Predictive maintenance system for transformer

Abstract

The invention belongs to the technical field of transformer maintenance, and discloses a predictive maintenance system for a transformer. Comprising the following steps: acquiring operation data of a transformer, and preprocessing the operation data to obtain operation characteristic data; clustering the operation characteristic data to obtain n operation state clusters; constructing a state directed graph by taking the n running state clusters as nodes; solving the shortest path probability from each node to all other nodes in the state directed graph; and constructing a constraint condition and an optimization function based on the state directed graph and the shortest path probability obtained by solving, and solving the optimization function by using an empire competition algorithm based on the constraint condition to obtain an optimal maintenance strategy of the transformer, so that the potential fault risk can be predicted, the distribution of maintenance resources can be optimized, and the maintenance efficiency is improved. The condition of unnecessary excessive maintenance or insufficient maintenance is avoided; and the scientificity and economy of transformer management are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer maintenance, and more particularly to a transformer predictive maintenance system. Background Art

[0002] The patent application with the publication number CN106643880A discloses a transformer early warning method for an operation and maintenance system, which solves the technical problem of identifying abnormal operating conditions of transformers. The method calculates the temperature alarm coefficient and current difference ratio of the transformer based on the rated parameters of the transformer, the transformer operating condition data measured by the operation and maintenance system, and the transformer operating condition data measured by manual inspection, and then calculates the transformer alarm value. The transformer's working status is judged based on the transformer alarm value, thereby timely identifying abnormal operating conditions of the transformer and abnormal data collection quality, which can help maintenance personnel discover equipment defects in a timely manner.

[0003] However, traditional maintenance strategies are often based on fixed time intervals or simple threshold triggers, and lack comprehensive consideration of the actual operating status of the transformer, which may lead to maintenance work being carried out too early or too late, resulting in waste of resources or equipment damage; secondly, in terms of data, data processing affects the accuracy of feature extraction, and the existing processing methods lead to deviations in the assessment of transformer status; thirdly, traditional methods cannot accurately reflect the complex operating characteristics of transformers under different operating conditions; in addition, it is difficult to accurately characterize the relationship between different operating states of the transformer, resulting in inaccurate predictions of future states; finally, there is a lack of optimization methods that comprehensively consider the operating status, maintenance costs and long-term benefits of the transformer, making it difficult for maintenance decisions to achieve the best balance between costs and benefits; these problems may lead to unnecessary power outages, shortened equipment life, increased maintenance costs, and even serious power accidents in actual operations.

[0004] In view of this, the present invention proposes a transformer predictive maintenance system to solve the above problems. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention provides the following technical solutions: A transformer predictive maintenance system, comprising: a data acquisition and processing module, for collecting transformer operation data, pre-processing the operation data, and obtaining operation characteristic data; A clustering module is used to cluster the operation characteristic data to obtain n operation status clusters; The preliminary fitting module is used to construct a state directed graph using n running state clusters as nodes; solve the shortest path probability from each node to all other nodes in the state directed graph; The strategy maintenance module constructs constraints and optimization functions based on the state directed graph and the shortest path probability obtained by solving the problem. Based on the constraints, the optimization function is solved using the Empire competition algorithm to obtain the optimal maintenance strategy for the transformer. Each module is connected by wire and / or wireless means.

[0006] Further, the operation data includes a temperature data sequence, a load data sequence and an insulating oil quality data sequence; The method of preprocessing the operating data includes: Standardizing or normalizing the operating data to obtain standard operating data, and removing trend items from the standard operating data to obtain preliminary processed data; Decompose the preliminary processed data to obtain m intrinsic modal components and residual sequences ; For The integrated energy of the intrinsic modal components is calculated , ;in, is the first intrinsic modal component The residual value, For the The weight of the residual value, is a real or complex number; is a real number parameter; is a positive real number parameter; Calculate the The proportion of the combined energy of an intrinsic modal component to the sum of the combined energies of all intrinsic modal components ; Preset importance threshold, if If it is greater than or equal to the importance threshold, Intrinsic modal components are retained if If the value is less than the importance threshold, The intrinsic modal components are removed, and all the retained intrinsic modal components constitute the effective component set; The intrinsic modal components in the effective component set are added and reconstructed to obtain a new sequence ; From the new sequence Extract running characteristic data.

[0007] Furthermore, the method of performing the trend item removal operation includes: Defines the window size of the moving window , and for the window size , set the adaptive window size adjustment policy; the setting methods of the adaptive window size adjustment policy include: Define the evaluation indicator function ; ;in, Indicates The high-frequency wavelet coefficients of the first-order wavelet decomposition, Indicates The low-frequency approximate component coefficients of the first-order wavelet decomposition, For the Data points The probability density of is the mean of the data points in the current moving window, is the total number of data points in the current moving window; is the standard deviation of the data points in the current moving window; , and is the weight parameter of the corresponding item; Calculate the value of the evaluation index function of the data point in the current moving window, recorded as the evaluation value ; Preset adjustment high threshold and adjust the high threshold Decrease the window size by the specified amount ; Preset adjustment low threshold and adjust the low threshold Increase the window size by the specified amount ; like Greater than , then the window size Reduce ;like Less than , then the window size Increase ; Otherwise, keep the window size constant; For time point , calculated with Centered on arrive , take the mean of the data points in the sequence as a small sequence of trend items , repeatedly calculate the trend item small sequence for the entire sequence, and splice it to get the moving average sequence ; Based on the original sequence and the moving average series , calculate the residual sequence after detrending ;in, For the original sequence The weight of is a moving average series weights; all residual sequences after detrending are the preliminary processed data.

[0008] Furthermore, the method of decomposing the preliminary processed data includes: Initially, the original sequence Assign to the residual sequence ; For the residual sequence Perform iterative processing to extract an intrinsic modal component ; The iterative processing method includes: finding All local maximum points and local minimum points are used to fit the upper envelope through the local maximum points. ; Fit the lower envelope through the local minimum point ; Calculate the upper envelope and lower envelope The mean ; From the residual sequence Remove the mean ; Repeat until Satisfy the intrinsic modal constraints; Denoted as intrinsic modal component; The intrinsic modal constraints include: The number of local maximum points and local minimum points is equal or the difference does not exceed 1, and The mean of the upper and lower envelopes of is close to 0; The extracted intrinsic modal components are extracted from the original residual sequence Remove it, that is, complete the residual sequence The updated residual sequence Repeat the acquisition of intrinsic modal components until Energy Less than a preset energy threshold; The energy calculation formula is: ;in, is the first Residual values.

[0009] Further, the new sequence The methods for extracting running feature data include: will new sequence Transform and obtain preliminary coefficients. The transformation formula is: ;in, is the preliminary coefficient, is the scale factor, is the translation factor, is the time factor index, is the wavelet basis function; is the time parameter; Calculate the preliminary energy based on the preliminary coefficients , based on the initial energy Calculate the initial entropy ; extract Along scale factor The maximum value line of the wavelet is taken as the ridge line, and the slope of the wavelet ridge line in the logarithmic coordinate is extracted as the slope; the running characteristic data include preliminary energy, preliminary entropy, ridge line and slope.

[0010] Furthermore, the method of clustering the operating characteristic data includes: The extracted operation feature data is constructed into a multi-dimensional feature space, and the feature space is divided into M1 rectangular cells; the rectangular cells are divided in the following ways: Determine the boundary of the feature space, that is, multiply the interval of the maximum and minimum values ​​of each dimension of the running feature data; take the entire feature space as the root node, and use a dimensional hyperrectangular representation, divide the nodes based on the root node, and for each node, calculate its variance ;in, For the The variance factor of each dimension; and Respectively The maximum and minimum values ​​of the dimensions; The partition threshold TH is preset. If the variance of a node is greater than the partition threshold TH, the node is recursively partitioned. The direction of the recursive partition is the node with the largest variance. dimension, move the node along the The midpoint of the interval of each dimension is divided into two nodes; this is repeated until the variance of all nodes is less than the division threshold TH, that is, M1 rectangular cells are constructed.

[0011] For each rectangular cell, calculate the statistical information of the cluster points inside it, including the number, mean and variance of data points; Traverse each rectangular cell and calculate the uniform parameter based on its statistical information. The uniform parameter is the weighted sum of the number, mean and variance of data points. Define a uniform threshold, take the rectangular cells whose uniform parameters are greater than the uniform threshold as cells to be segmented, further divide the cells to be segmented until all rectangular cells do not need to be further divided, and take each rectangular cell as a running state cluster; obtain n running state clusters; The further division method includes: For the first Cluster points , calculate its density, the density calculation formula is: ,in, is the total number of cluster points in the cell to be segmented, is the kernel function, is the bandwidth parameter, For the Cluster points, and Do not index the same cluster point at the same time. For the Cluster points density; Define the initial neighborhood radius and A cluster point builds a neighborhood based on the neighborhood radius, and the density of the cluster points in its neighborhood is summed to obtain the total density of the neighborhood; a density threshold is preset, and the neighborhood radius is continuously increased. When the calculated total neighborhood density is greater than the density threshold, the neighborhood constructed by the neighborhood radius at this time is recorded as a high-density area. For each high-density area, it is segmented out as a rectangular cell to complete further division.

[0012] Furthermore, the method of constructing the state directed graph includes: Each running state cluster is regarded as a corresponding node in the state directed graph, with a total of n nodes; for any two running state clusters and , calculate the similarity between them ; ;in, is the attenuation parameter; Running state cluster The mean vector of Running state cluster The mean vector of is the weight coefficient of the dispersion term, Running state cluster The dispersion of Running state cluster The dispersion of for and The Euclidean distance between ;in, is the weight adjustment parameter; Running state cluster The total number of points in the cluster, for The density of Running state cluster The cluster points within To disperse the weight adjustment parameters, Running state cluster The variance of internal cluster points; The calculation method and same; Preset similarity threshold ,like , then add a directed edge to the state directed graph , and set the directed edge The weight of , that is, the construction of the state directed graph is completed.

[0013] Furthermore, the shortest path probability is solved in the following manner: Create an empty minimum spanning tree MST, add any node to the minimum spanning tree MST, and for the remaining nodes, calculate their shortest distances to the nodes in the minimum spanning tree MST and record them; From the remaining nodes, select the node closest to the minimum spanning tree MST, add it to the minimum spanning tree MST, and update the shortest distance from the remaining nodes to the minimum spanning tree MST, repeat until all nodes are added to the minimum spanning tree MST; For each pair of nodes , find their nearest common ancestor node in the minimum spanning tree MST , then the node To Node The shortest path length ;in, Indicates that in the minimum spanning tree MST, from node To Node The shortest path length, Indicates that in the minimum spanning tree MST, from node To Node The shortest path length; Based on the shortest path length Calculate the node To Node The shortest path probability ,in, is the regulating factor; is the path attenuation parameter, is the smoothing parameter.

[0014] Furthermore, the formula of the optimization function is: ;in, For Node Determine a corresponding recommended maintenance strategy, For Node Determine a corresponding recommended maintenance strategy, Recommend maintenance strategies for execution The cost required, Recommend maintenance strategies for execution the costs required; For slave nodes Actual operation is transferred to the node time; For slave nodes Transfer to Node The probability of For operating losses; The constraints include a first constraint and a second constraint; the first constraint is that the out-degree probability of each node is 1, and the in-degree probability of each node is 1; the second constraint is that the probability of transferring from any node to other nodes is non-negative, and the shortest path probability is used as the upper limit of the probability of any node transferring to other nodes.

[0015] Furthermore, the method for solving the optimization function includes: Initialize a population, where each individual in the population represents a solution vector, i.e. the probability of transfer between nodes; randomly generate an initial population that satisfies the constraints; For each individual, calculate the value of its optimization function, recorded as the allocation degree; define the empire according to the individual's allocation degree, select the individual with the highest allocation degree as the emperor of the first empire, and calculate the difference between the allocation degrees of other individuals and the emperor. In the order of the difference in allocation degrees from small to large, allocate other individuals to the empire in turn until the power of the empire reaches the preset maximum power; repeat until all individuals are allocated to different empires; Define the weak threshold. Within each empire, treat individuals with a distribution degree less than the weak threshold as weak countries. Perform assimilation operations on weak countries to gradually make them closer to the emperor of the empire. The formula for assimilation operations is: ;in, A new position for small and weak countries, For the current position of weak countries, For the position of the emperor of the empire, is a random number in the interval [0, 1]. is a uniform random number in the interval [0.1, 0.9], for The distribution degree, for The distribution degree, is the difference adjustment parameter, is a positive number; is a random number that follows a standard normal distribution; Define a border distance threshold, calculate the distance between all individuals in the empire and the emperor, and mark individuals whose distance is greater than the border distance threshold as being at the border of the empire; For individuals marked as being at the border of the empire, a random number in the interval [0, 1] is generated according to the predefined revolution rate; if the random number is less than the revolution rate, the position of the individual is randomly initialized again; After the assimilation operation and the revolution, the sum of the distribution degrees of all individuals in the empire is calculated as the strength of the empire. In order of the strength of the empires from large to small, the smallest empire is merged into other empires; repeat until the preset maximum number of iterations is reached, and the individual with the highest distribution degree in all empires at this time is obtained as the optimal solution. Based on the optimal solution, the transition probability matrix between each node is obtained; Based on the obtained transition probability matrix and the maintenance cost corresponding to each node, the long-term total maintenance cost is calculated; different combinations of recommended maintenance strategies are enumerated, and the recommended maintenance strategy combination that can minimize the long-term total maintenance cost is found, which is the optimal maintenance strategy.

[0016] The technical effects and advantages of a transformer predictive maintenance system of the present invention are as follows: The present invention can effectively remove noise and trend items from original data, improve the accuracy of feature extraction, and use a method combining recursive partitioning and density clustering to achieve a fine division of transformer operating states. It can also accurately characterize the transfer relationship between different operating states by constructing a state directed graph and calculating the shortest path probability. This can comprehensively analyze the actual operating state of the transformer, realize the intelligence and refinement of the maintenance strategy, and greatly improve the pertinence and efficiency of maintenance. It can effectively reduce long-term maintenance costs, extend the service life of the transformer, and reduce the occurrence of unexpected failures, thereby improving the overall reliability and stability of the power system. In addition, its adaptive characteristics enable it to automatically adjust the maintenance strategy according to changes in the transformer operating environment to ensure that the optimal maintenance effect is always maintained. It can not only predict potential failure risks, but also optimize the allocation of maintenance resources to avoid unnecessary over-maintenance or under-maintenance, thereby improving the scientificity and economy of transformer management. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of a transformer predictive maintenance system of the present invention; Figure 2 It is a schematic diagram of a transformer predictive maintenance method of the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] Embodiment 1: See also Figure 1 As shown, a transformer predictive maintenance system described in this embodiment includes: The data acquisition and processing module is used to collect the operation data of the transformer, pre-process the operation data, and obtain the operation characteristic data; A clustering module is used to cluster the operation characteristic data to obtain n operation status clusters; The preliminary fitting module is used to construct a state directed graph using n running state clusters as nodes; solve the shortest path probability from each node to all other nodes in the state directed graph; The strategy maintenance module constructs constraints and optimization functions based on the state directed graph and the shortest path probability obtained by solving it. Based on the constraints, the optimization function is solved using the imperial competition algorithm to obtain the optimal maintenance strategy for the transformer. Each module is connected by wire and / or wireless means to achieve data transmission between modules.

[0020] Furthermore, the operating data includes a temperature data sequence, a load data sequence and an insulating oil quality data sequence; a temperature sensor is usually installed inside the transformer, and the temperature sensor samples and records the temperature value at a preset time interval to form a temperature data sequence, and a current transformer and a voltage transformer are used to detect the current and voltage, and the real-time load of the transformer is calculated based on the current and voltage, and the real-time load is sampled and recorded at a preset time interval to form a load data sequence.

[0021] Online monitoring sensors are installed in the insulating oil storage tank or oil circulation system of the transformer. These sensors detect the dielectric strength, acid value and particle content of the insulating oil in real time, and weighted calculations are performed on the dielectric strength, acid value and particle content to obtain the insulating oil quality data. The insulating oil quality data is sampled and recorded at preset time intervals to form an insulating oil quality data sequence.

[0022] Furthermore, the method of preprocessing the operating data includes: The operating data is standardized or normalized to have a zero mean, thereby obtaining standard operating data. The standard operating data is subjected to a trend item removal operation to obtain preliminary processed data.

[0023] Specifically, define the window size of the moving window (Initially set to 5%-25% of the sequence length), and for the window size , set the adaptive window size adjustment policy; the setting methods of the adaptive window size adjustment policy include: Define the evaluation indicator function ; ;in, Indicates High-frequency wavelet coefficients of the first-order wavelet decomposition (decompose the data in the current moving window by wavelet to obtain wavelet coefficients of different scales), Indicates The low-frequency approximate component coefficients of the first-order wavelet decomposition, For the Data points The probability density of is the mean of the data points in the current moving window, is the total number of data points in the current moving window; is the standard deviation of the data points in the current moving window; , and is the weight parameter of the corresponding item, and the weight parameter can be adaptively adjusted through methods such as Bayesian model and online learning.

[0024] It should be noted that the data value at each time point in the temperature data sequence, the load data sequence and the insulating oil quality data sequence is taken as a data point.

[0025] Calculate the value of the evaluation index function of the data point in the current moving window, recorded as the evaluation value ; Preset adjustment high threshold and adjust the high threshold Decrease the window size by the specified amount ; Preset adjustment low threshold and adjust the low threshold Increase the window size by the specified amount .

[0026] like Greater than , then the window size Reduce ;like Less than , then the window size Increase ; Otherwise, keep the window size constant.

[0027] For time point , calculated with Centered on arrive , take the mean of the data points in the sequence as a small sequence of trend items , repeatedly calculate the trend item small sequence for the entire sequence (temperature data sequence, load data sequence and insulating oil quality data sequence), and splice them to obtain the moving average sequence (trend item) ; It should be explained that the calculations are all for the corresponding sequences, for example, the temperature data sequence is calculated for itself.

[0028] Based on the original sequence and the moving average series , calculate the residual sequence after detrending ;in, For the original sequence The weight of is a moving average series The weight of .

[0029] All the residual series after detrending are the preliminary processed data; The residual series after detrending (preliminary processing data) Decompose to obtain m intrinsic modal components and residual sequences .

[0030] Specifically, initially, the original sequence Assign to the residual sequence ; For the residual sequence Perform iterative processing to extract an intrinsic modal component .

[0031] The iterative processing method includes: finding All local maximum points and local minimum points are used to fit the upper envelope through the local maximum points. ; Fit the lower envelope through the local minimum point Specifically, a cubic spline curve is used for interpolation fitting to obtain an upper envelope and a lower envelope.

[0032] Calculate the upper envelope and lower envelope The mean ; From the residual sequence Remove the mean ; Repeat until Satisfy the intrinsic modal constraints; are denoted as intrinsic modal components.

[0033] The intrinsic modal constraints include: The number of local maximum points and local minimum points is equal or the difference does not exceed 1, and The mean of the upper and lower envelopes of is close to 0; The extracted intrinsic modal components are extracted from the original residual sequence Remove it, that is, complete the residual sequence The updated residual sequence Repeatedly obtain the intrinsic mode components (also part of the residual sequence) until Energy Less than the preset energy threshold.

[0034] The energy calculation formula is: ;in, is the first Residual values.

[0035] Through the above iterative process, the original sequence is decomposed into m intrinsic modal components and the remaining residual sequence.

[0036] For The integrated energy of the intrinsic modal components is calculated , ;in, is the first intrinsic modal component The residual value, For the The weight of the residual value, is a real number or a complex number, thereby introducing a phase factor, making the energy calculation result have vector characteristics, which can reflect the oscillation and fluctuation characteristics of the residual; is a real number parameter used to control the curvature of the logarithmic function; is a positive real number parameter that controls the decay rate.

[0037] Calculate the The proportion of the combined energy of an intrinsic modal component to the sum of the combined energies of all intrinsic modal components .

[0038] Preset importance threshold, if If it is greater than or equal to the importance threshold, The intrinsic modal components are retained if If the value is less than the importance threshold, The intrinsic modal components are removed, and all the retained intrinsic modal components constitute the effective component set.

[0039] The intrinsic modal components in the effective component set are added and reconstructed to obtain a new sequence ;new sequence The original sequence is retained The main characteristic components are obtained by removing some noise and irrelevant components.

[0040] From the new sequence Extract the running feature data from Transform and obtain preliminary coefficients. The transformation formula is: ;in, is the preliminary coefficient, is the scale factor, is the translation factor, is the time factor index, used to change the effect of time shift, is a wavelet basis function (such as Haar wavelet, Coiflet wavelet, etc.); It is a time parameter that controls the degree of expansion of the function on the time axis.

[0041] Calculate the preliminary energy based on the preliminary coefficients , based on the initial energy Calculate the initial entropy .

[0042] extract Along scale factor The maximum value line of the wavelet is taken as the ridge line, and the slope of the wavelet ridge line in the logarithmic coordinate is extracted as the slope; the running characteristic data include preliminary energy, preliminary entropy, ridge line and slope.

[0043] Further, the method of clustering the running feature data includes: The extracted operation feature data is constructed into a multidimensional feature space, and the feature space is divided into M1 rectangular cells. Specifically, the boundary of the feature space is determined, that is, the interval of the maximum and minimum values ​​of each dimension of the operation feature data (preliminary energy, preliminary entropy, ridge line and slope) is multiplied; the entire feature space is used as the root node, and a dimensional hyperrectangular representation, divide the nodes based on the root node, and for each node, calculate its variance ;in, For the The variance factor of the dimension is generally taken as ; and Respectively The maximum and minimum values ​​of the dimensions.

[0044] The partition threshold TH is preset. If the variance of a node is greater than the partition threshold TH, the node is recursively partitioned. The direction of the recursive partition is the node with the largest variance. dimension, move the node along the The midpoint of the interval of each dimension is divided into two nodes; this is repeated until the variance of all nodes is less than the division threshold TH, that is, M1 rectangular cells are constructed, the root node corresponds to the entire feature space, and the other nodes correspond to rectangular cells.

[0045] For each rectangular cell, calculate the statistical information of the cluster points inside it, including the number, mean and variance of data points.

[0046] Each cluster point corresponds to a twelve-dimensional vector, which is a combination of vector representations of the four operating characteristic data of preliminary energy, preliminary entropy, ridge line and slope corresponding to the temperature data sequence, load data sequence and insulating oil quality data sequence respectively.

[0047] Traverse each rectangular cell and calculate the uniform parameter based on its statistical information. The uniform parameter is the weighted sum of the number, mean and variance of data points.

[0048] A uniform threshold is defined, and rectangular cells whose uniform parameters are greater than the uniform threshold are used as cells to be segmented. The cells to be segmented are further divided until all rectangular cells do not need to be divided any more, and each rectangular cell corresponding to the cells is used as an operating state cluster; n operating state clusters are obtained; and each operating state cluster represents an operating state of the transformer.

[0049] The further division method includes: For the first Cluster points , calculate its density, the density calculation formula is: ,in, is the total number of cluster points in the cell to be segmented, is a kernel function (such as a Gaussian kernel), is the bandwidth parameter, For the Cluster points, and Do not index the same cluster point at the same time. For the Cluster points density.

[0050] Define the initial neighborhood radius and A cluster point builds a neighborhood based on the neighborhood radius, and the density of the cluster points in its neighborhood is summed to obtain the total density of the neighborhood; a density threshold is preset, and the neighborhood radius is continuously increased. When the calculated total neighborhood density is greater than the density threshold, the neighborhood constructed by the neighborhood radius at this time is recorded as a high-density area. For each high-density area, it is segmented out as a rectangular cell to complete further division.

[0051] Further, the methods of constructing a state directed graph include: Each operating state cluster is regarded as a corresponding node in the state directed graph, with a total of n nodes (each corresponding to an operating state of the transformer); for any two operating state clusters and , calculate the similarity between them .

[0052] ;in, The decay parameter that controls the similarity decay rate takes a positive value; Running state cluster The mean vector of (the mean of all cluster points in), Running state cluster The mean vector of is the weight coefficient of the dispersion item, which is used to adjust the influence of the degree of data dispersion on the similarity. Running state cluster The dispersion of Running state cluster The dispersion of for and The Euclidean distance between .

[0053] ;in, is the weight adjustment parameter, and its value is in the range of [0, 1]; Running state cluster The total number of points in the cluster, for The density (calculated by The density is calculated in the same way as Running state cluster The cluster points within To disperse the weight adjustment parameters, Running state cluster The variance of internal cluster points; The calculation method and same.

[0054] Preset similarity threshold ,like , then add a directed edge to the state directed graph , and set the directed edge The weight of .

[0055] Create an empty minimum spanning tree MST, add any node (such as node A) to the minimum spanning tree MST, and for the remaining nodes, calculate their shortest distance to the nodes in the minimum spanning tree MST (obtained by adding the weight of the directed edge to the calculation of the Euclidean distance) and record it.

[0056] From the remaining nodes, select the node closest to the minimum spanning tree MST (such as node B), add it to the minimum spanning tree MST, and update the shortest distance from the remaining nodes to the minimum spanning tree MST, and repeat until all nodes are added to the minimum spanning tree MST.

[0057] For each pair of nodes , find their nearest common ancestor node in the minimum spanning tree MST , then the node To Node The shortest path length ;in, Indicates that in the minimum spanning tree MST, from node To Node The shortest path length, Indicates that in the minimum spanning tree MST, from node To Node The shortest path length.

[0058] Based on the shortest path length Calculate the node To Node The shortest path probability ,in, It is an adjustment factor used to control the overall size of the probability; it is usually set to a positive value less than 1, such as 0.9; is the path attenuation parameter, which is used to control the influence of path length on probability. The larger it is, the greater the impact of path length on probability, and the faster the probability value decreases; is a smoothing parameter used to smooth the changes in probability values.

[0059] Furthermore, the formula of the optimization function is: ;in, For Node Determine a corresponding recommended maintenance strategy (including maintenance operations that need to be performed in the corresponding state and their costs), For Node Determine a corresponding recommended maintenance strategy, Recommend maintenance strategies for execution The cost required, Recommend maintenance strategies for execution the costs required; For slave nodes Actual operation is transferred to the node time; For slave nodes Transfer to Node probability; is the running loss (at node After running for a while, transfer to the node Some operational losses may occur due to equipment aging, decreased efficiency, etc.) Operational losses include losses due to temperature, losses due to load, and losses due to insulating oil quality.

[0060] The losses due to temperature are estimated using the Arrhenius model, the losses due to load are estimated using the exponential model, and the losses due to the quality of the insulating oil are estimated using the linear model; for example, the losses due to load ;in, and is an empirical constant, For the node Running time The average load within is the rated load of the transformer.

[0061] The constraints include a first constraint and a second constraint; the first constraint is that the out-degree probability of each node (the sum of the probabilities of transferring from the node to all other nodes) is 1, and the in-degree probability of each node (the sum of the probabilities of transferring from all other nodes to the node) is 1; the second constraint is that the probability of transferring from any node to other nodes is non-negative, and the shortest path probability is used as the upper limit of the probability of transferring from any node to other nodes.

[0062] Ways to solve the optimization function include: Initialize a population, each individual in the population represents a solution vector, that is, the probability of transfer between nodes; randomly generate an initial population that meets the constraints.

[0063] For each individual (solution vector), calculate the value of its optimization function, recorded as the distribution degree; define the empire according to the individual's distribution degree, select the individual with the highest distribution degree as the emperor of the first empire, and calculate the difference between the distribution degrees of other individuals and the emperor. In the order of the difference in distribution degrees from small to large, distribute the other individuals to the empire in turn until the power of the empire (the sum of the distribution degrees of all individuals in the empire) reaches the preset maximum power; repeat until all individuals are assigned to different empires.

[0064] Define the weak threshold. Within each empire, treat individuals with a distribution degree less than the weak threshold as weak countries. Perform assimilation operations on weak countries to gradually move them closer to the center of the empire (the emperor). The formula for assimilation operations is: ;in, is the new position of the weak country (solution vector), is the current position of the weak country (solution vector), For the position of the emperor of the empire, is a random number in the interval [0, 1]. is a uniform random number in the interval [0.1, 0.9], used to control the degree of convergence. for The distribution degree, for The distribution degree, is a difference adjustment parameter used to control the effect of fitness differences on the degree of convergence. is a small positive number that controls the intensity of random perturbations; is a random number that follows a standard normal distribution, used to introduce random disturbances; in this way, weak countries will gradually get closer to the emperor, but they will not completely overlap.

[0065] Define a border distance threshold, calculate the distance (Euclidean distance) between all individuals in the empire and the emperor, and mark individuals whose distance is greater than the border distance threshold as being at the border of the empire; for individuals (countries) marked as being at the border of the empire, give them the opportunity to revolutionize, that is, randomly reinitialize the position of the individual (country).

[0066] For each border country, a random number in the interval [0, 1] is generated according to a predefined revolution rate (such as 0.1); if the random number is less than the revolution rate, the position (solution vector) of the individual (country) is randomly reinitialized.

[0067] After the assimilation operation and the revolution, the sum of the distribution degrees of all individuals (countries) in the empire is calculated as the strength of the empire. The smallest empire is merged into other empires in the order of the strength of the empires from large to small. Repeat until the preset maximum number of iterations is reached, and the individual with the highest distribution degree in all empires at this time is obtained as the optimal solution. Based on the optimal solution, the transition probability matrix between each node (operating state cluster) is obtained. The transition probability matrix represents the probability of the transformer transferring from the current operating state to other operating states.

[0068] For each node (operating status cluster), a recommended maintenance strategy is determined based on the characteristics of its corresponding operating status and combined with historical data. The recommended maintenance strategy corresponding to each node is substituted into the maintenance cost in the optimization function, and the cost required to execute the recommended maintenance strategy at the node can be calculated.

[0069] Based on the obtained transition probability matrix and the maintenance cost corresponding to each node, the long-term total maintenance cost is calculated; different combinations of recommended maintenance strategies are enumerated, and the recommended maintenance strategy combination that can minimize the long-term total maintenance cost is found, which is the optimal maintenance strategy.

[0070] The optimal maintenance strategy includes which specific maintenance operations should be performed under each possible operating state and the corresponding maintenance schedule.

[0071] This embodiment can effectively remove noise and trend items from the original data, improve the accuracy of feature extraction, and use a combination of recursive partitioning and density clustering to achieve a fine division of the transformer operating state. By constructing a state directed graph and calculating the shortest path probability, the transfer relationship between different operating states is accurately characterized; thereby comprehensively analyzing the actual operating state of the transformer, realizing the intelligence and refinement of the maintenance strategy, and greatly improving the pertinence and efficiency of maintenance; it can effectively reduce long-term maintenance costs, extend the service life of the transformer, and reduce the occurrence of unexpected failures, thereby improving the overall reliability and stability of the power system; in addition, the adaptive characteristics enable it to automatically adjust the maintenance strategy according to changes in the transformer operating environment to ensure that the optimal maintenance effect is always maintained; it can not only predict potential failure risks, but also optimize the allocation of maintenance resources to avoid unnecessary over-maintenance or under-maintenance; and improve the scientificity and economy of transformer management.

[0072] Embodiment 2: See also Figure 2 As shown, the part not described in detail in this embodiment is described in Example 1, which provides a transformer predictive maintenance method, including: S1. Collecting the operating data of the transformer, preprocessing the operating data, and obtaining operating characteristic data; S2, clustering the operation characteristic data to obtain n operation status clusters; S3, construct a state directed graph using n running state clusters as nodes; solve the shortest path probability from each node to all other nodes in the state directed graph; S4. Based on the state directed graph and the shortest path probability obtained, constraints and optimization functions are constructed. Based on the constraints, the optimization function is solved using the Imperial competition algorithm to obtain the optimal maintenance strategy for the transformer.

[0073] Embodiment 3: This embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the operation mode of the transformer predictive maintenance method provided above is implemented.

[0074] Since the electronic device introduced in this embodiment is an electronic device used to implement a transformer predictive maintenance method in the embodiment of this application, based on the transformer predictive maintenance method introduced in the embodiment of this application, the technical personnel of this field can understand the specific implementation of the electronic device of this embodiment and its various variations, so how the electronic device implements the method in the embodiment of this application is not described in detail here. As long as the technical personnel of this field implement the electronic device used in the transformer predictive maintenance method in the embodiment of this application, it belongs to the scope of protection of this application.

[0075] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0076] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technical users in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A transformer predictive maintenance system, characterized in that: include: The data acquisition and processing module is used to collect the operation data of the transformer, pre-process the operation data, and obtain the operation characteristic data; A clustering module is used to cluster the operation characteristic data to obtain n operation status clusters; A preliminary fitting module is used to construct a state directed graph using n running state clusters as nodes; Solve the shortest path probability from each node to all other nodes in the state directed graph; The strategy maintenance module constructs constraints and optimization functions based on the state directed graph and the shortest path probability obtained by solving it, and solves the optimization function based on the constraints using the Empire competition algorithm to obtain the optimal maintenance strategy for the transformer; each module is connected by wire and / or wireless means.

2. A transformer predictive maintenance system according to claim 1, characterized in that: The operation data includes a temperature data sequence, a load data sequence and an insulating oil quality data sequence; The method of preprocessing the operating data includes: Standardizing or normalizing the operating data to obtain standard operating data, and removing trend items from the standard operating data to obtain preliminary processed data; Decompose the preliminary processed data to obtain m intrinsic modal components and residual sequences ; For The integrated energy of the intrinsic modal components is calculated , ;in, is the first intrinsic modal component The residual value, For the The weight of the residual value, is a real or complex number; is a real number parameter; is a positive real number parameter; Calculate the The proportion of the combined energy of an intrinsic modal component to the sum of the combined energies of all intrinsic modal components ; Preset importance threshold, if If it is greater than or equal to the importance threshold, Intrinsic modal components are retained if If the value is less than the importance threshold, The intrinsic modal components are removed, and all the retained intrinsic modal components constitute the effective component set; The intrinsic modal components in the effective component set are added and reconstructed to obtain a new sequence ; From the new sequence Extract running characteristic data.

3. A transformer predictive maintenance system according to claim 2, characterized in that: The method of removing the trend item includes: Defines the window size of the moving window , and for the window size , set the adaptive window size adjustment policy; the setting methods of the adaptive window size adjustment policy include: Define the evaluation indicator function ; ;in, Indicates The high-frequency wavelet coefficients of the first-order wavelet decomposition, Indicates The low-frequency approximate component coefficients of the first-order wavelet decomposition, For the Data points The probability density of is the mean of the data points in the current moving window, is the total number of data points in the current moving window; is the standard deviation of the data points in the current moving window; , and is the weight parameter of the corresponding item; Calculate the value of the evaluation index function of the data point in the current moving window, recorded as the evaluation value ; Preset adjustment high threshold and adjust the high threshold Decrease the window size by the specified amount ; Preset adjustment low threshold and adjust the low threshold Increase the window size by the specified amount ; like Greater than , then the window size Reduce ;like Less than , then the window size Increase ; Otherwise, keep the window size constant; For time point , calculated with Centered on arrive , take the mean of the data points in the sequence as a small sequence of trend items , repeatedly calculate the trend item small sequence for the entire sequence, and splice it to get the moving average sequence ; Based on the original sequence and the moving average series , calculate the residual sequence after detrending ;in, For the original sequence The weight of is a moving average series weights; all residual sequences after detrending are the preliminary processed data.

4. A transformer predictive maintenance system according to claim 3, characterized in that: The method of decomposing the preliminary processed data includes: Initially, the original sequence Assign to the residual sequence ; For the residual sequence Perform iterative processing to extract an intrinsic modal component ; The iterative processing method includes: finding All local maximum points and local minimum points are used to fit the upper envelope through the local maximum points. ; Fit the lower envelope through the local minimum point ; Calculate the upper envelope and lower envelope The mean ; From the residual sequence Remove the mean ; Repeat until Satisfy the intrinsic modal constraints; Denoted as intrinsic modal component; The intrinsic modal constraints include: The number of local maximum points and local minimum points is equal or the difference does not exceed 1, and The mean of the upper and lower envelopes of is close to 0; The extracted intrinsic modal components are extracted from the original residual sequence Remove it, that is, complete the residual sequence The updated residual sequence Repeat the acquisition of intrinsic modal components until Energy Less than a preset energy threshold; The energy calculation formula is: ;in, is the first Residual values.

5. A transformer predictive maintenance system according to claim 4, characterized in that: The new sequence The methods for extracting running feature data include: will new sequence Transform and obtain preliminary coefficients. The transformation formula is: ;in, is the preliminary coefficient, is the scale factor, is the translation factor, is the time factor index, is the wavelet basis function; is the time parameter; Calculate the preliminary energy based on the preliminary coefficients , based on the initial energy Calculate the initial entropy ; extract Along scale factor The maximum value line of the wavelet is taken as the ridge line, and the slope of the wavelet ridge line in the logarithmic coordinate is extracted as the slope; the running characteristic data include preliminary energy, preliminary entropy, ridge line and slope.

6. A transformer predictive maintenance system according to claim 5, characterized in that: The method of clustering the operating characteristic data includes: The extracted operation feature data is constructed into a multi-dimensional feature space, and the feature space is divided into M1 rectangular cells; the rectangular cells are divided in the following ways: Determine the boundary of the feature space, that is, multiply the interval of the maximum and minimum values ​​of each dimension of the running feature data; take the entire feature space as the root node, and use a dimensional hyperrectangular representation, divide the nodes based on the root node, and for each node, calculate its variance ;in, For the The variance factor of each dimension; and Respectively The maximum and minimum values ​​of the dimensions; The partition threshold TH is preset. If the variance of a node is greater than the partition threshold TH, the node is recursively partitioned. The direction of the recursive partition is the node with the largest variance. dimension, move the node along the The midpoint of the interval of the dimension is divided into two nodes; repeat until the variance of all nodes is less than the division threshold TH, that is, M1 rectangular cells are constructed; For each rectangular cell, calculate the statistical information of the cluster points inside it, including the number, mean and variance of data points; Traverse each rectangular cell and calculate the uniform parameter based on its statistical information. The uniform parameter is the weighted sum of the number, mean and variance of data points. Define a uniform threshold, take the rectangular cells whose uniform parameters are greater than the uniform threshold as cells to be segmented, further divide the cells to be segmented until all rectangular cells do not need to be further divided, and take each rectangular cell as a running state cluster; obtain n running state clusters; The further division method includes: For the first Cluster points , calculate its density, the density calculation formula is: ,in, is the total number of cluster points in the cell to be segmented, is the kernel function, is the bandwidth parameter, For the Cluster points, and Do not index the same cluster point at the same time. For the Cluster points density; Define the initial neighborhood radius and A cluster point builds a neighborhood based on the neighborhood radius, and the density of the cluster points in its neighborhood is summed to obtain the total density of the neighborhood; a density threshold is preset, and the neighborhood radius is continuously increased. When the calculated total neighborhood density is greater than the density threshold, the neighborhood constructed by the neighborhood radius at this time is recorded as a high-density area. For each high-density area, it is segmented out as a rectangular cell to complete further division.

7. A transformer predictive maintenance system according to claim 6, characterized in that: The method of constructing the state directed graph includes: Each running state cluster is regarded as a corresponding node in the state directed graph, with a total of n nodes; for any two running state clusters and , calculate the similarity between them ; ;in, is the attenuation parameter; Running state cluster The mean vector of Running state cluster The mean vector of is the weight coefficient of the dispersion term, Running state cluster The dispersion of Running state cluster The dispersion of for and The Euclidean distance between ;in, is the weight adjustment parameter; Running state cluster The total number of points in the cluster, for The density of Running state cluster The cluster points within To disperse the weight adjustment parameters, Running state cluster The variance of internal cluster points; The calculation method and same; Preset similarity threshold ,like , then add a directed edge to the state directed graph , and set the directed edge The weight of , that is, the construction of the state directed graph is completed.

8. A transformer predictive maintenance system according to claim 7, characterized in that: The method for solving the shortest path probability includes: Create an empty minimum spanning tree MST, add any node to the minimum spanning tree MST, and for the remaining nodes, calculate their shortest distances to the nodes in the minimum spanning tree MST and record them; From the remaining nodes, select the node closest to the minimum spanning tree MST, add it to the minimum spanning tree MST, and update the shortest distance from the remaining nodes to the minimum spanning tree MST, repeat until all nodes are added to the minimum spanning tree MST; For each pair of nodes , find their nearest common ancestor node in the minimum spanning tree MST , then the node To Node The shortest path length ;in, Indicates that in the minimum spanning tree MST, from node To Node The shortest path length, Indicates that in the minimum spanning tree MST, from node To Node The shortest path length; Based on the shortest path length Calculate the node To Node The shortest path probability ,in, is the regulating factor; is the path attenuation parameter, is the smoothing parameter.

9. A transformer predictive maintenance system according to claim 8, characterized in that: The formula of the optimization function is: ;in, For Node Determine a corresponding recommended maintenance strategy, For Node Determine a corresponding recommended maintenance strategy, Recommend maintenance strategies for execution The cost required, Recommend maintenance strategies for execution the costs required; For slave nodes Actual operation is transferred to the node time; For slave nodes Transfer to Node probability; For operating losses; The constraints include a first constraint and a second constraint; the first constraint is that the out-degree probability of each node is 1, and the in-degree probability of each node is 1; the second constraint is that the probability of transferring from any node to other nodes is non-negative, and the shortest path probability is used as the upper limit of the probability of any node transferring to other nodes.

10. A transformer predictive maintenance system according to claim 9, characterized in that: The method for solving the optimization function includes: Initialize a population, where each individual in the population represents a solution vector, i.e. the probability of transfer between nodes; randomly generate an initial population that satisfies the constraints; For each individual, calculate the value of its optimization function, recorded as the allocation degree; define the empire according to the individual's allocation degree, select the individual with the highest allocation degree as the emperor of the first empire, and calculate the difference between the allocation degrees of other individuals and the emperor. In the order of the difference in allocation degrees from small to large, allocate other individuals to the empire in turn until the power of the empire reaches the preset maximum power; repeat until all individuals are allocated to different empires; Define the weak threshold. Within each empire, treat individuals with a distribution degree less than the weak threshold as weak countries. Perform assimilation operations on weak countries to gradually make them closer to the emperor of the empire. The formula for assimilation operations is: ;in, A new position for small and weak countries, For the current position of weak countries, For the position of the emperor of the empire, is a random number in the interval [0, 1]. is a uniform random number in the interval [0.1, 0.9], for The distribution degree, for The distribution degree, is the difference adjustment parameter, is a positive number; is a random number that follows a standard normal distribution; Define a border distance threshold, calculate the distance between all individuals in the empire and the emperor, and mark individuals whose distance is greater than the border distance threshold as being at the border of the empire; For individuals marked as being at the border of the empire, a random number in the interval [0, 1] is generated according to the predefined revolution rate; if the random number is less than the revolution rate, the position of the individual is randomly initialized again; After the assimilation operation and the revolution, the sum of the distribution degrees of all individuals in the empire is calculated as the strength of the empire. In order of the strength of the empires from large to small, the smallest empire is merged into other empires; repeat until the preset maximum number of iterations is reached, and the individual with the highest distribution degree in all empires at this time is obtained as the optimal solution. Based on the optimal solution, the transition probability matrix between each node is obtained; Based on the obtained transition probability matrix and the maintenance cost corresponding to each node, the long-term total maintenance cost is calculated; different combinations of recommended maintenance strategies are enumerated, and the recommended maintenance strategy combination that can minimize the long-term total maintenance cost is found, which is the optimal maintenance strategy.

Images