GIS device insulation aging judgment method and system based on Autoformer long time series modeling

Abstract

The application discloses a GIS device insulation aging judgment and system based on Autoformer long time sequence modeling, which comprises the following steps: collecting and preprocessing multi-source state quantities of GIS devices to construct historical input sequences; constructing an insulation aging field knowledge system and forming a field knowledge reasoning model, knowledge coding and multi-dimensional verification are performed on the field knowledge and device operation information to obtain knowledge coding results; a knowledge-enhanced time sequence prediction model is constructed and trained; the input sequence is trend-periodically decomposed; long-range dependence relationship modeling is combined with the knowledge coding results to output future prediction results and insulation aging trend representation quantities; and the insulation aging state grade of the GIS device is comprehensively judged in combination with the knowledge coding results. The application can improve the reliability of GIS device insulation aging trend prediction and early warning, provide decision basis for device state maintenance and operation maintenance, support continuous tracking of aging risk, risk early warning and maintenance optimization, and assist operation and maintenance personnel in accurately mastering the insulation deterioration process.

Description

Technical Field

[0001] This invention belongs to the field of high-voltage electrical equipment condition monitoring technology, and particularly relates to a GIS equipment insulation aging judgment method and system based on Autoformer long-time series modeling. Background Technology

[0002] Gas-insulated switchgear (GIS) is widely used in substations and power transmission and distribution systems due to its compact structure, reliable operation, and strong environmental adaptability. During long-term operation, the internal insulation of GIS equipment may be affected by a combination of factors, including partial discharge, accumulation of gas decomposition products, changes in moisture content, thermal stress, electrical stress, and external environmental disturbances, leading to gradual deterioration of insulation performance. Failure to promptly identify the progression of insulation aging may result in the expansion of equipment defects and compromise the safe operation of the power grid.

[0003] Current GIS insulation condition assessment methods typically employ threshold discrimination, single-state quantity analysis, or prediction methods based on short-term samples. Threshold discrimination can provide warnings for states that are approaching or exceeding limits, but its ability to identify early states that have not yet exceeded limits but have already shown a continuous deterioration trend is limited. Single-state quantity analysis struggles to comprehensively utilize multi-source information such as partial discharge, SF6 gas decomposition products, trace moisture content, temperature, and operating conditions. When using general short-sequence prediction models, it is also difficult to simultaneously distinguish between long-term aging trends and periodic fluctuations caused by factors such as environment, load, or maintenance cycles.

[0004] The multi-source state variables of GIS equipment exhibit significant long-term characteristics: on the one hand, the degradation of insulation materials and changes in gas state show a slow, cumulative long-term trend; on the other hand, factors such as equipment temperature, operating load, and ambient temperature and humidity can cause periodic fluctuations or short-term disturbances. Therefore, a method is needed to perform trend-period decomposition of multi-source state variables, extract long-term dependencies across periods, and predict future changes in insulation state accordingly, in order to improve the foresight and reliability of insulation aging assessment. Summary of the Invention

[0005] This invention addresses the aforementioned problems by providing a method, system, and storage medium for judging the insulation aging of GIS equipment based on Autoformer long-time series modeling. The method preprocesses the time series of multi-source state variables of GIS equipment, separates long-term trend terms and periodic terms using the sequence decomposition module in the Autoformer model, extracts long-term dependencies under different time delays using the Auto-Correlation mechanism, outputs the predicted sequence of future state variables and trend components through an encoder-decoder structure, and judges the insulation aging level by combining the trend slope, threshold proximity, and degree of persistent anomaly.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A method for judging the insulation aging of GIS equipment based on Autoformer long-time series modeling includes: Collect multi-source state time-series data of GIS equipment during continuous operation; and preprocess the time-series data to construct a historical input sequence of insulation status of GIS equipment; A knowledge base for the GIS insulation aging domain is constructed, and a corresponding domain knowledge reasoning model is formed. The domain knowledge reasoning model is used to perform semantic encoding and knowledge reasoning on the GIS insulation aging domain knowledge base, the current GIS equipment multi-source state variables, and equipment operation information to generate knowledge-encoded data. The knowledge-encoded data is then subjected to multi-dimensional consistency verification to obtain the verified GIS insulation aging knowledge encoding results. Based on the historical input sequence of the insulation status of the GIS equipment and the verified GIS insulation aging knowledge encoding result, a knowledge-enhanced insulation status prediction model is constructed and trained. The historical input sequence of the insulation state of the GIS equipment is input into the trained insulation state prediction model. Using the decomposition mechanism in the insulation state prediction model, and based on the trend constraint parameters in the knowledge encoding results, adaptive trend-period decomposition is performed on different types of state variables to obtain trend terms that characterize long-term degradation trends and periodic terms that characterize periodic fluctuations. The trend term representing the long-term degradation trend and the periodic term representing the periodic fluctuation are input into the time-series dependency extraction mechanism in the insulation state prediction model. The long-range dependency relationship in the sequence is modeled in combination with the knowledge encoding result, and the predicted values ​​of the multi-source state quantities and the insulation aging trend characterization quantities are output in the future prediction period. Based on the predicted values ​​of multi-source state quantities and insulation aging trend characterization quantities during the future prediction period, combined with the verified state quantity-fault mode correlation matrix and fault mode prior probability in the knowledge encoding results, the insulation aging state level of the GIS equipment is comprehensively determined.

[0007] Furthermore, the multi-source state time series data includes at least one or more of the following: partial discharge characteristic quantities, SF6 gas decomposition product content, trace water content, gas pressure, equipment temperature, operating voltage, operating current, and ambient temperature and humidity.

[0008] Furthermore, the construction of the GIS insulation aging knowledge base specifically includes: By integrating GIS equipment operation procedures, condition-based maintenance guidelines, test reports, historical fault cases, operation and maintenance records, and expert diagnostic rules, a knowledge base for the field of GIS insulation aging is constructed, which includes multi-source state variables, equipment types, equipment components, service life, operating conditions, historical maintenance records, influencing factors, insulation aging mechanisms, fault modes, and condition judgment rules.

[0009] Furthermore, the generated knowledge encoding data includes: Based on the retrieval enhancement generation mechanism, domain knowledge fragments related to the current equipment status are retrieved from the GIS insulation aging domain knowledge base; Based on a general pre-trained large model, GIS insulation aging domain adaptation is performed through instruction fine-tuning or low-rank adaptation fine-tuning constraints to obtain a domain knowledge reasoning model with the ability to call GIS insulation aging domain knowledge, reason about mechanisms, and output structured data. The large model parameter representation of the domain knowledge reasoning model is also obtained. Based on the retrieved domain knowledge fragments, GIS equipment multi-source state variables, equipment type, equipment components, service life, operating conditions, historical maintenance records, and structured output constraints, input prompts for the large model are constructed. Based on the large model parameter representation, the domain knowledge reasoning model performs semantic encoding on the input prompts of the large model to obtain a semantic hidden representation; and obtains the prior vector of insulation aging mechanism through pooling or mapping: Based on the prior vector of insulation aging mechanism and the structured output constraints of the large model, knowledge-encoded data is generated, including the prior vector of insulation aging mechanism of the large model, the importance weight of the state variables of the large model, the state variable-failure mode correlation matrix of the large model, the prior probability of the failure mode of the large model, the trend constraint parameters of the large model, and the time delay prior information of the large model.

[0010] Furthermore, the multi-dimensional consistency verification of the knowledge encoding data to obtain the verified GIS insulation aging knowledge encoding result includes: Based on the expert diagnosis rules and historical fault cases in the GIS insulation aging knowledge base, the expert rule benchmark parameters are constructed, including the expert rule state quantity importance score, the expert rule state quantity-fault mode correlation matrix, the expert rule fault mode prior score, the expert rule trend constraint parameters, and the expert rule time delay prior function. The knowledge-encoded data is subjected to dimensionality verification, range verification, normalization verification, and expert rule consistency verification. Conflict degree is calculated based on the large model state variable-failure mode association matrix and the expert rule state variable-failure mode association matrix, and retrieval relevance is also calculated. A consistency score is calculated based on the search relevance: Based on the confidence level, retrieval relevance, consistency score, and conflict level of the domain knowledge reasoning model, the trustworthy fusion coefficient is calculated: Based on the importance scores of the large model state variables, the importance scores of the expert rule state variables, and the reliable fusion coefficient, the validated state variable importance weights are generated: Based on the large model state-failure mode correlation matrix, the expert rule state-failure mode correlation matrix, and the reliable fusion coefficients, a verified state-failure mode correlation matrix is ​​generated: Based on the prior scores of the large model failure modes, the prior scores of the expert rule failure modes, and the reliability fusion coefficient, the validated prior probabilities of failure modes are generated: Based on the trend constraint parameters of the large model, the trend constraint parameters of the expert rules, and the reliable fusion coefficients, the validated trend constraint parameters are generated: Based on the GIS operation cycle, environmental cycle, maintenance cycle, and aging mechanism cycle identified by the domain knowledge reasoning model, a large-scale model time delay prior function is generated. This function is then fused with the expert rule time delay prior function to obtain the validated time delay prior information. When the conflict level exceeds the preset conflict threshold, the input prompt of the large model is reconstructed and the above steps are repeated; if the conflict level still exceeds the preset conflict threshold after the number of repetitions reaches the preset upper limit, the trusted fusion coefficient is reduced and the expert rule benchmark parameter is adopted.

[0011] Furthermore, constructing and training a knowledge-enhanced insulation state prediction model includes: Based on the sampling period of the multi-source state variables of GIS equipment, the state variable dimension, and the insulation aging judgment requirements, the historical input length, prediction length, state variable dimension, model hidden layer dimension, number of encoder layers, number of decoder layers, number of multi-heads, feedforward network dimension, moving average kernel size range, and Auto-Correlation delay selection coefficient of the insulation state prediction model are set. The historical input sequence of multi-source state variables of GIS equipment and the verified GIS insulation aging knowledge encoding result are jointly mapped to the latent space to obtain a knowledge-enhanced input representation. Based on the verified importance weights and trend constraint parameters of the state variables, different moving average kernel sizes or trend extraction weights are assigned to the state variables of different GIS devices, so that the insulation aging sensitive state variables in the multi-source state variables focus on long-term trend extraction, and the disturbance-related state variables in the multi-source state variables focus on periodic fluctuation separation. We construct linear mapping parameters for the query matrix, key matrix, and value matrix, and introduce the time delay prior information of the large model into the Auto-Correlation mechanism to enhance the knowledge of sequence correlation under different time delays; Build by An encoder formed by stacking coding layers and by A decoder is formed by stacking several decoding layers; each encoding layer includes a knowledge-enhancing auto-correlation module, a residual connection module, a knowledge-constrained sequence decomposition module, and a feedforward network; each decoding layer includes an internal knowledge-enhancing auto-correlation module, an encoder-decoder knowledge-enhancing auto-correlation module, a feedforward network, multiple knowledge-constrained sequence decomposition modules, and a trend term accumulation structure. The total training loss is constructed based on the prediction loss, trend consistency loss, mechanism consistency loss, and diagnostic distillation loss. The gradient descent optimization method is used to backpropagate and update the trainable parameters in the knowledge-enhanced insulation state prediction model until the total training loss converges or the preset number of training rounds is reached.

[0012] Furthermore, adaptive trend-period decomposition is performed on different types of state variables to obtain trend terms characterizing long-term evolution and periodic terms characterizing the impact of periodic fluctuations, including: Based on the verified importance weights and trend constraint parameters of the state variables, determine the size of the adaptive moving average kernel for each state variable. Perform a moving average pooling operation with boundary padding on each state variable sequence to extract the trend term of each state variable; The periodic term of each state variable is obtained by subtracting the original sequence of each state variable from the corresponding trend term. The trend terms of each state variable are weighted according to the importance weight of the state variable to obtain the comprehensive trend component of insulation aging of GIS equipment, which is used as the characterization of the trend term.

[0013] Furthermore, the temporal dependency extraction mechanism specifically includes: The hidden representation of GIS device status variables is mapped to query matrix, key matrix, and value matrix, respectively. Calculate the autocorrelation between the query matrix and the key matrix at different time delays; The autocorrelation relationship is fused with the verified time delay prior information to obtain the knowledge-enhanced autocorrelation result; Based on the knowledge, significant time delays are selected to enhance the autocorrelation results, and the autocorrelation values ​​corresponding to each significant time delay are normalized. The value matrix is ​​cyclically shifted and weighted according to the selected time delay to obtain the aggregation result; The aggregation results are processed by an encoder-decoder structure to generate future prediction sequences and future trend components; According to the importance weight of the verified state variables, the future trend components are weighted and summed to obtain the comprehensive trend component of future insulation aging, and the comprehensive trend component of future insulation aging is output as the characteristic quantity of insulation aging trend.

[0014] Furthermore, the comprehensive determination of the insulation aging status level of the GIS equipment specifically includes: For each state variable, the trend slope of its future prediction period is calculated based on the future trend component, which is used to characterize the rate of change of the state variable. For state quantities that indicate insulation aging, calculate the degree to which their predicted values ​​are close to the threshold values ​​of preset normal reference values ​​and alarm limits. The proportion of sampling points exceeding a preset attention threshold in the future prediction sequence of each state variable is used as the degree of continuous anomaly of that state variable. Based on the trend slope, threshold proximity, and persistent anomaly of each state variable, combined with the correlation strength between each state variable and each fault mode in the verified state variable-fault mode correlation matrix, and the verified prior probability of the fault mode, the comprehensive confidence of each type of fault mode is calculated. The comprehensive risk value of insulation aging of GIS equipment is calculated by weighting the comprehensive confidence level of various failure modes and the preset risk weights corresponding to each failure mode. Based on the preset multi-level threshold range where the comprehensive risk value of insulation aging is located, the insulation aging status level of the GIS equipment is determined. The aging status level includes at least insulation status normal, insulation aging concern, insulation aging warning, and insulation aging severe.

[0015] On the other hand, the present invention provides a GIS equipment insulation aging judgment system based on Autoformer long-time series modeling, comprising: The state quantity acquisition and preprocessing module is used to acquire multi-source state quantity time series data of GIS equipment during continuous operation; and to preprocess the time series data to construct a historical input sequence of insulation state of GIS equipment. The knowledge encoding module is used to construct a knowledge base for the GIS insulation aging domain and form a corresponding domain knowledge reasoning model. The domain knowledge reasoning model is used to perform semantic encoding and knowledge reasoning on the GIS insulation aging domain knowledge base, the current GIS equipment multi-source state variables and equipment operation information to generate knowledge encoding data. The knowledge encoding data is then subjected to multi-dimensional consistency verification to obtain the verified GIS insulation aging knowledge encoding result. The model building and training module is used to build and train a knowledge-enhanced insulation state prediction model based on the historical input sequence of the insulation state of the GIS equipment and the verified GIS insulation aging knowledge encoding result. The sequence decomposition module is used to input the historical input sequence of the insulation state of the GIS equipment into the trained insulation state prediction model. Using the decomposition mechanism in the insulation state prediction model, and based on the trend constraint parameters in the knowledge encoding results, it performs adaptive trend-period decomposition on different types of state variables to obtain trend terms that characterize long-term degradation trends and periodic terms that characterize periodic fluctuations. The temporal dependency extraction and prediction module is used to input the trend term representing the long-term degradation trend and the periodic term representing the periodic fluctuation into the temporal dependency extraction mechanism in the insulation state prediction model, combine the knowledge encoding results to model the long-range dependency relationship in the sequence, and output the predicted values ​​of the multi-source state quantities and the insulation aging trend characterization quantities in the future prediction period. The insulation aging comprehensive judgment module is used to comprehensively determine the insulation aging status level of GIS equipment based on the predicted values ​​of multi-source state quantities and insulation aging trend characterization quantities in the future prediction period, combined with the verified state quantity-fault mode association matrix and fault mode prior probability in the knowledge encoding results.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention, by collecting and fusing multiple source state quantities such as partial discharge characteristic quantities, SF6 gas decomposition product content, trace water content, gas pressure, equipment temperature, operating parameters, and ambient temperature and humidity, can avoid the problem of insufficient state description caused by relying on only a single indicator.

[0017] 2. This invention utilizes the Autoformer built-in sequence decomposition structure to perform layer-by-layer trend-periodic decomposition on the hidden sequence of multi-source state variables, thereby separating the slow cumulative evolution trend of insulation aging from periodic disturbances such as operating conditions and environmental changes, which facilitates insulation aging judgment based on trend components.

[0018] 3. This invention utilizes the Auto-Correlation mechanism to perform correlation calculation, time delay selection, and alignment aggregation under different time delays. It can extract the periodic dependence and cross-period evolution law of GIS equipment state variables over long time scales, and is suitable for prediction tasks where insulation state changes last for a long time and short-term fluctuations have a significant impact.

[0019] 4. This invention provides an insulation aging level based on the trend slope, threshold proximity, and degree of continuous anomaly within a future prediction period. It can identify the tendency of continuous degradation before the state quantity seriously exceeds the limit, providing an auxiliary basis for inspection, testing, and maintenance arrangements. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention 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 invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 is a flowchart of the GIS equipment insulation aging judgment method based on Autoformer long-time modeling according to an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0023] Example 1 like Figure 1 As shown in the figure, the GIS equipment insulation aging judgment method based on Autoformer long-time series modeling provided in this embodiment includes the following steps: Step 1: Collect multi-source state time-series data of GIS equipment during continuous operation; and preprocess the time-series data to construct a historical input sequence of insulation status of GIS equipment; Step 2: Construct a knowledge base for the GIS insulation aging domain and form a corresponding domain knowledge reasoning model. Use the domain knowledge reasoning model to perform semantic encoding and knowledge reasoning on the GIS insulation aging domain knowledge base, the current GIS equipment multi-source state variables, and equipment operation information to generate knowledge-encoded data. Perform multi-dimensional consistency verification on the knowledge-encoded data to obtain the verified GIS insulation aging knowledge encoding results. Step 3: Based on the historical input sequence of the insulation status of the GIS equipment and the verified GIS insulation aging knowledge encoding results, construct and train a knowledge-enhanced insulation status prediction model; Step 4: Input the historical input sequence of the insulation state of the GIS equipment into the trained insulation state prediction model. Using the decomposition mechanism in the insulation state prediction model, and based on the trend constraint parameters in the knowledge encoding results, perform adaptive trend-period decomposition on different types of state variables to obtain the trend term representing the long-term degradation trend and the periodic term representing the periodic fluctuation. Step 5: Input the trend term representing the long-term degradation trend and the periodic term representing the periodic fluctuation into the time-series dependency extraction mechanism in the insulation state prediction model, combine the knowledge encoding results to model the long-range dependency relationship in the sequence, and output the predicted values ​​of the multi-source state quantities and the insulation aging trend characterization quantities in the future prediction period. Step 6: Based on the predicted values ​​of the multi-source state quantities and the insulation aging trend characterization quantities during the future prediction period, and combined with the verified state quantity-fault mode association matrix and the prior probability of the fault mode in the knowledge encoding results, the insulation aging state level of the GIS equipment is comprehensively determined.

[0024] In step 1 of this embodiment, taking the GIS basin insulator to be monitored as the object, one or more state variables are continuously collected, including partial discharge amplitude, partial discharge pulse count, SF6 gas decomposition product content, trace moisture content, gas pressure, equipment temperature, operating voltage, operating current, and ambient temperature and humidity. The timestamps from the online monitoring device, sensors, and equipment operation records are unified, duplicate samples, obviously abnormal sampling points, and missing data are processed, and each state variable is formed into a multidimensional time series according to a unified sampling period. Specifically, the following sub-steps are included: Step 1.1, Construct the multi-source state vector of GIS equipment: Let the first... The multi-source state variables of the GIS basin insulator at each sampling time are:

[0025] In the formula, Indicates the first The state vector of the GIS basin insulator at each sampling time. Represents the dimension of state variables. They represent the first The state variable is in the first... The value at each sampling time; Step 1.2, Data Normalization Processing: Normalize each type of state variable separately to obtain the normalized GIS basin insulator state variable sequence:

[0026] In the formula, Represents the normalized th The sampling time of the first sampling moment A state quantity, and They represent the first and second training samples, respectively. Minimum and maximum values ​​of each state variable To prevent constants with a denominator of zero.

[0027] Step 1.3, Constructing the historical input window: A sliding window of length 1 is used to construct the historical input sequence of the insulation status of the GIS basin-type insulator.

[0028] In the formula, Indicates the length of the historical input. This represents the historical input sequence of the insulation status of GIS basin insulators; Step 2 of this embodiment includes the following sub-steps: Step 2.1, Construct a knowledge base for the GIS insulation aging domain: Structure the GIS equipment operation procedures, condition-based maintenance guidelines, test reports, historical fault cases, operation and maintenance records, and expert diagnosis rules to form a domain knowledge base that includes multi-source state variables, equipment types, equipment components, operating years, operating conditions, historical maintenance records, influencing factors, insulation aging mechanisms, fault modes, and state judgment rules. Step 2.2, Enhanced Domain Knowledge Retrieval: Based on the current GIS equipment's multi-source status variables, equipment type, equipment components, service life, operating conditions, and historical maintenance records, retrieve the most relevant knowledge fragments to the current equipment status from the GIS insulation aging domain knowledge base:

[0029] In the formula, This represents the set of domain knowledge fragments retrieved. This represents a knowledge base for the field of insulation aging of GIS basin insulators. The first in i Candidate knowledge fragments or knowledge documents. This represents a knowledge base for the field of insulation aging in GIS basin insulators. This indicates a search query consisting of the current status of GIS basin insulators and equipment operation information. and These represent the query encoder and the document encoder, respectively. This represents the similarity calculation function; Step 2.3, Domain Knowledge Reasoning Model Adaptation: Based on the general pre-trained large model, the domain of GIS basin insulator insulation aging is adapted through one or more methods, such as instruction fine-tuning or low-rank adaptation fine-tuning, to obtain a domain knowledge reasoning model with the ability to call domain knowledge, reason about mechanisms, and output structured data related to GIS basin insulator insulation aging.

[0030] In the formula, Representation of domain knowledge reasoning model, This represents a general pre-trained large model. This represents a set of instruction adaptation samples for the GIS insulation aging field. The set is compiled from GIS fault cases, expert Q&A, status evaluation samples, diagnostic rules, and maintenance records. Each instruction adaptation sample includes input prompts and target structured outputs. The input prompts include multi-source status variables of the GIS equipment, equipment information, relevant knowledge fragments, and structured output constraints. The target structured outputs include status variable importance scores, status variable-fault mode association matrices, fault mode prior scores, trend constraint parameters, delay prior parameters, and diagnostic conclusion labels. This indicates a domain adaptation operation, which includes one or more of the following: retrieval enhancement generation, prompt template constraint, instruction fine-tuning, and low-rank adaptation fine-tuning. When fine-tuning using instructions, the parameters of the domain-wide model are represented as follows:

[0031] In the formula, This represents the parameters of a general pre-trained large model. Represents the parameters of the large model in the domain. This indicates a supervisory instruction for fine-tuning operations.

[0032] When using low-rank adaptation fine-tuning, the large model parameters are expressed as:

[0033]

[0034] In the formula, and Represents a low-rank fitness matrix; Step 2.4, Construct Structured Output Hints: Based on the retrieved domain knowledge fragments, multi-source state variables of GIS equipment, equipment type, equipment components, service life, operating conditions, historical maintenance records, and structured output constraints, construct input hints for the large model:

[0035] In the formula, This represents the statistical and trend characteristics extracted from the current GIS device status data. This indicates the equipment type, years of service, operating conditions, and historical maintenance records. This represents structured output constraints, which include state quantity importance score dimensions. The dimensions of the state-fault mode correlation matrix are: The prior score dimension of the failure mode is .

[0036] Step 2.5, Semantic Encoding of Domain Knowledge Reasoning Model: The domain knowledge reasoning model is used to semantically encode the input prompts of the large model, resulting in a semantically hidden representation.

[0037] And the prior vector of insulation aging mechanism is obtained by pooling or mapping:

[0038] In the formula, The semantic hidden representation representing the output of a large model. Represents the prior vector of the insulation aging mechanism; Step 2.6, Generation of the original structured results of the large model: Based on the prior vector of the insulation aging mechanism and the constraints of the structured output of the large model, the original structured output of the large model is generated:

[0039] in,

[0040]

[0041]

[0042]

[0043] In the formula, This represents the importance score of the large model state variables. This represents the state-failure mode correlation matrix of a large model. This represents the prior score of the failure mode in the large model. This represents the trend constraint parameters of the large model. This represents the prior parameters of the large model delay. This indicates the confidence level of the large model output; , , , These represent the large model state quantity importance score mapping matrix, the large model state quantity-failure mode correlation matrix mapping matrix, the large model failure mode prior score mapping matrix, and the large model trend constraint parameter mapping matrix, respectively. This represents the Sigmoid activation function, used to map association strength to a preset range; This represents the hyperbolic tangent activation function, used to generate trend constraint parameters; Indicates the number of multi-source status variables of GIS devices; Indicates the number of failure modes; Let represent the real matrix space consisting of d state variables and m types of fault modes.

[0044] Step 2.7, Construction of Expert Rule Benchmark Parameters: Based on the expert diagnostic rules and historical fault cases in the knowledge base for GIS pot-type insulator insulation aging, construct the expert rule benchmark parameters:

[0045] In the formula, This represents the importance score of the expert rule state quantity. This represents the expert rule state quantity-failure mode correlation matrix. This represents the prior score of the failure mode in the expert rules. This represents the trend constraint parameters of the expert rule. Represents the expert rule delay prior function; Step 2.8, Output Consistency Verification: Perform dimensionality verification, range verification, normalization verification, and expert rule consistency verification on the original structured output of the large model, and calculate the conflict degree.

[0046] In the formula, This indicates the degree of conflict between the output of the large model and the expert rules. Represents the Frobenius norm; Calculate search relevance:

[0047] In the formula, Indicates the retrieval relevance; K represents the number of relevant knowledge fragments selected from the GIS pot-type insulator insulation aging knowledge base; i Indicates the sequence number of the knowledge segment; This indicates a search query consisting of the current status and operational information of the GIS basin insulator; This represents the query encoder, used for vectorized encoding of retrieval queries; This represents a document encoder, used to vectorize candidate knowledge fragments; This represents a similarity calculation function used to calculate the semantic relevance between a retrieval query and candidate knowledge fragments.

[0048] Calculate the consistency score:

[0049] Step 2.9, Calculation of the Trustworthy Fusion Coefficient from the Large Model Output: Based on the confidence score, retrieval relevance, consistency score, and conflict score of the large model output, calculate the trustworthy fusion coefficient:

[0050] In the formula, This indicates the reliable fusion coefficients output by the large model. Indicates the adjustment coefficient; Step 2.10, Generation of Verified State Importance Weights: Based on the state importance scores of the large model, the state importance scores of the expert rules, and the reliable fusion coefficients, generate the verified state importance weights.

[0051] In the formula, Indicates the first verifiable number j The importance weights of each state variable; j Indicates the current state variable number; r Indicates the index of the state variable; d Indicates the number of multi-source status variables of GIS devices; This represents the reliable fusion coefficient output by the large model, used to adjust the fusion ratio between the large model results and the expert rule results; The first generation of the domain large model represents the... j Importance score of each state variable; The first rule generated by the expert rules j Importance score of each state variable; The first generation of the domain large model represents the... r Importance score of each state variable; The first rule generated by the expert rules r Importance score of each state variable; Represents an exponential function; the denominator represents the function with respect to all... d The fusion scores of the state variables are exponentially normalized so that the weights of each state variable satisfy the following conditions: .

[0052] Step 2.11, Generation of the Verified State-Fault Mode Correlation Matrix: Based on the large model correlation matrix, the expert rule correlation matrix, and the reliable fusion coefficients, generate the verified state-fault mode correlation matrix:

[0053] in,

[0054] In the formula, This represents the verified state variables—fault mode correlation matrix; This represents the state-failure mode correlation matrix generated by the domain-wide model. This represents the state-failure mode correlation matrix generated by expert rules; it is used to determine the fusion ratio between the large-scale model correlation matrix and the expert rule correlation matrix in the control domain, and ; This represents the weight of the expert rule association matrix in the fusion process; This indicates a numerical truncation operation, used to limit the correlation strength after fusion to the range [0,1]; where 0 represents the lower limit of the correlation strength and 1 represents the upper limit of the correlation strength. Indicates the first verifiable number j The state quantity and the first q The strength of the association between different failure modes; j Indicates the state variable number; q Indicates the fault mode number; d Indicates the number of multi-source status variables of GIS devices; m Indicates the number of failure modes; Indicates by d Each state quantity and m The space of real-valued matrices formed by the fault modes.

[0055] Step 2.12, Generation of Verified Fault Mode Prior Probabilities: Based on the prior scores of fault modes in the large model, the prior scores of fault modes in the expert rules, and the reliability fusion coefficients, the verified prior probabilities of fault modes are generated.

[0056] In the formula, Indicates the first verifiable number q Prior probabilities of failure modes; q This represents the fault mode number whose prior probability is currently to be calculated; r This represents the failure mode index during the normalization summation process; m Indicates the number of failure modes; This represents the reliable fusion coefficients output by the large model, used to control the fusion ratio between the prior scores of fault modes in the domain-specific large model and the prior scores of fault modes in expert rules. ; This represents the weight of the prior score of the expert rule failure mode in the fusion process; The first generation of the domain large model represents the... q Prior score for each type of failure mode; The expert rule gives the first q Prior score for each type of failure mode; This represents the prior score of the r-th type of failure mode generated by the domain-wide large model. The expert rule gives the first r Prior score for each type of failure mode; Represents an exponential function; Indicates all m The exponential fusion scores of the fault modes are summed to implement Softmax normalization, ensuring that the prior probabilities of each fault mode satisfy the following: .

[0057] Step 2.13, Generation of validated trend constraint parameters: Based on the trend constraint parameters of the large model, the trend constraint parameters of the expert rules, and the credibility fusion coefficient, the validated trend constraint parameters are generated.

[0058] In the formula, , Indicates the first Trend constraint parameters corresponding to each state variable This represents the trend constraint parameters generated by the large model. This represents the trend constraint parameters of the expert rule. Step 2.14, Generation of Prior Delay Information after Verification: Based on the operating cycle, environmental cycle, maintenance cycle, and aging mechanism cycle of the GIS basin insulator identified by the large model, the prior delay function of the large model is generated:

[0059] Then, it is fused with the expert rule delay prior function to obtain the verified delay prior information:

[0060] In the formula, This represents the candidate time delay variable, used to characterize the number of sampling steps between the current GIS device state sequence and the historical state sequence; Indicates the first r The central delay corresponding to the class delay prior is used to characterize the typical operating cycle, environmental cycle, maintenance cycle or insulation aging evolution cycle identified by the domain large model; Indicates the first r The allowable fluctuation range of latency in the class center; Indicates the first r Importance weights of latency-related priors; Indicates the number of prior cycles of delay; Represents the time delay prior function of the large model; Represents the expert rule delay prior function; This represents the prior function for the time delay after verification.

[0061] Step 2.15, Conflict Regeneration Processing: When the conflict level... Greater than the preset conflict threshold If the input prompts for the large model are reconstructed, steps 2.4 to 2.14 are repeated. If the conflict degree still exceeds the preset conflict threshold after the number of repetitions reaches the preset upper limit, the reliable fusion coefficient is reduced. And the expert rule benchmark parameters are adopted first.

[0062] Step 3 of this embodiment includes the following sub-steps: Step 3.1, Prediction Task Parameter Setting: Based on the sampling period of the GIS basin insulator state variables, the dimension of the state variables, and the insulation aging judgment requirements, set the historical input length of the Autoformer insulation state prediction model. Predicted length State Dimension Model hidden layer dimension Number of encoder layers Number of decoder layers Multiple heads Feedforward network dimension Moving average kernel size range and the Auto-Correlation delay selection factor ; Step 3.2, Knowledge-enhanced input embedding layer construction: The historical input sequence of multi-source state variables of GIS devices and the validated domain large model knowledge encoding results are jointly mapped to the model latent space to obtain the knowledge-enhanced input representation:

[0063] In the formula, This indicates the operations of embedding state variable values ​​and embedding time stamps. This represents the input sequence of historical status data for GIS devices. This represents the weight vector indicating the importance of the verified state variables. and Represents a trainable mapping matrix. This indicates weighting based on the state variable dimension; Step 3.3, Knowledge Constraint Sequence Decomposition Module Construction: Based on the importance weights of the verified state variables and trend constraint parameters, different moving average kernel sizes or trend extraction weights are assigned to different GIS state variables, so that insulation aging sensitive state variables such as partial discharge characteristics, SF6 gas decomposition product content, and trace moisture content focus more on long-term trend extraction, and disturbance-related state variables such as equipment temperature, operating current, and ambient temperature and humidity focus more on periodic fluctuation separation. Step 3.4, Knowledge Enhancement Auto-Correlation Module Construction: Construct linear mapping parameters for the query matrix, key matrix, and value matrix, and introduce the verified time delay prior information into the Auto-Correlation mechanism to enhance the knowledge of sequence correlation under different time delays; Step 3.5, Encoder and Decoder Structure Construction: Construction of the encoder and decoder structure... An encoder formed by stacking coding layers and by A decoder is formed by stacking decoding layers; wherein each encoding layer includes a knowledge-enhancing auto-correlation module, a residual connection, a knowledge-constrained sequence decomposition module and a feedforward network, and each decoding layer includes a decoder-internal knowledge-enhancing auto-correlation module, an encoder-decoder knowledge-enhancing auto-correlation module, a feedforward network, multiple knowledge-constrained sequence decomposition modules and a trend term accumulation structure. Step 3.6, Model Training Loss Construction: Construct the total training loss based on the prediction loss, trend consistency loss, mechanism consistency loss, and diagnostic distillation loss.

[0064] In the formula, Indicates the predicted loss. This indicates a loss of trend consistency. Indicates the loss of consistency in mechanism. Indicates diagnostic distillation loss, , , Indicates the loss weight; Step 3.7, Model Parameter Update: The gradient descent optimization method is used to backpropagate and update the trainable parameters in the knowledge-enhanced Autoformer insulation state prediction model until the total training loss converges or the preset training rounds are reached.

[0065] The prediction loss, trend consistency loss, mechanism consistency loss, and diagnostic distillation loss in step 3.6 include: The predicted loss is:

[0066] In the formula, Indicates the first The predicted time point Predicted values ​​of each state variable Indicates the first The predicted time point The true value of each state variable; The trend consistency loss is:

[0067] In the formula, Indicates the predicted trend component. This represents the reference trend component obtained by decomposing the real state quantity sequence. Mechanism consistency loss is used to ensure that the state-fault mode associations learned by the knowledge-enhanced Autoformer remain consistent with the verified domain knowledge associations of the GIS basin insulator. This is based on the periodic state representation output from the final decoding layer in step 5.7. Generate the state variables—fault mode association matrix learned by the knowledge-enhanced Autoformer:

[0068] The mechanism consistency loss is:

[0069] In the formula, This indicates that the knowledge-enhanced Autoformer learns the state-fault mode correlation matrix based on the output of the final decoding layer. This represents the periodic state representation of the final decoding layer output; Indicates the parameters of the correlation matrix mapping; Indicates pooling operation; This represents a matrix reconstruction operation; This represents the verified state quantity-fault mode correlation matrix. Represents the Frobenius norm; Diagnostic distillation loss is used to ensure that the failure mode probabilities output by the knowledge-enhanced Autoformer are consistent with the prior failure mode probabilities of the validated large model. This is based on the periodic state representation output by the final decoding layer. The resulting failure mode probability vector from the knowledge-enhanced Autoformer is:

[0070] Diagnostic distillation loss is as follows:

[0071] In the formula, This represents the prior probability vector of the verified fault mode obtained in step 2.12. This indicates that the knowledge-enhancing Autoformer uses the fault mode probability vector obtained from the final decoding layer output. This represents the periodic state representation of the final decoding layer output. This represents the failure mode probability mapping parameter. This represents the KL divergence.

[0072] In step 4 of this embodiment, during model training and prediction, the hidden sequence input to the sequence decomposition module... First, the trend term is extracted using a moving average pooling method with fill. The periodic term is then obtained by subtracting the trend term from the original hidden sequence. The trend term is used to characterize the slow evolution direction of state variables over a long period of time; the periodic term is used to characterize the periodic or short-term fluctuations caused by operating load, equipment temperature, and environmental disturbances.

[0073] For periodic state representations, the Auto-Correlation module maps them to a query matrix, a key matrix, and a value matrix, uses Fast Fourier Transform to calculate the correlation at different time delays, and selects the first... A significant time delay is identified. Similar state subsequences corresponding to the selected time delay are cyclically shifted and weighted to align state changes in similar operational phases across different cycles, thereby obtaining a dependency representation over a long timescale.

[0074] Specifically, it includes the following sub-steps: Step 4.1, Determining the Adaptive Moving Average Kernel: Based on the importance weights of the verified state variables obtained in Step 2.10 and the verified trend constraint parameters obtained in Step 2.13, determine the kernel of the adaptive moving average. The moving average kernel size corresponding to each state variable:

[0075] In the formula, Indicates the first The moving average kernel size corresponding to each state variable Represents the normalized i-th The importance weight of each state variable Indicates the first Trend constraint parameters corresponding to each state variable and This represents the adjustment coefficient. Represents the Sigmoid function; Step 4.2, Trend Item Extraction: For the first... Perform a moving average pooling operation with padding on a sequence of state variables to extract the trend term:

[0076] In the formula, Indicates the first Trend terms of each state quantity This represents the historical input sequence corresponding to the j-th state variable. This indicates the boundary fill operation. This represents the adaptive moving average kernel size corresponding to the j-th state variable. Indicates the kernel size as Average pooling operation; Step 4.3, Periodic Term Extraction: Subtract the trend term from the original state quantity sequence to obtain the periodic term:

[0077] In the formula, Indicates the first Periodic terms of each state quantity; Step 4.4, Construction of Insulation Aging Trend Components: The trend terms of each state variable are weighted according to their importance to obtain the comprehensive insulation aging trend components of the GIS pot insulator.

[0078] In the formula, This represents the overall trend component of insulation aging in GIS equipment.

[0079] In step 5 of this embodiment, the encoder extracts features from the periodic state changes of the historical input window. The decoder input consists of the initial periodic term and the initial trend term obtained from the decomposition of the latter half of the historical sequence. The periodic term in the prediction segment uses zero-sequence placeholders, and the trend term in the prediction segment uses the mean of the historical sequence as placeholders. In each decoding layer, the decoder sequentially performs internal auto-correlation calculation, encoder-decoder auto-correlation calculation, feedforward transformation, and sequence decomposition, accumulating the trend components obtained from each decomposition layer by layer, and finally outputs the future... O The multi-source state variable prediction sequence and trend components within each sampling time point. Specifically, it includes the following sub-steps: Step 5.1, Construct the query matrix, key matrix, and value matrix: Map the hidden representation of GIS device status variables to the query matrix respectively. Key matrix Sum matrix ; Step 5.2, Calculate the sequence autocorrelation results: Use Fast Fourier Transform to calculate the autocorrelation relationship between the query matrix and the key matrix at different time delays:

[0080] In the formula, Indicates a time delay of The autocorrelation results of the sequence at time , Represents the Fast Fourier Transform. This represents the inverse fast Fourier transform. Indicates the conjugate operation; Step 5.3, Integrating the domain-wide model's time-delay prior: The sequence autocorrelation results are fused with the time-delay prior information of the large model to obtain the knowledge-enhanced autocorrelation results.

[0081] In the formula, This indicates knowledge-enhancing autocorrelation results. This represents the prior information about the time delay of a large model. Represents the time-delay prior fusion coefficient; Step 5.4, Select significant time delays: Select the first significant time delay based on the autocorrelation results of knowledge reinforcement. A significant time delay:

[0082] in, , The time delay selection factor is used. The length of the input sequence; Step 5.5, Time Delay Confidence Normalization: Softmax normalization is applied to the selected knowledge-enhanced autocorrelation values ​​to obtain the weights corresponding to different time delays.

[0083] Step 5.6, Delay Aggregation: Circularly shift the value matrix according to the selected time delay, and perform weighted aggregation on similar subsequences:

[0084] In the formula, This represents the knowledge-enhancing Auto-Correlation operation. This indicates that the value matrix is ​​adjusted according to time delay. Perform a circular shift; Step 5.7, Generate future state prediction sequence: The knowledge-enhanced Auto-Correlation output is processed through the encoder-decoder structure to obtain the final periodic state representation of the decoder and the future multi-state quantity trend matrix. The two are then fused to generate the future state prediction sequence of GIS basin insulator.

[0085] Specifically, in the knowledge-enhanced Autoformer model construction phase, the output mapping matrix is ​​set. And use these as trainable parameters for the model; during model training, The parameters of the encoder, decoder, and knowledge-enhancing Auto-Correlation module are updated via backpropagation.

[0086] Simultaneously, according to the importance weights of the state variables in step 4.4, the future multi-state variable trend matrix is ​​weighted to obtain the comprehensive trend components of future insulation aging:

[0087]

[0088] In the formula, This represents a sequence predicting the future state of a GIS basin insulator. This represents the periodic state representation of the final decoding layer output; The output mapping matrix represents the periodic state representation. It is a trainable parameter in the knowledge-enhanced Autoformer model and is used to map the periodic state representation output by the decoder to the GIS multi-source state space. This represents the trend matrix of future multi-state variables finally accumulated by the decoder; This represents the future trend component of the j-th state variable; d represents the importance weight of the j-th state variable after verification obtained in step 2.10; d represents the number of multi-source state variables of the GIS basin insulator. This represents the predicted comprehensive trend component of insulation aging in future GIS pot insulators.

[0089] In this embodiment, the calculation process of the encoder-decoder structure includes the following sub-steps: Encoder input initialization:

[0090] In the formula, This represents the input of the encoder at layer 0. Representation of knowledge-enhanced input embedding representation; Encoder knowledge enhancement correlation modeling and first decomposition: In the l-th encoding layer, the output of the previous encoder is input into the knowledge enhancement Auto-Correlation module, and the first periodic component and the first trend component are obtained through the residual connection and knowledge constraint sequence decomposition module.

[0091] In the formula, This represents the knowledge-constrained sequence decomposition operation. This represents the periodic component obtained from the first decomposition of the l-th coding layer. This represents the trend component obtained from the first decomposition of the l-th coding layer. Encoder feedforward transform and second decomposition:

[0092] And the second periodic component is used as the output of the l-th coding layer:

[0093] Encoder input initialization: The latter half of the encoder input sequence is decomposed using knowledge constraints to obtain the initial periodic term and the initial trend term.

[0094] And construct the decoder's periodic term input and trend term input:

[0095]

[0096] In the formula, This represents a zero-placement sequence of length O. This represents a trend placeholder sequence filled with the mean of a historical input sequence. Decoder internal knowledge enhancement relevance modeling and first decomposition:

[0097] Encoder-Decoder Knowledge Enhancement Relevance Modeling and Second Decomposition:

[0098] In the formula, This indicates the output of the last layer of the encoder; Decoder feedforward transform and third decomposition:

[0099] And the third periodic component is used as the output of the l-th decoding layer:

[0100] Decoder trend accumulation: In each layer, the decoder extracts multiple trend components through the knowledge-constrained sequence decomposition module, maps them, and accumulates them to obtain the future multi-state trend matrix of the l-th decoding layer.

[0101] In the formula, This represents the trend matrix of future multi-state variables accumulated by the l-th decoding layer. Indicates the lth The cumulative trend result obtained from one decoding layer. This represents the first trend component obtained by knowledge-constrained sequence decomposition after knowledge-enhanced Auto-Correlation within the decoder in the l-th decoding layer. This represents the second trend component obtained by knowledge-constrained sequence decomposition of the l-th decoding layer after encoder-decoder knowledge-enhanced Auto-Correlation. This represents the third trend component obtained by knowledge-constrained sequence decomposition after the feedforward network transformation of the l-th decoding layer. , , This represents the learnable mapping parameter corresponding to the trend component.

[0102] In step 6 of this embodiment, state variables that are positively correlated with the degree of insulation aging are directly evaluated based on the predicted sequence and predicted trend components; for state variables whose values ​​decrease indicating deterioration, directional consistency processing can be performed before evaluation so that subsequent indicators all indicate that the larger the value, the closer to the aging risk. For the first... Given a state variable, calculate the trend slope. Threshold proximity and the degree of persistent abnormality .

[0103] When all state variables are at a low threshold level, a low level of continuous abnormality, and the trend slope has not reached the threshold of concern, the output insulation status is normal; when at least one state variable approaches the threshold of concern or the trend continues to deteriorate, the output insulation aging is monitored; when at least one state variable simultaneously meets the conditions of approaching a high threshold and continuous abnormality, the output insulation aging is warned; when at least one state variable reaches or exceeds the alarm limit, or exceeds the alarm limit at multiple consecutive prediction times, the output insulation aging is severe.

[0104] In this embodiment, the judgment parameters can be preset and updated based on the corresponding equipment type, operating voltage level, testing specifications, historical fault samples, expert experience, or equipment operation and management requirements. Prediction results, trend curves, significant delays, and insulation aging levels can be output to the monitoring terminal for maintenance personnel to perform status verification, special testing, or maintenance decisions. Specifically, this includes: Step 6.1, calculate the trend slope of the state variables: for the first... For each GIS basin insulator insulation state quantity, the trend slope for the future prediction period is calculated based on the predicted trend components:

[0105] In the formula, Indicates the first The trend slope of each state quantity Indicates the first The predicted time point The trend component of each state quantity This represents the average of the predicted time sequence numbers. Indicates the first [number]th ... The mean of the trend components of each state quantity; Step 6.2, Calculate the proximity of the threshold: For state variables that indicate insulation aging, calculate the proximity of their predicted values ​​to preset limits:

[0106] In the formula, Indicates the first The degree to which the thresholds of the state variables are close Indicates the first The predicted time point Predicted values ​​of each state variable This indicates the normal reference value. Indicates the alarm limit; Step 6.3, Calculate the degree of persistent anomaly: Calculate the proportion of sampling points exceeding the preset attention threshold during the prediction period.

[0107] In the formula, Indicates the first The degree of persistent abnormality of each state variable Indicates the attention threshold. Indicates an indicator function; Step 6.4, Calculate the overall confidence of the failure mode: Based on the trend slope, threshold proximity, degree of persistent anomaly, state-failure mode correlation matrix, and prior probability of the failure mode, calculate the overall confidence of the q-th type of failure mode:

[0108] In the formula, Indicates the first The overall confidence level of the failure modes Indicates the first The state quantity and the first The strength of the association between different failure modes , , , Indicates the weighting coefficient. Represents the Sigmoid function; Step 6.5, calculate the comprehensive risk value of insulation aging for GIS pot-type insulators:

[0109] In the formula, This indicates the comprehensive risk value of insulation aging in GIS equipment. Indicates the first Risk weights corresponding to different failure modes; Step 6.6, Determine the insulation aging assessment result: Based on the comprehensive risk value of insulation aging of GIS pot-type insulators. Determine the insulation aging status of GIS basin insulators: when When the insulation condition is normal, it is determined that the insulation condition is normal. when At that time, it was determined to be a concern due to insulation aging; when When this occurs, it is determined to be an insulation aging warning; when At that time, it was determined that the insulation was severely aged; in, , , To preset risk thresholds, and .

[0110] Example 2 This embodiment provides a GIS equipment insulation aging judgment system based on Autoformer long-time series modeling, including: The status quantity acquisition and preprocessing module is used to acquire multi-source status quantity time-series data of GIS equipment during continuous operation; and to preprocess the time-series data to construct a historical input sequence of insulation status of GIS basin insulators. The knowledge encoding module is used to construct a knowledge base for the field of GIS insulation aging and form a corresponding domain knowledge reasoning model. The domain knowledge reasoning model is used to perform semantic encoding and knowledge reasoning on the GIS pot insulator insulation aging knowledge base, the current GIS equipment multi-source state variables and equipment operation information to generate knowledge encoding data. The knowledge encoding data is then subjected to multi-dimensional consistency verification to obtain the verified GIS pot insulator insulation aging knowledge encoding result. The model building and training module is used to build and train a knowledge-enhanced insulation state prediction model based on the historical input sequence of the insulation state of the GIS equipment and the verified insulation aging knowledge encoding result of the GIS basin insulator. The sequence decomposition module is used to input the historical input sequence of the insulation state of the GIS equipment into the trained insulation state prediction model. Using the decomposition mechanism in the insulation state prediction model, and based on the trend constraint parameters in the knowledge encoding results, it performs adaptive trend-period decomposition on different types of state variables to obtain trend terms that characterize long-term degradation trends and periodic terms that characterize periodic fluctuations. The temporal dependency extraction and prediction module is used to input the trend term representing the long-term degradation trend and the periodic term representing the periodic fluctuation into the temporal dependency extraction mechanism in the insulation state prediction model, combine the knowledge encoding results to model the long-range dependency relationship in the sequence, and output the predicted values ​​of the multi-source state quantities and the insulation aging trend characterization quantities in the future prediction period. The insulation aging comprehensive judgment module is used to comprehensively determine the insulation aging status level of GIS pot insulators based on the predicted values ​​of multi-source state quantities and insulation aging trend characterization quantities in the future prediction period, combined with the verified state quantity-fault mode correlation matrix and fault mode prior probability in the knowledge encoding results.

[0111] It should be understood that any parts not described in detail in this specification belong to the prior art.

[0112] It should be understood that the above description of the preferred embodiments is quite detailed, but this should not be construed as limiting the scope of protection of this invention. It is neither necessary nor possible to exhaustively describe all possible implementations. Those skilled in the art, guided by this invention, can make substitutions or modifications without departing from the scope of the claims, all of which fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for judging the insulation aging of GIS equipment based on Autoformer long-time series modeling, characterized in that, The method includes: Collect multi-source state time-series data of GIS equipment during continuous operation; and preprocess the time-series data to construct a historical input sequence of insulation status of GIS equipment; A knowledge base for the GIS insulation aging domain is constructed, and a corresponding domain knowledge reasoning model is formed. The domain knowledge reasoning model is used to perform semantic encoding and knowledge reasoning on the GIS insulation aging domain knowledge base, the current GIS equipment multi-source state variables, and equipment operation information to generate knowledge-encoded data. The knowledge-encoded data is then subjected to multi-dimensional consistency verification to obtain the verified GIS insulation aging knowledge encoding results. Based on the historical input sequence of the insulation status of the GIS equipment and the verified GIS insulation aging knowledge encoding result, a knowledge-enhanced insulation status prediction model is constructed and trained. The historical input sequence of the insulation state of the GIS equipment is input into the trained insulation state prediction model. Using the decomposition mechanism in the insulation state prediction model, and based on the trend constraint parameters in the knowledge encoding results, adaptive trend-period decomposition is performed on different types of state variables to obtain trend terms that characterize long-term degradation trends and periodic terms that characterize periodic fluctuations. The trend term representing the long-term degradation trend and the periodic term representing the periodic fluctuation are input into the time-series dependency extraction mechanism in the insulation state prediction model. The long-range dependency relationship in the sequence is modeled in combination with the knowledge encoding result, and the predicted values ​​of the multi-source state quantities and the insulation aging trend characterization quantities are output in the future prediction period. Based on the predicted values ​​of multi-source state quantities and insulation aging trend characterization quantities during the future prediction period, combined with the verified state quantity-fault mode correlation matrix and fault mode prior probability in the knowledge encoding results, the insulation aging state level of the GIS equipment is comprehensively determined.

2. The method for judging the insulation aging of GIS equipment based on Autoformer long-term modeling according to claim 1, characterized in that, The multi-source state time series data includes at least one or more of the following: partial discharge characteristic quantities, SF6 gas decomposition product content, trace water content, gas pressure, equipment temperature, operating voltage, operating current, and ambient temperature and humidity.

3. The method for judging the insulation aging of GIS equipment based on Autoformer long-term modeling according to claim 1, characterized in that, The construction of the GIS insulation aging knowledge base specifically includes: By integrating GIS equipment operation procedures, condition-based maintenance guidelines, test reports, historical fault cases, operation and maintenance records, and expert diagnostic rules, a knowledge base for the field of GIS insulation aging is constructed, which includes multi-source state variables, equipment types, equipment components, service life, operating conditions, historical maintenance records, influencing factors, insulation aging mechanisms, fault modes, and condition judgment rules.

4. The method for judging the insulation aging of GIS equipment based on Autoformer long-time series modeling according to claim 1, characterized in that, The generated knowledge encoding data includes: Based on the retrieval enhancement generation mechanism, domain knowledge fragments related to the current equipment status are retrieved from the GIS insulation aging domain knowledge base; Based on a general pre-trained large model, GIS insulation aging domain adaptation is performed through instruction fine-tuning or low-rank adaptation fine-tuning constraints to obtain a domain knowledge reasoning model with the ability to call GIS insulation aging domain knowledge, reason about mechanisms, and output structured data. The large model parameter representation of the domain knowledge reasoning model is also obtained. Based on the retrieved domain knowledge fragments, GIS equipment multi-source state variables, equipment type, equipment components, service life, operating conditions, historical maintenance records, and structured output constraints, input prompts for the large model are constructed. Based on the large model parameter representation, the domain knowledge reasoning model performs semantic encoding on the input prompts of the large model to obtain a semantic hidden representation; and obtains the prior vector of insulation aging mechanism through pooling or mapping: Based on the prior vector of insulation aging mechanism and the structured output constraints of the large model, knowledge-encoded data is generated, including the prior vector of insulation aging mechanism of the large model, the importance weight of the state variables of the large model, the state variable-failure mode correlation matrix of the large model, the prior probability of the failure mode of the large model, the trend constraint parameters of the large model, and the time delay prior information of the large model.

5. The method for judging the insulation aging of GIS equipment based on Autoformer long-term modeling according to claim 4, characterized in that, The multi-dimensional consistency verification of the knowledge encoding data, resulting in the verified GIS insulation aging knowledge encoding result, includes: Based on the expert diagnosis rules and historical fault cases in the GIS insulation aging knowledge base, the expert rule benchmark parameters are constructed, including the expert rule state quantity importance score, the expert rule state quantity-fault mode correlation matrix, the expert rule fault mode prior score, the expert rule trend constraint parameters, and the expert rule time delay prior function. The knowledge-encoded data is subjected to dimensionality verification, range verification, normalization verification, and expert rule consistency verification. Conflict degree is calculated based on the large model state variable-failure mode association matrix and the expert rule state variable-failure mode association matrix, and retrieval relevance is also calculated. A consistency score is calculated based on the search relevance: Based on the confidence level, retrieval relevance, consistency score, and conflict level of the domain knowledge reasoning model, the trustworthy fusion coefficient is calculated: Based on the importance scores of the large model state variables, the importance scores of the expert rule state variables, and the reliable fusion coefficient, the validated state variable importance weights are generated: Based on the large model state-failure mode correlation matrix, the expert rule state-failure mode correlation matrix, and the reliable fusion coefficients, a verified state-failure mode correlation matrix is ​​generated: Based on the prior scores of the large model failure modes, the prior scores of the expert rule failure modes, and the reliability fusion coefficient, the validated prior probabilities of failure modes are generated: Based on the trend constraint parameters of the large model, the trend constraint parameters of the expert rules, and the reliable fusion coefficients, the validated trend constraint parameters are generated: Based on the GIS operation cycle, environmental cycle, maintenance cycle, and aging mechanism cycle identified by the domain knowledge reasoning model, a large-scale model time delay prior function is generated. This function is then fused with the expert rule time delay prior function to obtain the validated time delay prior information. When the conflict level exceeds the preset conflict threshold, the input prompt of the large model is reconstructed and the above steps are repeated; if the conflict level still exceeds the preset conflict threshold after the number of repetitions reaches the preset upper limit, the trusted fusion coefficient is reduced and the expert rule benchmark parameter is adopted.

6. The method for judging the insulation aging of GIS equipment based on Autoformer long-time-series modeling according to claim 1, characterized in that, Building and training a knowledge-enhanced insulation state prediction model includes: Based on the sampling period of the multi-source state variables of GIS equipment, the state variable dimension, and the insulation aging judgment requirements, the historical input length, prediction length, state variable dimension, model hidden layer dimension, number of encoder layers, number of decoder layers, number of multi-heads, feedforward network dimension, moving average kernel size range, and Auto-Correlation delay selection coefficient of the insulation state prediction model are set. The historical input sequence of multi-source state variables of GIS equipment and the verified GIS insulation aging knowledge encoding result are jointly mapped to the latent space to obtain a knowledge-enhanced input representation. Based on the verified importance weights and trend constraint parameters of the state variables, different moving average kernel sizes or trend extraction weights are assigned to the state variables of different GIS devices, so that the insulation aging sensitive state variables in the multi-source state variables focus on long-term trend extraction, and the disturbance-related state variables in the multi-source state variables focus on periodic fluctuation separation. We construct linear mapping parameters for the query matrix, key matrix, and value matrix, and introduce the time delay prior information of the large model into the Auto-Correlation mechanism to enhance the knowledge of sequence correlation under different time delays; Build by An encoder formed by stacking coding layers and by A decoder is formed by stacking several decoding layers; each encoding layer includes a knowledge-enhancing auto-correlation module, a residual connection module, a knowledge-constrained sequence decomposition module, and a feedforward network; each decoding layer includes an internal knowledge-enhancing auto-correlation module, an encoder-decoder knowledge-enhancing auto-correlation module, a feedforward network, multiple knowledge-constrained sequence decomposition modules, and a trend term accumulation structure. The total training loss is constructed based on the prediction loss, trend consistency loss, mechanism consistency loss, and diagnostic distillation loss. The gradient descent optimization method is used to backpropagate and update the trainable parameters in the knowledge-enhanced insulation state prediction model until the total training loss converges or the preset number of training rounds is reached.

7. The method for judging the insulation aging of GIS equipment based on Autoformer long-time-series modeling according to claim 5, characterized in that, Adaptive trend-period decomposition is performed on different types of state variables to obtain trend terms characterizing long-term evolution and periodic terms characterizing the impact of periodic fluctuations, including: Based on the verified importance weights and trend constraint parameters of the state variables, determine the size of the adaptive moving average kernel for each state variable. Perform a moving average pooling operation with boundary padding on each state variable sequence to extract the trend term of each state variable; The periodic term of each state variable is obtained by subtracting the original sequence of each state variable from the corresponding trend term. The trend terms of each state variable are weighted according to the importance weight of the state variable to obtain the comprehensive trend component of insulation aging of GIS equipment, which is used as the characterization of the trend term.

8. The method for judging the insulation aging of GIS equipment based on Autoformer long-term modeling according to claim 7, characterized in that, The time-series dependency extraction mechanism specifically includes: The hidden representation of GIS device status variables is mapped to query matrix, key matrix, and value matrix, respectively. Calculate the autocorrelation between the query matrix and the key matrix at different time delays; The autocorrelation relationship is fused with the verified time delay prior information to obtain the knowledge-enhanced autocorrelation result; Based on the knowledge, significant time delays are selected to enhance the autocorrelation results, and the autocorrelation values ​​corresponding to each significant time delay are normalized. The value matrix is ​​cyclically shifted and weighted according to the selected time delay to obtain the aggregation result; The aggregation results are processed by an encoder-decoder structure to generate future prediction sequences and future trend components; According to the importance weight of the verified state variables, the future trend components are weighted and summed to obtain the comprehensive trend component of future insulation aging, and the comprehensive trend component of future insulation aging is output as the characteristic quantity of insulation aging trend.

9. The method for judging the insulation aging of GIS equipment based on Autoformer long-time series modeling according to claim 8, characterized in that, The comprehensive determination of the insulation aging status level of GIS equipment specifically includes: For each state variable, the trend slope of its future prediction period is calculated based on the future trend component, which is used to characterize the rate of change of the state variable. For state quantities that indicate insulation aging, calculate the degree to which their predicted values ​​are close to the threshold values ​​of preset normal reference values ​​and alarm limits. The proportion of sampling points exceeding a preset attention threshold in the future prediction sequence of each state variable is used as the degree of continuous anomaly of that state variable. Based on the trend slope, threshold proximity, and persistent anomaly of each state variable, combined with the correlation strength between each state variable and each fault mode in the verified state variable-fault mode correlation matrix, and the verified prior probability of the fault mode, the comprehensive confidence of each type of fault mode is calculated. The comprehensive risk value of insulation aging of GIS equipment is calculated by weighting the comprehensive confidence level of various failure modes and the preset risk weights corresponding to each failure mode. Based on the preset multi-level threshold range where the comprehensive risk value of insulation aging is located, the insulation aging status level of the GIS equipment is determined. The aging status level includes at least insulation status normal, insulation aging concern, insulation aging warning, and insulation aging severe.

10. A GIS equipment insulation aging judgment system based on Autoformer long-time series modeling, characterized in that, include: The state quantity acquisition and preprocessing module is used to acquire multi-source state quantity time series data of GIS equipment during continuous operation; and to preprocess the time series data to construct a historical input sequence of insulation state of GIS equipment. The knowledge encoding module is used to construct a knowledge base for the GIS insulation aging domain and form a corresponding domain knowledge reasoning model. The domain knowledge reasoning model is used to perform semantic encoding and knowledge reasoning on the GIS insulation aging domain knowledge base, the current GIS equipment multi-source state variables and equipment operation information to generate knowledge encoding data. The knowledge encoding data is then subjected to multi-dimensional consistency verification to obtain the verified GIS insulation aging knowledge encoding result. The model building and training module is used to build and train a knowledge-enhanced insulation state prediction model based on the historical input sequence of the insulation state of the GIS equipment and the verified GIS insulation aging knowledge encoding result. The sequence decomposition module is used to input the historical input sequence of the insulation state of the GIS equipment into the trained insulation state prediction model. Using the decomposition mechanism in the insulation state prediction model, and based on the trend constraint parameters in the knowledge encoding results, it performs adaptive trend-period decomposition on different types of state variables to obtain trend terms that characterize long-term degradation trends and periodic terms that characterize periodic fluctuations. The temporal dependency extraction and prediction module is used to input the trend term representing the long-term degradation trend and the periodic term representing the periodic fluctuation into the temporal dependency extraction mechanism in the insulation state prediction model, combine the knowledge encoding results to model the long-range dependency relationship in the sequence, and output the predicted values ​​of the multi-source state quantities and the insulation aging trend characterization quantities in the future prediction period. The insulation aging comprehensive judgment module is used to comprehensively determine the insulation aging status level of GIS equipment based on the predicted values ​​of multi-source state quantities and insulation aging trend characterization quantities in the future prediction period, combined with the verified state quantity-fault mode correlation matrix and fault mode prior probability in the knowledge encoding result. The GIS equipment insulation aging judgment system based on Autoformer long-time-series modeling is used to perform the steps in the GIS equipment insulation aging judgment method based on Autoformer long-time-series modeling as described in any one of claims 1-9.

Images