge is realized, and the explainability fusion and adaptive optimization are performed, so that the accuracy and reliability of fault diagnosis are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of power grid technology, and more specifically, to a method for diagnosing power grid equipment faults based on time-series modeling and knowledge-enhanced reasoning. Background Technology

[0002] With the accelerated construction of new power systems, the operating loads of key power grid equipment are becoming increasingly variable and the working environment increasingly complex. Therefore, higher demands are placed on the accurate monitoring of their health status and fault diagnosis. Once such equipment malfunctions, it can easily trigger large-scale power outages or even safety accidents. Therefore, achieving real-time monitoring of equipment status, early risk warnings, and rapid fault assessment has become a crucial foundation for ensuring the safe and stable operation of the power grid.

[0003] Currently, the power industry is actively promoting the construction of intelligent equipment operation and maintenance systems. By introducing technologies such as artificial intelligence, big data, and image recognition, it is driving the transition of operation and maintenance models from periodic inspections to intelligent diagnosis based on equipment status, thereby improving the timeliness of fault detection and the efficiency of fault handling. However, although some power grids have attempted to use data-driven methods for fault diagnosis, the following problems still exist in practical applications:

[0004] (1) Single data source: Most systems rely on a single type of data, such as gas in oil, vibration signals or thermal imaging information, which makes it difficult to comprehensively and accurately characterize the overall operating status of the equipment.

[0005] (2) Limitations of diagnostic methods: Existing methods are mostly based on simple empirical rules or "black box" models, lacking the ability to fuse multimodal information and perform temporal modeling, resulting in insufficient diagnostic sensitivity and robustness.

[0006] (3) The reasoning process is not transparent: the output classification results lack clear fault inference basis and explanatory explanation, making it difficult for maintenance personnel to understand the diagnostic logic and support actual decision-making.

[0007] In summary, existing technologies have not fully addressed key issues such as the fusion and utilization of multi-source heterogeneous data, dynamic state evolution modeling, and interpretability of the diagnostic process, which restricts the accuracy, reliability, and practical value of intelligent diagnostic systems in actual operation and maintenance. Summary of the Invention

[0008] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a fault diagnosis method for power grid equipment based on time-series modeling and knowledge-enhanced reasoning, which solves the problem that the status and fault risks of power grid equipment in new power systems are not easily known quickly and accurately, thus affecting the safe and stable operation of the power grid.

[0009] To achieve the above objectives, embodiments of the present invention provide a fault diagnosis method for power grid equipment based on time-series modeling and knowledge-enhanced reasoning, comprising: extracting multimodal features from multi-source monitoring data of the equipment and fusing them to obtain a fused feature vector; based on the time series of the fused feature vector and combined with the static structured parameters of the equipment, jointly modeling to obtain a state evolution coding vector; based on the fused feature vector and the state evolution coding vector, employing knowledge-enhanced multi-path reasoning to obtain the fault confidence of each reasoning path; and performing interpretability fusion and adaptive optimization on each fault confidence to obtain an optimal fault diagnosis strategy.

[0010] In a preferred embodiment, obtaining the fused feature vector includes: extracting representative features of key diagnostic information of the device and mapping them to a unified feature space; dynamically weighting and fusing the representative features in the feature space to obtain the fused feature vector.

[0011] In a preferred embodiment, the dynamic weighted fusion includes: uniformly encoding representative features of the multimodalities, and obtaining corresponding output feature vectors through modality adaptive representation learning, modality relation graph convolutional network, and dynamic modality credibility scoring; and fusing the output feature vectors with their contextual information features to obtain the fused feature vector.

[0012] In a preferred embodiment, the joint modeling to obtain the state evolution encoding vector includes: obtaining a short-term failure risk score based on the static structured parameters of the equipment using a gradient boosting decision tree model; jointly modeling the nonlinear time-series dynamic evolution of the equipment operating state by combining the multimodal feature time series of the fused feature vector to obtain the dynamic features of the equipment operating state and its key indicators; normalizing the short-term failure risk score and each of the key indicators, and fusing and splicing them to obtain the state evolution encoding vector.

[0013] In a preferred embodiment, obtaining the dynamic characteristics and key indicators of the device's operating state further includes: modeling the dynamic evolution process of the device's operating state using a long short-term memory network to obtain a hidden state sequence for each time step; and extracting key indicators characterizing the dynamic characteristics of the device's operating state based on the hidden state sequence.

[0014] In a preferred embodiment, the key indicators include: the evolution trend slope, which characterizes the overall deterioration or improvement trend of the equipment operating status, obtained by weighted linear regression based on the hidden state sequence; the abrupt change points, their timing, and intensity, identified by the second difference of the hidden state sequence; and the fluctuation intensity, which quantifies the fluctuation amplitude of the equipment operating status, obtained by the mean square deviation of the hidden state sequence.

[0015] In a preferred embodiment, the normalization process of the short-term failure risk score and each of the key indicators, and the fusion and splicing to obtain the state evolution coding vector, specifically involves: normalizing the short-term failure risk score, the evolution trend slope, and the fluctuation intensity; and fusing and splicing the result of the normalization process with the curvature response intensity of the mutation point to generate the state evolution coding vector.

[0016] In a preferred embodiment, the step of using knowledge-enhanced multi-path reasoning to obtain the fault confidence of each reasoning path includes: fusing the fused feature vector and the state evolution encoding vector to obtain an enhanced diagnostic feature vector; establishing a multi-path reasoning system, and obtaining the fault confidence of each reasoning path output based on the enhanced diagnostic feature vector; wherein, the multi-path reasoning system includes: graph neural network enhanced reasoning path, fuzzy logic rule reasoning path, dynamic Bayesian network reasoning path, and historical case database matching reasoning path.

[0017] In a preferred embodiment, obtaining the fault confidence of each inference path output specifically involves: based on the enhanced diagnostic feature vector, enhancing the inference path via a modal graph neural network to obtain the feature representation of each parallel inference path input; based on the feature representation, obtaining the fault confidence vector of the corresponding inference path according to the independent diagnosis of each parallel inference path; wherein, the parallel inference path includes: fuzzy logic rule inference path, dynamic Bayesian network inference path, and historical case database matching inference path; the feature representation is obtained by fusing and concatenating the graph features output by the graph neural network inference with the state coding features.

[0018] In a preferred embodiment, interpretability fusion and adaptive optimization are performed on each of the fault confidence scores to obtain the optimal fault diagnosis strategy. This includes: normalizing each of the fault confidence scores and weighted fusion to obtain a fusion confidence score representing the final diagnosis result; generating an interpretable diagnosis report based on the reasoning basis of each of the fault confidence scores; and introducing a reasoning mechanism that balances fault category diagnosis accuracy and fault handling decision interpretability based on the fusion confidence score and the diagnosis report, and using hyperparameter adaptive optimization based on a multi-objective genetic algorithm to obtain the optimal fault diagnosis strategy.

[0019] The beneficial effects of this invention are:

[0020] By extracting multimodal features and fusing them to obtain a fused feature vector to achieve a unified representation of multimodal features, the multidimensional expressive capability of multi-source data of power grid equipment is significantly improved. By jointly modeling the time series of equipment structured parameters and fused feature vectors, the entire dynamic evolution of the operating state of power grid equipment is accurately depicted, enhancing the foresight and adaptability of equipment fault diagnosis. By employing knowledge-enhanced multi-path reasoning, the effective integration between data-driven and expert experience knowledge is achieved, significantly improving the credibility of equipment fault diagnosis. By performing interpretability fusion and adaptive optimization on the confidence of each fault, the optimal fault diagnosis strategy that meets the dual requirements of interpretability and transparency in power equipment fault diagnosis scenarios is obtained, greatly improving the accuracy and reliability of equipment fault diagnosis. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a fault diagnosis method for power grid equipment based on time-series modeling and knowledge-enhanced reasoning.

[0022] Figure 2 A flowchart illustrating the process of obtaining the fused feature vector;

[0023] Figure 3 A schematic diagram illustrating the process of obtaining state evolution encoding vectors through joint modeling;

[0024] Figure 4 This is a schematic diagram of the framework of a multi-path reasoning system;

[0025] Figure 5 A flowchart illustrating the process of obtaining the optimal fault diagnosis strategy. Detailed Implementation

[0026] 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.

[0027] Example

[0028] Please see Figure 1A fault diagnosis method for power grid equipment based on time-series modeling and knowledge-enhanced reasoning includes: extracting multimodal features from multi-source monitoring data of the equipment and fusing them to obtain a fused feature vector; based on the time series of the fused feature vector and combined with the static structured parameters of the equipment, jointly modeling to obtain a state evolution coding vector; based on the fused feature vector and the state evolution coding vector, using knowledge-enhanced multi-path reasoning to obtain the fault confidence of each reasoning path; and performing interpretability fusion and adaptive optimization on each fault confidence to obtain the optimal fault diagnosis strategy.

[0029] Multimodal features of key diagnostic information of power grid equipment are extracted from multi-source monitoring data, including dissolved gas concentration in oil, vibration signals, thermal imaging images, inspection records, and environmental parameters. These extracted multimodal features are then fused to obtain a fused feature vector. Based on the time series of this fused feature vector, combined with static structured parameters such as the equipment number, voltage, current, and ambient temperature, joint time series and structured modeling is used to obtain the dynamic evolution law of the power grid equipment's operating state over time, thus obtaining a state evolution encoding vector for the power grid equipment's operating state. Based on this fused feature vector and the state evolution encoding vector, knowledge-enhanced multi-path reasoning is used to obtain the fault confidence of each independent diagnostic output path. Finally, based on each fault confidence, while ensuring interpretability, the data is fused and spliced ​​to determine the fault type of the power grid equipment and obtain an interpretable diagnostic report corresponding to the fault type. Simultaneously, through adaptive optimization and iterative updates, the optimal combination of global fault type identification and interpretable diagnostic report is selected and output as the optimal fault diagnosis strategy.

[0030] By extracting multimodal features and fusing them to obtain a fused feature vector, a unified representation of multimodal features is achieved. This significantly improves the multi-dimensional expressive capability of multi-source data for power grid equipment, enhances fault identification capabilities during equipment fault diagnosis, and effectively avoids the problem of single-modal data features being susceptible to noise interference. Simultaneously, by jointly modeling the time series of equipment structured parameters and fused feature vectors, a precise characterization of the entire dynamic evolution of power grid equipment operating status is achieved, providing a reliable basis for early warning and trend prediction, and enhancing the foresight and adaptability of equipment fault diagnosis. Through knowledge-enhanced multi-path reasoning, effective integration between data-driven approaches and expert experience is achieved. The fault confidence output through multi-path reasoning not only helps identify the fault type of the equipment but also simultaneously provides the reasoning basis for the reasoning path, significantly improving the credibility of equipment fault diagnosis. Based on the fault confidence of each reasoning path, interpretability fusion and adaptive optimization are performed to obtain the optimal fault diagnosis strategy that meets the dual requirements of interpretability and transparency in power equipment fault diagnosis scenarios, greatly improving the accuracy and reliability of equipment fault diagnosis.

[0031] Please see Figure 2 To improve the expressive power of multimodal features in multi-source monitoring data across time, space, and semantic dimensions, and to reduce noise interference with multimodal features, this embodiment further elaborates on obtaining the fused feature vector: Representative features of key diagnostic information in the multi-source monitoring data of power grid equipment are extracted and mapped to a unified feature space; then, the representative features in the feature space are dynamically weighted and fused to obtain the fused feature vector. Specifically, as follows:

[0032] S1. Extract multimodal features from multi-source monitoring data of the equipment and fuse them to obtain a fused feature vector;

[0033] A multimodal feature extraction and adaptive fusion method is adopted to obtain key diagnostic information from dissolved gases in oil, vibration, thermal imaging, text records and environmental parameters. Representative features are extracted and mapped to a unified space. Subsequently, the features are dynamically weighted and fused through a modal adaptive gating mechanism, a modal relation graph convolutional network and a credibility scoring mechanism to form a fusion feature vector with high expressive power, providing comprehensive information support for subsequent state modeling and reasoning.

[0034] S11. Obtain key diagnostic information and extract its representative features;

[0035] The key diagnostic information obtained came from data across different modalities, including dissolved gas concentration in oil, vibration signals, thermal imaging images, inspection records, and environmental parameters. Features were extracted from this data, specifically as follows:

[0036] S111, Characteristics of dissolved gas concentration in extracted oil;

[0037] Dissolved gas concentration data in oil can provide important clues for fault diagnosis of power grid equipment. By analyzing time-series data of gas concentrations, multiple statistical features, trend features, and ratio features are extracted. Dissolved gases in oil are an important basis for transformer fault diagnosis; common gases include hydrogen. methane ethane ethylene propane Nitrogen carbon monoxide ,carbon dioxide and oxygen Changes in the concentration of these gases can reflect problems such as electrical discharges, localized overheating, or oil decomposition inside the transformer.

[0038] S1111, Statistical characteristics;

[0039] The mean for each gas type represents the average concentration level of that gas over a time series:

[0040]

[0041] In the formula, For the first The mean of the gas type; For the first Gas-like substances in time The concentration; This represents the total length of the time series.

[0042] S1112, Standard Deviation:

[0043] The standard deviation describes the degree of fluctuation in gas concentration, that is, the range of change in gas concentration relative to the mean:

[0044]

[0045] S1113, Trend Characteristics;

[0046] First-order differences are used to capture trends in gas concentration over time, helping to detect abrupt changes in gas concentration.

[0047]

[0048] In the formula, For the first Gas-like substances in time The first-order difference; For the first Gas-like substances in time The concentration.

[0049] The second-order difference is used to capture the acceleration of changes in gas concentration, that is, the rate of change of gas concentration:

[0050]

[0051] In the formula, For the first Gas-like substances in time The second-order difference.

[0052] Calculate the ratios between different gases to reveal the relationships between gas concentrations. For example, methane. With hydrogen The ratio between them is:

[0053]

[0054] In the formula, for The concentration of the gas; This represents the concentration of the gas. The calculation of the concentration of other gases is similar.

[0055] S1114. Membership degree calculation;

[0056] The purpose of fuzzy membership is to introduce uncertainties into gas concentration analysis in order to better handle changes in transformers in complex operating environments.

[0057] In the analysis of dissolved gases in oil, fuzzy membership functions are used to quantify the uncertainty of gas concentration changes, thereby enhancing the system's adaptability to fuzzy or noisy data. Fuzzy membership functions Used to calculate the concentration of a gas The membership value, which is usually expressed as... Function form:

[0058]

[0059] In the formula, It is the gas concentration. The membership degree value; Indicates gas concentration; The shape parameter represents the fuzzy membership function and controls the steepness of the function; The offset parameter represents the fuzzy membership function and determines the position of the function's midpoint.

[0060] By calculating the membership degrees of different gas concentrations, the system can assess the reliability and impact of gas concentration changes, thus providing a more stable basis for fault diagnosis. This is crucial for handling uncertainties, dynamic changes, and complex environments in transformer operation, improving the accuracy and robustness of the fault diagnosis system.

[0061] S112, Vibration signal analysis;

[0062] Vibration signals can reflect the mechanical health status of equipment. By analyzing the time and frequency domain characteristics of equipment vibration signals, abnormal mechanical vibration modes can be identified, thereby inferring potential equipment faults.

[0063] S1121, Temporal characteristics;

[0064] The root mean square (RMS) is used to measure the energy intensity of a vibration signal, reflecting the overall level of equipment vibration.

[0065]

[0066] In the formula, It is the vibration signal in time The amplitude; It is the total number of time steps of the signal.

[0067] S1122, kurtosis;

[0068] Kurtosis measures the sharpness of a signal and reflects abnormal abrupt changes in vibration signals.

[0069]

[0070] In the formula, This represents the mean value of the vibration signal. It represents the standard deviation of the vibration signal.

[0071] S1123, Frequency domain characteristics;

[0072] Perform a Fourier transform on the vibration signal to obtain its spectrum:

[0073]

[0074] In the formula, Represents the spectrum of a frequency domain signal; This indicates the Fourier transform operation.

[0075] Find the frequency component with the largest amplitude in the signal:

[0076]

[0077] S1124, Multi-band Energy;

[0078] Calculate frequency range Internal signal energy:

[0079]

[0080] In the formula, , Indicates the upper and lower bounds of the frequency band; This represents a spectrum signal.

[0081] S1125, Wavelet packet decomposition energy;

[0082] In vibration signal analysis, besides the extraction of time-domain and frequency-domain features, wavelet packet decomposition energy and short-time Fourier transform time-frequency plots also play crucial roles. Wavelet packet decomposition energy... It is extracted by performing wavelet packet decomposition on the signal and calculating the energy of each sub-band. The formula is as follows:

[0083]

[0084] In the formula: Indicates the first Layer wavelet packet division is the energy; Indicates the first Layer and first Each sub-frequency component represents the wavelet packet coefficients. Wavelet packet decomposition allows for a more precise capture of the energy distribution of vibration signals across different frequency bands, especially in multi-frequency band scenarios, enabling in-depth analysis of the time-varying characteristics of vibration signals.

[0085] The short-time Fourier transform is used to analyze the time-frequency characteristics of a signal. By dividing the signal into small segments and performing a Fourier transform on each segment, a time-frequency diagram is obtained, which reflects the frequency components of the vibration signal at different time points. The formula is:

[0086]

[0087] In the formula, Indicates time and frequency The time spectrum on; Original signal; It is a window function; It is frequency; It is a time variable. By combining wavelet energy decomposition and short-time Fourier transform, the energy and frequency characteristics of vibration signals can be effectively captured, which helps to reveal the mechanical state of the transformer and potential fault risks.

[0088] S113, Thermal imaging image analysis;

[0089] Thermal imaging images can help detect overheating on the surface of equipment. In this step, the system calculates the thermal center offset of the thermal imaging image and extracts spatial relationship features of the image using a graph convolutional network. The calculated thermal center position of the image is:

[0090]

[0091] In the formula, Indicates the thermal image in coordinates Temperature value at that location; This represents the location of the thermal center of a thermal imaging image, assuming the image size is [size missing]. The location of the thermal center is .

[0092] S114. Text inspection record analysis;

[0093] The text in the inspection logs can provide crucial information about equipment health. Using natural language processing techniques, the system extracts sentence vectors from the text and calculates their similarity to knowledge graph nodes, thereby analyzing potential equipment malfunctions.

[0094] use The model extracts sentence vectors:

[0095]

[0096] In the formula, It is the first Inspection record text; It is text The corresponding sentence vector.

[0097] Calculate the similarity between text and knowledge graph nodes:

[0098]

[0099] In the formula, A vector representing a node in a knowledge graph.

[0100] S115. Analysis of operating and environmental parameters;

[0101] Environmental parameters of equipment have a significant impact on its operating status. By analyzing the fluctuations in these parameters, the system can detect equipment failures caused by environmental factors.

[0102] Calculate environmental parameters The volatility is:

[0103]

[0104] Calculate environmental parameters The offset is:

[0105]

[0106] In the formula, It represents the value of a certain environmental parameter (such as current or voltage); This represents the standard deviation of the parameter; This represents the mean of the parameter.

[0107] S12. Encode and map to a unified feature space;

[0108] To ensure effective fusion of data from different modalities within the same feature space and to provide input for subsequent weighted fusion, representative features of the multimodal data are uniformly encoded, and the data from each modality are mapped using a mapping function. This is mapped to a unified feature space, preparing for subsequent weighted fusion. The mapping is as follows:

[0109]

[0110] In the formula, Indicates the first The original characteristics of class data; Indicates the first Features after class data mapping; This represents the mapping function designed for each modality.

[0111] S13. Dynamically weight and fuse representative features in the feature space to obtain a fused feature vector;

[0112] After mapping features from different modalities to a unified feature space, complex feature fusion is performed on representative features in the feature space, including: modal adaptive representation learning, modal graph convolutional networks, and dynamic modal credibility scoring. After this complex weighted fusion, corresponding output feature vectors are obtained. Then, based on each output feature vector and combined with the contextual information features of the power grid equipment, a fused feature vector is obtained, which generates the fused feature vector of fault parameters, providing accurate input for equipment fault diagnosis and condition prediction. Specifically, modal feature vectors from different data sources such as oil chromatography, vibration, and infrared are aligned and preprocessed, and then multiple fusion methods such as modal adaptive representation learning, modal graph convolutional networks, and dynamic modal credibility scoring are used to integrate data from different modalities into a high-dimensional fused feature vector through weighted fusion technology.

[0113] S131, Modal Adaptive Representation Learning;

[0114] A dynamic gating weight score is calculated for the feature vector of each modality. This score is automatically generated by learning the correlation between modality features and the current task, and can reflect the importance of different modalities at a specific time or state.

[0115] S1311, Calculate the table for each mode The original gating score :

[0116]

[0117] In the formula, It is the first Modal gating score; and The first The learnable weight matrix and bias terms corresponding to the modal gating score; for Activation function.

[0118] S1312. Then, normalize the original scores of all obtained modalities to obtain the final fusion weighting coefficients. :

[0119]

[0120] In the formula, This represents the total number of all modalities obtained.

[0121] S1313 Finally, an adaptive representation is formed through weighted fusion. :

[0122]

[0123] In the formula, For the first The weighting coefficients of the modes.

[0124] S132, Modal Relationship Graph Convolutional Network;

[0125] Node updates and aggregation:

[0126]

[0127]

[0128] In the formula, It is the first The set of neighbor nodes of a modality; It is the first Modality and the first Edge weights between modes; For nodes and nodes The degree.

[0129] S133, Dynamic Modal Reliability Score;

[0130] Quality scoring and weighting adjustment:

[0131] ,

[0132]

[0133] In the formula, For the first Modal quality score.

[0134] S134, Final fusion output;

[0135] By combining the outputs of all modes with contextual information, a final fused feature vector is generated, providing accurate support for intelligent diagnosis and fault prediction of transformers.

[0136] The outputs and context information of all modalities are combined as follows:

[0137]

[0138] In the formula, These are contextual information features, such as device ID.

[0139] Please see Figure 3To accurately depict the dynamic evolution of the operating state of power grid equipment, this embodiment further elaborates on the joint modeling to obtain the state evolution encoding vector: Based on the static structured parameters of the power grid equipment, a gradient boosting decision tree model is used to obtain a short-term fault risk score; combined with the multimodal feature time series of the fused feature vector, the nonlinear time-series dynamic evolution of the equipment operating state is jointly modeled to obtain the dynamic features and key indicators of the equipment operating state; the short-term fault risk score and each of the key indicators are normalized and fused together to obtain the state evolution encoding vector. Specifically:

[0140] S2. Based on the time series of fused feature vectors, combined with the static structured parameters of the device, the state evolution encoding vector is obtained through joint modeling of time series and structured parameters;

[0141] A joint modeling framework for the temporal characteristics of static structured parameters and fused feature vectors of power grid equipment is constructed. The CatBoost (Gradient Boosting Decision Trees) model is used to score the short-term fault risk of the structured features of power grid equipment. Simultaneously, a Long Short-Term Memory (LSTM) network is employed to dynamically model the multimodal time-series features of the fused feature vectors, extracting key dynamic features of the power grid equipment's operating state evolution over time, such as trend slope, second-order difference abrupt change points, and fluctuation intensity. Finally, based on the obtained short-term fault risk scores and key dynamic features of the operating state, normalization is performed and the data is fused to generate a unified state evolution encoding vector. This provides a structurally clear, semantically interpretable, and temporally logical intermediate representation for subsequent knowledge reasoning, serving as a crucial input for the subsequent knowledge reasoning module.

[0142] S21. Based on the static structured parameters of power grid equipment, a gradient boosting decision tree model is used to obtain a short-term fault risk score.

[0143] Based on the nonlinear fault risk information in the static structured parameters of power grid equipment, a gradient boosting decision tree model (CatBoost) is used to obtain a short-term fault risk score characterizing the fault risk of power grid equipment; the short-term fault risk score is related to the inherent attributes and operating state of the power grid equipment. That is:

[0144] To capture the fault risk information implicit in the static structured parameters of power grid equipment, this embodiment introduces the CatBoost gradient boosting decision tree model. While maintaining the interpretability of this model, it efficiently extracts nonlinear fault risk information from the static structured parameters of power grid equipment, specifically extracting the nonlinear correlations between various static structured features of the power grid equipment to obtain an accurate short-term fault risk score characterizing the fault risk of the power grid equipment. The static structured parameters mainly consist of static structured features of the power grid equipment, such as its number, voltage, current, and ambient temperature, within the current timeframe and recent historical periods. Specifically:

[0145] S211, Sequence representation of static structured parameters of power grid equipment;

[0146] Assume that the power grid equipment is in continuous operation The structured feature observation sequence for each historical time step is as follows:

[0147]

[0148] In the formula, Indicates the first The structured feature vector at any given time, also known as the structured observation vector, includes features such as equipment model, years of operation, voltage, current, load rate, ambient temperature, and humidity. This refers to the historical time step used for risk modeling, i.e., the length of the historical observation window; For feature dimensions.

[0149] S212, CatBoost model structure;

[0150] The CatBoost model uses an additive model of gradient decision trees, combining multiple weak learners in a forward distribution, as shown below:

[0151]

[0152] In the formula, The total number of decision trees; The learning rate controls the contribution of each tree to prevent overfitting. For the first The output value of each tree.

[0153] For each sample in the input sequence The model outputs its corresponding structured risk score as follows:

[0154]

[0155] In practical applications, this score can be mapped to the Sigmoid function. The interval represents the estimated probability of a fault occurring.

[0156] S213, Objective Function and Model Training;

[0157] The goal of the CatBoost model is to minimize the loss function with regularization, i.e.:

[0158]

[0159] In the formula, The loss function; For the first The complexity regularization term for a tree is used to control the complexity of the tree structure, such as the depth of the tree and the number of leaf nodes, in order to improve the generalization ability of the model.

[0160] The training process employs a gradient boosting framework. In the... In each iteration, the current model is calculated. The negative gradient of the loss function, i.e., the pseudo-residual, is:

[0161]

[0162] Fit a new decision tree The pseudo-residual can be approximated as follows:

[0163]

[0164] S214, Category Feature Processing;

[0165] One of the advantages of the CatBoost model is its natively efficient processing of categorical features, such as the model number and manufacturer code of power grid equipment. The CatBoost model effectively prevents the leakage of target information by employing an ordered encoding strategy based on target statistics.

[0166] For a certain category feature In the sample The value in the range is calculated as follows:

[0167]

[0168] In the formula, This represents a random permutation of the training data; For indicator functions; This represents a smoothing coefficient to prevent the denominator from being zero. This is the prior probability, typically the global mean of the target variable.

[0169] Encoding is performed under ordered data arrangement to ensure that each sample is encoded using only its historical data, thus preventing the leakage of target information.

[0170] S215, Risk Score Output;

[0171] After training, the entire structured feature sequence Inputting the data into the CatBoost model yields the following risk score sequence:

[0172]

[0173] Risk score at the final moment in the risk scoring sequence This is extracted and used as the output of S21, namely the short-term failure risk score. This score... As a key short-term static risk indicator, it is integrated with the dynamic characteristics of the operating status output in S22 to jointly constitute the state evolution coding vector of power grid equipment.

[0174] By employing the gradient boosting decision tree model CatBoost, leveraging its excellent ability to handle category features and prevent overfitting, nonlinear fault risk information is efficiently extracted from the static structured features of power grid equipment. While maintaining model interpretability, it enhances the responsiveness to fault occurrence probabilities. The short-term fault risk score output after processing by the gradient boosting decision tree model CatBoost can accurately reflect the prior probability of faults based on the inherent attributes of power grid equipment and its recent operating state, providing reliable static prior support for subsequent dynamic evolution modeling of power grid equipment operating states.

[0175] S22. Combine the multimodal feature time series of the fused feature vector to jointly model the nonlinear time-series dynamic evolution of the equipment operating state;

[0176] Based on the multimodal feature time series of the fused feature vectors, a Long Short-Term Memory (LSTM) network is used to model the nonlinear temporal dynamic evolution of the operating state of power grid equipment, obtaining a hidden state sequence for each time step. Based on the hidden state sequence, key indicators characterizing the dynamic features of the equipment operating state are extracted. These key indicators include: the slope of the evolution trend, representing the overall deterioration or improvement trend of the equipment operating state, obtained through weighted linear regression based on the hidden state sequence; the second-order difference of the hidden state sequence, identifying abrupt changes in the operating state of power grid equipment, along with their timing and intensity; and the mean square deviation of the hidden state sequence, obtaining the fluctuation intensity that quantifies the amplitude of operating state fluctuations.

[0177] The Long Short-Term Memory (LSTM) neural network is used to model the multimodal feature time series of fused feature vectors. The multimodal features include gas concentration changes, vibration spectrum fluctuations, and image thermal features. By modeling, the dynamic evolution law of the operating state of power grid equipment is extracted, and the hidden state sequence output by LSTM is analyzed. Further quantification is used to extract the global evolution trend slope, state fluctuation intensity, and key abrupt change response of the operating state of power grid equipment, providing key dynamic features for the assessment of the operating state of power grid equipment.

[0178] The time-dependent and nonlinear evolution patterns of power grid equipment during operation are captured as follows:

[0179] S221. Define the time series of multimodal features fused from feature vectors as follows:

[0180]

[0181] In the formula, For the first The input feature vector at each time step; This represents the total number of time steps, i.e., the total length of the time series. The fusion feature dimension is the output of S1.

[0182] S222 and LSTM network modeling;

[0183] LSTM cells at each time step Perform the following calculations:

[0184] S2221, Input Gate: Controls the current input. The extent to which something is written into a memory cell.

[0185]

[0186] S2222, Forget Gate: Determines which memory units from the previous moment will be retained. The degree of.

[0187]

[0188] S2223, Candidate Memory Update Unit: Generates new candidate information for memory units to refer to.

[0189]

[0190] S2224, Memory Unit Update: The memory unit at the current time step integrates both forgetting old information and adding new information.

[0191]

[0192] S2225, Output Gate: Controls the generation of the current hidden state.

[0193]

[0194] S2226, Hidden state generation;

[0195] The hidden state of the final output for:

[0196]

[0197] In the above formula, Use the Sigmoid activation function; Given a hyperbolic tangent function, restrict the output to... ; It is element-wise multiplication; , , , These are the corresponding weight matrices; , , , These are the corresponding bias vectors; For the previous time step The hidden state; Outsourced forgetting gate activation value; Candidate memory cell values; It is the current state of the memory cell; For the current time step Input features.

[0198] Finally, the LSTM network outputs the hidden states of the entire sequence as follows:

[0199]

[0200] In the formula, Let be the dimension of the LSTM hidden state.

[0201] The final output sequence of hidden states is the sequence of hidden states calculated internally by the LSTM network. It is the internal feature representation of the deep learning model and is mainly used to extract dynamic features such as evolution trend and fluctuation intensity.

[0202] S223. Extract and obtain the dynamic characteristics of the evolution of the equipment's operating status;

[0203] Based on the hidden state sequence finally output by S222, three key dynamic features are extracted during the evolution of the operating state of power grid equipment. Specifically:

[0204] S2231, Evolutionary Trend Slope ;

[0205] The evolution trend slope is a key indicator for measuring the overall trend of power grid equipment operating status over time. It is used to identify whether the operating status tends to deteriorate or improve within the observation time window. The value of the evolution trend slope is calculated by performing a weighted linear regression on the hidden state sequence output by the LSTM, with the weights being the offset of each time step relative to the time center. The calculation is as follows:

[0206]

[0207] in, Center of the time series mean ; The slope represents the evolutionary trend, indicating the global degradation or recovery rate. If... A value greater than 0 indicates that the overall operating condition of the power grid equipment is deteriorating; if... A value less than 0 indicates that the overall operating status of the power grid equipment is showing a recovery trend; The larger the absolute value, the more drastic the overall change and the more significant the fluctuations in the operating status of the power grid equipment. By extracting the evolution trend slope as a key indicator, the long-term evolution direction of the power grid equipment during operation can be intuitively reflected, which is an important input factor for subsequent diagnosis and risk assessment.

[0208] S2232, Fluctuation Intensity ;

[0209] Fluctuation intensity is used to assess the dynamic stability of power grid equipment during operation, reflecting whether its operating state is within a stable, slightly disturbed, or drastically fluctuating range. This index is obtained by calculating the dispersion of the LSTM hidden state sequence relative to its overall mean; that is, by calculating the mean square deviation of the hidden states relative to its mean, the fluctuation amplitude of the operating state is quantified. The calculation is as follows:

[0210] ,

[0211] In the formula, This represents the mean of all hidden states; Indicates the intensity of the fluctuation; This represents the total length of the time series. Indicates the index of the current time step; No. The LSTM hidden state vector at each time step. When the fluctuation intensity... When the fluctuation is relatively small, it indicates that the power grid equipment is operating relatively stably; while when the fluctuation intensity is high... A large value indicates that the operating status of the power grid equipment is fluctuating drastically and may be in an unstable or fault-prone stage.

[0212] By extracting the fluctuation intensity as a statistical feature in time series modeling It can effectively help identify the health and stability level of power grid equipment and improve the ability to respond to abnormal fluctuations during subsequent diagnosis.

[0213] S2233, Mutation point detection

[0214] Abrupt point detection aims to identify the timing and intensity of abrupt changes in the operating state sequence of power grid equipment, which typically indicate an impending potential anomaly or fault. Abrupt point detection identifies these changes by calculating the second-order difference of the hidden state sequence to measure the instantaneous change in the curvature of the operating state, as calculated below:

[0215] ,

[0216] In the formula, Indicates the first The curvature response intensity of the power grid equipment operating status at the time step.

[0217] When the curvature response intensity at a certain moment A point significantly exceeding the threshold of its overall distribution is considered a mutation point. If the preset threshold for the overall distribution of curvature response intensity is... When the curvature response intensity Above the threshold If the threshold is reached, then the point is determined to be a potential mutation point. It can be set based on the historical curvature response intensity, for example, to the historical curvature response intensity. times, .

[0218] Among all abrupt change points, the point with the strongest curvature response was identified as the critical abrupt change point, i.e.:

[0219] ,like

[0220] By extracting mutation points, key anomalies can be accurately detected in dynamic modeling. This not only captures sudden risks that emerge during the gradual evolution of the operating state, but also helps to build an early warning mechanism that is sensitive to system disturbances, thereby improving the response capability of power grid equipment when its operating state changes abruptly.

[0221] S23. Normalize the short-term fault risk score and the key indicators of each dynamic feature, fuse and splice them to generate the state evolution coding vector.

[0222] The short-term failure risk score, the evolution trend slope, and the fluctuation intensity are normalized; the result of the normalization is then fused and concatenated with the curvature response intensity of the abrupt change point to generate a state evolution encoding vector. That is, the short-term failure risk score... The slope of the evolutionary trend and the intensity of the fluctuation Normalization was performed, and the result was compared with the curvature response intensity at the abrupt change point. The vectors are fused and spliced ​​to generate a state evolution encoding vector.

[0223] S231, Normalization process:

[0224] , ,

[0225] In the formula, It is a very small quantity used to prevent the denominator from being zero.

[0226] S232. Generate state evolution encoding vector:

[0227]

[0228] The final output state evolution encoding vector It includes static risk scores, evolution trends, dynamic stability, and anomaly detection information, and serves as the structured input for subsequent fault reasoning and expert interpretation modules.

[0229] By constructing a structured and temporal joint modeling mechanism, we can accurately identify the characteristics of the evolution of the operating status of power grid equipment over time. Then, we can extract key dynamic feature information of the operating status evolution, including static risk score, evolution trend change, mutation symptoms and fluctuation intensity, and obtain a state evolution coding vector that comprehensively and accurately represents the multi-source fusion information of the operating status of power grid equipment. This provides comprehensive and accurate information input for subsequent fault reasoning and intermediate explanatory explanation.

[0230] Please see Figure 4 To improve the reliability of equipment fault diagnosis, this embodiment further elaborates on the use of knowledge-enhanced multi-path reasoning to obtain the fault confidence of each reasoning path: Based on the fused feature vector and the state evolution encoding vector, an enhanced diagnostic feature vector is obtained; a multi-path reasoning system is established, and based on the enhanced diagnostic feature vector, the fault confidence of each reasoning path is obtained; wherein, the multi-path reasoning system includes: graph neural network enhanced reasoning path, fuzzy logic rule reasoning path, dynamic Bayesian network reasoning path, and historical case database matching reasoning path. Specifically:

[0231] S3. Based on the fusion feature vector and state evolution coding vector, knowledge-enhanced multi-path reasoning is adopted to obtain the fault confidence of each reasoning path;

[0232] Using the fused feature vector from S1 and the state evolution encoding vector from S2 as inputs, and combining the multimodal sensing information of power grid equipment, a reasoning system containing four knowledge-enhanced reasoning paths is constructed to assess operational faults of power grid equipment from different perspectives. Specifically, based on the state evolution encoding vector generated in S2, heterogeneous feature information from multimodal sensing data is further fused. Based on graph neural networks, fuzzy logic rules, dynamic Bayesian networks, and a historical case library, a knowledge-enhanced reasoning system is constructed to achieve multi-path, multi-mechanism diagnostic reasoning and confidence output.

[0233] S31. Preprocess the fused feature vector and the state evolution encoding vector;

[0234] The system receives the multimodal fusion feature vector output by S1 and the state evolution coding vector output by S2, and then fuses and concatenates the fusion feature vector and the state evolution coding vector to form an enhanced diagnostic feature vector containing static risk, dynamic evolution and multimodal information, which is the final input vector of the diagnostic module's multi-path reasoning.

[0235] The multimodal fusion feature vector output by S1 and the state evolution coding vector output by S2 are concatenated to form the input vector of the diagnostic module, i.e., the enhanced diagnostic feature vector:

[0236]

[0237] in, It is the multimodal fusion feature vector output by S1; It is the state evolution encoding vector output by S2; It is the enhanced diagnostic feature vector that is ultimately unified as the input for reasoning along each path in the diagnostic module.

[0238] The enhanced diagnostic feature vector obtained after preprocessing will be used as an input vector in subsequent graph neural network fusion, fuzzy logic rules and dynamic Bayesian networks.

[0239] S32. Constructing a reasoning system;

[0240] Construct a reasoning system with graph neural networks as the core structure, fuzzy logic rules as the knowledge carrier, dynamic Bayesian networks as the causal modeling mechanism, and historical cases as experience references.

[0241] S321, Modal Graphical Neural Network (GNN) Enhances Inference;

[0242] In multimodal diagnostic tasks, single-modal information often lacks sufficient contextual semantics. Therefore, using graph neural network structures to explicitly express the connectivity between modalities can improve the capture of correlations and contextual influences in complex scenarios.

[0243] Based on the enhanced diagnostic feature vector, the feature representations of each parallel inference path input are obtained through modal graph neural network-enhanced inference paths. Specifically, this involves explicitly modeling the intrinsic relationships between different modalities in the enhanced diagnostic feature vector using graph neural networks, enhancing the expressive power of modality fusion, and achieving knowledge enhancement at the feature level.

[0244] S3211. Construct a modal relationship diagram;

[0245] By treating each modality feature in the enhanced diagnostic feature vector as a node in a graph neural network, semantic associations between modalities are established through edges. That is: The feature subsets of different modes are nodes, such as hydrogen. Average concentration, hotspot temperature, etc. Edges between nodes and their weights. It is learned through knowledge or data-driven learning from fields such as physical laws, and is used to characterize the strength of the correlation between features.

[0246] S3212, Graph Convolution Propagation;

[0247] Based on the adjacency structure of a graph neural network, a multi-layer graph convolutional network is used to perform multiple rounds of information propagation and aggregation on node features, so that each node feature incorporates the information of its neighboring nodes, and outputs a fused graph feature representation. This graph feature representation will be used to enhance the expressiveness of modality fusion and will be combined with state-encoded features for subsequent inference.

[0248] The process of propagating and aggregating multi-layer graph neural networks after constructing modality graphs is as follows:

[0249] S32121, Symmetric Normalized Adjacency Matrix:

[0250]

[0251] In the formula, Represents the adjacency matrix of the modal graph; Represents a symmetric normalized adjacency matrix; The degree matrix of the graph nodes , .

[0252] S32122, During the propagation of the convolutional layer, the first... Layer propagation is as follows:

[0253]

[0254] In the formula, The graph convolution weight matrix; For example, ReLU activation functions; Indicates the first Layer node features, definition These are the initial node features.

[0255] S32123, Output and Re-fusion;

[0256] through After layer propagation, the features of the final layer nodes are... Global average pooling is performed to obtain a high-order feature representation with graph structure enhancement that incorporates modal information, i.e., a graph feature representation. for:

[0257]

[0258] Representing graph features The state evolution encoding vector output by S2 After further fusion and splicing, a richer feature representation is obtained as follows:

[0259]

[0260] In the formula, The feature representation of the final input obtained by fusing and splicing the graph features of the graph neural network inference output with the state evolution coding features will serve as the main input for subsequent fuzzy logic rule inference, dynamic Bayesian network DBN inference, and matching historical cases.

[0261] Based on the obtained feature representations, and according to the independent diagnosis of each parallel inference path, the fault confidence vector of the corresponding inference path is obtained. The parallel inference paths include: fuzzy logic rule inference path, dynamic Bayesian network inference path, and historical case matching inference path.

[0262] S322, Fuzzy logic rule-based reasoning;

[0263] Fuzzy logic rule systems are reasoning paths built on the experiential knowledge of human experts. By transforming expert experience into interpretable fuzzy logic judgments, they are suitable for handling scenarios with high uncertainty and fuzziness.

[0264] S3221, Blurring;

[0265] The core advantage of leveraging fuzzy logic rules lies in explaining the origin of rules. It maps continuous changes in key features to semantically fuzzy judgments, such as logical expressions like "too high" or "too low," thereby enhancing the explanatory power and credibility of fault assessment. Through a pre-defined membership function, key features are mapped to predefined fuzzy linguistic values. For example, regarding... The first in Key features The Sigmoid membership function is used to map it to fuzzy linguistic values, as follows:

[0266]

[0267] In the formula, Indicates the first Key features; Indicates the kurtosis of the function; Indicates the center of the function; Indicates the first The membership degree of a key feature, that is, the degree to which it belongs to a certain fuzzy linguistic value.

[0268] S3222, Rule Evaluation;

[0269] Each fuzzy logic rule consists of multiple feature conditions, corresponding to the diagnostic conclusion after specific fault assessment; based on the key input features, the matching degree of each fuzzy logic rule is judged to form a preliminary diagnostic output confidence level.

[0270] The expert's experience knowledge base contains strip Rules. Among them, the first... The rule is expressed as follows:

[0271]

[0272] Based on the Key features The corresponding membership degrees are multiplied or the minimum value is taken to obtain the rule matching degree:

[0273] or

[0274] In the formula, Indicates the first The matching degree of each rule output.

[0275] S3223, Deblurring and Output: For each fault type The conclusions are summarized as follows: The rule matching degree is obtained and output as a fault confidence vector based on fuzzy logic rules through methods such as weighted averaging. :

[0276]

[0277] In the formula, This represents the total number of fault categories.

[0278] Fuzzy logic rules are suitable for human-controlled scenarios that incorporate empirical knowledge. By obtaining the fault confidence vector of power grid equipment through fuzzy logic rules, it can be used for subsequent fusion and decision-making.

[0279] S323, Dynamic Bayesian Network DBN Inference;

[0280] Dynamic Bayesian networks (DNBs) are probabilistic graphical models for modeling time-series states, suitable for depicting the evolution of equipment operating states over time. By modeling the prior, observational, and posterior probabilities of the time-series evolution of power grid equipment operating states, DNBs can perform causal and probabilistic inferences, outputting posterior fault probabilities based on observational evidence, which can be used to identify the operating states of power grid equipment and provide early warnings.

[0281] S3231, State transition modeling;

[0282] Dynamic Bayesian Networks (DBNs) contain sequences of hidden states that characterize the internal health state of a device. and characterize input features Observation sequence of slices at corresponding time steps Among them, the hidden state sequence Unlike the hidden state sequence calculated internally by the LSTM network Hidden state sequence defined in Dynamic Bayesian Network (DBN) In a probabilistic graphical model, a latent random variable represents the health state of a device that cannot be directly observed. It follows the assumptions of a state-space model, has a probability distribution, and is used to describe the uncertainty of the device's operating state.

[0283] The state transition model is as follows:

[0284] ,

[0285] In the formula, , They represent the first Time, Number The hidden state at any given moment; Represents the state transition matrix; Uncertainties or random disturbances during state transitions, i.e., process noise or system noise; express It follows a pattern with a mean of 0 and a covariance matrix of . The state transition model is based on a multivariate Gaussian distribution and is used to describe the evolution of the device's operating state over time.

[0286] The observation model is:

[0287] ,

[0288] In the formula, Indicates the first The observation sequence input at any given time; Represents the observation mapping matrix; This refers to observation noise, which represents errors and uncertainties during the observation process. express It follows a pattern with a mean of 0 and a covariance matrix of . The model uses a multivariate Gaussian distribution to describe how the hidden states in a Dynamic Bayesian Network (DBN) generate observation data.

[0289] S3232, Joint Probability Distribution; Based on the state transition model, the joint probability distribution of the complete observation sequence and the hidden state sequence is decomposed into:

[0290]

[0291] In the formula, Let the initial prior distribution of the hidden states be assumed to be... ; Define the state transition equation; Define the observation equation.

[0292] S3233, Reasoning and Calculation;

[0293] Given a sequence of observations over a period of time Using the forward-backward algorithm, the posterior probability distribution of the hidden state at any time step is calculated. Proportional to the forward variable With backward variables The product of these is expressed as follows:

[0294]

[0295] The posterior probability distribution of the hidden state is still a Gaussian distribution, and its mean and covariance can be accurately calculated using Kalman filtering and smoothing algorithms.

[0296] S3234, Fault classification output;

[0297] During the diagnostic phase, attention should be paid to the final state judgment based on the observed sequence. Therefore, the posterior probability distribution of the obtained hidden states should be considered. Take the expected posterior probability of the hidden state at the final time. As a state summary, inputting a Softmax classifier yields the following fault probability distribution based on the probabilistic graphical model:

[0298]

[0299]

[0300] In the formula, , These represent the weights and biases of the classifier, respectively.

[0301] By inputting state evolution encoding features containing time evolution information and constructing the state transition relationship between them and the hidden states through overtraining, the distribution of device states is estimated using a forward-backward algorithm. In the diagnosis stage, based on the various fault posterior probabilities output in the dynamic Bayesian network DBN, a fault probability distribution based on a probabilistic graphical model is obtained, which is used to participate in the final decision together with fuzzy logic rules and historical case library.

[0302] S324, Reasoning by matching historical case database;

[0303] Based on the analogical reasoning idea that "similar problems have similar solutions," this method provides analogical support for input samples by comparing them with historical case scenarios in a typical historical case library. This historical case matching method is particularly suitable for scenarios with incomplete data, ambiguous features, or rules that cannot cover them, helping to improve the model's fault tolerance and generalization ability. Specifically:

[0304] S3241. Construct a historical case library;

[0305] Pre-built includes A database of typical historical cases, each historical case State feature vector containing its typical states And the verified corresponding fault type labels .

[0306] S3242. Calculate similarity;

[0307] For the current input features The sample is used to calculate the state feature vector of each historical case in the historical case database. The weighted distance is:

[0308]

[0309] In the formula, For the first The weights of the features reflect the importance of those features in diagnosis. This weighted distance is either a weighted Manhattan distance or a Euclidean distance.

[0310] S3243, Similarity Conversion and Retrieval;

[0311] Convert weighted distance into weighted similarity score The conversion is as follows:

[0312]

[0313] The weighted similarity score obtained through calculation measures the current input features. The degree of similarity between the sample and the corresponding historical cases in the historical case library.

[0314] Then, the top few historical cases with the highest similarity scores and their fault type labels are selected as references, and the corresponding fault types and similarities are returned as auxiliary diagnostic outputs. That is, based on the selected historical cases, a fault confidence vector based on historical case matching is generated through voting or weighted voting. .

[0315] S33, Multi-path reasoning output;

[0316] The four parallel inference paths in S32 independently completed the independent diagnosis of power grid equipment faults and output their fault confidence vectors as follows:

[0317] The fault confidence vector output by the inference path based on fuzzy logic rules is: ;

[0318] The fault confidence vector output by the inference path based on the dynamic Bayesian network is: ;

[0319] The fault confidence vector output by the inference path based on the matching historical case library is: .

[0320] Among them, the input feature representation obtained by enhancing the inference path based on the modal graphical neural network (GNN) The input feature representation, which is the fusion of graph features and state evolution coding features, serves as the input to the inference paths of fuzzy logic rules, dynamic Bayesian networks, and historical case matching databases. Therefore, it does not directly output a fault confidence vector, but it is already integrated into the fault confidence vectors output by the three inference paths. Combining these fault confidence vectors forms... This is then used as input for the subsequent S4, for final weighted fusion and decision-making.

[0321] Please see Figure 5 To obtain an optimal fault diagnosis strategy with high accuracy in fault oscillation and strong interpretability of fault diagnosis decisions, this embodiment further elaborates on the interpretability fusion and adaptive optimization of the confidence scores of each fault to obtain the optimal fault diagnosis strategy: The confidence scores of each fault are normalized, and weighted fusion is introduced using weighted coefficients to obtain a fused confidence score representing the final diagnosis result; this fused confidence score is the comprehensive confidence score for each type of fault. Simultaneously, an interpretable diagnosis report is generated based on the reasoning basis of each fault confidence score; and based on the fused confidence score and the diagnosis report, a reasoning mechanism that balances fault category diagnosis accuracy and fault handling decision interpretability is introduced, employing hyperparameter adaptive optimization based on a multi-objective genetic algorithm to obtain the optimal fault diagnosis strategy. Specifically:

[0322] S4. Perform interpretable fusion and adaptive optimization on the confidence levels of each fault to obtain the optimal fault diagnosis strategy;

[0323] Based on the fault confidence vectors output by S3, an interpretable weighted fusion mechanism is introduced to generate the final diagnostic results. Simultaneously, to ensure long-term system optimality, a genetic optimization strategy that integrates diagnostic information and interpretability is introduced to automatically optimize key parameters.

[0324] S41. Normalize and weight the confidence scores of multi-source faults to generate a fused confidence score and output the fault diagnosis decision.

[0325] Based on the fault confidence vectors output by each inference path of S3, a weighted fusion mechanism is introduced to overcome the limitations of a single inference mechanism. The diagnostic information of the modal graph neural network (GNN) enhanced inference path, the fuzzy logic rule inference path, the dynamic Bayesian network inference path, and the inference path matching the historical case library is integrated. Through normalization and weight configuration, a global comprehensive confidence is generated and the final fault type is determined.

[0326] S411, Normalization processing;

[0327] Fault confidence vector output by inference path based on fuzzy logic rules Fault confidence vector output from inference path based on dynamic Bayesian network And the fault confidence vector output by the inference path based on the matching historical case library. After normalization, the following normalized values ​​were obtained:

[0328] , ,

[0329] In the formula, This represents the normalized fault confidence value output by the fuzzy logic rule; This represents the normalized fault confidence value output from the historical case library. This represents the posterior output probability of a dynamic Bayesian network.

[0330] S412, Weighted Fusion: The normalized fault confidence scores are linearly weighted to calculate the final comprehensive score for each fault category:

[0331]

[0332] In the formula, , respectively, represent the fusion weights of the fuzzy logic rules, the matching historical case library, and the dynamic Bayesian network, satisfying ; Indicates the final fusion confidence level;

[0333] S413, Output fault diagnosis decision;

[0334] The fault type with the highest comprehensive score is selected as the final diagnostic result output. By fusing the results of multi-source reasoning, the final fault type and interpretable path are output as fault diagnosis decisions, which can ensure that the diagnostic results are accurate, transparent and adaptive.

[0335] The final diagnosis was:

[0336]

[0337] In the formula, This indicates the final diagnosed fault category.

[0338] In summary, considering that a single inference mechanism often has coverage blind spots in complex diagnostic tasks, the confidence outputs from fuzzy logic rule inference, historical case matching inference, and dynamic Bayesian network inference are normalized and weighted. Then, a linear weighted average is used to calculate the comprehensive confidence score for each fault type, and the fault type with the highest comprehensive score is selected as the final diagnostic fault category. By introducing a weighted fusion strategy, the outputs from the three inference paths—fuzzy logic rules, historical case matching, and dynamic Bayesian network—are combined, improving the overall diagnostic accuracy, robustness, and consistency. Furthermore, this output can support further strategy optimization, enabling on-demand emphasis on rules or learning models.

[0339] S42. Based on the reasoning basis of each of the aforementioned fault confidence levels, generate an interpretable diagnostic report;

[0340] To enhance the transparency and credibility of diagnostic results, the system simultaneously outputs detailed reasoning explanations, generating an interpretable diagnostic report. This report includes:

[0341] S421. Fuzzy logic rule triggering conditions: List all rules whose matching degree exceeds the threshold, including the triggered rule number, the precondition of fuzzy conditions, and the specific matching degree. Among them, the preconditions are ambiguous, such as hydrogen gas. The concentration is considered "high".

[0342] S422, Historical Case Matching Details: Displays the top few historical cases with the highest similarity, including case number and similarity score. And analysis of key feature differences.

[0343] S423. Dynamic Bayesian Network (DBN) Causal Evolution Path: Based on the trajectory of the key hidden states obtained by the inference path of the Dynamic Bayesian Network (DBN) over time, and visualize it as a graph, it shows the dynamic development process of power grid equipment faults, that is, the most likely causal chain of posterior probability change over time inferred by the Dynamic Bayesian Network (DBN).

[0344] S424, Fusion Decision Details: Clearly lists the current fusion weight configuration used in this diagnosis. Normalized values ​​of the confidence scores for each type of fault output by each inference path and the final calculated fusion confidence level .

[0345] S43, Adaptive Genetic Optimization;

[0346] The parameters of the fusion reasoning mechanism have a crucial impact on the final diagnostic results, while manual parameter tuning introduces significant uncertainty. Therefore, to address the limitations of manual parameter tuning and improve the system's adaptability and long-term reliability across different scenarios and power grid equipment, a reasoning mechanism that balances fault category diagnostic accuracy with the interpretability of fault handling decisions is introduced. Furthermore, a hyperparameter adaptive optimization strategy based on a multi-objective genetic algorithm is employed to enhance the system's adaptability. Specifically:

[0347] S431, Chromosome Encoding: The key parameters to be optimized, i.e., the fusion parameters, are encoded into a single chromosome as follows:

[0348]

[0349] In the formula, These represent the fusion weights of the fuzzy logic rules, the matching historical case library, and the dynamic Bayesian network, respectively, satisfying... ; Other optimizable parameters include fuzzy logic rule thresholds and similarity thresholds for historical case matching.

[0350] S432. Design the fitness function: Considering both the accuracy of fault diagnosis and the interpretability of decision-making, guide the algorithm to find the optimal balance. The fitness function is constructed as follows:

[0351]

[0352] In the formula, This represents the vector of key parameters to be optimized. Indicates the accuracy of fault diagnosis; This represents an interpretability penalty term, applied when the system relies excessively on factors such as the weights of a Dynamic Bayesian Network (DBN). When the black-box model is too high or the decision-making process is too complex, the value of this penalty term increases; This represents the explanatory penalty coefficient. This is used to adjust the intensity of the requirement for interpretability.

[0353] S433, Genetic Operations:

[0354] S4331. Selection: Based on fitness levels, use roulette or tournament selection methods to retain high-quality individuals.

[0355] S4332, Crossover: Perform single-point or multi-point crossover on the selected parent chromosomes to produce new individuals.

[0356] S4333, Mutation: Randomly perturbing genes in a chromosome with a certain probability, such as using Gaussian mutation as follows:

[0357]

[0358] In the formula, Indicates the variance of Gaussian noise; This represents the updated key parameter vector.

[0359] After mutation, the fusion weights need to be normalized to ensure... .

[0360] S4334. Evaluation: Evaluation criteria include the accuracy of fault diagnosis and the interpretability of decisions.

[0361] By encoding the fusion parameters into chromosome individuals and searching for the optimal combination in the parameter space using a genetic algorithm, and then evaluating the obtained optimal combination based on evaluation criteria, it can be ensured that the optimization results are not only highly accurate but also engineering-acceptable.

[0362] S434. Optimization Process: The genetic algorithm proceeds iteratively using a "selection-crossover-mutation-evaluation" process until the preset number of generations or convergence criterion is reached. Its optimization process can be as follows:

[0363] Offline operation: Before system deployment, a set of globally optimal parameters is optimized using historical data.

[0364] Periodic online operation: After the system has been running for a period of time, it uses newly accumulated data to start optimization, realizing dynamic self-calibration and continuous evolution of parameters.

[0365] In summary, this invention extracts multi-source diagnostic data from various devices and constructs a unified feature representation framework that integrates multimodal perception information. This significantly improves the expressive power of input data in the temporal, spatial, and semantic dimensions. Furthermore, by employing a fusion mechanism combining modal adaptive representation learning and graph convolutional networks, it can flexibly adapt to differences in modal information quality under various operating conditions, enhancing the diagnostic model's ability to identify faults under complex conditions. This effectively avoids the problem of single-modal input being susceptible to noise interference, thereby improving the overall diagnostic accuracy and operational stability of the system. By introducing a structured and temporal joint modeling mechanism in the state modeling stage, combining the nonlinear modeling capability of the CatBoost algorithm for static structural features with the modeling advantages of LSTM networks for time dependencies, it achieves a precise characterization of the entire process of equipment state evolution. Analyzing changes in equipment operating status not only outputs short-term risk scores but also extracts key indicators such as trend slope, abrupt change response, and fluctuation intensity. This systematically reflects the pattern characteristics of equipment operating status changes over time, providing a reliable basis for early warning and trend prediction, and enhancing the foresight and adaptability of the fault diagnosis system. By constructing a multi-inference path system supported by graph neural networks, using fuzzy logic as the knowledge carrier, dynamic Bayesian networks as the causal modeling tool, and historical cases as a reference, this system effectively integrates data-driven methods with expert experience and knowledge. Through fuzzy logic rule triggering conditions, details of similar historical cases, and the evolution process of dynamic Bayesian causal chains, it intelligently determines the final fault type of power grid equipment, greatly improving the accuracy and reliability of equipment fault diagnosis. The multi-inference path system not only outputs various fault confidence levels but also simultaneously outputs the reasoning basis for the inference paths, including the triggered rule number, similar case number, and Bayesian propagation chain, significantly improving the traceability and credibility of the diagnostic process and meeting the dual requirements of interpretability and transparency in power equipment fault diagnosis scenarios. Based on the fault confidence levels output by the multi-path reasoning system, an adaptive optimization genetic algorithm is used to optimize the fusion weights and threshold parameters, achieving automatic adjustment of the diagnostic strategy and optimization of fault diagnosis decisions. This ensures online system adjustment and / or scenario adaptation, enabling a clear judgment of the current power grid equipment fault category.

[0366] 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.

[0367] 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.

[0368] 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.

[0369] 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.

[0370] 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.

[0371] 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 fault diagnosis method for power grid equipment based on time-series modeling and knowledge-enhanced reasoning, characterized in that, include: Extract multimodal features from multi-source monitoring data of equipment and fuse them to obtain a fused feature vector; Based on the static structured parameters of the equipment, a gradient boosting decision tree model is used to obtain a short-term failure risk score. Combined with the multimodal feature time series of the fused feature vector, the nonlinear temporal dynamic evolution of the equipment's operating state is jointly modeled. A long short-term memory network is used to model the dynamic evolution process of the equipment's operating state, obtaining a hidden state sequence for each time step. Based on the hidden state sequence, key indicators characterizing the dynamic features of the equipment's operating state are extracted. These key indicators include the evolution trend slope, abrupt change points, and fluctuation intensity. The short-term failure risk score, the evolution trend slope, and the fluctuation intensity are normalized. The normalized result is fused and concatenated with the curvature response intensity of the abrupt change points to generate a state evolution encoding vector. Based on the fused feature vector and the state evolution encoding vector, knowledge-enhanced multi-path reasoning is used to obtain the fault confidence of each reasoning path; the multi-path reasoning system includes: graph neural network-enhanced reasoning path, fuzzy logic rule reasoning path, dynamic Bayesian network reasoning path, and historical case database matching reasoning path. The interpretability fusion and adaptive optimization of the various fault confidence scores are performed to obtain the optimal fault diagnosis strategy.

2. The power grid equipment fault diagnosis method based on time-series modeling and knowledge-enhanced reasoning according to claim 1, characterized in that, The process of obtaining the fused feature vector includes: Extract representative features of key diagnostic information of the equipment and map them to a unified feature space; The representative features in the feature space are dynamically weighted and fused to obtain a fused feature vector.

3. The power grid equipment fault diagnosis method based on time-series modeling and knowledge-enhanced reasoning according to claim 2, characterized in that, The dynamic weighted fusion includes: Representative features of multiple modalities are uniformly encoded and then processed through modal adaptive representation learning, modal relation graph convolutional network, and dynamic modal credibility scoring to obtain the corresponding output feature vectors. Based on each of the output feature vectors, and combined with their contextual information features, the fused feature vector is obtained.

4. The power grid equipment fault diagnosis method based on time-series modeling and knowledge-enhanced reasoning according to claim 1, characterized in that, The key indicators include: Based on the hidden state sequence, a weighted linear regression is performed to obtain the evolution trend slope that characterizes the overall deterioration or improvement trend of the equipment's operating status. Based on the second-order difference of the hidden state sequence, identify the abrupt change points where the device operating state changes drastically, as well as the timing and intensity of the abrupt change. Based on the mean square deviation of the hidden state sequence, the fluctuation intensity of the quantization device operating state fluctuation amplitude is obtained.

5. The power grid equipment fault diagnosis method based on time-series modeling and knowledge-enhanced reasoning according to claim 1, characterized in that, The multi-path reasoning method employing knowledge enhancement obtains the fault confidence of each reasoning path, including: Based on the fused feature vector and the state evolution coding vector, an enhanced diagnostic feature vector is obtained by fusing them; A multi-path reasoning system is established, and based on the enhanced diagnostic feature vector, the fault confidence of each reasoning path output is obtained.

6. The power grid equipment fault diagnosis method based on time-series modeling and knowledge-enhanced reasoning according to claim 5, characterized in that, The specific steps for obtaining the fault confidence of each inference path output are as follows: Based on the enhanced diagnostic feature vector, the feature representations of each parallel inference path input are obtained by enhancing the inference path through the modal graph neural network; Based on the aforementioned feature representation, and according to the independent diagnosis of each parallel inference path, the fault confidence vector of the corresponding inference path is obtained. The parallel inference paths include: fuzzy logic rule inference path, dynamic Bayesian network inference path, and historical case matching inference path. The feature representation is obtained by fusing and concatenating the graph features output by the graph neural network inference with the state coding features.

7. The power grid equipment fault diagnosis method based on time-series modeling and knowledge-enhanced reasoning according to claim 1, characterized in that, The optimal fault diagnosis strategy is obtained by interpretable fusion and adaptive optimization of the aforementioned fault confidence scores, including: The confidence scores of each fault are normalized and then weighted and fused to obtain the fusion confidence score representing the final diagnostic result. Based on the reasoning basis of each of the aforementioned fault confidence levels, an interpretable diagnostic report is generated; Based on the fusion confidence and the diagnostic report, a reasoning mechanism that balances the accuracy of fault category diagnosis and the interpretability of fault handling decisions is introduced. Hyperparameter adaptive optimization based on a multi-objective genetic algorithm is used to obtain the optimal fault diagnosis strategy.

Images