Transformer state early warning method and system based on data analysis

Abstract

The invention provides a mutual inductor state early warning method and system based on data analysis, and the method comprises the steps: collecting a high-frequency electric signal and basic electric quantity of a mutual inductor, optimizing a wavelet basis and the number of layers through wavelet packet decomposition, extracting the characteristic frequency band energy of the high-frequency signal, and forming a characteristic vector with the basic electric quantity; constructing a probabilistic neural network: calculating an activation value by a mode layer neuron based on a multivariate Gaussian kernel of an anisotropic adaptive covariance matrix; the summation layer performs weighted summation on the activation values according to state categories to obtain probability density; based on the evidence theory, the probability density is converted into basic probability distribution, and an early warning result is output through confidence function comparison.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of state early warning, and particularly relates to a mutual inductor state early warning method and system based on data analysis. BACKGROUND

[0002] The mutual inductor is a device for measurement and protection in the power system. Common mutual inductor state monitoring methods such as dissolved gas analysis (DGA) in oil and partial discharge detection rely on offline tests and manual experience interpretation, and have problems such as poor real-time monitoring and insensitivity to early latent faults. By analyzing the electrical quantity data collected during the operation of the mutual inductor, a machine learning model is used to automatically identify the operating state. The probabilistic neural network (PNN) has a simple structure, fast training, strong nonlinear approximation ability, and a Bayesian decision theory basis, and has been widely used in the field of pattern recognition. However, the existing mutual inductor state evaluation method based on the probabilistic neural network usually uses a fixed parameter signal processing method in the feature extraction stage, such as a fixed wavelet basis and a fixed decomposition level, which cannot adaptively extract the most relevant feature information of the fault state, affecting the effectiveness of the model input. In the model construction, the mode layer of the standard probabilistic neural network usually uses a global and unified smoothing factor, which is equivalent to assuming that the data of all classes is distributed in a spherical shape in the feature space, and has the same degree of dispersion, which does not match the complex distribution characteristics of the actual operating data, limiting the classification accuracy of the model. In the decision output stage, the traditional PNN uses the maximum a posteriori probability criterion for decision-making, without considering the misjudgment cost difference between different states, and cannot effectively handle the uncertainty problem when the to-be-tested sample does not match all known patterns, which may lead to misjudgment or omission, especially in the reliability of the early warning of the key state. SUMMARY

[0003] In order to improve the accuracy of the mutual inductor state early warning, the application provides a mutual inductor state early warning method based on data analysis, which comprises the following steps:

[0004] Obtain mutual inductor operating data, the operating data comprising at least one high-frequency electrical signal and multiple basic electrical quantities; process the high-frequency electrical signal by using wavelet packet decomposition, determine the optimal wavelet basis and decomposition level by calculating the Fisher discriminant of the energy features under the combination of the preset wavelet basis and the decomposition level array, extract the energy features of the corresponding frequency band, and construct a feature vector with the basic electrical quantities;

[0005] A probabilistic neural network model is constructed, which comprises a pattern layer and a summation layer; in the pattern layer, each neuron corresponds to a historical training sample, and an activation value of the neuron is calculated according to an input to-be-tested sample and the historical training sample corresponding to the neuron by a multivariate Gaussian kernel function, a covariance matrix of the multivariate Gaussian kernel function being anisotropic and being determined according to a feature covariance of a local neighborhood sample set of the historical training sample corresponding to the neuron; in the summation layer, neuron activation values of the pattern layer belonging to a same state category are weighted and summed to obtain a probability density of each state category, and a weighting weight is determined by a preset transformer state misjudgment cost matrix and a time decay factor of the training sample corresponding to the neuron;

[0006] A decision rule based on evidence theory is adopted to convert the probability densities of the state categories into basic probability assignment functions, wherein a basic probability assignment value of any state category as a single subset is obtained by multiplying a corresponding probability density by a belief discount factor, the belief discount factor being inversely proportional to a Euclidean distance from the to-be-tested sample to a center of the state category training sample, and the remaining probability assignment values are assigned to a universal set, and a pre-warning result is output based on a confidence function value of each single subset.

[0007] Preferably, the determination of the optimal wavelet basis and the decomposition level comprises:

[0008] The preset wavelet basis set is {db4, db8, sym5, sym8}, and the preset decomposition level set is {3, 4, 5};

[0009] All combinations of the wavelet basis and the decomposition level are traversed, and Fisher discrimination of the energy features extracted under each combination between different state category samples is calculated;

[0010] The combination that makes the Fisher discrimination reach a maximum value is selected as the optimal wavelet basis and the decomposition level.

[0011] Preferably, the determination of the feature covariance of the local neighborhood sample set of the historical training sample corresponding to the neuron comprises:

[0012] For the neuron in the pattern layer representing any historical training sample, a k-neighbor algorithm is adopted to find k samples closest in Euclidean distance from other historical training samples of the same category as the historical training sample, to constitute a local neighborhood sample set of the neuron;

[0013] A covariance matrix of feature vectors of the k samples in the local neighborhood sample set is calculated as the covariance matrix of the Gaussian kernel function corresponding to the neuron, wherein k is a positive integer.

[0014] Preferably, the weighting weight is determined by a preset transformer state misjudgment cost matrix and a time decay factor of the neuron corresponding to the training sample, and includes:

[0015] The time decay factor , wherein t is a current time, is a collection time of the training sample, and λ is a preset time decay coefficient.

[0016] The cost factor is calculated from the real state category k of the training sample according to a preset cost matrix M, and the value of the cost factor is the sum of the cost of misjudging category k as all other categories , wherein M(i, j) is the cost value of misjudging the real state i as state j.

[0017] Preferably, the conversion of the probability density of each state category into the basic probability assignment function includes:

[0018] The probability density of each state category is normalized to obtain the normalized probability of each state category.

[0019] The singleton basic probability assignment value m(i) of any state category i is obtained by multiplying the normalized probability P(i|x) and the credibility discount factor .

[0020] The credibility discount factor is calculated by the formula , wherein β is a preset positive coefficient, is the Euclidean distance from the to-be-tested sample to the center of the state category i training sample; and the basic probability assignment value of the universal set is .

[0021] Preferably, the output of the pre-warning result based on the confidence function value of each singleton includes:

[0022] The confidence function value of any state category i is equal to the basic probability assignment value m(i) of the state category i.

[0023] If there is a certain fault state category , the confidence function value of which is the largest among all state categories, and is greater than a preset decision threshold, the fault state category is output as the pre-warning result.

[0024] In addition, the present application proposes a data analysis-based transformer state pre-warning system, characterized in that it comprises the following units:

[0025] ​​The feature extraction unit acquires the operating data of the transformer, the operating data including at least one high-frequency electrical signal and a plurality of basic electrical quantities; the high-frequency electrical signal is processed by using wavelet packet decomposition, the Fisher discrimination of the energy features under the combination of a preset wavelet basis and a decomposition level is calculated, the optimal wavelet basis and the decomposition level are determined, the energy features of the corresponding frequency band are extracted, and the basic electrical quantities are combined to form a feature vector;

[0026] The model construction unit constructs a probabilistic neural network model, the probabilistic neural network model including a mode layer and a summation layer; in the mode layer, each neuron corresponds to a historical training sample, and the activation value of the neuron is calculated by using a multivariate Gaussian kernel function according to an input to-be-tested sample and the historical training sample corresponding to the neuron, the covariance matrix of the multivariate Gaussian kernel function being anisotropic and being determined according to the feature covariance of a local neighborhood sample set of the historical training sample corresponding to the neuron; in the summation layer, the activation values of the mode layer neurons belonging to the same state category are weighted and summed to obtain the probability density of each state category, and the weighting weight is determined by a preset transformer state misjudgment cost matrix and a time decay factor of the training sample corresponding to the neuron;

[0027] The result output unit converts the probability densities of the state categories into basic probability assignment functions by using a decision rule based on evidence theory, wherein the basic probability assignment value of any state category as a singleton set is obtained by multiplying the corresponding probability density by a belief discount factor, the belief discount factor being inversely proportional to the Euclidean distance from the to-be-tested sample to the center of the state category training sample, the remaining probability assignment values being assigned to the universal set, and an early warning result being output based on the belief function values of the singletons.

[0028] Preferably, the determination of the optimal wavelet basis and the decomposition level includes:

[0029] The preset wavelet basis set is {db4, db8, sym5, sym8}, and the preset decomposition level set is {3, 4, 5};

[0030] All combinations of the wavelet basis and the decomposition level are traversed, and the Fisher discrimination of the extracted energy features under each combination between different state category samples is calculated;

[0031] The combination that makes the Fisher discrimination reach the maximum value is selected as the optimal wavelet basis and the decomposition level.

[0032] Preferably, the determination of the feature covariance of the local neighborhood sample set of the historical training sample corresponding to the neuron includes:

[0033] For a neuron representing any historical training sample in the pattern layer, a k-neighbor algorithm is used to find the k nearest samples in other historical training samples of the same category as the historical training sample, to form a local neighborhood sample set of the neuron;

[0034] A covariance matrix of the k sample feature vectors in the local neighborhood sample set is calculated as a covariance matrix of a Gaussian kernel function corresponding to the neuron, where k is a positive integer.

[0035] Preferably, the weighting weight is determined by a preset misjudgment cost matrix of the transformer state and a time decay factor of the training sample corresponding to the neuron, and includes:

[0036] The time decay factor , wherein is the current time, is the collection time of the training sample, and λ is a preset time decay coefficient.

[0037] The cost factor is calculated from a preset cost matrix M according to the real state category k of the training sample, and the value of the cost factor is the sum of the costs of misjudging category k as all other categories , wherein M(i, j) is the misjudgment cost value of misjudging the real state i as state j.

[0038] Preferably, the conversion of the probability density of each state category into the basic probability assignment function includes:

[0039] The probability density of each state category is normalized to obtain the normalized probability of each state category.

[0040] The singleton basic probability assignment value m(i) of any state category i is obtained by multiplying the normalized probability P(i|x) and the confidence discount factor ; the confidence discount factor is calculated by the formula , wherein β is a preset positive coefficient, is the Euclidean distance from the sample to be measured to the center of the state category i training sample.

[0041] The basic probability assignment value of the universal set is .

[0042] Preferably, the pre-warning result is output based on the confidence function value of each singleton, and includes:

[0043] The confidence function value of any state category i is equal to the basic probability assignment value m(i) of the state category i.

[0044] If there is a fault state category , the confidence function value is the largest among all state categories, and is greater than a preset decision threshold, the fault state category is output as a pre-warning result.

[0045] The present application provides a kind of mutual inductor state pre-warning method.In feature extraction link, by optimizing selection to wavelet packet analysis parameter, it is ensured that the signal feature extracted has the highest correlation degree with the specific state of mutual inductor, and the capture ability of early weak fault feature is improved.In model construction, the probability neural network model used configures unique kernel function for each neuron of its mode layer, which can reflect the distribution form of adjacent data, overcome the limitation that traditional model cannot accurately fit complex, non-spherical data distribution using uniform smoothing factor, and improve the accuracy of state recognition.The mutual inductor state pre-warning method fuses the misjudgment cost of different fault types and the time decay effect of sample data in decision-making process, so that the pre-warning conclusion is closer to the real health level of mutual inductor at present;Using evidence theory for decision-making can reasonably handle uncertain information, and when the sample to be tested does not match all known fault modes, it can be identified as unknown state, avoiding misjudgment of new type fault or interference signal. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 is the flow chart of specific embodiment;

[0047] Figure 2 is the schematic diagram of mode layer adaptive covariance;

[0048] Figure 3 is the schematic diagram of weighted weight of summation layer;

[0049] Figure 4 is the schematic diagram of basic probability assignment. DETAILED DESCRIPTION

[0050] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0051] The plurality involved in the present application refers to two or more. In addition, it should be understood that in the description of the present application, the words "first", "second", etc. are only used for the purpose of distinguishing description, and cannot be understood as indicating or implying relative importance, nor can it be understood as indicating or implying order.

[0052] In one embodiment, the application provides a data analysis-based mutual inductor state early warning method, as shown in Figure 1 comprising the following steps:

[0053] S1, obtaining mutual inductor operation data, the operation data including at least one high-frequency electrical signal and multiple basic electrical quantities; processing the high-frequency electrical signal by wavelet packet decomposition, determining the optimal wavelet basis and decomposition level by calculating the Fisher discriminant of the energy features under the combination of the preset wavelet basis and decomposition level, extracting the energy features of the corresponding frequency band, and constructing a feature vector with the basic electrical quantities;

[0054] The high-frequency electrical signal is collected by a high-frequency current sensor, and the basic electrical quantities such as voltage, current, active power, and reactive power are obtained by the existing SCADA system. To determine the optimal wavelet basis and decomposition level, a candidate parameter set is established, for example, the wavelet basis candidate set includes Daubechies series db2 to db10 and Symlets series sym2 to sym8, and the decomposition level candidate set includes 3, 4, and 5 layers. For historical fault data, each combination of wavelet basis and decomposition level in the candidate set is traversed, the high-frequency signal is decomposed by wavelet packet, and the energy of each frequency band node is calculated. The Fisher discriminant criterion is used to calculate the discrimination ability of the energy features under the parameter combination for different state categories. The Fisher discriminant criterion measures by maximizing the ratio of inter-class dispersion to minimizing intra-class dispersion, and the wavelet basis and decomposition level that maximize the Fisher discriminant value are selected as the optimal parameters. The new high-frequency signal is processed using the optimal parameter combination, the energy features are extracted, and the basic electrical quantities of the same time section are spliced to form a high-dimensional feature vector.

[0055] S2, constructing a probabilistic neural network model, the probabilistic neural network model including a mode layer and a summation layer; in the mode layer, each neuron corresponds to a historical training sample, and the activation value of the neuron is calculated by a multivariate Gaussian kernel function according to the input test sample and the historical training sample corresponding to the neuron, the covariance matrix of the multivariate Gaussian kernel function is anisotropic and is determined according to the feature covariance of the local neighborhood sample set of the historical training sample corresponding to the neuron; in the summation layer, the activation values of the mode layer neurons belonging to the same state category are weighted and summed to obtain the probability density of each state category, and the weighting weight is determined by the preset mutual inductor state misjudgment cost matrix and the time decay factor of the training sample corresponding to the neuron;

[0056] Assume there are N historical training samples, the pattern layer sets N neurons, each neuron i stores the feature vector Xi and its state category of a historical training sample. When a feature vector X of a to-be-tested sample is input, the activation value calculated by the i-th neuron of the pattern layer is a value based on a multivariate Gaussian kernel function, which reflects the similarity between X and Xi. The covariance matrix of the kernel function Σi is determined for each neuron i. The determination method is: for each training sample Xi, use the k-nearest neighbor algorithm to find the k samples closest to Xi in the entire training set, and construct a local neighborhood sample set with these k samples and Xi itself. Calculate the feature covariance matrix of the local neighborhood sample set, which is used as the kernel function covariance matrix Σi of the neuron i. The anisotropic covariance matrix can better describe the ellipsoidal distribution characteristics of the data in the local region of the training sample, as shown in Figure 2 .

[0057] The summation layer groups and sums the activation values of the pattern layer neurons according to the state categories of the transformer, such as normal, inter-turn short circuit, core overheating, etc. The summation preferably uses weighted summation. The weight Wi of each neuron i is obtained by multiplying two parts. The first part is the misjudgment cost, which predefines a cost matrix C, where the element Cjk represents the cost of misjudging a sample of class j as class k, for example, the cost of misjudging a serious fault as normal is much higher than the cost of misjudging normal as a slight fault. The cost weight of neuron i is related to the c-th row of the cost matrix, where c is the true category of the sample corresponding to neuron i. The second part is the time decay factor, which is a function decreasing with time, for example, an exponential decay function, so that the training samples collected in the near future have greater weights than the old training samples. Multiply the two parts to obtain the weight Wi, and perform weighted accumulation on the activation values belonging to the same state category to obtain the posterior probability density estimate of the state category.

[0058] S3, using the decision rule based on evidence theory, converts the probability density of each state category into a basic probability assignment function, where any state category is the basic probability assignment value of a singleton, which is obtained by multiplying the corresponding probability density and a belief discount factor, which is inversely proportional to the Euclidean distance from the to-be-tested sample to the center of the state category training sample. The remaining probability assignment values are assigned to the universal set, and based on the confidence function values of each singleton, the pre-warning result is output, as shown in Figure 4 .

[0059] The probability density P c of each state category c output by the summation layer is taken as the initial evidence. A credibility discount factor α c is calculated for each category c. The credibility discount factor is calculated as follows: the feature mean of all samples belonging to category c in the training data is calculated as the center point μ c of category c; the Euclidean distance d between the sample to be measured X and the center point μ c is calculated; the discount factor α c is inversely proportional to the distance d, for example, α c equals 1 ÷ (1 + a constant × distance d). Each state category is taken as a singleton set, and the basic probability assignment value m{c} of the singleton set is equal to the probability density P c multiplied by the corresponding discount factor α c. The sum of the basic probability assignment values of all singleton sets is usually less than 1, and the difference, i.e. 1 minus the sum of all m{c}, is assigned to the universal set Θ, denoted as m{Θ}, representing the degree of uncertainty or unknown for all known categories. The belief function Bel{c} of each singleton set is calculated based on the Dempster combination rule, and the value is m{c}. The state category with the largest belief function value is selected as the final early warning result. If the value of m{Θ} exceeds a preset threshold, an unknown state can be output to cope with new types of failures.

[0060] In an optional embodiment, the determining of the optimal wavelet basis and the number of decomposition levels comprises:

[0061] The preset set of wavelet bases is {db4, db8, sym5, sym8}, and the preset set of numbers of decomposition levels is {3, 4, 5};

[0062] All combinations of wavelet bases and numbers of decomposition levels are traversed, and the Fisher discrimination of the energy features extracted under each combination between different state category samples is calculated.

[0063] The combination that makes the Fisher discrimination reach the maximum value is selected as the optimal wavelet basis and the number of decomposition levels.

[0064] In order to make the extracted features most effectively represent the differences between the equipment in different health states, the most suitable wavelet basis and the number of decomposition levels need to be selected. This process is an optimization process, which limits the range of optimization, for example, the candidate wavelet basis functions are set to db4, db8, ym5, and sym8, and the number of decomposition levels is set to 3, 4, and 5. A total of 12 different combination schemes are formed, such as db4 wavelet basis 3-layer decomposition, db4 wavelet basis 4-layer decomposition, and sym8 wavelet basis 5-layer decomposition. Those skilled in the art should know that the above-mentioned preset wavelet set and preset number of decomposition levels are limited to the above-mentioned one.

[0065] For each of the 12 combinations, wavelet packet decomposition is performed on the historical fault sample signals and energy features of each frequency band are extracted. Fisher discriminant is used to measure the classification performance of the extracted features under the 12 combinations. The greater the Fisher discriminant value, the greater the inter-class dispersion and the smaller the intra-class dispersion of the feature data in the feature space, i.e. the better the class distinguishability, for different state categories such as normal, bearing wear, rotor imbalance, etc. For example, the Fisher discriminant under the combination of db8 wavelet basis and 5-layer decomposition is calculated to be 5.8, and the Fisher discriminant under the combination of sym5 wavelet basis and 4-layer decomposition is calculated to be 4.9. After all 12 combinations are traversed and the corresponding Fisher discriminants are calculated, the combination with the maximum value is selected. If 5.8 is the maximum value among all combinations, the optimal wavelet basis is determined to be db8 and the optimal number of decomposition layers is determined to be 5.

[0066] In an optional embodiment, the determining according to the feature covariance of the local neighborhood sample set of the neuron corresponding to the historical training sample comprises:

[0067] For a neuron in the pattern layer representing any historical training sample, a k- nearest neighbor algorithm is used to find the k nearest samples in other historical training samples of the same category as the historical training sample, to form a local neighborhood sample set of the neuron.

[0068] A covariance matrix of the feature vectors of the k samples in the local neighborhood sample set is calculated as the covariance matrix of the Gaussian kernel function corresponding to the neuron, where k is a positive integer.

[0069] In the pattern layer of the probabilistic neural network, each neuron stores the feature information of a historical training sample, and a Gaussian kernel function is used to calculate the similarity between a to-be-tested sample and a historical sample. The shape of the Gaussian kernel function is determined by the covariance matrix. In order to make it more accurately reflect the distribution characteristics of the local data, the covariance matrix is determined in an adaptive manner in this embodiment. This process is performed independently for each neuron, i.e. each historical training sample. For example, for a historical sample A representing a normal state, a positive integer k is preset, such as k equal to 5.

[0070] Among all other historical training samples which are also labeled as normal state, calculate the Euclidean distance between their respective feature vectors and the feature vector of sample A. According to the calculated distance values from small to large, select the 5 samples with the closest distance, and these 5 samples constitute the local neighborhood sample set of sample A. Calculate the covariance matrix of their feature vectors in the set containing 5 samples. The covariance matrix represents the distribution shape and dispersion degree of normal state data in a small range near sample A. Set the calculated covariance matrix as the covariance matrix of the Gaussian kernel function used by the neuron in the pattern layer representing sample A. Repeat this process for all neurons, so that the Gaussian kernel function of each neuron can adapt to its local data characteristics.

[0071] In an optional embodiment, the weighting weight is determined by a preset transformer state misjudgment cost matrix and a time decay factor of the neuron corresponding to the training sample; the time decay factor is , wherein is the current time, is the collection time of the training sample, and λ is a preset time decay coefficient; the cost factor is calculated from a preset cost matrix M according to the real state category k of the training sample, and the value of the cost factor is the sum of the costs of misjudging category k as all other categories , wherein M(i, j) is the cost value of misjudging the real state i as state j.

[0072] In order to improve the diagnostic accuracy of the model and the sensitivity to key faults, different importance, i.e. weighting weight, is given to each historical training sample. The weight is determined by two parts, one is time recency, and the other is misjudgment cost, as shown in Figure 3 . The newer data can better reflect the current condition of the equipment, so the weight should be higher. For example, set the time decay coefficient λ as 0.05, the current time as the 100th day, a sample A collected on the 98th day and a sample B collected on the 20th day, and their time difference is 2 days and 80 days respectively. The time decay factor of sample A is the exp function, and its parameter is about 0.90, while the time decay factor of sample B is the exp function, and its parameter is about 0.018. The weight of the new sample is higher than that of the old sample.

[0073] The cost factor takes into account the risk difference caused by different misjudgments. A cost matrix M is preset, for example, for normal state 1, slight fault state 2, and serious fault state 3, the cost of misjudging a serious fault as normal M(3, 1) can be as high as 100, and the cost of misjudging normal as slight fault M(1, 2) can be only 5. The calculation method of the cost factor is that for a sample of a known class, the cost of misjudging it into all other classes is added. For example, a sample belonging to the serious fault state 3, its cost factor is M(3, 1) plus M(3, 2). If M(3, 2) is 50, then its cost factor is 150. The cost factor of a normal sample can be only M(1, 2) plus M(1, 3), and the value can be much smaller than 150. The total weight of a sample is the product of the time decay factor and the cost factor, and then the fault samples with high misjudgment cost and recent time will be focused on.

[0074] In an optional embodiment, the converting the probability density of each state class into a basic probability assignment function comprises:

[0075] normalizing the probability density of each state class to obtain a normalized probability of each state class;

[0076] The singleton basic probability assignment value m(i) of any state class i is obtained by multiplying the normalized probability P(i|x) and a credibility discount factor .

[0077] The credibility discount factor is calculated by the formula , where β is a preset positive coefficient, is the Euclidean distance from the to-be-tested sample to the center of the training samples of the state class i;

[0078] The basic probability assignment value of the universal set is .

[0079] The network outputs the original probability density of each state for a to-be-tested sample x, and after normalization processing, the normalized probability is obtained, for example, for state A, state B and state C, the probabilities are 0.6, 0.3 and 0.1 respectively, and the sum of these probabilities is 1. In order to introduce the reliability measure of the evidence, a credibility discount factor is calculated. The credibility discount factor is related to the distance from the to-be-tested sample to each class of training sample center. The closer the distance, the greater the possibility that the to-be-tested sample belongs to the class, the more reliable the evidence, and the less the discount.

[0080] Suppose the preset coefficient β is 0.1, the distance from the to-be-tested sample x to the center of state A is 2, and the distance to the center of state B is 8. Then the credibility discount factor of state A is is the exp function with parameter approximately equal to 0.82. The discount factor for state B is the exp function with parameter approximately equal to 0.45. The basic probability assignment values for each singleton are obtained by multiplying the normalized probability by the corresponding discount factor. For example, the basic probability assignment value for state A, m({A}), is 0.492. The portion of the probability that is discounted, i.e., 1 minus the discount factor times the original probability, does not disappear, but is lumped into the uncertainty. The difference between the sum of the basic probability assignment values for all singleton classes and 1 is assigned to the empty set Θ, representing the extent to which the model is not certain about the classification of the sample into any single state, i.e., the model's own ignorance or uncertainty, as shown in Figure 4

[0081] In an alternative embodiment, the output of the early warning result is based on the belief function values of each singleton, including:

[0082] The belief function value of any state class i is equal to the basic probability assignment value of state class i, m(i);

[0083] If there exists a fault state class whose belief function value is the largest among all state classes, and is greater than a pre-set decision threshold, then the fault state class is output as the early warning result.

[0084] After obtaining the basic probability assignment values of each state, a final diagnostic decision needs to be made based on these values. In evidence theory, for a subset containing only one element, i.e., a singleton, its belief function value is equal to the basic probability assignment value of state class i. For example, the belief of the normal state, Bel(normal), is calculated to be 0.25, the belief of the bearing fault, Bel(bearing fault), is 0.60, and the belief of the rotor fault, Bel(rotor fault), is 0.05.

[0085] The decision rule contains two conditions that must be satisfied simultaneously to issue an early warning. The first condition is that the belief of a certain fault state must be the highest among all states. In the above example, the belief of the bearing fault, 0.60, is the highest, satisfying the first condition. The second condition is that the highest belief must also exceed a pre-set decision threshold, such as a threshold of 0.5. The role of the threshold is to ensure the reliability of the decision, avoiding making misleading judgments in cases where the overall belief is low, i.e., the uncertainty is high. In the example, 0.60 is greater than the pre-set threshold of 0.5, so the second condition is also satisfied. After both conditions are satisfied, the early warning result of the bearing fault is output. If the highest belief is 0.45, although it is still the highest, it does not exceed the threshold of 0.5, so no early warning of a specific fault is output. ​

[0086] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, a system or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, and the like) embodying computer readable program code.

[0087] The present application is described in reference to the flowchart illustrations and / or block diagrams of the methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processing machine, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in one or more of the flowchart illustrations and / or block diagrams. Figure 1 means for carrying out one or more functions specified in one or more of the flowchart illustrations and / or block diagrams.

[0088] The embodiments of the present application described above are intended to be exemplary only. The application can be modified and varied as numerous apparent to those skilled in the art. Accordingly, it is intended that all requirements broadly interpreted under the spirit and principles of the application be within the scope of the claims appended hereto.

Claims

1. A method for early warning of transformer status based on data analysis, characterized in that, Includes the following steps: Acquire the operating data of the current transformer, which includes at least one high-frequency electrical signal and multiple basic electrical quantities; process the high-frequency electrical signal using wavelet packet decomposition, determine the optimal wavelet basis and decomposition level by calculating the Fisher discriminant of the energy characteristics under the combination of a preset wavelet basis and a number of decomposition levels, extract the energy characteristics of the corresponding frequency band, and form a feature vector with the basic electrical quantities; A probabilistic neural network model is constructed, comprising a pattern layer and a summation layer. In the pattern layer, each neuron corresponds to a historical training sample. Based on the input test sample and the historical training sample corresponding to the neuron, the activation value of the neuron is calculated using a multivariate Gaussian kernel function. The covariance matrix of the multivariate Gaussian kernel function is anisotropic and is determined based on the feature covariance of the local neighborhood sample set of the historical training sample corresponding to the neuron. In the summation layer, the activation values ​​of pattern layer neurons belonging to the same state category are weighted and summed to obtain the probability density of each state category. The weighting weights are jointly determined by a preset transformer state misjudgment cost matrix and the time decay factor of the training sample corresponding to the neuron. The decision rule based on evidence theory is adopted to transform the probability density of each state category into a basic probability assignment function. The basic probability assignment value of any state category as a subset is obtained by multiplying the corresponding probability density by a confidence discount factor. The confidence discount factor is inversely proportional to the Euclidean distance from the test sample to the center of the state category training sample. The remaining probability assignment values ​​are assigned to the whole set, and the warning result is output based on the confidence function value of each subset.

2. The method according to claim 1, characterized in that, Determining the optimal wavelet basis and the number of decomposition layers includes: The preset wavelet basis set is {db4, db8, sym5, sym8}, and the preset decomposition level set is {3, 4, 5}; Iterate through all combinations of wavelet basis and decomposition level, and calculate Fisher's discrimination of the extracted energy features between samples of different state categories under each combination; The combination that maximizes the Fisher discriminant is selected as the optimal wavelet basis and the number of decomposition layers.

3. The method according to claim 1, characterized in that, The determination based on the feature covariance of the local neighborhood sample set corresponding to the historical training samples of the neuron includes: For a neuron representing any historical training sample in the pattern layer, the k-nearest neighbor algorithm is used to find the k samples with the closest Euclidean distance from other historical training samples of the same category as the historical training sample, which constitute the local neighborhood sample set of the neuron; Calculate the covariance matrix of the feature vectors of k samples in the local neighborhood sample set, and use it as the covariance matrix of the Gaussian kernel function corresponding to the neuron, where k is a positive integer.

4. The method according to claim 1, characterized in that, The weighting weights are determined by a preset mutual inductor state misjudgment cost matrix and the time decay factor of the training samples corresponding to the neurons, including: The time decay factor ,in For the current time, λ represents the collection time of the training samples, and λ is the preset time decay coefficient. The cost factor is calculated from a preset cost matrix M based on the true state category k of the training samples. The value of the cost factor is the sum of the costs of misclassifying category k as all other categories. , where M(i, j) is the cost of misclassifying the true state i as state j.

5. The method according to claim 1, characterized in that, The process of transforming the probability density of each state category into a basic probability assignment function includes: The probability density of each state category is normalized to obtain the normalized probability of each state category; The basic probability assignment value m(i) of a single subset of any state class i is composed of the normalized probability P(i|x) and the confidence discount factor. Multiplying them together yields the result. The credibility discount factor Through formula Calculate, where β is a preset positive coefficient. The Euclidean distance is the distance from the test sample to the center of the training sample of state class i. The basic probability assignment value of the entire set is .

6. The method according to claim 1, characterized in that, The warning results, based on the confidence function values ​​of each subset, are output, including: Confidence function value for any state class i It is equal to the basic probability assignment value m(i) of state category i; If a certain fault state category exists Its confidence function value It is the largest among all state categories, and If the value exceeds the preset decision threshold, the fault state category is output as a warning result.

7. A current transformer status early warning system based on data analysis, characterized in that, Includes the following units: The feature extraction unit acquires the operating data of the transformer, which includes at least one high-frequency electrical signal and multiple basic electrical quantities; it processes the high-frequency electrical signal using wavelet packet decomposition, determines the optimal wavelet basis and decomposition level by calculating the Fisher discriminant of the energy features under the combination of a preset wavelet basis and a number of decomposition levels, extracts the energy features of the corresponding frequency band, and forms a feature vector with the basic electrical quantities. The model building unit constructs a probabilistic neural network model, which includes a pattern layer and a summation layer. In the pattern layer, each neuron corresponds to a historical training sample. Based on the input test sample and the historical training sample corresponding to the neuron, the activation value of the neuron is calculated using a multivariate Gaussian kernel function. The covariance matrix of the multivariate Gaussian kernel function is anisotropic and is determined based on the feature covariance of the local neighborhood sample set of the historical training sample corresponding to the neuron. In the summation layer, the activation values ​​of pattern layer neurons belonging to the same state category are weighted and summed to obtain the probability density of each state category. The weighting weights are jointly determined by a preset mutual inductor state misjudgment cost matrix and the time decay factor of the training sample corresponding to the neuron. The output unit uses decision rules based on evidence theory to transform the probability density of each state category into a basic probability assignment function. The basic probability assignment value of any state category as a subset is obtained by multiplying the corresponding probability density by a confidence discount factor. The confidence discount factor is inversely proportional to the Euclidean distance from the test sample to the center of the state category training samples. The remaining probability assignment values ​​are assigned to the entire set, and the warning result is output based on the confidence function value of each subset.

8. The system according to claim 7, characterized in that, Determining the optimal wavelet basis and the number of decomposition layers includes: The preset wavelet basis set is {db4, db8, sym5, sym8}, and the preset decomposition level set is {3, 4, 5}; Iterate through all combinations of wavelet basis and decomposition level, and calculate Fisher's discrimination of the extracted energy features between samples of different state categories under each combination; The combination that maximizes the Fisher discriminant is selected as the optimal wavelet basis and the number of decomposition layers.

9. The system according to claim 7, characterized in that, The determination based on the feature covariance of the local neighborhood sample set corresponding to the historical training samples of the neuron includes: For a neuron representing any historical training sample in the pattern layer, the k-nearest neighbor algorithm is used to find the k samples with the closest Euclidean distance from other historical training samples of the same category as the historical training sample, which constitute the local neighborhood sample set of the neuron; Calculate the covariance matrix of the feature vectors of k samples in the local neighborhood sample set, and use it as the covariance matrix of the Gaussian kernel function corresponding to the neuron, where k is a positive integer.

10. The system according to claim 7, characterized in that, The weighting weights are determined by a preset mutual inductor state misjudgment cost matrix and the time decay factor of the training samples corresponding to the neurons, including: The time decay factor ,in For the current time, λ represents the collection time of the training samples, and λ is the preset time decay coefficient. The cost factor is calculated from a preset cost matrix M based on the true state category k of the training samples. The value of the cost factor is the sum of the costs of misclassifying category k as all other categories. , where M(i, j) is the cost of misclassifying the true state i as state j.

Images