A method for intelligently generating an operation and inspection strategy of a power transformation device

Abstract

This invention discloses a method for intelligently generating operation and maintenance strategies for substation equipment, relating to the field of substation equipment operation and maintenance technology. The method involves collecting multi-dimensional data from the equipment and standardizing its processing; constructing a static feature baseline by combining fault causal relationships; optimizing the feature set through dynamic calibration and accurately determining the equipment's operating status; constructing and training a multi-constraint health assessment model, outputting health levels and feature contribution degrees; generating differentiated operation and maintenance strategies and performing double verification before accurately pushing them to the appropriate operation and maintenance terminals; and achieving full-process traceability through blockchain. This method overcomes the shortcomings of existing technologies, such as poor adaptability and insufficient accuracy, achieving precise and intelligent operation and maintenance, improving efficiency, and ensuring standardized operation and maintenance.

Description

Technical Field

[0001] This invention belongs to the field of power equipment operation and maintenance technology, specifically a method for intelligently generating power equipment operation and maintenance strategies. Background Technology

[0002] Power substation equipment is the core support for the safe and stable operation of the power system, and its operating status directly determines the reliability and security of the power grid. As the power system develops towards intelligence and large scale, the types of power substation equipment are constantly increasing and their structures are becoming more complex. Traditional operation and maintenance strategies rely heavily on manual experience, which has problems such as poor pertinence, insufficient adaptability, and delayed response, and can no longer meet the operation and maintenance needs of modern power substation equipment.

[0003] In existing technologies, the generation of operation and maintenance strategies for substation equipment is mostly based on fixed feature thresholds or simple data analysis, failing to fully consider the dynamic changes in equipment operating status and the deep causal relationships between faults and status data. This leads to inaccurate matching of operation and maintenance strategies, easily resulting in over-maintenance or under-maintenance. Furthermore, the feature extraction process in existing technologies often fails to incorporate fault causal relationships, feature calibration is lagging, and it cannot adapt to changes in equipment operating conditions in real time. Equipment health assessment models lack dynamic constraints from actual operating scenarios, resulting in insufficient accuracy. The generation of operation and maintenance strategies does not fully incorporate fault causes, and the full-process traceability mechanism after strategy execution is imperfect, making data susceptible to tampering and unable to achieve traceable management of the entire fault lifecycle.

[0004] Furthermore, existing technologies have relatively simple processing methods for discrete and continuous state data, lack a close correlation with fault types, and have insufficient effectiveness in feature selection; the push of operation and maintenance strategies does not fully consider the fault handling qualifications and responsibility areas of operation and maintenance personnel, resulting in low push accuracy and affecting operation and maintenance efficiency. Summary of the Invention

[0005] The purpose of this invention is to provide a method for intelligently generating operation and maintenance strategies for substation equipment. In view of the shortcomings of the prior art, there is an urgent need for an intelligent operation and maintenance strategy generation method that can accurately match the real-time status of equipment, dynamically adapt to changes in operating conditions, and realize full-process traceability, so as to improve the intelligence level and reliability of substation equipment operation and maintenance.

[0006] A method for intelligently generating operation and maintenance strategies for substation equipment includes the following steps: Step 1: Collect the condition evaluation results, operating status data and historical stable operation data of the power equipment. Perform hierarchical decomposition of fault causes and classification of fault types on the historical fault records. Establish a unique association relationship between fault types and condition evaluation results, operating status data and historical stable operation data through data association algorithms to form a standardized equipment status dataset with unified fields and unique mapping. Based on the inherent attributes of power equipment and the deep causal relationship between different fault types, a feature extraction mechanism is constructed that separates continuous and discrete data processing and associates fault causal relationships. Through this mechanism, state features that are directly causally related to faults are screened from the standardized equipment state dataset. Combined with the statistical analysis results of historical stable operation data, a static feature baseline that can be dynamically iterated is constructed, rather than a fixed feature baseline. Step 2: Continuously collect and dynamically track the various state features within the static feature baseline. By combining the time series mutation detection algorithm with the fault precursor features, locate the time node and cause of the mutation of the state parameters. Compare the real-time collected parameters with the static feature baseline dimension by dimension. Mark the abnormal data and abnormal level by the multi-threshold anomaly detection algorithm. Combine the distribution density, impact range and mutation cause of the abnormal data to comprehensively determine the current operating status of the equipment. The mutual information gradient integral algorithm is used to quantify the correlation sensitivity between various state features and equipment operating status and the influence of parameter changes. Based on the correlation sensitivity calculation results, features within the static feature baseline are screened, weights are redistributed, and structures are reconstructed to generate a dynamic calibration feature set that can adapt to changes in operating conditions in real time, thus solving the technical defects of feature calibration lag and inability to adapt to real-time operating conditions. Step 3: Use the static feature baseline, dynamic calibration feature set and corresponding equipment operating status as training samples, introduce multi-parameter dynamic constraints of the actual equipment operating scenario, including real-time changing parameters of operating environment, power grid load and commissioning years, construct a multi-constraint equipment health assessment model, and perform feature mapping two-stage training and output logic adaptive correction of the model through sample data. The real-time collected substation status data is input into the corrected health assessment model, and the output is the equipment health level judgment result and the contribution data of each status feature to the health judgment. The contribution data is obtained through the feature importance hierarchical calculation algorithm, which is different from the single-dimensional health assessment and feature importance calculation method. Step 4: Based on the technical specifications for operation and maintenance of substation equipment, construct an operation and maintenance strategy matching system based on fault causal correlation and real-time data-driven approach. The matching coefficient between the real-time status of the equipment and the operation and maintenance requirements is obtained through a dynamic calculation algorithm of the matching coefficient. The equipment health level judgment result, feature contribution data and matching coefficient are collaboratively input into the operation and maintenance strategy matching system to generate a differentiated initial operation and maintenance strategy that accurately corresponds to the real-time status of the equipment and the causes of the fault. The initial operation and maintenance strategy is sequentially verified for logical compliance and on-site operating condition simulation. The strategy that passes the dual verification is determined as an executable operation and maintenance strategy. The executable operation and maintenance strategy is pushed to the corresponding operation and maintenance terminal through a precise matching algorithm. At the same time, the blockchain association storage algorithm is used to perform tamper-proof association storage and full-process traceability of the entire process data from the generation to execution of the operation and maintenance strategy, which solves the technical problems of inaccurate strategy matching and easy tampering of traceability data.

[0007] The technical solutions provided by the embodiments of this disclosure have at least the following beneficial effects: By constructing a feature extraction and dynamic calibration mechanism based on fault causal correlation, combined with a multi-constraint equipment health assessment model, the shortcomings of existing technologies, such as lagging feature calibration, insufficient health assessment accuracy, and inaccurate matching of operation and maintenance strategies, are effectively solved. This enables accurate determination of equipment operating status and early warning of fault risks. It can generate differentiated operation and maintenance strategies that are highly adapted to the real-time status of equipment and fault causes, avoiding over-maintenance and under-maintenance, and significantly improving the scientific nature and accuracy of operation and maintenance decisions.

[0008] By using digital twin verification, precise push notifications from operation and maintenance terminals, and blockchain full-process traceability technology, the problems of insufficient feasibility of operation and maintenance strategies, inaccurate push notifications, and imperfect traceability mechanisms in existing technologies have been solved. This has effectively improved operation and maintenance efficiency, reduced operation and maintenance costs, and achieved full-process traceability from the generation to execution of operation and maintenance strategies as well as full lifecycle traceability of faults, ensuring the standardization and verifiability of the operation and maintenance process. Attached Figure Description

[0009] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0010] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0011] Please see Figure 1 This application provides a method for intelligently generating operation and maintenance strategies for substation equipment, comprising the following steps: Step 1: Collect the condition evaluation results, operating status data and historical stable operation data of the power equipment. Perform hierarchical decomposition of fault causes and classification of fault types on the historical fault records. Establish a unique association relationship between fault types and condition evaluation results, operating status data and historical stable operation data through data association algorithms to form a standardized equipment status dataset with unified fields and unique mapping. Based on the inherent attributes of power equipment and the deep causal relationship between different fault types, a feature extraction mechanism is constructed that separates continuous and discrete data processing and associates fault causal relationships. Through this mechanism, state features that are directly causally related to faults are screened from the standardized equipment state dataset. Combined with the statistical analysis results of historical stable operation data, a static feature baseline that can be dynamically iterated is constructed, rather than a fixed feature baseline. Step 2: Continuously collect and dynamically track the various state features within the static feature baseline. By combining the time series mutation detection algorithm with the fault precursor features, locate the time node and cause of the mutation of the state parameters. Compare the real-time collected parameters with the static feature baseline dimension by dimension. Mark the abnormal data and abnormal level by the multi-threshold anomaly detection algorithm. Combine the distribution density, impact range and mutation cause of the abnormal data to comprehensively determine the current operating status of the equipment. The mutual information gradient integral algorithm is used to quantify the correlation sensitivity between various state features and equipment operating status and the influence of parameter changes. Based on the correlation sensitivity calculation results, features within the static feature baseline are screened, weights are redistributed, and structures are reconstructed to generate a dynamic calibration feature set that can adapt to changes in operating conditions in real time, thus solving the technical defects of feature calibration lag and inability to adapt to real-time operating conditions. Step 3: Use the static feature baseline, dynamic calibration feature set and corresponding equipment operating status as training samples, introduce multi-parameter dynamic constraints of the actual equipment operating scenario, including real-time changing parameters of operating environment, power grid load and commissioning years, construct a multi-constraint equipment health assessment model, and perform feature mapping two-stage training and output logic adaptive correction of the model through sample data. The real-time collected substation status data is input into the corrected health assessment model, and the output is the equipment health level judgment result and the contribution data of each status feature to the health judgment. The contribution data is obtained through the feature importance hierarchical calculation algorithm, which is different from the single-dimensional health assessment and feature importance calculation method. Step 4: Based on the technical specifications for operation and maintenance of substation equipment, construct an operation and maintenance strategy matching system based on fault causal correlation and real-time data-driven approach. The matching coefficient between the real-time status of the equipment and the operation and maintenance requirements is obtained through a dynamic calculation algorithm of the matching coefficient. The equipment health level judgment result, feature contribution data and matching coefficient are collaboratively input into the operation and maintenance strategy matching system to generate a differentiated initial operation and maintenance strategy that accurately corresponds to the real-time status of the equipment and the causes of the fault. The initial operation and maintenance strategy is sequentially verified for logical compliance and on-site operating condition simulation. The strategy that passes the dual verification is determined as an executable operation and maintenance strategy. The executable operation and maintenance strategy is pushed to the corresponding operation and maintenance terminal through a precise matching algorithm. At the same time, the blockchain association storage algorithm is used to perform tamper-proof association storage and full-process traceability of the entire process data from the generation to execution of the operation and maintenance strategy, which solves the technical problems of inaccurate strategy matching and easy tampering of traceability data.

[0012] The data association algorithm adopts an improved Apriori association rule algorithm, which determines the association threshold through dynamic calculation of minimum support and minimum confidence. Minimum support = frequency of occurrence of fault type in historical data / total number of data entries, and minimum confidence = co-occurrence frequency of fault and status data / total frequency of fault occurrence. The mutual information gradient integral algorithm calculates the association sensitivity through double integration of the joint probability density function and the logarithm of mutual information, which is specifically divided into two parts of integration and summation. The multi-constraint health assessment model employs an improved BP neural network, using the ReLU function as the activation function, the cross-entropy loss function as the loss function, and the Adam optimizer as the optimizer. The dynamic calculation algorithm for the adaptation coefficient uses a weighted summation formula, with weights determined by combining the analytic hierarchy process (AHP) with the impact of faults. The logical compliance verification algorithm includes truth table calculation for logical consistency verification, missing rate calculation for data integrity verification, and similarity calculation for historical matching verification. The precise matching algorithm for maintenance terminals uses cosine similarity matching combined with hash mapping association. All key parameters of the algorithms are obtained through training with historical fault data and calibration with on-site measured data. As an optional embodiment, step 1 specifically includes: Operational status data is divided into electrical operation data, equipment body status data, and equipment environment data. The Z-score standardization method, combined with fault correlation characteristics, is used to standardize the fields and dimensions of these three types of data. The calculation process involves taking each original data value in each data type, subtracting the mean of that data type, and then dividing by the standard deviation of that data type to obtain the standardized data value. This eliminates the influence of dimensions while retaining fault correlation characteristics. Based on data change characteristics and the degree of fault correlation, a data type determination algorithm is used to classify the standardized data into continuous and discrete status data, forming standardized status data that can be directly used for feature extraction. This differs from designs that only use dimensional standardization without considering fault correlation characteristics. Historical stable operation data consists of statistical data generated from long-term normal operation of the equipment. It is calculated using a sliding window statistical method combined with fault boundary analysis, including the average value of various state data, normal fluctuation range, and change rate threshold. The normal fluctuation range is calculated using the mean and standard deviation of this type of data. The difference between the mean and twice the standard deviation is taken as the lower limit of the fluctuation range, and the sum of the mean and twice the standard deviation is taken as the upper limit of the fluctuation range. At the same time, the fluctuation range is corrected by combining fault boundary data to ensure that the range can accurately distinguish between normal and abnormal states. The change rate threshold is determined by statistical analysis of the change rate of historical normal operation data combined with fault mutation data to determine the critical value that can distinguish between normal and abnormal changes. This is different from the design that only uses historical normal data to calculate the threshold without combining fault data for correction. This study employs fault tree analysis combined with causal inference algorithms to perform hierarchical decomposition of historical fault records. Fault types are categorized according to three dimensions: fault location, external manifestation, and scope of impact. Each fault type is assigned a unique identifier using a hash algorithm, which includes the fault cause, fault characteristics, fault scope of impact, and fault development pattern. By combining association rule mining algorithms with fault causal relationships, a two-way hierarchical association relationship is constructed, linking fault types to identifier information, identifier information to standardized status data, standardized status data to historical stable operation data, and historical stable operation data to status evaluation results. This forms a unified equipment status dataset with unique mapping relationships, addressing the shortcomings of weak correlation between faults and data and the simplistic nature of such associations.

[0013] The Z-score standardization formula is Z=(X-μ) / σ, where X is the original data, μ is the mean of the data, and σ is the standard deviation of the data. The calculation process uses a fault correlation coefficient to correct the result. The fault correlation coefficient is the absolute value of the correlation coefficient between the data and historical faults. The data type determination algorithm uses the K-nearest neighbor classification algorithm, calculating the probability of continuous change of data points P=the amount of continuously changing data / the total amount of data. When P≥0.8, the data is considered continuous; when P<0.8, it is considered discrete. The hash algorithm uses the SHA-256 algorithm, where the hash value of the identifier information = SHA-256(fault type code + fault cause code + timestamp). The association strength of the association rule mining algorithm is calculated using three dimensions: confidence, support, and lift. Lift = confidence / (probability of fault type occurring alone × probability of state data occurring alone). Only associations with a lift ≥1.2 are retained. As an optional embodiment, the feature extraction mechanism in step 1 is implemented as follows: For continuous state data, the time series is first arranged and fault-related nodes are labeled. Then, the mean, variance, kurtosis, skewness and rate of change of the series data are calculated by combining the characteristics of the fault nodes with the sliding window algorithm. The mean is calculated by dividing the sum of all data in the series by the number of data. The variance is calculated by dividing the sum of the squares of the differences between each data in the series and the mean by the number of data minus one. Kurtosis is calculated as the sum of the fourth power of the differences between each data point in the sequence and the mean, divided by the product of the number of data points and the fourth power of the standard deviation; skewness is calculated as the sum of the cube power of the differences between each data point in the sequence and the mean, divided by the product of the number of data points and the cube power of the standard deviation. The rate of change is calculated as the difference between the parameters of two adjacent acquisition nodes, divided by the time interval between the two adjacent acquisition nodes. The above calculation results are compared with the corresponding indicators in the historical stable operation data item by item. The correlation coefficient between each indicator and the fault type is calculated by combining the correlation analysis algorithm with the fault causal relationship. The indicators whose absolute values ​​of the correlation coefficients meet the preset correlation conditions and exceed the normal fluctuation range are extracted as continuous key state features. The preset correlation conditions are determined by combining the correlation analysis of historical fault data and state indicators with the fault causal relationship. This is different from the design that only selects features by correlation without combining the fault causal relationship. For discrete state data, a frequency statistics algorithm is used in conjunction with the co-occurrence pattern of faults to calculate the frequency and proportion of occurrence, as well as the co-occurrence frequency with various types of faults. A correlation degree calculation algorithm is used in conjunction with the causal relationship of faults to calculate the correlation degree between discrete data and fault types. Indicators whose correlation degree meets the correlation threshold are extracted as discrete key state features. The correlation threshold is determined by combining the historical co-occurrence data of discrete data and fault types with the causal relationship of faults, thus solving the problem of weak correlation between discrete feature extraction and faults. By fusing continuous critical state features with discrete critical state features, and combining the standard range of historical stable operation data with fault boundary data, a static feature baseline that can be dynamically iterated is constructed by combining fault causal relationships through feature fusion algorithms. The inherent attributes of power equipment include equipment structure type, rated operating parameters, service life and environmental adaptability parameters. The causal relationship between these attributes and fault types is determined by combining causal inference algorithms with historical fault data, which is different from designs that determine causal relationships solely through experience.

[0014] The window size for the sliding window algorithm is determined through fault cycle statistics. The window size is calculated as: window size = average fault duration / data acquisition interval, expressed as W = round(T_avg / Δt), where T_avg is the historical average fault duration and Δt is the data acquisition interval. The correlation analysis algorithm uses the Pearson correlation coefficient, calculated as r = Σ[(X_i-μ_X)(Y_i-μ_Y)] / √[Σ(X_i-μ_X)²Σ(Y_i-μ_Y)²], where X represents the status index data, Y... The fault occurrence is identified (occurrence = 1, non-occurrence = 0); the correlation is calculated using the mutual information method, mutual information I(X,Y) = ΣΣP(X,Y)log[P(X,Y) / (P(X)P(Y))], and the correlation threshold = 0.3 + 0.7 × (fault occurrence probability); the feature fusion algorithm uses weighted average fusion, weight = correlation coefficient between feature and fault / sum of correlation coefficients of all features, static feature baseline parameter = Σ (feature value × weight), and baseline fluctuation range = ±2 × Σ (feature standard deviation × weight); As an optional embodiment, step 2 specifically includes: Key state features within the static feature baseline are arranged in a time series and associated with fault precursor information. Real-time feature parameters are collected at fixed time intervals. The parameter change rate and cumulative change between adjacent acquisition nodes are calculated by combining the differential algorithm with the fault mutation law. The change rate is calculated as the parameter difference between two adjacent acquisition nodes, divided by the time interval between the two adjacent acquisition nodes. The cumulative change is calculated as the difference between the parameter value of each acquisition node and the baseline parameter value. All differences are summed sequentially. At the same time, the real-time parameters are compared with the static feature baseline parameters point by point, the absolute difference between the two is calculated and associated with the fault risk level, which is different from the design that only calculates the parameter difference without associating it with the fault risk. The 3σ criterion combined with the fault anomaly level classification is used as the basis for parameter anomaly judgment. When the real-time parameter change rate exceeds the baseline change rate threshold, the cumulative change exceeds the normal fluctuation range, or the absolute difference is greater than three times the standard deviation, it is marked as an abnormal parameter and the anomaly level is classified. By combining the analytic hierarchy process (AHP) with fault impact weights, the influence weights of various key state characteristics on equipment operation are calculated. The weight calculation is completed sequentially through the construction of a judgment matrix, consistency check, and weight normalization. Combining the type, quantity, distribution density, and anomaly level of abnormal parameters, the current operating status of the equipment is comprehensively determined by the fuzzy comprehensive evaluation method. The equipment operating status is divided into three types: normal operation, anomaly warning, and fault pending investigation. The evaluation matrix is ​​constructed by the influence weights, distribution density, and anomaly level of abnormal parameters, which solves the defects of anomaly judgment without level classification and inaccurate status judgment.

[0015] The time series mutation detection algorithm adopts the Cumulative Sum Control Chart (CUSUM) algorithm, with an upper control limit UCL = μ + 3σ, a lower control limit LCL = μ - 3σ, and a cumulative sum S_i = max(0, S_{i-1} + (X_i - μ - δ)), where δ is the minimum detectable offset, δ = 0.5σ. The multi-threshold anomaly detection algorithm has three thresholds: Level 1 threshold = baseline value ± σ (anomaly warning), Level 2 threshold = baseline value ± 2σ (minor fault), and Level 3 threshold = baseline value ± 3σ (serious fault). The thresholds are weighted and corrected by the degree of fault loss, with the correction coefficient = fault economic loss / average economic loss of all faults. The fault risk level is quantified as R = α × anomaly level + β × distribution density + γ × impact range, where α, β, and γ are weight coefficients, determined by the analytic hierarchy process as α = 0.5, β = 0.3, and γ = 0.2. Where S_{i-1} represents the cumulative sum calculated for the (i-1)th data point (the previous time step); As an optional embodiment, step 2 uses the mutual information gradient integral algorithm combined with the fault parameter change pattern to calculate the influence of parameter changes on the sensitivity of the correlation between key state characteristics and equipment operating status. The calculation process is divided into two parts: The first part is the joint probability density function after taking the partial derivative with respect to the parameter changes, and the product of the logarithmic term of the mutual information is double-integrated; The second part is the joint probability density function, which is double-integrated with the result of the partial derivative of the mutual information logarithm term with respect to the parameter change. The two integral results are added together to obtain the final calculated value. This calculation process combines the change law of fault parameters to ensure that the correlation sensitivity calculation is more in line with the actual fault scenario of the equipment, which is different from the design of simple mutual information calculation without combining the fault scenario. Mutual information values ​​represent the degree of dependency between key state features in the static feature baseline and the current operating state of the device; The partial derivative of the mutual information value with respect to the change in parameters indicates the degree to which the dependency is affected by the change in parameters, i.e., the degree of correlation sensitivity. The key state feature is any one of the static feature baselines, and the difference between the feature and the baseline value is the difference between the real-time collected value of the feature and the baseline value, and is associated with the fault risk coefficient; The quantitative representation value of the current operating status of the equipment is obtained through the operating status classification standard, which is determined based on the substation equipment operation and maintenance technical specifications and the degree of fault impact. The marginal probability density function is calculated by combining the kernel density estimation method with the distribution characteristics of fault data. The marginal probability density function of the current equipment state is calculated by combining the frequency statistics method with the frequency of fault occurrence. The joint probability density function is calculated by combining the two-dimensional kernel density estimation method with the fault correlation characteristics. This is different from the design that does not combine fault characteristics in the probability density function calculation. The calculated correlation sensitivity is sorted from high to low. A threshold screening algorithm combined with fault impact weights is used to retain features whose correlation sensitivity meets the screening criteria. The screening criteria are determined by training with historical data and combining fault causal relationships. Invalid features whose correlation sensitivity does not meet the screening criteria are removed. For the retained features, a normalization algorithm combined with fault impact weights is used to redistribute the impact weights. The calculation process is to take the correlation sensitivity value of each retained feature and divide it by the sum of the correlation sensitivity values ​​of all retained features to obtain the weight of each retained feature. This completes the dynamic reconstruction of the static feature baseline, forming a real-time dynamic calibration feature set, and solves the technical defect of feature weight allocation not taking into account fault impacts.

[0016] The specific calculation formula for the mutual information gradient integral algorithm is as follows: S=∫∫[ f(X,Y) / ΔX×log(P(X,Y) / (P(X)P(Y)))]dXdY+∫∫[f(X,Y)× log(P(X,Y) / (P(X)P(Y))) / ΔX]dXdY; Where f(X,Y) is the joint probability density function of X (state feature) and Y (operational state), ΔX is the parameter change, and P(X) and P(Y) are the marginal probability density functions of X and Y, respectively; the joint probability density function is calculated by kernel density estimation, f(X,Y)=1 / (nh²)ΣK((X-X_i) / h)K((Y-Y_i) / h), where K is the Gaussian kernel function, h is the bandwidth, and h=1.06×min(σ,IQR / 1.34)×n^(-1 / 5); the normalization algorithm adopts min-max normalization, the weight w_i=S_i / ΣS_i, where S_i is the calculated value of the correlation sensitivity of the i-th feature, and the corrected weight w'_i=w_i×ω_i, where ω_i is the fault impact weight, which is determined by fault tree analysis; As an optional embodiment, step 3 specifically includes: Using static feature baselines and real-time dynamic calibration feature sets as the model input dimensions, and equipment operating status combined with fault risk levels as the model output dimensions, a feature mapping matrix between input and output is established through matrix factorization algorithm combined with fault feature mapping relationship. Multi-parameter dynamic constraints of actual equipment operating scenarios are introduced, including real-time changing parameters of operating environment, power grid load, and years of operation. The constraints are integrated into the feature mapping matrix after parameter normalization to build the basic framework of equipment health assessment model. The model adopts a BP neural network structure, with the number of input layer nodes consistent with the feature dimensions, the number of hidden layer nodes determined by trial and error combined with model convergence speed, and the number of output layer nodes consistent with the number of equipment operating status types and fault risk levels. This is different from the design of models without scenario dynamic constraints and without associated fault risk levels. By combining the K-means clustering algorithm with the distribution characteristics of fault types, the static feature baseline, real-time dynamic calibration feature set, and corresponding operating status data are clustered. The clustering results are adaptively divided into training and testing data to ensure uniform distribution of various fault types and operating status data in both sets. Simultaneously, the sample proportion of each fault type is guaranteed to be consistent with the actual fault occurrence proportion. The division ratio is determined by cluster purity calculation combined with fault sample distribution. The training data is input into the model's basic framework. Based on the feature mapping matrix, gradient descent combined with the convergence law of fault features is used to iteratively adjust the feature matching weights. The iteration process involves taking the current weight value and subtracting the product of the learning rate and the partial derivative of the loss function with respect to the current weight value to obtain the weight value for the next iteration. After each iteration, the consistency between the model's judgment result and the actual state is calculated. The consistency is calculated by dividing the number of correct judgments by the total number of judgments. The weights are then adjusted in reverse based on the consistency until the consistency meets the training criteria. These criteria are determined by the model training requirements combined with the fault judgment accuracy requirements. This completes the model feature mapping training, distinguishing it from designs that do not incorporate fault distribution during clustering or fail to consider fault features during iterative training. The test data is input into the trained model, and the model output is compared with the actual state and fault risk level. Data with judgment errors are extracted and the causes of errors are analyzed, including feature weight deviation and fault feature mapping deviation. The corresponding feature weights are adjusted by combining the error backpropagation algorithm with fault feature correction logic, and the model output logic is optimized. The test and correction are repeated until the judgment accuracy meets the model calibration requirements. The model calibration requirements are determined based on the substation equipment operation and maintenance technical specifications and the fault judgment accuracy requirements, and the overall model calibration is completed. The feature contribution data is calculated by combining the fault impact weight with the Gini coefficient method. The calculation process is to take 1 and subtract the sum of squares of the probabilities of each feature corresponding to each type of operating state. The Gini coefficient value is positively correlated with the feature contribution and the fault impact weight, which is different from the design of feature contribution calculation without considering the fault impact.

[0017] The feature mapping matrix is ​​constructed using a matrix factorization algorithm, M = U × V^T, where U is the input feature matrix and V is the output label matrix. During the factorization process, a dynamic constraint term C = λ × (E × W_E + L × W_L + Y × W_Y) is added, where E is the operating environment parameter matrix, L is the power grid load parameter matrix, Y is the service life parameter matrix, W_E, W_L, and W_Y are constraint weight matrices, and λ is the constraint coefficient, λ = 0.1 × (1 - model training accuracy). The number of nodes in the model input layer = the number of static features + the number of dynamic calibration features; the number of nodes in the hidden layer = 2 × the number of input layer nodes + 1, corrected by trial and error combined with model convergence speed; the number of nodes in the output layer = the number of operating state types × the number of fault risk levels. The feature mapping undergoes a two-stage training process: the first stage is forward propagation, with the output value Y = σ(W × X + B); the second stage is backward propagation, with weight updates W = W - η × Loss / W, bias update B = B - η × Loss / B, where η is the learning rate, η = 0.001 × (1 + model iterations / 1000). As an optional embodiment, the division between training data and test data is determined by the principle of maximizing cluster purity combined with the requirement of balanced distribution of fault samples. Cluster purity is calculated by taking the reciprocal of the total amount of data and multiplying it by the sum of the data of the most numerous category in each cluster. When the cluster purity reaches the maximum value and the distribution of fault samples is balanced, the division ratio between training data and test data is determined, which is different from the design that only pursues cluster purity without considering the distribution of fault samples. The training data is split into batches. The amount of data in each batch is determined by a comprehensive calculation of the model memory capacity, convergence speed, and fault sample balance. After each batch of data is input into the model, the weight adjustment range is calculated by combining the feature mapping matrix with the fault feature weight rules. The adjustment range is calculated by combining the step size of the gradient descent method with the fault feature sensitivity. The model parameters are adjusted according to the adjustment range until the accuracy of the fault judgment in that batch of data meets the standard and the fault judgment accuracy meets the requirements before the next batch of data is input. For data with incorrect judgments, a correlation analysis algorithm is used to combine the causal relationship of the fault, compare the feature weights with the actual correlation, and calculate the deviation between the feature weights and the actual correlation. The deviation is calculated as the difference between the absolute values ​​of the feature weights in the model and the actual correlation. When the difference meets the weight adjustment conditions, the weights of features with high correlation but low weights are increased, and the weights of features with low correlation but high weights are decreased. The weight adjustment conditions are determined by the deviation analysis between the feature weights and the actual correlation and the fault judgment accuracy requirements. The weight adjustment range is proportional to the deviation and the weight of the fault impact. The model is retested until the judgment accuracy meets the operation and maintenance technical requirements and the fault judgment accuracy requirements.

[0018] The clustering algorithm used is K-means, with the number of clusters K = √(total number of samples / 2), and the cluster purity P = (1 / N) × Σmax(C_ik), where N is the total data volume and C_ik is the number of samples in the k-th class of the i-th cluster. The distribution balance of fault samples is determined by variance testing, with the balance index V = σ² (the proportion of fault samples in each class). A balance is considered achieved when V ≤ 0.01. The batch size B = min(round(M / 10), round(memory capacity / (bytes per sample × feature dimension))), where M is the total amount of training data. The weight adjustment magnitude ΔW = η × ( Loss / W)×S_f, where S_f is the fault feature sensitivity, S_f = the correlation coefficient between the feature and the fault / the maximum correlation coefficient of all features; judgment fit Acc = number of correctly judged samples / total number of samples, with a threshold of Acc ≥ 0.92; fault judgment accuracy Prec = number of correctly judged faults / total number of fault judgments, with a threshold of Prec ≥ 0.95. As an optional embodiment, the specific implementation of the operation and maintenance strategy matching system in step 4 is as follows: By combining a feature contribution ranking algorithm with fault impact weights, key state features are ranked, and core features that meet the core feature selection criteria are extracted. The core feature selection criteria are determined by feature contribution analysis combined with fault causal relationships. A feature-based operation and maintenance strategy association model based on fault causal relationships is constructed. The model inputs are core feature parameter values, health level judgment results, adaptation coefficients, and fault risk levels. The output is an operation and maintenance strategy that accurately corresponds to the fault cause and health status. The adaptation coefficient is calculated by combining a weighted summation algorithm with the degree of fault impact. The calculation process is to take the product of the weight of each core feature and the standardized parameter value of that feature, sum all the products, and then add the product of the fault impact range weight and the standardized value of the fault impact range to obtain the adaptation coefficient. The fault impact range weight is determined by the quantitative calculation of the degree of fault impact combined with the operation and maintenance priority, which is different from the design of the adaptation coefficient calculation that does not combine fault impact and operation and maintenance priority. The core feature parameter values, health level judgment results, adaptation coefficients, and fault risk levels are collaboratively input into the feature operation and maintenance strategy association model. Using a cosine similarity algorithm combined with fault cause matching logic, the similarity between the input data and the preset operation and maintenance strategy template is calculated. The calculation process involves taking the sum of the products of corresponding items in the input data and template data, dividing it by the product of the square root of the sum of squares of each item in the input data and the square root of the sum of squares of each item in the template data. Templates that meet the matching criteria for similarity with the input data and match the fault cause are selected. Combined with real-time equipment status parameters and fault causes to supplement operational details, an initial differentiated operation and maintenance strategy corresponding to the real-time equipment status and fault cause is generated. This process achieves accurate matching through data similarity, fault cause matching, and quantitative calculation. It has no hierarchical matching rules, unlike strategy generation methods that rely on manually defined hierarchical rules and do not incorporate fault cause matching.

[0019] Feature contribution is calculated using the Gini coefficient method: Gini = 1 - Σ(P_i)², where P_i is the probability of the feature corresponding to each operating state. The dynamic calculation formula for the adaptation coefficient is K = Σ(w_i × Z_i) + w_r × Z_r, where w_i is the core feature weight, Z_i is the standardized value of the core feature, w_r is the fault impact range weight, Z_r is the standardized value of the fault impact range, w_r = number of devices affected by the fault / total number of devices, Z_r = fault impact duration / average fault handling time. The cosine similarity calculation formula is Sim = (A B) / (|A|×|B|), where A is the input data vector, B is the strategy template vector, the matching standard Sim≥0.85, when multiple templates satisfy the condition, the template corresponding to the maximum value of Sim is selected and combined with the fault cause correction, the correction coefficient = fault cause matching degree ×0.3 + similarity ×0.7; As an optional embodiment, the specific implementation method of the logical compliance verification in step 4 is as follows: A logical consistency algorithm is used in conjunction with fault execution logic to verify the logical relationship between the execution conditions of the operation and maintenance strategy. By constructing logical expressions and calculating the truth table of the expressions, if there is a truth value contradiction, it is determined to be a logical conflict. The conflicting conditions are eliminated by combining the fault handling logic and the corresponding execution steps are regenerated to ensure that the execution steps are consistent with the fault handling logic. A data integrity verification algorithm is adopted in combination with fault data requirements. A dataset is generated based on key status features and health judgment results through comparison strategy. The missing rate is calculated by dividing the number of missing data items by the total number of data items. When the missing rate is greater than 0, the missing data is supplemented and the correlation between the data and the fault is verified. Then, the data is re-entered into the matching system for matching to ensure that the supplemented data can support accurate fault matching. A historical matching algorithm is used in conjunction with fault handling patterns to compare the similarity between the strategy execution requirements and the historical operating status and historical fault handling methods. When the similarity does not meet the historical matching criteria, the execution requirements are adjusted by combining the causal analysis algorithm with the fault handling patterns to ensure consistency with historical data and continuity of fault handling. The historical matching criteria are determined by combining historical operation and maintenance data analysis with fault handling effects. Actual operating condition simulation verification is achieved through digital twin technology combined with fault scenario reconstruction. A 1:1 digital twin model is constructed according to the actual operating environment, operating conditions, and typical fault scenarios of the equipment. The initial operation and maintenance strategy that has passed logical verification is transformed into executable instructions in the model. The complete execution process under different fault scenarios is simulated in the digital twin model. Data on the completion status, time consumption, resource consumption, and equipment status changes during the execution process are collected through sensors. When the step completion rate reaches 100%, the time consumption does not exceed the time consumption threshold, the resource consumption does not exceed the resource consumption threshold, the standardized values ​​of the equipment status parameters change towards normal operation, and the fault is effectively controlled, the simulation verification is deemed to have passed and become an executable operation and maintenance strategy. The time consumption threshold is obtained by statistical calculation of historical operation and maintenance data combined with the fault handling timeliness requirements, and the resource consumption threshold is obtained by calculation of operation and maintenance costs combined with fault handling priority. This is different from the design of simulation verification that does not recreate fault scenarios and does not consider the fault control effect.

[0020] The logical expression is constructed as L=∧(C_i→A_i), where C_i is the execution condition and A_i is the execution action. The truth table is calculated by traversing all combinations of conditions (true / false). When the same combination of conditions corresponds to different action results, it is determined to be a conflict. The missing rate M=number of missing data items / total number of data items. The supplementary data is calculated using the KNN interpolation method, and the supplementary value X=Σ(X_j×Sim_ij) / ΣSim_ij, where X_j is the data corresponding to the neighboring sample and Sim_ij is the similarity. The historical matching algorithm uses Jaccard similarity calculation, J=|S∩T| / |S∪T|, where S is the current policy execution requirement set and T is the historical policy execution requirement set, and the matching standard J≥0.75. The causal analysis algorithm uses a Bayesian network, and the conditional probability P(A|B)=P(B|A)×P(A) / P(B). The execution requirements are adjusted so that P(fault resolution|B)≥0.9. As an optional embodiment, the specific implementation method of pushing from the operation and maintenance terminal in step 4 is as follows: A parameter association model is constructed between operation and maintenance terminals and operation and maintenance responsibility areas, operation permissions, and fault handling qualifications. The model inputs are the hardware parameters, permission codes, equipment responsibility area codes, and fault handling qualification information of operation and maintenance personnel of the operation and maintenance terminals. By combining the hash mapping algorithm with the operation and maintenance responsibility division rules, the association relationship between terminals and responsibility areas, permissions, and fault handling qualifications is established. The equipment responsibility area codes, operation permission codes, and required fault handling qualifications corresponding to the executable operation and maintenance strategies are extracted and input into the parameter association model. The matching operation and maintenance terminals are filtered out by a similarity matching algorithm. The filtering conditions are that the responsibility area codes are consistent, the permission codes contain the operation permissions required by the strategy, and the operation and maintenance personnel have the corresponding fault handling qualifications. The executable operation and maintenance strategies are pushed to the filtered terminals through a message push protocol, which is different from the design that does not consider fault handling qualifications and has inaccurate push. The entire process of operation and maintenance strategy traceability is achieved through blockchain storage technology combined with fault lifecycle data. A blockchain storage node is constructed, and the strategy generation time, generation basis, health assessment results, feature contribution data, adaptation coefficient, push information, execution information, execution results, and fault lifecycle data are hashed and associated using the device's unique identifier. The association process involves combining the device's unique identifier with data from each stage and the fault lifecycle data, calculating a hash value using a hash function, and storing the associated data on the blockchain node. Each node stores complete data for the entire process and fault lifecycle data. Data from the corresponding blockchain node can be retrieved using the device's unique identifier or fault identifier, enabling seamless traceability of the entire operation and maintenance strategy process and fault lifecycle data. This ensures data immutability and full traceability, addressing the deficiency of only tracing the strategy process without associating it with fault lifecycle data.

[0021] The hash mapping algorithm adopts the consistent hashing algorithm. The terminal hash value = Hash(terminal ID) mod 2^32, and the responsibility area hash interval = Hash(responsibility area code) ± Δ, where Δ = 2^28 / total number of responsibility areas. The similarity calculation of the parameter association model adopts the weighted Euclidean distance, Dist = √[w1×(Z1-Z2)²+w2×(Z3-Z4)²+w3×(Z5-Z6)²], where Z1-Z6 are the standardized values ​​of responsibility area, authority, and qualification of terminal and policy, w1=0.4, w2=0.3, w3=0.3, and the screening threshold Dist≤0.2. The blockchain association storage algorithm adopts the Merkle tree structure. The root hash = Hash(Hash(policy data) + Hash(execution data) + Hash(fault data)). Each block contains the hash value of the previous block. Data association is realized through the unique device identifier. The association key = Hash(device ID + timestamp) to ensure that the data is tamper-proof and traceable throughout the entire process. Working principle First, three types of core data from the substation equipment are collected comprehensively: 1) Equipment status evaluation results, reflecting the overall health level of the equipment; 2) Real-time operating status data, which are divided into electrical operating data (such as winding temperature, oil level, etc.), equipment body status data (such as core grounding current, bushing dielectric loss, etc.), and equipment environmental data (such as ambient temperature, humidity, etc.), covering all dimensions of equipment operation; 3) Historical stable operating data and historical fault records, providing historical references for characteristic baseline construction and fault correlation analysis.

[0022] Secondly, the collected operational status data is standardized: the Z-score standardization method is adopted, and combined with the fault correlation characteristics, the fields of different types and dimensions are uniformly named and the dimensions are eliminated to ensure the comparability of the data; at the same time, according to the data change characteristics and the degree of fault correlation, the standardized data is divided into two categories: continuous (such as continuously changing data such as temperature and current) and discrete (such as discrete data such as equipment operation mode and fault type), laying the foundation for subsequent targeted feature extraction.

[0023] Finally, historical fault records are processed hierarchically: fault tree analysis combined with causal inference algorithms is used to decompose and classify historical faults by cause, assigning a unique identifier to each type of fault (including information such as fault causes, characteristics, and scope of impact). Furthermore, through association rule mining algorithms, a bidirectional hierarchical association relationship is constructed between fault types and standardized operational data, historical stable data, and status evaluation results, forming a standardized equipment status dataset with unified fields and unique mappings, achieving deep binding between data and faults. This step, through standardization and fault association mining, addresses the shortcomings of the background technology, such as the simplistic processing of discrete and continuous data and the weak correlation between faults and data. It provides data support for subsequent efforts to strengthen the correlation between faults and data and improve the effectiveness of feature extraction, echoing the core objective of precise operation and maintenance in the beneficial effects.

[0024] Static feature baselines are the core benchmarks for determining whether the operating status of equipment is abnormal. The core of their construction is to combine the causal relationship of faults to select key features that can accurately reflect the operating status of equipment and are directly related to faults, while ensuring the rationality and iterability of the baseline, thus solving the problem of insufficient effectiveness of feature selection in existing technologies.

[0025] Based on the inherent attributes of power equipment (such as equipment structure, rated parameters, and years of operation) and the deep causal relationship between different fault types, a targeted feature extraction mechanism is constructed to screen features for both continuous and discrete standardized data: For continuous data, statistical characteristics (mean, variance, etc.) of the data are calculated by time series arrangement, fault node labeling, and sliding window algorithm, and then features directly related to faults and with high correlation are selected as key continuous features by correlation analysis combined with fault causal relationships; For discrete data, the correlation between the data and various types of faults is calculated by frequency statistics and fault co-occurrence pattern analysis, and features with high correlation are selected as key discrete features.

[0026] The selected continuous and discrete key features are fused and integrated, and combined with the standard range of historical stable operation data and fault boundary data, a static feature baseline that can be dynamically iterated is constructed through feature fusion algorithms. This baseline can reflect both the parameter range of normal equipment operation and the correlation with fault features, providing a benchmark for subsequent real-time status determination and dynamic calibration. This step, through fault correlation-oriented feature screening, solves the shortcomings of the previous technology, such as the weak correlation between feature extraction and faults and insufficient screening effectiveness. It supports the goal of improving the accuracy of operation and maintenance strategy matching, and lays the foundation for subsequent dynamic calibration and status determination.

[0027] The core of this step is to solve the problem that static baselines cannot adapt to real-time changes in equipment operating conditions. By dynamically calibrating features and accurately determining the equipment operating status, it provides real-time basis for subsequent health assessments and strategy generation, while also enabling early warning of fault risks. This addresses the pain points of lagging feature calibration and inaccurate status determination in existing technologies.

[0028] First, key features within the static feature baseline are collected and dynamically tracked in real time: real-time feature parameters of the equipment are collected at fixed time intervals, combined with fault precursor features, and the rate of change and cumulative change of parameters are calculated through a differential algorithm. At the same time, the real-time parameters are compared point by point with the static feature baseline, the absolute difference between the two is calculated and associated with the fault risk level, so as to realize the real-time capture of changes in the operating status of the equipment.

[0029] Secondly, the 3σ criterion, combined with fault anomaly classification, is used to determine anomalies in real-time parameters: when the parameter change rate or cumulative change exceeds the baseline range, or the absolute difference is greater than three times the standard deviation, it is marked as an abnormal parameter and classified into different anomaly levels according to severity, providing a basis for status determination; at the same time, the influence weight of each key feature on equipment operation is calculated by the analytic hierarchy process, and combined with the type, quantity, distribution density and anomaly level of abnormal parameters, the equipment operating status is divided into three categories, namely normal operation, anomaly warning and fault pending investigation, by the fuzzy comprehensive evaluation method, so as to achieve accurate status determination.

[0030] Finally, the static feature baseline is dynamically calibrated: a mutual information gradient integral algorithm is used to quantify the correlation sensitivity between each key feature and the equipment operating status, and to analyze the impact of parameter changes on the correlation between features and status. Based on the correlation sensitivity, effective features are selected, invalid features are eliminated, feature weights are reallocated, and the static feature baseline is dynamically reconstructed to generate a real-time dynamic calibration feature set. This ensures that the features can accurately adapt to real-time changes in equipment operating conditions, improving the accuracy of subsequent health assessments and strategy generation. This step, through dynamic calibration and accurate status determination, solves the shortcomings of the background technology, such as lagging feature calibration and inability to adapt to real-time operating conditions. It supports the beneficial effects of dynamically adapting to changes in operating conditions and providing early warning of fault risks, thus providing accurate real-time data support for equipment health assessment.

[0031] The core of this step is to accurately assess the health level of equipment by constructing a multi-constraint health assessment model and combining it with dynamic calibration features. At the same time, it outputs the contribution of each feature to the health determination, providing core support for the differentiated generation of operation and maintenance strategies, and solving the problems of insufficient accuracy and lack of scenario adaptability in existing health assessment models.

[0032] Using static feature baselines and dynamic calibration feature sets as model input dimensions, and equipment operating status combined with fault risk level as model output dimensions, a basic framework of BP neural network is built. Multi-parameter dynamic constraints (such as ambient temperature, power grid load, and years of operation) of actual equipment operating scenarios are introduced, and the constraints are integrated into the feature mapping matrix to solve the problem of lack of scenario adaptability of existing models and improve the model evaluation accuracy.

[0033] The K-means clustering algorithm is used, combined with the distribution characteristics of fault types, to cluster static features, dynamic calibration features, and corresponding operating status data. This clustering is adaptively divided into training and testing data to ensure uniform distribution and that the proportion of fault samples matches the actual situation. The model undergoes a two-stage training and adaptive correction using gradient descent: the first stage involves feature mapping training, iteratively adjusting feature matching weights until the model's judgment accuracy meets the requirements; the second stage verifies the model using test data, analyzes the causes of judgment errors, and adjusts feature weights and output logic until the model's judgment accuracy meets the requirements, thus completing model calibration.

[0034] After model calibration, inputting real-time equipment status data will output the equipment health level assessment result. Simultaneously, the contribution of each key feature to the health assessment is calculated using the Gini coefficient method, clarifying the core factors affecting equipment health and providing a basis for subsequent targeted generation of operation and maintenance strategies. This step, through multi-constraint model construction and two-stage training, addresses the shortcomings of previous health assessment models, such as lack of scenario constraints and insufficient accuracy, supporting the core objectives of accurately determining equipment status and improving the matching degree of operation and maintenance strategies.

[0035] The core of this step is to generate accurate and feasible differentiated operation and maintenance strategies based on equipment health level, feature contribution, and fault causes. Through verification and precise push, the strategies are ensured to be effectively implemented, improving operation and maintenance efficiency and solving the problems of inaccurate matching, push deviation, and insufficient feasibility of operation and maintenance strategies in existing technologies.

[0036] First, we construct an operation and maintenance strategy matching system based on fault causal correlation and real-time data-driven approach: We extract core features according to feature contribution ranking, construct a feature-operation and maintenance strategy correlation model, and input the core feature parameter values, health level judgment results, adaptation coefficient (calculated by combining the degree of fault impact and operation and maintenance priority) and fault risk level into the model. The output is an initial differentiated operation and maintenance strategy that accurately corresponds to the real-time status of the equipment and the fault causes, ensuring the relevance of the strategy.

[0037] Secondly, the initial operation and maintenance strategy undergoes dual verification to ensure feasibility and compliance: First, logical compliance verification is performed by using a logical consistency algorithm to verify the logic of the strategy execution conditions, identify logical conflicts, supplement missing data, and adjust the strategy execution requirements based on historical fault handling patterns; Second, on-site operating condition simulation verification is performed by constructing a 1:1 twin model of the equipment using digital twin technology to simulate the strategy execution process, collecting data on execution steps, time consumption, resource consumption, and equipment status changes. Only when the step completion rate, time consumption, and resource consumption all meet the requirements, and the equipment status changes in a normal direction, is the strategy deemed to have passed the verification and become an executable operation and maintenance strategy.

[0038] Finally, precise delivery of operation and maintenance strategies is achieved: a correlation model is constructed between operation and maintenance terminals and their respective responsibility areas, operating permissions, and fault handling qualifications. The responsibility areas, permissions, and qualification requirements corresponding to executable strategies are extracted. A similarity matching algorithm is used to filter matching operation and maintenance terminals, and the strategies are then pushed to the corresponding operation and maintenance personnel, avoiding push errors and improving operation and maintenance efficiency. This step, through precise matching, double verification, and precise delivery, solves the shortcomings of the background technology, such as inaccurate matching and delivery of operation and maintenance strategies, and insufficient feasibility. It directly supports the beneficial effects of improving operation and maintenance efficiency, reducing operation and maintenance costs, and achieving differentiated operation and maintenance.

[0039] This step is the closed-loop guarantee of the entire solution. Its core is to use blockchain storage technology to achieve seamless traceability of the entire process of operation and maintenance strategy from generation to execution. At the same time, it links the entire life cycle data of the fault to ensure that the operation and maintenance process is verifiable and the data is tamper-proof, thus solving the problems of imperfect traceability mechanism and easy data tampering in the existing technology.

[0040] A blockchain storage node is constructed to hash and associate the operation and maintenance strategy generation time, generation basis (such as feature data, health assessment results, etc.), adaptation coefficient, push information, execution information, execution results, and fault lifecycle data (such as fault causes, early warning information, handling process, etc.) through the unique device identifier, generating immutable associated data. Each blockchain node stores complete full-process data, and relevant data can be retrieved through the unique device identifier or fault identifier, realizing full-process traceability of operation and maintenance strategy generation, verification, push, and execution, as well as full lifecycle traceability of faults from early warning to handling, ensuring the standardization and verifiability of the operation and maintenance process. This step, through blockchain traceability technology, solves the defects of imperfect traceability mechanism and easy data tampering in the background technology, supporting the goal of achieving full-process traceability and ensuring the standardization of operation and maintenance, forming a closed-loop guarantee for the entire operation and maintenance system.

[0041] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligently generating an operation and inspection strategy for a power transformation device, characterized in that, Includes the following steps: Step 1: Collect the condition evaluation results, operating status data and historical stable operation data of the power equipment. Perform hierarchical decomposition of fault causes and classification of fault types on the historical fault records. Establish a unique association relationship between fault types and condition evaluation results, operating status data and historical stable operation data through data association algorithms to form a standardized equipment condition dataset. Based on the inherent attributes of power equipment and the deep causal relationship between different fault types, a feature extraction mechanism is constructed that separates continuous and discrete data processing and associates fault causal relationships. Step 2: Continuously collect and dynamically track the various state features within the static feature baseline. By combining the time series mutation detection algorithm with the fault precursor features, locate the time node and cause of the mutation of the state parameters. Compare the real-time collected parameters with the static feature baseline dimension by dimension. Mark the abnormal data and abnormal level by the multi-threshold anomaly detection algorithm. Combine the distribution density, impact range and mutation cause of the abnormal data to comprehensively determine the current operating status of the equipment. The mutual information gradient integral algorithm is used to quantify the correlation sensitivity between various state features and equipment operating status and the influence of parameter changes. Based on the correlation sensitivity calculation results, features within the static feature baseline are screened, weights are redistributed, and structures are reconstructed to generate a dynamic calibration feature set that can adapt to changes in operating conditions in real time. Step 3: Use the static feature baseline, dynamic calibration feature set and corresponding equipment operating status as training samples, introduce multi-parameter dynamic constraints of the actual equipment operating scenario, including real-time changing parameters of operating environment, power grid load and commissioning years, construct a multi-constraint equipment health assessment model, and perform feature mapping two-stage training and output logic adaptive correction of the model through sample data. The real-time collected substation status data is input into the corrected health assessment model, and the output is the equipment health level judgment result and the contribution data of each status characteristic to the health judgment. Step 4: Based on the technical specifications for operation and maintenance of substation equipment, construct an operation and maintenance strategy matching system based on fault causal correlation and real-time data-driven approach. The matching coefficient between the real-time status of the equipment and the operation and maintenance requirements is obtained through a dynamic calculation algorithm of the matching coefficient. The equipment health level judgment result, feature contribution data and matching coefficient are collaboratively input into the operation and maintenance strategy matching system to generate a differentiated initial operation and maintenance strategy that accurately corresponds to the real-time status of the equipment and the causes of the fault. The initial operation and maintenance strategy is sequentially checked for logical compliance and on-site operating condition simulation. The strategy that passes both checks is determined to be an executable operation and maintenance strategy. The executable operation and maintenance strategy is then pushed to the corresponding operation and maintenance terminal through a precise matching algorithm.

2. The method for intelligently generating an operation and inspection strategy for a power transformation device according to claim 1, characterized in that, Step 1 specifically includes: The operational status data is divided into electrical operation data, equipment body status data, and equipment environment data. The Zscore standardization method is used in combination with fault correlation characteristics to perform unified field naming and dimensional standardization on the three types of data. Based on the characteristics of data change and the degree of correlation with faults, the standardized data is divided into continuous state data and discrete state data by using a data type determination algorithm, forming standardized state data that can be directly used for feature extraction. This is different from the design of simple dimensional standardization that does not consider fault correlation characteristics. The fault tree analysis method combined with the causal inference algorithm is used to decompose the historical fault records into causes in a hierarchical manner. The fault types are classified according to the location of the fault occurrence, external manifestations, and scope of impact. A unique identifier is configured for each type of fault using a hash algorithm. The identifier includes the fault cause, fault characteristics, fault scope of impact, and fault development pattern. By combining association rule mining algorithms with fault causal relationships, a two-way hierarchical association relationship is constructed, in which fault type points to identification information, identification information points to standardized status data, standardized status data points to historical stable operation data, and historical stable operation data points to status evaluation results, forming a device status dataset with unified fields and unique mapping relationships.

3. The method for intelligently generating operation and maintenance strategies for substation equipment according to claim 2, characterized in that, The specific implementation method of the feature extraction mechanism in step 1 is as follows: For continuous state data, the time series is first arranged and fault-related nodes are labeled. Then, the mean, variance, kurtosis, skewness and rate of change of the series data are calculated by combining the characteristics of fault nodes with the sliding window algorithm. The above calculation results are compared with the corresponding indicators in the historical stable operation data one by one. The correlation coefficient between each indicator and the fault type is calculated by combining the correlation analysis algorithm with the fault causal relationship. The indicators whose absolute value of the correlation coefficient meets the preset correlation conditions and exceeds the normal fluctuation range are extracted as continuous key state characteristics. The preset correlation conditions are determined by combining the correlation analysis of historical fault data and state indicators with the fault causal relationship. For discrete state data, a frequency statistics algorithm combined with fault co-occurrence patterns is used to calculate its occurrence frequency, proportion, and co-occurrence frequency with various fault types. A correlation degree calculation algorithm combined with fault causality is used to calculate the correlation degree between discrete data and fault types. Indicators whose correlation degree meets the correlation threshold are extracted as discrete key state features. By fusing continuous and discrete key state features, and combining the standard range of historical stable operation data with fault boundary data, a static feature baseline that can be dynamically iterated is constructed by using a feature fusion algorithm to combine fault causal relationships.

4. The method for intelligently generating operation and maintenance strategies for substation equipment according to claim 1, characterized in that, Step 2 specifically includes: Key state features within the static feature baseline are arranged in a time series and associated with fault precursor information. Real-time feature parameters are collected at fixed time intervals. The parameter change rate and cumulative change between adjacent acquisition nodes are calculated by combining the differential algorithm with the fault mutation law. The change rate is calculated as the parameter difference between two adjacent acquisition nodes, divided by the time interval between the two adjacent acquisition nodes. The cumulative change is calculated as the difference between the parameter value of each acquisition node and the baseline parameter value. All differences are summed sequentially. Simultaneously, the real-time parameters are compared point by point with the static feature baseline parameters, the absolute difference between the two is calculated, and the fault risk level is associated with it.

5. The method for intelligently generating operation and maintenance strategies for substation equipment according to claim 4, characterized in that, Step 2 uses the mutual information gradient integral algorithm combined with the fault parameter variation pattern to calculate the influence of parameter changes on the sensitivity of the correlation between key state characteristics and equipment operating status. The calculation process is divided into two parts: The first part is the joint probability density function after taking the partial derivative with respect to the parameter changes, and the product of the logarithmic term of the mutual information is double-integrated; The second part is the joint probability density function, which is double-integrated with the result of the partial derivative of the mutual information logarithm term with respect to the parameter change. The two integral results are added together to obtain the final calculated value. This calculation process combines the change law of fault parameters to ensure that the correlation sensitivity calculation is more in line with the actual fault scenario of the equipment, which is different from the design of simple mutual information calculation without combining the fault scenario. The calculated correlation sensitivity is sorted from high to low. A threshold screening algorithm combined with fault impact weights is used to retain features whose correlation sensitivity meets the screening criteria. The screening criteria are determined by training with historical data and combining fault causal relationships. Invalid features whose correlation sensitivity does not meet the screening criteria are removed. The retained features are then normalized using a normalization algorithm combined with fault impact weights, and the impact weights are redistributed.

6. The method for intelligently generating operation and maintenance strategies for substation equipment according to claim 1, characterized in that, Step 3 specifically includes: Using static feature baseline and real-time dynamic calibration feature set as the model input dimension, and equipment operating status combined with fault risk level as the model output dimension, a feature mapping matrix between input and output is established, and multi-parameter dynamic constraints of actual equipment operating scenarios are introduced.

7. The method for intelligently generating operation and maintenance strategies for substation equipment according to claim 6, characterized in that, The division between training data and test data is determined by the principle of maximizing cluster purity combined with the requirement of balanced distribution of fault samples. Cluster purity is calculated by taking the reciprocal of the total amount of data and multiplying it by the sum of the data of the most numerous category in each cluster. When the cluster purity reaches the maximum value and the distribution of fault samples is balanced, the division ratio between training data and test data is determined, which is different from the design that only pursues cluster purity without considering the distribution of fault samples. The training data is split into batches. The amount of data in each batch is determined by a comprehensive calculation of the model memory capacity, convergence speed, and fault sample balance. After each batch of data is input into the model, the weight adjustment range is calculated by combining the feature mapping matrix with the fault feature weight rules. The adjustment range is calculated by combining the step size of the gradient descent method with the fault feature sensitivity. The model parameters are adjusted according to the adjustment range until the accuracy of the fault judgment in that batch of data meets the standard. Then, the next batch of data is input.

8. The method for intelligently generating operation and maintenance strategies for substation equipment according to claim 1, characterized in that, The specific implementation method of the operation and maintenance strategy matching system in step 4 is as follows: By combining the feature contribution ranking algorithm with the fault impact weight, the key state features are ranked, and the core features that meet the core feature selection criteria are extracted. The core feature selection criteria are determined by combining feature contribution analysis with fault causal correlation. A feature operation and maintenance strategy association model based on fault causal correlation is constructed. The model is input with core feature parameter values, health level judgment results, adaptation coefficients and fault risk levels, and outputs operation and maintenance strategies that accurately correspond to fault causes and health status. The core feature parameter values, health level judgment results, adaptation coefficients, and fault risk levels are collaboratively input into the feature operation and maintenance strategy association model. The similarity between the input data and the preset operation and maintenance strategy template is calculated by combining the cosine similarity algorithm with the fault cause matching logic.

9. The method for intelligently generating operation and maintenance strategies for substation equipment according to claim 8, characterized in that, The specific implementation method of logical compliance verification in step 4 is as follows: A logical consistency algorithm is used in conjunction with fault execution logic to verify the logical relationship between the execution conditions of the operation and maintenance strategy. By constructing logical expressions and calculating the truth table of the expressions, if there is a truth value contradiction, it is determined to be a logical conflict. The conflicting conditions are eliminated by combining the fault handling logic and the corresponding execution steps are regenerated to ensure that the execution steps are consistent with the fault handling logic. A data integrity verification algorithm is adopted in combination with fault data requirements. A dataset is generated based on key status features and health judgment results through comparison strategy. The missing rate is calculated by dividing the number of missing data items by the total number of data items. When the missing rate is greater than 0, the missing data is supplemented and the correlation between the data and the fault is verified. Then, the data is re-entered into the matching system for matching to ensure that the supplemented data can support accurate fault matching. A historical matching algorithm is used in conjunction with fault handling patterns to compare the similarity between the strategy execution requirements and historical operating states and fault handling methods. When the similarity does not meet the historical matching criteria, the execution requirements are adjusted by combining the causal analysis algorithm with fault handling patterns to ensure consistency with historical data and continuity of fault handling. The historical matching criteria are determined by analyzing historical operation and maintenance data and combining it with the fault handling effect.

10. The method for intelligently generating operation and maintenance strategies for substation equipment according to claim 9, characterized in that, The specific implementation method of push notifications from the maintenance terminal in step 4 is as follows: A parameter association model is constructed between the operation and maintenance terminal and the operation and maintenance responsibility area, operation permissions, and fault handling qualifications. The model input includes the hardware parameters, permission code, equipment responsibility area code, and fault handling qualification information of the operation and maintenance personnel. By combining the hash mapping algorithm with the operation and maintenance responsibility division rules, the association relationship between the terminal and the responsibility area, permission, and fault handling qualifications is established. The equipment responsibility area code, operation permission code, and required fault handling qualifications corresponding to the executable operation and maintenance strategy are extracted and input into the parameter association model. The matching operation and maintenance terminal is filtered out by a similarity matching algorithm. The filtering conditions are that the responsibility area code is consistent, the permission code contains the operation permission required by the strategy, and the operation and maintenance personnel have the corresponding fault handling qualifications. The executable operation and maintenance strategy is pushed to the filtered terminal through a message push protocol.

Images