Transformer operation and maintenance management method and system based on digital twinning

Abstract

The invention provides a transformer operation and maintenance management method and system based on digital twinning, and relates to the technical field of power equipment operation and maintenance, and the method comprises the steps: carrying out the multi-source integration of transformer sampling data, and obtaining traceable state monitoring data through combining with a distributed block chain technology; adaptively decomposing the monitoring data according to frequency bands and extracting time-varying features, and introducing a cross-frequency domain feature fusion device based on mutual information maximization to analyze relevance of different frequency bands to obtain a state index matrix; constructing a state evolution diagram of the transformer parts, analyzing state propagation characteristics among the parts by adopting dynamic causal reasoning, extracting multi-scale time sequence characteristics, and performing layered interactive coupling to realize space-time conjoint analysis; and calculating the reliability index of each component based on probability distribution, tracing degradation time nodes and influence factors, and formulating a maintenance strategy. According to the invention, omnibearing monitoring and accurate prediction of the state of the transformer are realized, and the operation and maintenance efficiency and the equipment reliability are improved.

Description

Technical Field

[0001] This invention relates to the field of power equipment operation and maintenance technology, and in particular to a transformer operation and maintenance management method and system based on digital twins. Background Technology

[0002] As a core piece of equipment in the power system, the operating status of transformers directly affects the safe and stable operation of the power grid. With the continuous expansion of power system scale and the deepening of smart grid construction, traditional transformer operation and maintenance management models are no longer sufficient to meet the higher requirements of modern power systems for equipment reliability and operational efficiency. In recent years, digital twin technology, as an emerging digital means, has provided a new technical path for transformer operation and maintenance management by constructing digital mappings of physical entities in virtual space, enabling real-time monitoring, diagnostic analysis, and predictive maintenance of physical equipment.

[0003] In existing technologies, transformer monitoring data comes from diverse sources and formats, making effective integration between different systems difficult and resulting in severe data silos. Furthermore, the lack of effective traceability mechanisms during data acquisition, transmission, and storage makes it difficult to trace the complete data chain and historical changes when data anomalies or equipment failures occur, affecting the accuracy and reliability of fault cause analysis. Most existing transformer condition analysis methods employ single-frequency or time-domain feature extraction, neglecting the inherent correlation and complementarity between signals from different frequency bands. The multi-source signals generated during transformer operation have complex frequency distribution characteristics; signals from different frequency bands reflect different physical processes and fault mechanisms. Single-dimensional feature extraction cannot comprehensively depict the true state of the equipment, leading to one-sided and inaccurate condition assessments. Existing technologies often evaluate each component as an independent object when analyzing transformer condition, lacking in-depth analysis of the state propagation mechanisms and spatiotemporal evolution laws between components. In reality, a transformer is a complex system with close functional coupling and state correlations between its components. Deterioration of one component affects other components through complex propagation paths, ultimately leading to systemic failures. Traditional methods struggle to capture this dynamic, multi-dimensional state evolution process, making it difficult to accurately predict potential cascading failure risks and impacting the scientific rigor and foresight of operational decisions. Summary of the Invention

[0004] This invention provides a transformer operation and maintenance management method and system based on digital twins, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a transformer operation and maintenance management method based on digital twins, comprising: Multi-source data integration of transformer sampling data, combined with distributed blockchain technology, yields traceable status monitoring data; The traceable status monitoring data is adaptively decomposed by frequency band, and the time-varying features of each frequency band are extracted. A cross-frequency domain feature fusion fusion based on mutual information maximization is introduced to extract the correlation between the time-varying features of different frequency bands, and the status index matrix of the transformer is obtained. A state evolution diagram of transformer components is constructed, and the state index matrix is ​​mapped to the graph structure as nodes. Dynamic causal reasoning is used to analyze the state propagation between nodes to obtain the state propagation characteristics between components. At the same time, feature representations at different time scales are dynamically extracted to obtain multi-scale time series features. The state propagation characteristics between components and the multi-scale time series features are coupled in a hierarchical interactive manner to obtain the spatiotemporal joint analysis results of the transformer state. Based on the spatiotemporal joint analysis results, the probability distribution of the transformer's health status is calculated. Based on the probability distribution, the reliability index of each component of the transformer is calculated. Components with reliability indexes lower than a preset reliability threshold are screened. The historical operating status is traced and analyzed in conjunction with the traceable status monitoring data to determine the time nodes and influencing factors of reliability degradation and to formulate corresponding maintenance strategies.

[0006] Multi-source data integration of transformer sampling data, combined with distributed blockchain technology, yields traceable status monitoring data, including: A blockchain network node is constructed. The transformer sampling data is divided into multiple feature sub-matrices according to a preset time window. Based on the feature sub-matrices, the response latency, consensus accuracy, and data verification pass rate of each blockchain network node in processing the feature sub-matrices are calculated and combined to obtain a node credibility score. Blockchain network nodes whose node credibility scores exceed a preset credibility threshold are selected to obtain a set of trusted nodes. The set of trusted nodes receives verification requests from other blockchain network nodes and returns confirmation messages based on the Byzantine fault tolerance mechanism. When more than a preset number of blockchain network nodes in the set of trusted nodes return confirmation messages for the same feature sub-matrices, the current feature sub-matrices are marked as verified. The equipment identification information of the transformer is collected. The feature submatrix marked as verified is concatenated with the corresponding time window information and equipment identification information, and then hashed to obtain an index key value. Based on the index key value, the feature submatrix, the time window information, and the equipment identification information are packaged and stored in the blockchain network to generate traceable status monitoring data.

[0007] The traceable status monitoring data is adaptively decomposed by frequency band, and the time-varying features of each frequency band are extracted. A cross-frequency domain feature fusion fusion device based on maximizing mutual information is introduced to extract the correlation between the time-varying features of different frequency bands, resulting in the transformer status index matrix, which includes: The extreme points of the traceable status monitoring data are calculated, and the envelope is constructed using the cubic spline interpolation method for the extreme points. The difference between the traceable status monitoring data and the envelope is used as the frequency band component. The construction process is iteratively executed until the envelope is less than the preset residual ratio threshold, and multiple frequency band components of different frequency bands are obtained. For each frequency band component, a phase analysis transform is performed to extract time-varying phase features and time-varying amplitude features. Time-varying frequency features are calculated based on the time-varying phase features, and energy density features are calculated based on the time-varying amplitude features. The time-varying frequency features and energy density features are combined to form the time-varying features corresponding to each frequency band. Calculate the joint probability distribution and marginal probability distribution of different frequency bands. Based on the logarithmic ratio of the joint probability distribution and the marginal probability distribution, obtain the mutual information value between the time-varying features. Use the gradient ascent method to iteratively optimize the mutual information value until convergence to obtain the maximum mutual information value. Construct feature fusion weights based on the maximum mutual information value. Perform interactive operations between the feature fusion weights and the time-varying features of the corresponding frequency bands and arrange them according to the frequency band correlation to construct the state index matrix.

[0008] Dynamic causal reasoning is used to analyze the state propagation between nodes to obtain the state propagation characteristics between components. Simultaneously, feature representations at different time scales are dynamically extracted to obtain multi-scale temporal features, including: Calculate the partial correlation coefficient and Granger causality test statistic between node pairs. When both the partial correlation coefficient and the Granger causality test statistic are greater than the corresponding preset significance threshold, it is determined that there is a state propagation relationship between the node pairs. Based on the state propagation relationship, construct a component state propagation network. In the component state propagation network, calculate the propagation centrality by the proportion of the shortest path between nodes, and calculate the propagation efficiency by the reciprocal of the shortest distance between nodes. Combine the propagation centrality and the propagation efficiency to obtain the state propagation characteristics between components. An adaptive gating mechanism including a reset gate and an update gate is constructed. The weight matrix of the reset gate is obtained through adaptive learning of the temporal correlation of the state propagation features between the components. The gating parameters of the update gate are dynamically adjusted based on the changing trend of the state propagation features between the components. Features at multiple time scales are captured by setting different time window lengths. Candidate features are obtained by selectively forgetting the historical features at each time scale according to the weight matrix of the reset gate. The candidate features and the current features are adaptively fused with the gating parameters of the update gate to obtain the multi-scale temporal features.

[0009] An adaptive gating mechanism including a reset gate and an update gate is constructed. The weight matrix of the reset gate is obtained through adaptive learning of the temporal correlation of the state propagation features between the components. The gating parameters of the update gate are dynamically adjusted based on the changing trends of the state propagation features between the components, including: The state propagation features between the components are expanded in the time dimension to obtain a temporal state propagation sequence; the covariance of the feature vectors of adjacent time points in the temporal state propagation sequence is calculated to obtain a temporal covariance matrix and eigenvalue decomposition is performed to extract the feature vectors corresponding to the feature values ​​as the dominant correlation direction. The temporal state propagation sequence is projected onto the subspace corresponding to the dominant correlation direction to obtain the correlation enhancement sequence; the inner product operation is performed on the features at different times in the correlation enhancement sequence to calculate the correlation intensity distribution between times, and the weight matrix of the reset gate is constructed based on the correlation intensity distribution; The time-series state propagation sequence is subjected to time-by-time differential operation and sliding window statistics. Based on the rate of change within the sliding window, a trend statistic is obtained. A trend sensitivity coefficient is constructed based on the trend statistic. The trend sensitivity coefficient is directly proportional to the variance of the trend statistic and inversely proportional to the mean of the trend statistic. The trend sensitivity coefficient is used to modulate a preset gating reference parameter to obtain the gating parameter of the update gate.

[0010] By hierarchically and interactively coupling the inter-component state propagation characteristics with the multi-scale temporal characteristics, the spatiotemporal joint analysis results of the transformer state are obtained, including: The inter-component state propagation characteristics and the multi-scale temporal characteristics are decomposed in the frequency domain. Based on a preset frequency threshold, the frequency domain decomposition results are divided into slowly varying feature components and rapidly varying feature components. Calculate the time-delay correlation sequence between the slowly varying feature components, construct a slowly varying domain coupling strength matrix based on the peak position and peak amplitude of the time-delay correlation sequence, and perform convolution operation to obtain the slowly varying domain interaction features; perform wavelet transform on the rapidly varying feature components to obtain state propagation domain wavelet coefficients and time-series domain wavelet coefficients, calculate the scale-wise Hadamard product of the state propagation domain wavelet coefficients and the time-series domain wavelet coefficients, and perform wavelet reconstruction to obtain the rapidly varying domain interaction features; Fourier transforms are performed on the slowly varying domain interaction features and the rapidly varying domain interaction features respectively to obtain the slowly varying domain spectral features and the rapidly varying domain spectral features; the slowly varying domain spectral features and the rapidly varying domain spectral features are dynamically fused in the frequency domain to obtain the fused spectral features; the fused spectral features, the inter-component state propagation features, and the multi-scale time series features are residually connected to obtain the spatiotemporal joint analysis results of the transformer state.

[0011] Based on the probability distribution, the reliability index of each component of the transformer is calculated. Components with reliability indices lower than a preset reliability threshold are screened. Historical operating status is traced and analyzed using the traceable status monitoring data to determine the time points and influencing factors of reliability degradation, including: The failure frequency and operating time of each component of the transformer are obtained from the probability distribution. The failure rate per unit time is calculated and combined with the component importance weight to obtain the failure index. The historical and current values ​​of the transformer performance parameters are obtained, the degradation rate of the performance parameters is calculated, and the performance index is obtained. The failure probability of each component under different state levels is calculated and the fitness value of each component is calculated in combination with the importance weight of the component. The non-dominated ranking is performed by comparing the failure index and the performance index of each component. The congestion degree is obtained by calculating the Euclidean distance between each component in the same non-dominated layer. Based on the fitness value and the congestion degree, selection and crossover operations are performed to obtain the optimal solution set. The reliability index of each component is selected from the optimal solution set. For components whose reliability index is lower than the preset reliability threshold, the traceable status monitoring data is segmented into time series, the status characteristic values ​​in each time period are calculated, the correspondence between the status characteristic values ​​and the equipment performance is established, the status change points and performance degradation inflection points are identified, the status transition time is determined based on the status change points, the correlation degree between each influencing factor and the performance degradation inflection point is calculated using the grey relational analysis method, and the time node and influencing factors of reliability degradation are determined according to the status transition time and the correlation degree.

[0012] A second aspect of this invention provides a transformer operation and maintenance management system based on digital twins, comprising: The first unit is used to integrate multi-source data from transformer sampling data and combine it with distributed blockchain technology to obtain traceable status monitoring data. The second unit is used to adaptively decompose the traceable status monitoring data according to frequency bands, extract the time-varying features of each frequency band data, introduce a cross-frequency domain feature fusion device based on mutual information maximization, extract the correlation between the time-varying features of different frequency bands, and obtain the transformer status index matrix. The third unit is used to construct the state evolution diagram of transformer components. The state index matrix is ​​mapped to the graph structure as nodes. Dynamic causal reasoning is used to analyze the state propagation between nodes to obtain the state propagation characteristics between components. At the same time, feature representations at different time scales are dynamically extracted to obtain multi-scale time series features. The state propagation characteristics between components and the multi-scale time series features are coupled in a hierarchical interactive manner to obtain the spatiotemporal joint analysis results of the transformer state. The fourth unit is used to calculate the probability distribution of the transformer's health status based on the spatiotemporal joint analysis results, calculate the reliability index of each component of the transformer based on the probability distribution, screen the components whose reliability index is lower than the preset reliability threshold, trace and analyze the historical operating status in conjunction with the traceable status monitoring data, determine the time node and influencing factors of reliability degradation, and formulate corresponding maintenance strategies.

[0013] A third aspect of the present invention, An electronic device is provided, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0014] Fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] The beneficial effects of this application are as follows: This invention integrates multi-source data from transformer sampling data and combines it with distributed blockchain technology to obtain traceable status monitoring data. This solves the problems of scattered data sources, low data reliability, and difficulty in tracing historical status in traditional transformer operation and maintenance management. It ensures the integrity, authenticity, and traceability of status monitoring data, provides a reliable data foundation for subsequent status analysis and fault diagnosis, and improves the accuracy and reliability of operation and maintenance decisions.

[0016] This invention adaptively decomposes traceable condition monitoring data by frequency band and extracts the time-varying features of each frequency band. It then introduces a cross-frequency domain feature fusion fusion device based on maximizing mutual information to extract the correlation between time-varying features of different frequency bands, constructs a transformer component state evolution diagram, and uses dynamic causal reasoning to analyze the state propagation between nodes. Simultaneously, it dynamically extracts feature representations at different time scales and obtains the spatiotemporal joint analysis results of transformer state through hierarchical interactive coupling. This solves the problem that traditional methods only focus on single frequency domain features or time features and ignore the correlation between components and the state propagation mechanism, thus realizing a comprehensive and in-depth analysis of transformer operating state and improving the accuracy and comprehensiveness of state assessment.

[0017] This invention calculates the probability distribution of transformer health status and the reliability indicators of each component based on spatiotemporal joint analysis results, screens components with reliability indicators below a preset threshold, and combines traceable status monitoring data to trace and analyze historical operating status, determine the time nodes and influencing factors of reliability degradation, and formulate corresponding maintenance strategies. This solves the problems of lack of initiative and pertinence in traditional operation and maintenance methods and insufficient precision in maintenance strategies, realizing the transformation from passive maintenance to proactive preventive maintenance, improving the scientificity and pertinence of maintenance decisions, reducing operation and maintenance costs, and extending the service life of transformers. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the transformer operation and maintenance management method based on digital twins according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the process for generating traceable status monitoring data for blockchain. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0021] Figure 1 This is a flowchart illustrating the transformer operation and maintenance management method based on digital twins according to an embodiment of the present invention. Figure 1 As shown, the method includes: Multi-source data integration of transformer sampling data, combined with distributed blockchain technology, yields traceable status monitoring data; The traceable status monitoring data is adaptively decomposed by frequency band, and the time-varying features of each frequency band are extracted. A cross-frequency domain feature fusion fusion based on mutual information maximization is introduced to extract the correlation between the time-varying features of different frequency bands, and the status index matrix of the transformer is obtained. A state evolution diagram of transformer components is constructed, and the state index matrix is ​​mapped to the graph structure as nodes. Dynamic causal reasoning is used to analyze the state propagation between nodes to obtain the state propagation characteristics between components. At the same time, feature representations at different time scales are dynamically extracted to obtain multi-scale time series features. The state propagation characteristics between components and the multi-scale time series features are coupled in a hierarchical interactive manner to obtain the spatiotemporal joint analysis results of the transformer state. Based on the spatiotemporal joint analysis results, the probability distribution of the transformer's health status is calculated. Based on the probability distribution, the reliability index of each component of the transformer is calculated. Components with reliability indexes lower than a preset reliability threshold are screened. The historical operating status is traced and analyzed in conjunction with the traceable status monitoring data to determine the time nodes and influencing factors of reliability degradation and to formulate corresponding maintenance strategies.

[0022] In one optional implementation, multi-source data integration of transformer sampling data, combined with distributed blockchain technology, yields traceable status monitoring data, including: A blockchain network node is constructed. The transformer sampling data is divided into multiple feature sub-matrices according to a preset time window. Based on the feature sub-matrices, the response latency, consensus accuracy, and data verification pass rate of each blockchain network node in processing the feature sub-matrices are calculated and combined to obtain a node credibility score. Blockchain network nodes whose node credibility scores exceed a preset credibility threshold are selected to obtain a set of trusted nodes. The set of trusted nodes receives verification requests from other blockchain network nodes and returns confirmation messages based on the Byzantine fault tolerance mechanism. When more than a preset number of blockchain network nodes in the set of trusted nodes return confirmation messages for the same feature sub-matrices, the current feature sub-matrices are marked as verified. The equipment identification information of the transformer is collected. The feature submatrix marked as verified is concatenated with the corresponding time window information and equipment identification information, and then hashed to obtain an index key value. Based on the index key value, the feature submatrix, the time window information, and the equipment identification information are packaged and stored in the blockchain network to generate traceable status monitoring data.

[0023] like Figure 2 As shown, the method includes: For the transformer operation and maintenance management system, a blockchain network of nodes needs to be constructed, which consists of multiple distributed nodes. Transformer sampling data is divided into multiple feature sub-matrices according to a preset time window. The preset time window can be set to 15 minutes, that is, the operating parameters of the transformer, such as temperature, voltage, current, and load rate, collected every 15 minutes form a feature sub-matrix. The feature sub-matrix contains the key status information of the transformer, and its dimension is the number of parameters multiplied by the number of sampling points.

[0024] For each feature submatrix, a trustworthiness score is calculated for each blockchain network node. This score is based on three key metrics: response latency, consensus accuracy, and data verification pass rate. Response latency, measured in milliseconds, refers to the time a node requires to process the feature submatrix; consensus accuracy refers to the proportion of nodes providing correct results during the consensus process; and data verification pass rate refers to the success rate of nodes verifying the validity of data. A weighted average of these three metrics yields a comprehensive score, with weights set as follows: response latency 0.3, consensus accuracy 0.4, and data verification pass rate 0.3.

[0025] Specifically, response latency is calculated by recording the time interval from when a node receives a feature submatrix to when it completes processing; the shorter the time, the higher the score. Consensus accuracy is calculated by comparing the degree of agreement between the consensus results proposed by the nodes and the final consensus results. Data verification pass rate is calculated by counting the proportion of times a node successfully verifies data out of the total number of verifications. When the response latency of a node processing a feature submatrix is ​​50 milliseconds, the consensus accuracy is 95%, and the data verification pass rate is 90%, its credibility score is 0.3×(1-50 / 200)+0.4×0.95+0.3×0.9=0.3×0.75+0.4×0.95+0.3×0.9=0.225+0.38+0.27=0.875.

[0026] Based on the calculated node credibility scores, nodes with scores exceeding a preset credibility threshold of 0.8 are selected to form a set of trusted nodes; that is, nodes with scores greater than 0.8 are considered trusted nodes. These trusted nodes handle verification requests based on a Byzantine fault tolerance mechanism. This mechanism ensures data consistency and reliability even if some nodes fail or engage in malicious behavior. When the trusted node set receives verification requests for feature submatrices from other nodes, each trusted node independently verifies the matrix and returns a confirmation message. The preset number can be set to two-thirds of the total number of trusted nodes. When more than two-thirds of the trusted nodes return confirmation messages for the same feature submatric, the matrix is ​​marked as verified. For example, when the trusted node set contains 9 nodes, at least 6 nodes need to confirm before a feature submatric can be marked as verified.

[0027] For the verified feature submatrix, the corresponding transformer's equipment identification information is collected, including equipment number, installation location, production date, and specifications. The verified feature submatrix is ​​concatenated with the corresponding time window information and equipment identification information to form a complete data packet. The data packet structure is: feature submatrix data + time window information + equipment identification information.

[0028] The resulting complete data packet is hashed to generate an index key. The hash calculation uses the SHA-256 algorithm to convert the data packet into a fixed-length hash value, which serves as the index identifier in the blockchain. For example, the index key generated after hashing the feature submatrix, time information, and device identifier of a transformer within a specific time window is "8f7d56a1e2b3c4d5e6f7g8h9i0j1k2l3m4n5o6p7q8r9s0t1u2v3w4x5y6z7".

[0029] Based on the generated index key, the feature submatrix data, time window information, and device identification information are packaged as transaction data. This data is then confirmed and stored in a block through the consensus mechanism of the blockchain network. Each block in the blockchain network contains transaction data, the hash value of the previous block, a timestamp, and the block height. Transaction data is stored in key-value pairs, with the index key serving as a retrieval identifier and the corresponding value representing the packaged data content.

[0030] When it is necessary to trace the operating status of a transformer at a specific time, a query index can be generated using the equipment identifier and time information to retrieve the corresponding status monitoring data from the blockchain network. Due to the immutability of the blockchain, the retrieved data has high credibility and can accurately reflect the historical operating status of the transformer.

[0031] The digital twin-based transformer operation and maintenance management method implemented through the aforementioned technologies ensures the authenticity and traceability of transformer monitoring data, effectively preventing data tampering or loss. The introduction of blockchain technology enhances data security, the Byzantine fault tolerance mechanism improves fault tolerance capabilities, and the dynamic node credibility assessment mechanism ensures the reliability of network nodes. This method achieves full lifecycle management of transformer operating status, providing reliable data support for equipment fault analysis, lifespan prediction, and maintenance decisions. It improves the safe and stable operation of the power system, reduces maintenance costs, and extends the service life of transformer equipment.

[0032] In one optional implementation, the traceable status monitoring data is adaptively decomposed by frequency band, and the time-varying features of each frequency band are extracted. A cross-frequency domain feature fusion fusion device based on maximizing mutual information is introduced to extract the correlation between the time-varying features of different frequency bands, resulting in a transformer status index matrix including: The extreme points of the traceable status monitoring data are calculated, and the envelope is constructed using the cubic spline interpolation method for the extreme points. The difference between the traceable status monitoring data and the envelope is used as the frequency band component. The construction process is iteratively executed until the envelope is less than the preset residual ratio threshold, and multiple frequency band components of different frequency bands are obtained. For each frequency band component, a phase analysis transform is performed to extract time-varying phase features and time-varying amplitude features. Time-varying frequency features are calculated based on the time-varying phase features, and energy density features are calculated based on the time-varying amplitude features. The time-varying frequency features and energy density features are combined to form the time-varying features corresponding to each frequency band. Calculate the joint probability distribution and marginal probability distribution of different frequency bands. Based on the logarithmic ratio of the joint probability distribution and the marginal probability distribution, obtain the mutual information value between the time-varying features. Use the gradient ascent method to iteratively optimize the mutual information value until convergence to obtain the maximum mutual information value. Construct feature fusion weights based on the maximum mutual information value. Perform interactive operations between the feature fusion weights and the time-varying features of the corresponding frequency bands and arrange them according to the frequency band correlation to construct the state index matrix.

[0033] After acquiring the traceable status monitoring data of the transformer, the extreme points of the data are first calculated. These extreme points include local maximum and local minimum points. For a data point at a given time t, if its value is greater than the adjacent data points on either side, it is determined to be a local maximum point; if its value is less than the adjacent data points on either side, it is determined to be a local minimum point. For example, for the sampling point sequence [1.2, 2.5, 1.8, 0.9, 1.5, 3.2, 2.1], the identified extreme points are the 2nd point (2.5), the 4th point (0.9), and the 6th point (3.2).

[0034] For the calculated extreme points, a cubic polynomial function is constructed between adjacent extreme points to ensure the continuity of the first and second derivatives at the connection points, thus obtaining a smooth curve. The upper envelope connects all local maxima and the lower envelope connects all local minima. For the example above, cubic spline interpolation can be used to obtain the upper envelope of the connection points (2, 2.5) and (6, 3.2), and the lower envelope of the connection point (4, 0.9).

[0035] The difference between the original traceable status monitoring data and the average value of the constructed envelope is calculated to obtain a frequency band component. The average envelope is the arithmetic mean of the upper and lower envelopes, which represents the local mean trend of the signal. Subtracting this average envelope from the original data yields the residual signal after removing the low-frequency trend. This residual signal is a frequency band component.

[0036] The algorithm determines whether the fluctuation amplitude of the obtained envelope is less than a preset residual ratio threshold. The residual ratio is calculated as the ratio of the standard deviation of the envelope to the standard deviation of the original data. The preset residual ratio threshold can be set to 0.05, meaning that the iteration process terminates when the fluctuation amplitude of the envelope drops below 5% of the fluctuation amplitude of the original data. If it is greater than the threshold, the frequency band component obtained in the previous step is used as new input data, and the process of constructing the envelope and extracting the frequency band component is repeated until the termination condition is met.

[0037] Through the above iterative process, multiple frequency band components of different frequency ranges can be obtained, arranged from high to low frequency. For example, for the vibration signal of a transformer, components representing high-frequency mechanical vibration, medium-frequency load fluctuation, and low-frequency temperature change can be decomposed. Each frequency band component reflects the operating characteristics of the transformer in different frequency domains.

[0038] For each frequency band component, a phase analytical transform is performed. The real signal is converted into an analytic signal using the Hilbert transform. The analytic signal consists of the original real part and the imaginary part obtained through the Hilbert transform. The analytic signal can be represented as a combination of amplitude and phase, from which time-varying amplitude and phase features can be directly extracted. For a specific frequency band component of the transformer, the time-varying amplitude feature reflects changes in signal strength, and the time-varying phase feature reflects changes in signal phase.

[0039] Time-varying frequency features, representing the rate of phase change, are obtained by calculating the time derivative of the phase characteristics. For example, for the extracted phase sequence [0.2π, 0.5π, 0.9π, 1.4π, 2.0π], the phase differences between adjacent sampling points are 0.3π, 0.4π, 0.5π, and 0.6π, respectively. Dividing these differences by the sampling interval yields the corresponding instantaneous frequency features.

[0040] The energy density characteristic is obtained by squaring the time-varying amplitude characteristic and integrating it within a time window. This energy density characteristic reflects the energy distribution of the signal over different time periods and can indicate changes in the transformer's operating status. For example, a sudden increase in the vibration signal's energy density indicates equipment loosening or abnormal friction.

[0041] The calculated time-varying frequency characteristics and energy density characteristics are combined to form the time-varying characteristics corresponding to each frequency band. The combination method is direct cascading of the two characteristics, and the resulting feature vector more comprehensively describes the time-frequency characteristics of the frequency band signal. For each frequency band component of the transformer, its time-varying characteristics contain the dominant frequency and energy distribution information of that frequency band.

[0042] To extract the correlation between time-varying features in different frequency bands, the joint probability distribution and marginal probability distribution of different frequency bands are calculated. Histogram statistics are performed on the time-varying feature data to estimate their probability density distribution. For any two time-varying features in different frequency bands, a two-dimensional histogram is constructed to obtain their joint probability distribution; a one-dimensional histogram is constructed for each feature individually to obtain its marginal probability distribution. For example, the numerical distributions of high-frequency vibration features and mid-frequency load fluctuation features are statistically analyzed to obtain their respective marginal probability distributions; simultaneously, the numerical combinations of both are statistically analyzed to obtain their joint probability distribution.

[0043] Based on the calculated joint probability distribution and marginal probability distribution, the mutual information value is calculated by taking the logarithmic ratio of the product of the joint probability distribution and the marginal probability distribution, which measures the degree of interdependence between the two features. For example, the mutual information value of high-frequency vibration features and mid-frequency load fluctuation features reflects the degree of influence of load changes on equipment vibration.

[0044] The gradient ascent method is used to iteratively optimize the mutual information value. In each iteration, the gradient of the mutual information with respect to the feature fusion parameters is calculated, and the parameters are updated along the gradient direction, gradually increasing the mutual information value. The iterative process continues until the mutual information value converges, that is, the change in mutual information value between two adjacent iterations is less than a preset convergence threshold, such as 0.001. The maximum mutual information value obtained through iterative optimization reflects the strongest correlation between features in different frequency bands.

[0045] Based on the maximum mutual information value, feature fusion weights are constructed. The weight value is proportional to the mutual information value, and the frequency band combination with the higher the mutual information value receives a larger fusion weight. For example, when the mutual information value between high-frequency vibration features and low-frequency temperature change features is high, these two frequency bands will receive a larger weight in the fusion process, highlighting their correlation.

[0046] The feature fusion weights and time-varying features of the corresponding frequency bands are interactively calculated using a weighted summation method to achieve cross-frequency domain feature fusion. The features of each frequency band are linearly combined according to their corresponding weights. The fused features are arranged according to the frequency band correlation to construct the transformer's state index matrix. The rows of the state index matrix represent different frequency bands, the columns represent different time points, and the matrix element values ​​reflect the state index values ​​of the corresponding frequency bands at specific time points.

[0047] The aforementioned cross-frequency domain feature fusion method based on maximizing mutual information enables in-depth feature extraction and correlation analysis of transformer traceable condition monitoring data. This method fully leverages the transformer operating status information contained in data from different frequency bands, overcoming the limitations of traditional single-frequency band analysis. The status index matrix comprehensively reflects the transformer's operating status, providing high-precision data support for fault diagnosis and predictive maintenance. This method effectively identifies early fault symptoms in transformers, improving the sensitivity and accuracy of fault detection and reducing false alarms and missed alarms. Simultaneously, by extracting cross-frequency domain feature correlations based on the principle of maximizing mutual information, the anti-interference capability of condition monitoring is enhanced, adapting to the transformer operation and maintenance management needs in complex power environments.

[0048] In one optional implementation, dynamic causal reasoning is used to analyze the state propagation between nodes to obtain the state propagation characteristics between components. Simultaneously, feature representations at different time scales are dynamically extracted to obtain multi-scale temporal features, including: Calculate the partial correlation coefficient and Granger causality test statistic between node pairs. When both the partial correlation coefficient and the Granger causality test statistic are greater than the corresponding preset significance threshold, it is determined that there is a state propagation relationship between the node pairs. Based on the state propagation relationship, construct a component state propagation network. In the component state propagation network, calculate the propagation centrality by the proportion of the shortest path between nodes, and calculate the propagation efficiency by the reciprocal of the shortest distance between nodes. Combine the propagation centrality and the propagation efficiency to obtain the state propagation characteristics between components. An adaptive gating mechanism including a reset gate and an update gate is constructed. The weight matrix of the reset gate is obtained through adaptive learning of the temporal correlation of the state propagation features between the components. The gating parameters of the update gate are dynamically adjusted based on the changing trend of the state propagation features between the components. Features at multiple time scales are captured by setting different time window lengths. Candidate features are obtained by selectively forgetting the historical features at each time scale according to the weight matrix of the reset gate. The candidate features and the current features are adaptively fused with the gating parameters of the update gate to obtain the multi-scale temporal features.

[0049] After acquiring the state data of transformer components, the partial correlation coefficients and Granger causality test statistics between component node pairs are calculated to determine the state propagation relationship between components. During the calculation of the partial correlation coefficient, time series data of two component nodes are selected, and the influence of other nodes is controlled before calculating the residual correlation. For example, when calculating the partial correlation coefficient between the transformer oil temperature node and the winding temperature node, the influence of other nodes such as ambient temperature and load rate needs to be controlled. When the partial correlation coefficient between two components is 0.75, and the preset significance threshold is 0.6, it indicates a strong correlation between the two components.

[0050] The Granger causality test statistic is calculated based on a time series prediction model. For two component nodes A and B, a model containing only historical data of A and a model containing historical data of both A and B are established respectively. The Granger causality test statistic is calculated by comparing the prediction errors of the two models. If the statistic is greater than a preset significance threshold, it indicates that node B has predictive power for node A, i.e., a causal relationship exists. For example, when the Granger causality test statistic for oil temperature on winding temperature is 8.3 and the preset significance threshold is 5.0, it is considered that changes in oil temperature have a significant impact on winding temperature.

[0051] When both the partial correlation coefficient and the Granger causality test statistic are greater than their respective preset significance thresholds, a state propagation relationship is determined between the node pairs. Taking the transformer oil temperature node and the winding temperature node as an example, when their partial correlation coefficient is 0.75, which is greater than the threshold of 0.6, and the Granger causality test statistic is 8.3, which is greater than the threshold of 5.0, a state propagation relationship is determined between these two nodes.

[0052] A component state propagation network is constructed based on defined state propagation relationships. In this network, nodes represent various components of the transformer, such as the oil tank, windings, cooling system, and bushings; directed edges represent state propagation relationships, with the direction of the edges pointing from the source of influence to the affected component. For example, there exists a directed edge from the oil temperature node to the winding temperature node, indicating that changes in oil temperature will propagate and affect the winding temperature.

[0053] In the constructed component state propagation network, the propagation centrality of each node is calculated. The propagation centrality is obtained by calculating the proportion of the shortest paths passing through the node. Specifically, for each node in the network, the ratio of the number of shortest paths passing through that node to the total number of shortest paths in the network is calculated. For example, if there are a total of 120 shortest paths in the transformer state propagation network, and 30 of them pass through the oil temperature node, then the propagation centrality of the oil temperature node is 30 / 120 = 0.25.

[0054] The propagation efficiency of a network is calculated by taking the reciprocal of the shortest distance between nodes. For each pair of nodes in the network, the distance of the shortest path between them is calculated, and the average of the reciprocals is taken. For example, in a transformer state propagation network, the shortest distance from the oil temperature node to the winding temperature node is 1, and the shortest distance from the oil temperature node to the bushing temperature node is 2. The propagation efficiency for this part is calculated as (1 / 1 + 1 / 2) / 2 = 0.75.

[0055] The calculated propagation centrality and propagation efficiency are combined and concatenated into a feature vector to obtain the state propagation characteristics between components. For example, the propagation centrality of the oil temperature node is 0.25, and the propagation efficiency of the related node pair is 0.75, so the corresponding state propagation characteristics are [0.25, 0.75]. These characteristics reflect the importance and influence range of each component of the transformer in the state propagation network.

[0056] The state propagation characteristics between components are expanded over time to form sequence data, such as recording the propagation centrality and propagation efficiency of each component of a transformer over 24 hours; the covariance of the feature vectors at adjacent time points is calculated to obtain the time-series covariance matrix; eigenvalue decomposition is performed on the covariance matrix, and the main feature vectors (such as v1 and v2 corresponding to eigenvalues ​​5.2 and 3.8) are selected as the dominant correlation directions; the original sequence is projected onto these dominant directions to obtain the correlation-enhanced sequence; the correlation strength distribution is constructed by calculating the inner product of features at different time points, thereby forming the reset gate weight matrix; at the same time, the sequence is differentially processed, and the mean and variance of the rate of change are statistically analyzed within a sliding window (such as a 3-hour window); based on these statistics, the trend sensitivity coefficient (such as 0.012 / |0.08|=0.15) is calculated, and the gating baseline parameters (such as 0.5×0.15=0.075) are modulated by this coefficient to obtain the gating parameters of the update gate.

[0057] Different time window lengths are set to capture features across multiple time scales. For example, time windows of 3 hours, 6 hours, and 12 hours are set to capture short-term, medium-term, and long-term variation features, respectively. Within each time window, feature representations of the transformer component states are extracted.

[0058] Candidate features are obtained by selectively forgetting historical features at each time scale based on the weight matrix of the reset gate. Specifically, the weight matrix of the reset gate is multiplied element-wise with the historical features, and historical features with small weights are significantly reduced, achieving selective forgetting. For example, when the reset gate weight of a certain historical feature is 0.15, the influence of that feature in the candidate features will be reduced by 85%.

[0059] The candidate features and the current features are adaptively weighted and summed based on the gating parameter of the update gate to obtain multi-scale temporal features, where the weights are determined by the gating parameter of the update gate. For example, when the gating parameter of the update gate is 0.075, the fused features consist of 7.5% candidate features and 92.5% current features.

[0060] Through the aforementioned dynamic causal reasoning and adaptive gating mechanism, accurate extraction of state propagation characteristics of transformer components and dynamic fusion of multi-scale time-series features are achieved. This method can adaptively capture the state transmission relationships between different transformer components, identifying key propagation paths and core influencing components. The extraction of multi-scale time-series features effectively integrates information from different time scales, enhancing the model's ability to perceive short- and long-term changes. This approach significantly improves the accuracy and robustness of state assessment, effectively responding to dynamic changes in the transformer's operating environment and load conditions, providing reliable data support for predictive maintenance, greatly reducing false alarm and false negative rates, and improving the safe and stable operation of the power system.

[0061] In one optional implementation, an adaptive gating mechanism including a reset gate and an update gate is constructed. The weight matrix of the reset gate is obtained through adaptive learning of the temporal correlation of the inter-component state propagation features. The gating parameters of the update gate are dynamically adjusted based on the changing trend of the inter-component state propagation features, including: The state propagation features between the components are expanded in the time dimension to obtain a temporal state propagation sequence; the covariance of the feature vectors of adjacent time points in the temporal state propagation sequence is calculated to obtain a temporal covariance matrix and eigenvalue decomposition is performed to extract the feature vectors corresponding to the feature values ​​as the dominant correlation direction. The temporal state propagation sequence is projected onto the subspace corresponding to the dominant correlation direction to obtain the correlation enhancement sequence; the inner product operation is performed on the features at different times in the correlation enhancement sequence to calculate the correlation intensity distribution between times, and the weight matrix of the reset gate is constructed based on the correlation intensity distribution; The time-series state propagation sequence is subjected to time-by-time differential operation and sliding window statistics. Based on the rate of change within the sliding window, a trend statistic is obtained. A trend sensitivity coefficient is constructed based on the trend statistic. The trend sensitivity coefficient is directly proportional to the variance of the trend statistic and inversely proportional to the mean of the trend statistic. The trend sensitivity coefficient is used to modulate a preset gating reference parameter to obtain the gating parameter of the update gate.

[0062] Constructing an adaptive gating mechanism that includes reset and update gates requires expanding the state propagation characteristics between components along a time dimension to obtain a temporal state propagation sequence. For example, recording the state propagation characteristics of transformer components hourly throughout a day forms a sequence of data at 24 time points. The feature vector at each time point contains the propagation centrality and propagation efficiency of each component node at that moment.

[0063] The covariance of the eigenvectors of adjacent moments in the time-series state propagation sequence is calculated to obtain the time-series covariance matrix. Taking 24-hour transformer condition monitoring as an example, the covariance of the eigenvectors of adjacent moments such as hour 1 and hour 2, hour 2 and hour 3 is calculated to form the time-series covariance matrix.

[0064] Eigenvalue decomposition is performed on the time-series covariance matrix, and the eigenvectors corresponding to the eigenvalues ​​are extracted as the dominant correlation directions. Eigenvectors with larger eigenvalues ​​represent the main patterns of change in the time-series data. For example, if the decomposition yields eigenvalues ​​[5.2, 3.8, 1.5, 0.6], and the corresponding eigenvectors are v1, v2, v3, and v4, then v1 and v2 are considered the dominant correlation directions because their corresponding eigenvalues ​​are significantly larger than the other eigenvalues.

[0065] Projecting the temporal state propagation sequence onto the subspace corresponding to the dominant correlation direction, and performing an inner product operation between the original feature vector and the feature vector of the dominant correlation direction, yields a correlation-enhanced sequence. For example, projecting the original 24-hour state propagation sequence onto the plane formed by v1 and v2 yields a correlation-enhanced sequence.

[0066] The inner product operation is performed on the features at different times in the correlation enhancement sequence to calculate the correlation strength distribution between time points. For example, the inner product values ​​of features at different times such as hour 1 and hour 5, hour 2 and hour 8 are calculated to obtain the correlation strength matrix between time points.

[0067] The weight matrix of the reset gate is constructed based on the calculated correlation strength distribution. Time points with high correlation strength correspond to larger weights, and time points with weak correlation correspond to smaller weights. For example, if the correlation strength between the 1st hour and the 5th hour is 0.85, then the corresponding element in the reset gate weight matrix is ​​set to a larger value, such as 0.85; if the correlation strength between the 2nd hour and the 12th hour is only 0.15, then the corresponding weight is set to a smaller value, such as 0.15.

[0068] Perform time-by-time difference operations on the time-series state propagation sequence to calculate the changes in features at adjacent time points. For example, calculate the difference between the features of the 2nd hour and the 1st hour, the difference between the features of the 3rd hour and the 2nd hour, and so on, to obtain the sequence of feature changes.

[0069] A sliding window statistical analysis is performed on the differencing sequence. Based on the rate of change within the sliding window, a trend statistic is obtained. The sliding window size can be set to 3 hours, shifting forward by 1 hour each time. Within each window, the mean and variance of the rate of change are calculated as the trend statistic for that window. For example, for a window containing hours 1 to 3, the mean of the characteristic rate of change within these 3 hours is calculated to be 0.08, and the variance is 0.012.

[0070] A trend sensitivity coefficient is constructed based on the trend statistic. The trend sensitivity coefficient is directly proportional to the variance of the trend statistic and inversely proportional to the mean of the trend statistic. Specifically, the trend sensitivity coefficient is equal to the absolute value of the variance of the rate of change divided by the mean of the rate of change. For example, when the mean of the rate of change is 0.08 and the variance is 0.012, the trend sensitivity coefficient is 0.012 / |0.08| = 0.15.

[0071] The gate parameters for the updated gate are obtained by modulating the preset gate reference parameters using a trend sensitivity coefficient. The preset gate reference parameter can be set to 0.5, and the modulation method is to multiply the trend sensitivity coefficient by the reference parameter. For example, when the trend sensitivity coefficient is 0.15 and the gate reference parameter is 0.5, the gate parameters for the updated gate are 0.15 × 0.5 = 0.075.

[0072] In one optional implementation, the inter-component state propagation characteristics and the multi-scale temporal characteristics are hierarchically and interactively coupled to obtain the spatiotemporal joint analysis results of the transformer state, including: The inter-component state propagation characteristics and the multi-scale temporal characteristics are decomposed in the frequency domain. Based on a preset frequency threshold, the frequency domain decomposition results are divided into slowly varying feature components and rapidly varying feature components. Calculate the time-delay correlation sequence between the slowly varying feature components, construct a slowly varying domain coupling strength matrix based on the peak position and peak amplitude of the time-delay correlation sequence, and perform convolution operation to obtain the slowly varying domain interaction features; perform wavelet transform on the rapidly varying feature components to obtain state propagation domain wavelet coefficients and time-series domain wavelet coefficients, calculate the scale-wise Hadamard product of the state propagation domain wavelet coefficients and the time-series domain wavelet coefficients, and perform wavelet reconstruction to obtain the rapidly varying domain interaction features; Fourier transforms are performed on the slowly varying domain interaction features and the rapidly varying domain interaction features respectively to obtain the slowly varying domain spectral features and the rapidly varying domain spectral features; the slowly varying domain spectral features and the rapidly varying domain spectral features are dynamically fused in the frequency domain to obtain the fused spectral features; the fused spectral features, the inter-component state propagation features, and the multi-scale time series features are residually connected to obtain the spatiotemporal joint analysis results of the transformer state.

[0073] Frequency domain decomposition is performed on the state propagation characteristics and multi-scale time-series characteristics between components. A Fast Fourier Transform (FFT) is used to convert the time-domain signal into a frequency-domain representation. Based on a preset frequency threshold, such as 0.05Hz, the frequency domain decomposition result is divided into slowly varying characteristic components and rapidly varying characteristic components. The slowly varying characteristic components correspond to the low-frequency part, reflecting the long-term evolution trend of the transformer state; the rapidly varying characteristic components correspond to the high-frequency part, reflecting short-term fluctuations and transient changes.

[0074] Cross-correlation analysis was performed on the slowly varying components of the state propagation characteristics and the slowly varying components of the time series characteristics at different time lags to obtain a correlation coefficient sequence that varies with time lag. For example, the correlation coefficients between the slowly varying components of oil temperature and winding temperature were calculated at different time lags such as 0 hours, 1 hour, and 2 hours, forming a time lag correlation sequence [0.65, 0.83, 0.72, 0.51, 0.32].

[0075] The peak position of the time-delay correlation sequence reflects the time delay of state propagation, and the peak amplitude represents the coupling strength. For example, when the time-delay correlation sequence reaches a peak of 0.83 at 1 hour, it indicates that the time delay for the oil temperature change to propagate to the winding temperature is approximately 1 hour, and the coupling strength is 0.83. The coupling strength and time delay of each component pair are organized into a matrix form.

[0076] A convolution operation is performed on the coupling strength matrix of the slowly varying domain, and a Gaussian kernel or an exponentially decaying kernel is used to smooth the coupling strength, enhancing the representation of the coupling relationship between adjacent components and obtaining the interactive features of the slowly varying domain. The convolution kernel size can be selected as 3×3 to balance local information extraction and computational efficiency.

[0077] For the rapidly changing characteristic components, wavelet transform processing is performed. A wavelet basis function suitable for the transformer state signal characteristics is selected, such as the Daubechies wavelet db4. Wavelet decomposition is performed on the rapidly changing components with state propagation characteristics and the rapidly changing components with time-series characteristics, respectively, to obtain wavelet coefficients at different scales. For example, a 5-level wavelet decomposition is performed on the rapidly changing oil temperature component to obtain detailed coefficients at 5 scales and one approximate coefficient.

[0078] The scale-wise Hadamard product of wavelet coefficients in the state propagation domain and the time-series domain is calculated to achieve interactive fusion of features from the two domains in the wavelet space. The Hadamard product operation preserves the common salient features in both domains and suppresses noise and interference in a single domain.

[0079] Wavelet reconstruction is performed on the interactive wavelet coefficients obtained from the Hadamard product, and the signals are converted back to the time domain to obtain the fast variable domain interactive features. The reconstruction process uses the same wavelet basis functions as the decomposition to ensure consistent information restoration.

[0080] Fourier transforms are performed on the slowly varying domain interaction features and the fast varying domain interaction features respectively to convert the time-domain signal into a frequency-domain representation, resulting in slowly varying domain spectral features and fast varying domain spectral features. The spectral features represent the frequency composition of the signal in the form of amplitude and phase, which facilitates subsequent frequency domain fusion processing.

[0081] A frequency-adaptive weighting strategy is used to dynamically fuse the spectral features of the slowly varying domain and the rapidly varying domain. The weighting coefficients are dynamically adjusted according to the energy distribution of each frequency point. The slowly varying domain spectral features are mainly retained in the low frequency band, the rapidly varying domain spectral features are mainly retained in the high frequency band, and the mid frequency band undergoes a smooth transition to obtain the fused spectral features.

[0082] The fused spectral features are transformed back to the time domain and then residually connected with the original inter-component state propagation features and multi-scale temporal features. This residual connection preserves the original feature information while enhancing the information after interactive fusion, preventing information loss. The residual connection method involves a weighted sum of the fused features and the original features, with weights set to 0.6 and 0.4 to balance the old and new information.

[0083] Through the aforementioned hierarchical interactive coupling process, the spatiotemporal joint analysis results of the transformer condition are obtained. These results comprehensively integrate the spatial relationships of state propagation between components and the temporal patterns of state evolution, capturing both slowly changing long-term trends and retaining rapidly fluctuating short-term characteristics, thus providing a more comprehensive and accurate representation of transformer condition.

[0084] In one optional implementation, reliability indices for each component of the transformer are calculated based on the probability distribution. Components with reliability indices below a preset reliability threshold are selected. Historical operating conditions are then analyzed using traceable status monitoring data to determine the time points and influencing factors of reliability degradation, including: The failure frequency and operating time of each component of the transformer are obtained from the probability distribution. The failure rate per unit time is calculated and combined with the component importance weight to obtain the failure index. The historical and current values ​​of the transformer performance parameters are obtained, the degradation rate of the performance parameters is calculated, and the performance index is obtained. The failure probability of each component under different state levels is calculated and the fitness value of each component is calculated in combination with the importance weight of the component. The non-dominated ranking is performed by comparing the failure index and the performance index of each component. The congestion degree is obtained by calculating the Euclidean distance between each component in the same non-dominated layer. Based on the fitness value and the congestion degree, selection and crossover operations are performed to obtain the optimal solution set. The reliability index of each component is selected from the optimal solution set. For components whose reliability index is lower than the preset reliability threshold, the traceable status monitoring data is segmented into time series, the status characteristic values ​​in each time period are calculated, the correspondence between the status characteristic values ​​and the equipment performance is established, the status change points and performance degradation inflection points are identified, the status transition time is determined based on the status change points, the correlation degree between each influencing factor and the performance degradation inflection point is calculated using the grey relational analysis method, and the time node and influencing factors of reliability degradation are determined according to the status transition time and the correlation degree.

[0085] The failure frequency and operating time data of each component of the transformer are obtained from the probability distribution. For example, the transformer core experienced two failures in the past 5 years, with a total operating time of 43,800 hours; the windings experienced three failures during the same period, with a total operating time of 43,800 hours. The failure rate per unit time is calculated, such as the core failure rate being 2 / 43800 = 4.57 × 10⁻⁶. -5 The winding failure rate is 3 / 43800 = 6.85 × 10⁻⁶ times per hour. -5 times / hour.

[0086] Fault indicators are calculated by combining the importance weights of components. The importance weights of components are determined based on their criticality within the transformer system; for example, the importance weight of the winding is 0.95, and the importance weight of the core is 0.9. The fault indicator is calculated by multiplying the fault rate by the importance weight; for example, the winding fault indicator is 6.85 × 10⁻⁶. -5 ×0.95=6.51×10 -5 The core failure index is 4.57 × 10⁻⁶. -5 ×0.9=4.11×10-5 .

[0087] Obtain historical and current values ​​of transformer performance parameters, and calculate the degradation rate of these parameters. For example, if the initial degree of polymerization of the insulating paper is 1200 and the current degree of polymerization is 850, after 5 years of operation, the degradation rate is (1200-850) / (5×365×24)=8.00×10⁻¹⁰. -3 / hour; the initial excitation current of the iron core was 0.5A, currently it is 0.8A, and the degradation rate is (0.8-0.5) / (5×365×24)=6.85×10 -6 A / hour.

[0088] Performance indicators are calculated based on the degradation rate. These indicators can be determined by the ratio of the degradation rate to the standard degradation rate. For example, if the degradation rate of the standard insulating paper's degree of polymerization is 6.00 × 10⁻⁶... -3 / hour, the standard growth rate of the core excitation current is 4.00 × 10 -6 If A / hour, then the performance index of the insulating paper is 8.00×10. -3 / 6.00×10 -3 =1.33, the core performance index is 6.85×10 -6 / 4.00×10 -6 =1.71.

[0089] Calculate the failure probability of each component under different state levels. Estimate the failure probability of components under different states using a reliability model, such as the Weibull distribution model. For example, when the degree of polymerization of the insulating paper drops to 850, the failure probability is 0.15; when the core excitation current rises to 0.8A, the failure probability is 0.18.

[0090] The fitness value of each component is calculated by combining the component importance weights. Fitness value = (1 - failure probability) × importance weight. For example, the fitness value of insulating paper is (1 - 0.15) × 0.95 = 0.808, and the fitness value of iron core is (1 - 0.18) × 0.9 = 0.738.

[0091] Non-dominated ranking is performed based on fault indicators and performance indicators. This ranking divides all components into different levels. The first level contains components that are not dominated by any other components, meaning that their fault indicators and performance indicators are no worse than those of other components. For example, if the main components of a transformer are ranked into three non-dominated levels: the first level includes the cooling system and bushings; the second level includes the windings and core; and the third level includes the on-load tap changer and insulation system.

[0092] The congestion degree is obtained by calculating the Euclidean distance between components in the same non-dominated layer. The congestion degree reflects the distribution density of components in the target space. The greater the congestion degree, the higher the distinction between components and other components. For example, in the second layer, if the Euclidean distance between the winding and the core in the fault index and performance index space is 0.12, then the congestion degree of both the winding and the core is 0.12.

[0093] Based on fitness values ​​and crowding levels, selection and crossover operations are performed to obtain the optimal solution set. The selection operation prioritizes components with lower non-dominated levels, and within the same level, it prioritizes components with higher crowding levels. The crossover operation generates new solutions by combining the characteristics of different components. After multiple iterations, a solution set containing the optimal reliability assessments for each component is obtained.

[0094] The reliability index of each component is selected from the optimal solution set. The reliability index usually depends on a comprehensive evaluation of the component's failure probability, importance weight, fault index, and performance index. For example, the reliability index of the winding is 0.82, the reliability index of the iron core is 0.80, and the reliability index of the on-load tap changer is 0.68.

[0095] The reliability index of each component is compared with a preset reliability threshold, such as 0.70, and components with reliability indices below the threshold are selected. In the example above, the reliability index of the on-load tap changer is 0.68, which is lower than the threshold of 0.70, requiring further analysis.

[0096] For components with reliability indicators below the threshold, traceable condition monitoring data is segmented over time, with segmentation based on key events such as changes in operating conditions, load level adjustments, or maintenance activities. For example, five years of monitoring data for on-load tap changers can be divided into 10 time periods, each approximately six months long.

[0097] Calculate the state characteristic values ​​for each time period, such as the switching time of the on-load tap changer, contact resistance, and dissolved gas content in the oil. For example, in the first time period: switching time 50ms, contact resistance 300μΩ, acetylene content in the oil 5ppm; in the fifth time period: switching time 65ms, contact resistance 450μΩ, acetylene content in the oil 28ppm.

[0098] Establish the correspondence between state characteristic values ​​and equipment performance, and construct a mapping model between characteristic values ​​and performance indicators through methods such as regression analysis or neural networks. For example, establish a regression model between switching time, contact resistance and voltage regulator performance score, where performance score = (100 - switching time / 100 × 30) × 0.5 + (100 - contact resistance / 600 × 40) × 0.5.

[0099] Identify state change points and performance degradation inflection points. A state change point is the moment when the state characteristic value changes significantly, while a performance degradation inflection point is the moment when the rate of performance decline changes significantly. For example, analysis revealed that at the end of time period 4, the acetylene content in the oil of the on-load regulator suddenly increased from 12 ppm to 28 ppm, which was a state change point; and in time period 6, the rate of performance score decline increased from 0.5% per month to 1.2%, which was a performance degradation inflection point.

[0100] The state transition time is determined based on the state abrupt change point. The state transition time is the point in time when the equipment changes from a normal state to an abnormal state, and it is usually close to the state abrupt change point. For example, the state transition time of an on-load tap changer is determined to be after 2 years and 1 month of operation.

[0101] Grey relational analysis was used to calculate the correlation between each influencing factor and the performance degradation inflection point. Influencing factors included operating environment conditions, load level, and operating frequency. First, the correlation coefficient between the reference sequence (performance indicators) and each comparison sequence (influencing factor) was calculated, and then the average correlation degree was calculated. For example, the analysis results showed that the correlation degree between voltage regulation operating frequency and the performance degradation inflection point was 0.92, the correlation degree between ambient temperature was 0.75, and the correlation degree between load level was 0.88.

[0102] The time point and influencing factors of reliability degradation are determined based on the state transition time and correlation degree. The time point is usually selected as the earlier of the state transition time or the performance degradation inflection point, and the influencing factors are sorted from high to low correlation degree. For example, the time point for reliability degradation of the on-load tap changer is determined to be 2 years and 1 month of operation, and the main influencing factors are the frequency of tap changer operation (correlation degree 0.92) and the load level (correlation degree 0.88).

[0103] Through the above analysis process, not only were components with low reliability identified, but the specific time and main influencing factors of reliability degradation were also determined. This provides a scientific basis for predictive maintenance and operation strategy optimization of transformers, effectively reducing the risk of failure and extending the service life of equipment.

[0104] This invention relates to a transformer operation and maintenance management system based on digital twins, the system comprising: The first unit is used to integrate multi-source data from transformer sampling data and combine it with distributed blockchain technology to obtain traceable status monitoring data. The second unit is used to adaptively decompose the traceable status monitoring data according to frequency bands, extract the time-varying features of each frequency band data, introduce a cross-frequency domain feature fusion device based on mutual information maximization, extract the correlation between the time-varying features of different frequency bands, and obtain the transformer status index matrix. The third unit is used to construct the state evolution diagram of transformer components. The state index matrix is ​​mapped to the graph structure as nodes. Dynamic causal reasoning is used to analyze the state propagation between nodes to obtain the state propagation characteristics between components. At the same time, feature representations at different time scales are dynamically extracted to obtain multi-scale time series features. The state propagation characteristics between components and the multi-scale time series features are coupled in a hierarchical interactive manner to obtain the spatiotemporal joint analysis results of the transformer state. The fourth unit is used to calculate the probability distribution of the transformer's health status based on the spatiotemporal joint analysis results, calculate the reliability index of each component of the transformer based on the probability distribution, screen the components whose reliability index is lower than the preset reliability threshold, trace and analyze the historical operating status in conjunction with the traceable status monitoring data, determine the time node and influencing factors of reliability degradation, and formulate corresponding maintenance strategies.

[0105] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0106] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0107] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A transformer operation and maintenance management method based on digital twins, characterized in that, include: Multi-source data integration of transformer sampling data, combined with distributed blockchain technology, yields traceable status monitoring data; The traceable status monitoring data is adaptively decomposed by frequency band, and the time-varying features of each frequency band are extracted. A cross-frequency domain feature fusion fusion based on mutual information maximization is introduced to extract the correlation between the time-varying features of different frequency bands, and the status index matrix of the transformer is obtained. A state evolution diagram of transformer components is constructed, and the state index matrix is ​​mapped to the graph structure as nodes. Dynamic causal reasoning is used to analyze the state propagation between nodes to obtain the state propagation characteristics between components. At the same time, feature representations at different time scales are dynamically extracted to obtain multi-scale time series features. The state propagation characteristics between components and the multi-scale time series features are coupled in a hierarchical interactive manner to obtain the spatiotemporal joint analysis results of the transformer state. Based on the spatiotemporal joint analysis results, the probability distribution of the transformer's health status is calculated. Based on the probability distribution, the reliability index of each component of the transformer is calculated. Components with reliability indexes lower than a preset reliability threshold are screened. The historical operating status is traced and analyzed in conjunction with the traceable status monitoring data to determine the time nodes and influencing factors of reliability degradation and to formulate corresponding maintenance strategies.

2. The method according to claim 1, characterized in that, Multi-source data integration of transformer sampling data, combined with distributed blockchain technology, yields traceable status monitoring data, including: A blockchain network node is constructed. The transformer sampling data is divided into multiple feature sub-matrices according to a preset time window. Based on the feature sub-matrices, the response latency, consensus accuracy, and data verification pass rate of each blockchain network node in processing the feature sub-matrices are calculated and combined to obtain a node credibility score. Blockchain network nodes whose node credibility scores exceed a preset credibility threshold are selected to obtain a set of trusted nodes. The set of trusted nodes receives verification requests from other blockchain network nodes and returns confirmation messages based on the Byzantine fault tolerance mechanism. When more than a preset number of blockchain network nodes in the set of trusted nodes return confirmation messages for the same feature sub-matrices, the current feature sub-matrices are marked as verified. The equipment identification information of the transformer is collected. The feature submatrix marked as verified is concatenated with the corresponding time window information and equipment identification information, and then hashed to obtain an index key value. Based on the index key value, the feature submatrix, the time window information, and the equipment identification information are packaged and stored in the blockchain network to generate traceable status monitoring data.

3. The method according to claim 1, characterized in that, The traceable status monitoring data is adaptively decomposed by frequency band, and the time-varying features of each frequency band are extracted. A cross-frequency domain feature fusion fusion device based on maximizing mutual information is introduced to extract the correlation between the time-varying features of different frequency bands, resulting in the transformer status index matrix, which includes: The extreme points of the traceable status monitoring data are calculated, and the envelope is constructed using the cubic spline interpolation method for the extreme points. The difference between the traceable status monitoring data and the envelope is used as the frequency band component. The construction process is iteratively executed until the envelope is less than the preset residual ratio threshold, and multiple frequency band components of different frequency bands are obtained. For each frequency band component, a phase analysis transform is performed to extract time-varying phase features and time-varying amplitude features. Time-varying frequency features are calculated based on the time-varying phase features, and energy density features are calculated based on the time-varying amplitude features. The time-varying frequency features and energy density features are combined to form the time-varying features corresponding to each frequency band. Calculate the joint probability distribution and marginal probability distribution of different frequency bands. Based on the logarithmic ratio of the joint probability distribution and the marginal probability distribution, obtain the mutual information value between the time-varying features. Use the gradient ascent method to iteratively optimize the mutual information value until convergence to obtain the maximum mutual information value. Construct feature fusion weights based on the maximum mutual information value. Perform interactive operations between the feature fusion weights and the time-varying features of the corresponding frequency bands and arrange them according to the frequency band correlation to construct the state index matrix.

4. The method according to claim 1, characterized in that, Dynamic causal reasoning is used to analyze the state propagation between nodes to obtain the state propagation characteristics between components. Simultaneously, feature representations at different time scales are dynamically extracted to obtain multi-scale temporal features, including: Calculate the partial correlation coefficient and Granger causality test statistic between node pairs. When both the partial correlation coefficient and the Granger causality test statistic are greater than the corresponding preset significance threshold, it is determined that there is a state propagation relationship between the node pairs. Based on the state propagation relationship, construct a component state propagation network. In the component state propagation network, calculate the propagation centrality by the proportion of the shortest path between nodes, and calculate the propagation efficiency by the reciprocal of the shortest distance between nodes. Combine the propagation centrality and the propagation efficiency to obtain the state propagation characteristics between components. An adaptive gating mechanism including a reset gate and an update gate is constructed. The weight matrix of the reset gate is obtained through adaptive learning of the temporal correlation of the state propagation features between the components. The gating parameters of the update gate are dynamically adjusted based on the changing trend of the state propagation features between the components. Features at multiple time scales are captured by setting different time window lengths. Candidate features are obtained by selectively forgetting the historical features at each time scale according to the weight matrix of the reset gate. The candidate features and the current features are adaptively fused with the gating parameters of the update gate to obtain the multi-scale temporal features.

5. The method according to claim 1, characterized in that, Dynamic causal reasoning is used to analyze the state propagation between nodes to obtain the state propagation characteristics between components. Simultaneously, feature representations at different time scales are dynamically extracted to obtain multi-scale temporal features, including: Calculate the partial correlation coefficient and Granger causality test statistic between node pairs. When both the partial correlation coefficient and the Granger causality test statistic are greater than the corresponding preset significance threshold, it is determined that there is a state propagation relationship between the node pairs. Based on the state propagation relationship, construct a component state propagation network. In the component state propagation network, calculate the propagation centrality by the proportion of the shortest path between nodes, and calculate the propagation efficiency by the reciprocal of the shortest distance between nodes. Combine the propagation centrality and the propagation efficiency to obtain the state propagation characteristics between components. An adaptive gating mechanism including a reset gate and an update gate is constructed. The weight matrix of the reset gate is obtained through adaptive learning of the temporal correlation of the state propagation features between the components. The gating parameters of the update gate are dynamically adjusted based on the changing trend of the state propagation features between the components. Features at multiple time scales are captured by setting different time window lengths. Candidate features are obtained by selectively forgetting the historical features at each time scale according to the weight matrix of the reset gate. The candidate features and the current features are adaptively fused with the gating parameters of the update gate to obtain the multi-scale temporal features.

6. The method according to claim 1, characterized in that, By hierarchically and interactively coupling the inter-component state propagation characteristics with the multi-scale temporal characteristics, the spatiotemporal joint analysis results of the transformer state are obtained, including: The inter-component state propagation characteristics and the multi-scale temporal characteristics are decomposed in the frequency domain. Based on a preset frequency threshold, the frequency domain decomposition results are divided into slowly varying feature components and rapidly varying feature components. Calculate the time-delay correlation sequence between the slowly varying feature components, construct a slowly varying domain coupling strength matrix based on the peak position and peak amplitude of the time-delay correlation sequence, and perform convolution operation to obtain the slowly varying domain interaction features; perform wavelet transform on the rapidly varying feature components to obtain state propagation domain wavelet coefficients and time-series domain wavelet coefficients, calculate the scale-wise Hadamard product of the state propagation domain wavelet coefficients and the time-series domain wavelet coefficients, and perform wavelet reconstruction to obtain the rapidly varying domain interaction features; Fourier transforms are performed on the slowly varying domain interaction features and the rapidly varying domain interaction features respectively to obtain the slowly varying domain spectral features and the rapidly varying domain spectral features; the slowly varying domain spectral features and the rapidly varying domain spectral features are dynamically fused in the frequency domain to obtain the fused spectral features; the fused spectral features, the inter-component state propagation features, and the multi-scale time series features are residually connected to obtain the spatiotemporal joint analysis results of the transformer state.

7. The method according to claim 1, characterized in that, Based on the probability distribution, the reliability index of each component of the transformer is calculated. Components with reliability indices lower than a preset reliability threshold are screened. Historical operating status is traced and analyzed using the traceable status monitoring data to determine the time points and influencing factors of reliability degradation, including: The failure frequency and operating time of each component of the transformer are obtained from the probability distribution. The failure rate per unit time is calculated and combined with the component importance weight to obtain the failure index. The historical and current values ​​of the transformer performance parameters are obtained, the degradation rate of the performance parameters is calculated, and the performance index is obtained. The failure probability of each component under different state levels is calculated and the fitness value of each component is calculated in combination with the importance weight of the component. The non-dominated ranking is performed by comparing the failure index and the performance index of each component. The congestion degree is obtained by calculating the Euclidean distance between each component in the same non-dominated layer. Based on the fitness value and the congestion degree, selection and crossover operations are performed to obtain the optimal solution set. The reliability index of each component is selected from the optimal solution set. For components whose reliability index is lower than the preset reliability threshold, the traceable status monitoring data is segmented into time series, the status characteristic values ​​in each time period are calculated, the correspondence between the status characteristic values ​​and the equipment performance is established, the status change points and performance degradation inflection points are identified, the status transition time is determined based on the status change points, the correlation degree between each influencing factor and the performance degradation inflection point is calculated using the grey relational analysis method, and the time node and influencing factors of reliability degradation are determined according to the status transition time and the correlation degree.

8. A transformer operation and maintenance management system based on digital twins, used to implement the method as described in any one of claims 1-7, characterized in that, include: The first unit is used to integrate multi-source data from transformer sampling data and combine it with distributed blockchain technology to obtain traceable status monitoring data. The second unit is used to adaptively decompose the traceable status monitoring data according to frequency bands, extract the time-varying features of each frequency band data, introduce a cross-frequency domain feature fusion device based on mutual information maximization, extract the correlation between the time-varying features of different frequency bands, and obtain the transformer status index matrix. The third unit is used to construct the state evolution diagram of transformer components. The state index matrix is ​​mapped to the graph structure as nodes. Dynamic causal reasoning is used to analyze the state propagation between nodes to obtain the state propagation characteristics between components. At the same time, feature representations at different time scales are dynamically extracted to obtain multi-scale time series features. The state propagation characteristics between components and the multi-scale time series features are coupled in a hierarchical interactive manner to obtain the spatiotemporal joint analysis results of the transformer state. The fourth unit is used to calculate the probability distribution of the transformer's health status based on the spatiotemporal joint analysis results, calculate the reliability index of each component of the transformer based on the probability distribution, screen the components whose reliability index is lower than the preset reliability threshold, trace and analyze the historical operating status in conjunction with the traceable status monitoring data, determine the time node and influencing factors of reliability degradation, and formulate corresponding maintenance strategies.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images