A method and system for determining key characteristic quantities representing the health state of a transformer

Abstract

The application relates to a key characteristic quantity determination method and system for characterizing the health state of a transformer, comprising the following steps: determining a characteristic component analysis matrix of a to-be-detected transformer according to a characteristic quantity normalization matrix of the to-be-detected transformer; determining a characteristic component correlation analysis matrix between characteristic quantities of the to-be-detected transformer according to the characteristic component analysis matrix of the to-be-detected transformer; and determining key characteristic quantities for characterizing the health state of the to-be-detected transformer according to the characteristic component correlation analysis matrix between the characteristic quantities of the to-be-detected transformer. The technical scheme provided by the application solves the problems of insufficient fitting precision of a traditional feature extraction method and slow training speed of an artificial intelligence algorithm based on a neural network, provides theoretical support for the estimation of the health state of the transformer, and has immeasurable engineering practical value for the timely elimination of future transformer faults.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of feature engineering, in particular to a method and system for determining key feature quantities representing the health state of a transformer. BACKGROUND

[0002] As one of the important operating equipment of the power distribution network, the safe and reliable operation of the transformer is directly related to the safety and stability of the power distribution network. However, the internal structure of the transformer is complex, and the accident rate is on the rise, so certain technical means must be taken to ensure the safe and stable operation of the transformer.

[0003] Currently, the power distribution network mainly carries out planned maintenance on the transformer to ensure the safe and stable operation of the transformer, but planned maintenance actually sacrifices the economic effect of the power distribution network under the premise of ensuring the safety and reliability of the transformer. With the advancement of the electricity reform and the reduction of electricity prices, ensuring the safe and stable operation of the transformer in the form of planned maintenance has brought great pressure to the power grid enterprises, so accurate maintenance and state-based maintenance of the transformer become particularly important.

[0004] Traditional transformer fault diagnosis methods, whether they are the dissolved gas analysis (DGA) method for the transformer, the diagnosis method for the aging of the transformer oil-paper insulation, or other transformer fault diagnosis methods, only consider a specific fault of the transformer, such as insulation failure or winding deformation, and do not consider all factors affecting the health state of the transformer. There is a lack of a complete set of feature quantities representing the fault state of the transformer, and the current method of comprehensively considering multiple factors for transformer state evaluation also does not form a complete set of feature quantities representing the fault state of the transformer. In addition, the transformer feature quantities involved are determined based on expert experience and are not guided by scientific theories.

[0005] On the contrary, the technical personnel represent the health state of the transformer by setting a transformer health index, so as to realize accurate prediction of the state of the transformer. When analyzing the health state of the transformer, there is also a lack of a complete set of feature quantities representing the health state of the transformer, so it is very important to establish an index system representing the health state of the transformer.

[0006] And the transformer is a multi-element complex device, and there are many factors affecting its operating condition. When the transformer has an abnormal accident, there are many corresponding feature information, and such features may be highly redundant. Therefore, which features are key features representing the health state of the transformer, and can maximize the retention of key fault information, eliminate the correlation between fault feature quantities, and make the final obtained feature quantities effectively represent the health state of the transformer, is a key step for the health state evaluation of the transformer, and is also a hot and difficult problem to be solved for the health diagnosis of the transformer.

[0007] At present, the traditional feature extraction methods include independent component analysis (ICA), principal component analysis (PCA), kernel independent principal component analysis (KICA) and kernel principal component analysis (KPCA), but the above methods cannot guarantee the accuracy of feature extraction. In recent years, with the wide application of artificial intelligence methods, the traditional feedforward neural network can directly approximate complex nonlinear mapping through input samples, and provides a solution to a large number of natural and artificial phenomena that cannot be solved by traditional methods. However, this method needs to adjust all parameters, and has poor adaptability to high-dimensional feature quantities, slow training speed, and sometimes needs several hours, several days or even more time to train the neural network, which greatly restricts the practical engineering application prospect of this kind of algorithm. SUMMARY

[0008] In view of the shortcomings of the prior art, the purpose of the present application is to provide a method for determining key feature quantities representing the health state of a transformer, which solves the problems of insufficient fitting accuracy of traditional feature extraction methods and slow training speed of artificial intelligence algorithms based on neural networks, and provides theoretical support for the estimation of the health state of the transformer, and has immeasurable engineering practical value for the timely elimination of future transformer faults.

[0009] The purpose of the present application is achieved by using the following technical solutions:

[0010] The present application provides a method for determining key feature quantities representing the health state of a transformer, which improves in that the method comprises:

[0011] determining a feature component analysis matrix of the transformer to be tested according to a feature quantity normalization matrix of the transformer to be tested;

[0012] determining a feature component correlation analysis matrix between the feature quantities of the transformer to be tested according to the feature component analysis matrix of the transformer to be tested;

[0013] determining key feature quantities representing the health state of the transformer to be tested according to the feature component correlation analysis matrix between the feature quantities of the transformer to be tested.

[0014] Preferably, the feature component analysis matrix of the transformer to be tested is determined according to the feature quantity normalization matrix of the transformer to be tested, comprising:

[0015] obtaining a key feature extraction model with an output layer node number of 1 to S τ ;

[0016] obtaining S τ transformer health state calculation models respectively corresponding to the key feature extraction model with 1 to S τ ;

[0017] The characteristic quantity normalization matrix of the transformer to be tested is taken as an input of an input layer of a transformer health state calculation model with S τ output layers, and the HI value of the transformer output by the transformer health state calculation model is obtained. τ

[0018] The difference between the HI value of the transformer output by the transformer health state calculation model corresponding to the key feature extraction model with λ output layer nodes and the standard HI value of the transformer corresponding to the characteristic quantity normalization matrix of the transformer to be tested is taken as the λth element in the HI error sequence corresponding to the characteristic quantity normalization matrix of the transformer to be tested, and the HI error sequence corresponding to the characteristic quantity normalization matrix of the transformer to be tested is generated.

[0019] The difference between each element in the HI error sequence corresponding to the characteristic quantity normalization matrix of the transformer to be tested and the first preset value is obtained, and the HI standard deviation sequence corresponding to the characteristic quantity normalization matrix of the transformer to be tested is obtained.

[0020] Starting from the first element in the HI standard deviation sequence corresponding to the characteristic quantity normalization matrix of the transformer to be tested, a sequence segment with consecutive element values less than 0 is found, and a key feature extraction model corresponding to any element value in the sequence segment is obtained.

[0021] The characteristic quantity normalization matrix of the transformer to be tested is substituted into the key feature extraction model, and a characteristic component analysis matrix of the transformer to be tested is obtained.

[0022] The input layer structure of the transformer health state calculation model corresponding to the key feature extraction model with λ output layer nodes is an output layer node number λ key feature extraction model constructed in advance, and λ∈(1~S τ ), S τ is the number of characteristic quantity types of the transformer.

[0023] Preferably, the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be tested is determined according to the characteristic component analysis matrix of the transformer to be tested, and the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be tested comprises:

[0024] The least square generalized inverse matrix of the transpose matrix of the characteristic component analysis matrix of the transformer to be tested is calculated

[0025] The transpose matrix of the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be tested is determined according to the matrix

[0026] Wherein, X T is the transpose matrix of the characteristic quantity normalization matrix of the transformer to be tested, and T is the transpose symbol. ​​

[0027] Preferably, the key characteristic quantity representing the health state of the transformer to be measured is determined according to the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be measured, and the method comprises the following steps.

[0028] If the number of elements in the τth row of the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be measured is greater than the second preset threshold value, and the number of elements whose values are greater than the second preset threshold value exceeds the third preset threshold value, the characteristic quantity corresponding to the element in the τth row of the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be measured is the key characteristic quantity representing the health state of the transformer to be measured.

[0029] Wherein, τ ∈ (1 ~ S τ ), S τ is the number of characteristic quantities of the transformer.

[0030] The application provides a system for determining a key characteristic quantity representing the health state of a transformer, and the improvement lies in that the system comprises:

[0031] A first determining module is configured to determine a characteristic component analysis matrix of the transformer to be measured according to a characteristic quantity normalization matrix of the transformer to be measured.

[0032] A second determining module is configured to determine a characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be measured according to the characteristic component analysis matrix of the transformer to be measured.

[0033] A third determining module is configured to determine a key characteristic quantity representing the health state of the transformer to be measured according to the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be measured.

[0034] Preferably, the first determining module comprises:

[0035] A first obtaining unit is configured to obtain a key characteristic extraction model with an output layer node number of 1 to S τ ;

[0036] A second obtaining unit is configured to obtain S τ transformer health state calculation models respectively corresponding to the key characteristic extraction models with the output layer node number of 1 to S τ ;

[0037] A third obtaining unit is configured to take the characteristic quantity normalization matrix of the transformer to be measured as the input quantity of an input layer of S τ transformer health state calculation models, and obtain the HI value of the transformer output by the S τ transformer health state calculation models.

[0038] The generating unit is configured to generate a HI error sequence corresponding to the feature quantity normalization matrix of the transformer to be measured by taking a difference between an HI value of the transformer output by a transformer health state calculation model corresponding to a key feature extraction model with an output layer node number of λ and a standard HI value of the transformer corresponding to the feature quantity normalization matrix of the transformer to be measured as a λth element in the HI error sequence corresponding to the feature quantity normalization matrix of the transformer to be measured.

[0039] The fourth obtaining unit is configured to obtain a HI standard deviation value sequence corresponding to the feature quantity normalization matrix of the transformer to be measured by taking a difference between each element in the HI error sequence corresponding to the feature quantity normalization matrix of the transformer to be measured and a first preset value.

[0040] The fifth obtaining unit is configured to find a sequence segment with first continuous element values less than 0 from a first element in the HI standard deviation value sequence corresponding to the feature quantity normalization matrix of the transformer to be measured, and obtain a key feature extraction model corresponding to any element value in the sequence segment.

[0041] The sixth obtaining unit is configured to obtain a feature component analysis matrix of the transformer to be measured by substituting the feature quantity normalization matrix of the transformer to be measured into the key feature extraction model.

[0042] The input layer structure of the transformer health state calculation model corresponding to the key feature extraction model with the output layer node number of λ is an output layer node number of λ of the key feature extraction model constructed in advance, and λ∈(1~S τ ), S τ is a feature quantity type number of the transformer.

[0043] Preferably, the second determining module comprises:

[0044] The calculating unit is configured to calculate a least square generalized inverse matrix of a transpose matrix of the feature component analysis matrix of the transformer to be measured.

[0045] The first determining unit is configured to determine a transpose matrix of a feature component correlation analysis matrix between the feature quantities of the transformer to be measured according to the matrix

[0046] wherein, X T is a transpose matrix of the feature quantity normalization matrix of the transformer to be measured, and T is a transpose symbol.

[0047] Preferably, the third determining module is configured to:

[0048] ​If the number of elements in the first τ row of the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be measured is greater than the second preset threshold value, the characteristic quantity corresponding to the element in the first τ row of the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be measured is the key characteristic quantity representing the health state of the transformer to be measured.

[0049] Wherein, τ ∈ (1 ~ S τ ), S τ is the number of characteristic quantities of the transformer.

[0050] Compared with the closest prior art, the present application has the beneficial effects of:

[0051] The technical scheme provided by the present application determines the characteristic component analysis matrix of the transformer to be measured according to the characteristic quantity normalization matrix of the transformer to be measured; determines the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be measured according to the characteristic component analysis matrix of the transformer to be measured; and determines the key characteristic quantity representing the health state of the transformer to be measured according to the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be measured. The present application solves the problems of insufficient fitting accuracy of traditional feature extraction methods and slow training speed of artificial intelligence algorithms based on neural networks, and provides theoretical support for the estimation of the health state of the transformer, and has immeasurable engineering practical value for the timely elimination of transformer faults in the future. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 It is a flow chart of a method for determining a key characteristic quantity representing the health state of a transformer;

[0053] Figure 2 It is a structure diagram of a randomly configured network model in the embodiment of the present application;

[0054] Figure 3 It is a structure diagram of a randomly configured network model when a pth hidden layer node is newly added in the embodiment of the present application;

[0055] Figure 4 It is a structure diagram of a randomly configured network of a randomly configured network model in the form of an autoregressive in the embodiment of the present application;

[0056] Figure 5 It is a structure diagram of a transformer health state calculation model in the embodiment of the present application;

[0057] Figure 6 It is a comparison diagram of the simulation results of SCN, ELM and PCA in the embodiment of the present application;

[0058] Figure 7 It is a representation diagram of the correlation between the key characteristic quantity and the actual index in the embodiment of the present application;

[0059] Figure 8 is a simulation structure display diagram in the embodiment of the present application;

[0060] Figure 9 is a system structure diagram for determining a key characteristic quantity representing the health state of a transformer. DETAILED DESCRIPTION

[0061] The specific embodiments of the present application will be further described in detail below with reference to the accompanying drawings.

[0062] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only some of the embodiments of the present application, but not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts are within the scope of the present application.

[0063] The present application provides a method for determining a key characteristic quantity representing the health state of a transformer, as shown in Figure 1 , comprising:

[0064] Step 101. Determine a characteristic component analysis matrix of the transformer to be measured according to the characteristic quantity normalization matrix of the transformer to be measured.

[0065] Step 102. Determine a characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be measured according to the characteristic component analysis matrix of the transformer to be measured.

[0066] Step 103. Determine a key characteristic quantity representing the health state of the transformer to be measured according to the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be measured.

[0067] Specifically, the step 101 comprises:

[0068] Step 101-1. Obtain a key characteristic extraction model with an output layer node number of 1 to S τ previously constructed;

[0069] Step 101-2. Obtain S τ transformer health state calculation models respectively corresponding to the key characteristic extraction model with 1 to S τ previously constructed;

[0070] Step 101-3. Take the characteristic quantity normalization matrix of the transformer to be measured as the input quantity of the input layer of the S τ transformer health state calculation models respectively, and obtain the HI value of the transformer output by the S τ transformer health state calculation models;

[0071] Step 101-4. Take the difference between the HI value of the transformer output by the transformer health state calculation model corresponding to the key feature extraction model with the output layer node number λ and the standard HI value of the transformer corresponding to the feature quantity normalization matrix of the transformer to be measured as the λth element in the HI error sequence corresponding to the feature quantity normalization matrix of the transformer to be measured, and generate the HI error sequence corresponding to the feature quantity normalization matrix of the transformer to be measured;

[0072] Step 101-5. Subtract each element in the HI error sequence corresponding to the feature quantity normalization matrix of the transformer to be measured from the first preset value to obtain the HI standard deviation value sequence corresponding to the feature quantity normalization matrix of the transformer to be measured;

[0073] Step 101-5. From the first element in the HI standard deviation value sequence corresponding to the feature quantity normalization matrix of the transformer to be measured, find the first sequence segment of continuous element values less than 0, and obtain the key feature extraction model corresponding to any element value in the sequence segment.

[0074] Step 101-6. Substitute the feature quantity normalization matrix of the transformer to be measured into the key feature extraction model to obtain the feature component analysis matrix of the transformer to be measured.

[0075] The input layer structure of the transformer health state calculation model corresponding to the key feature extraction model with the output layer node number λ is the key feature extraction model with the output layer node number λ constructed in advance, and λ∈(1~S τ ), S τ is the feature quantity type number of the transformer.

[0076] In constructing the key feature extraction model, the traditional method adopts a single-hidden layer feedforward neural network for training. In the training process, each parameter needs to be adjusted. The adjustment method is mainly based on the gradient descent method. However, due to inappropriate step length, the training time will be too long, and there may also be a local minimum value phenomenon, so a large amount of iterative learning is required. When the parameter dimension is too high and the network is complex, the network training speed will be very slow, greatly limiting the application of the algorithm in engineering practice. To avoid the above problems, the present application adopts a random configuration network method to construct the key feature extraction model.

[0077] The random configuration network (Stochastis Configuration Networks, SCN) has no difference in network structure from the traditional single-hidden layer feedforward neural network (single-hidden layer feedforward neural networks, SLFNs), that is, it still adopts a three-layer structure of input layer, hidden layer and output layer, and adopts full connection between layers. The network structure is as follows:Figure 2 The random configuration network is an improvement of an extreme learning machine (ELM) algorithm, and the random parameter allocation monitoring mechanism with inequality constraints is set, the range of the random parameters is adaptively selected, and the condition that the network cannot approximate the target function with a high probability due to improper setting of the random parameters is avoided.

[0078] The random configuration network is gradually constructed by using the incremental construction method of the hidden layer nodes. Figure 3 The network structure of the SCN is given, and it is assumed that a single-layer feedforward network (SLFN) with P-1 hidden nodes has been constructed, that is, for the first P-1 hidden layer nodes, the weights ω i between the input nodes and the hidden layer nodes and the bias values b i of the hidden layer nodes are randomly selected, and the weights of the output layer are calculated.

[0079] The feature extraction method based on the random configuration network fully utilizes the nonlinearity of the hidden layer activation function, effectively solves the problem of insufficient fitting precision of the traditional principal component analysis method, and effectively solves the problem of long training time of the traditional neural network due to the high-dimensional characteristics of the transformer health feature quantity.

[0080] The following are specific steps of the method based on the random configuration network and the feature quantity normalization matrix of the sample transformer for obtaining a key feature extraction model with a pre-constructed output layer node number λ:

[0081] Step A: initialize θ = 1, and set the initial number of hidden layer nodes of the initial random configuration network model to λ;

[0082] Step B: during the θth training, the input weight matrix ω λθ and the hidden layer node bias matrix b λθ of the initial random configuration network model satisfying the first preset constraint condition are randomly generated;

[0083] Step C: Use the normalized feature matrix X of the sample transformer as the input layer sample data of the initial randomized network model, and set the feature component analysis matrix of the sample transformer output by the hidden layer of the initial randomized network model. Initially, the output weight matrix of the network model is randomly configured. Train the initial randomized network model to obtain the randomized network model with λ hidden layer nodes constructed during the θ-th training period and its output matrix Y of the output layer. θ ;

[0084] Step D: Calculate matrix X and matrix Y θ Error between

[0085] Step E: If θ = S ε When, then the sequence will be... minimum value The model structure between the input layer and the hidden layer of the randomly configured network model with λ hidden layer nodes built during the xth training period is used as the model structure between the input layer and the output layer of the key feature extraction model with λ output layer nodes, and the key feature extraction model with λ output layer nodes is obtained; otherwise, let θ = θ + 1, and return to step B;

[0086] Among them, b' λθ Let b' be a λ×N matrix. λθ The values ​​of each element in each row of the matrix are the same as those in matrix b. λθ The values ​​of each row of elements in the middle, Let θ be the bias of the λth hidden layer node randomly generated in the initial randomized configuration of the network model during the θth training iteration. For matrix H λθ The least-squares generalized inverse of the transpose of a matrix, where T is the transpose sign.

[0087]

[0088] x τi Let be the value of the τth type of feature quantity of the i-th sample transformer. Let be the value of the τ-th class feature of the i-th sample transformer output by a randomly configured network model with λ hidden layer nodes built during the θ-th training period. Let be the weight between the λth hidden layer node randomly generated in the initial random configuration of the network model during the θth training period and the τth type of feature of the sample transformer. Let X and Y be matrices θ The characteristic difference value of the i-th sample transformer. N is the number of sample transformers, Sε is a first preset training number.

[0089] Further, the first preset constraint condition is determined according to the following formula:

[0090]

[0091] In the formula, The difference between the output matrix of the key feature extraction model with the input layer node number of λ-1 and the characteristic quantity normalized matrix X of the sample transformer of the i-th sample transformer, and r is any value in the interval (0, 1).

[0092] H λθ,λi is a matrix H λθ is the value of the element in the i-th column of the λ-th row, b g is any positive real number, μ p,L is a non-negative real number sequence is the L-th element in the sequence, and μ p,L satisfies the constraint condition lim λ→+∞ μ p,L = 0 and μ p,L ≤ 1-r, L∈(1~S L ), S L is the number of elements in the non-negative real number sequence.

[0093] In order to use the random configuration network structure described above to perform feature extraction, by observing the SCN network, it can be seen that the hidden layer node output data of the SCN network can be regarded as a certain characteristic representation of the transformer input feature quantity. We regard the hidden layer node output data of the SCN network as the key feature quantity corresponding to the transformer input feature quantity. Then, whether the obtained key feature quantity is reasonable should be judged by whether the key feature quantity can be restored to the original initial feature (transformer input feature quantity). Therefore, we need to set a random configuration network model in the form of self-recurrence as shown in Figure 4 ; that is, by setting the number of hidden layer nodes and the input weight and output weight of the random configuration network, the error between the output layer data of the random configuration network and its input layer data is as small as possible, that is, in the ideal case, the error between the output layer data of the random configuration network and its input layer data is 0. In this way, if the number of hidden layer nodes is less than the number of input nodes, the network can achieve the purpose of reducing the dimension of the input feature, thereby realizing feature extraction. At the same time, by gradually increasing the number of hidden layer nodes, the process of gradually increasing the feature components can be realized.

[0094] Because the output of the hidden layer nodes in the trained random configuration network structure is the key feature extracted by us, the structure between the input layer nodes and the hidden layer nodes of the trained random configuration network structure is taken as the structure between the input layer nodes and the output layer nodes of the key feature extraction model, and the key feature extraction model needed by us is obtained;

[0095] Therefore, the key feature extraction model with different numbers of output layer nodes can be trained by the above method. When the number of output layer nodes is a certain value, the trained key feature extraction model is more effective in representing the health state of the transformer. We need to take each trained key feature extraction model as the input layer structure of the initial extreme learning machine network model, construct the transformer health state calculation model corresponding to each key feature extraction model, as shown in Figure 5 , and calculate the error between the standard HI value corresponding to the transformer feature quantity input into the transformer health state calculation model corresponding to each key feature extraction model and the output transformer HI value, respectively. From the error value, we select a transformer health state calculation model and its corresponding key feature extraction model with a suitable error value. The selected key feature extraction model can better represent the health state of the transformer.

[0096] The training process of the transformer health state calculation model corresponding to the key feature extraction model with the number of output layer nodes λ is as follows:

[0097] Step 1: initialize δ = 1, and set the number of hidden layer nodes of the initial extreme learning machine network model as P;

[0098] Step 2: randomly generate the input weight matrix ω λ,Pδ and the hidden layer node bias matrix b λ,Pδ of the initial extreme learning machine network model during the δth training;

[0099] Step 3: take the key feature extraction model with the number of output layer nodes λ as the input layer structure of the initial extreme learning machine network model, and take the feature quantity normalization matrix X of the sample transformer as the input layer sample data of the initial extreme learning machine network model;

[0100] Set the hidden layer output matrix of the initial extreme learning machine network model, the output weight matrix of the initial extreme learning machine network model, and train the initial extreme learning machine network model to obtain the extreme learning machine network model constructed during the δth training and the output matrix HI λ,δ of the output layer thereof;

[0101] Step 4: calculate the error λ between the matrix HI λ,δ and the matrix HI

[0102] Step 5: when delta is not equal to S δ , let delta = delta + 1, and return to step 2; otherwise, the minimum value in the sequence The extreme learning machine network model constructed during the vth training is taken as the transformer health state calculation model corresponding to the pre-constructed key feature extraction model with the output layer node number of lambda.

[0103] Wherein, is the weight between the Pth hidden layer node in the initial extreme learning machine network model during the dth training and the feature component in the lambda row of matrix H λx is the feature quantity matrix of the ith sample transformer output by the extreme learning machine network model constructed during the dth training, and HI λx is the output matrix of the key feature extraction model after the feature quantity matrix X of the sample transformer is input into the key feature extraction model with the output layer node number of lambda, is the HI value corresponding to the feature quantity matrix of the ith sample transformer, i is the HI value corresponding to the feature quantity matrix of the ith sample transformer, T is a transpose symbol, and b' λ,Pδ is a P*N matrix, and the values of the elements in each row of matrix b' λ,Pδ are equal to the values of the elements in each row of matrix b λ,Pδ , is the bias of the Pth hidden layer node randomly generated in the initial extreme learning machine network model during the dth training, is the least square generalized inverse matrix of the transpose matrix of matrix H λ,Pδ , i is an element in (1~N), N is the number of sample transformers, S δ is the second preset training number.

[0104] The key feature extraction model provided by the application can only be used for dimension reduction to achieve the purpose of simplifying calculation. After the feature quantity normalization matrix of the transformer to be tested is input into the selected key feature extraction model, the feature component analysis matrix of the transformer to be tested output by the model can be used to a certain extent to represent the key feature quantity of the transformer health state, but it does not have actual physical significance, so the transpose matrix of the feature component correlation analysis matrix of the transformer to be tested (which gives the correlation between the key feature quantity representing the transformer health state and the original feature quantity of the transformer) is obtained by using the feature component analysis matrix of the transformer to be tested, and the specific steps are as follows.

[0105] ​Step 102-1. Calculate the least square generalized inverse matrix of the transpose matrix of the characteristic component analysis matrix of the transformer to be measured

[0106] Step 102-2. Determine the transpose matrix of the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be measured according to the matrix

[0107] Wherein, X T is the transpose matrix of the characteristic quantity normalization matrix of the transformer to be measured, and T is a transpose symbol.

[0108] The transpose matrix of the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be measured gives the correlation between the key characteristic quantities representing the health state of the transformer and the original characteristic quantities of the transformer, so that the key characteristic quantities representing the health state of the transformer to be measured can be determined, and the specific process is as shown below:

[0109] If the number of elements with values greater than the second preset threshold in the τth row elements of the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be measured exceeds the third preset threshold, the characteristic quantity corresponding to the τth row elements of the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be measured is the key characteristic quantity representing the health state of the transformer to be measured.

[0110] Wherein, τ ∈ (1 ~ S τ ), S τ is the number of characteristic quantities of the transformer.

[0111] In view of the problem that the reduced features of the traditional feature extraction algorithm do not have physical meaning, the present application finally obtains a simplified index system representing the health state of the transformer with physical meaning by analyzing the correlation and contribution degree of the transformation matrix.

[0112] ​The present application carries out the research on the transformer health index index system, since different key characteristic quantities reflect different aspects of the transformer health state, aiming at the power distribution transformer health state evaluation problem, we need to classify the transformer health state influencing factors, and establish a unified, systematic and reasonable power distribution transformer evaluation index system, aiming at covering all factors affecting the transformer health state. Therefore, the present application summarizes and analyzes the existing transformer health state evaluation index, taking the actual operation state data and the field preventive test data as the reference, on the basis of considering the traditional transformer health state evaluation insulation oil test data, comprehensively considering the importance of other influencing factors, also increasing the load information, the equipment accessory health condition and the natural factors and other influencing factors. The characteristic quantities affecting the power transformer health are divided into four first-level indexes, which are electrical performance, physical and chemical performance, appearance and accessory performance and natural factors. And according to (QGDW645-2011) 'Distribution network equipment state evaluation guide', (QGDW644-2011) 'Distribution network state maintenance guide', (DLT596-2005) 'Electric power equipment preventive test regulation' and other relevant policies, regulations, procedures and standards of the electric power industry, and referring to the research results of many experts and scholars and the research results of relevant projects of State Grid Corporation, the four first-level indexes are classified to obtain the second-level indexes under each index, and the transformer health state index system is established.

[0113] Among them, the electrical performance indexes of the transformer include: winding direct current resistance phase difference of the transformer, winding insulation resistance of the transformer, winding absorption ratio of the transformer, winding polarization index of the transformer, winding dielectric loss of the transformer, winding capacitance variation level of the transformer, capacitive bushing dielectric loss of the transformer, capacitive bushing insulation resistance of the transformer, capacitive bushing capacitance value variation level of the transformer, capacitive bushing end screen insulation resistance variation level of the transformer, transformer core grounding current, transformer core insulation resistance, transformer no-load loss variation level, transformer no-load current variation level, transformer load loss variation level, transformer impedance voltage variation level, transformer oil breakdown voltage, transformer load rate level, transformer low voltage level, transformer three-phase imbalance rate;

[0114] The physical and chemical performance indexes of the transformer include: oil appearance grade of the transformer, insulating oil color grade of the transformer, furfural content grade in the oil of the transformer, mass fraction grade of oil sludge and precipitate of the transformer, temperature of the bushing lead joint of the transformer, micro-water content in the oil of the transformer, flash point of the transformer, oil temperature of the transformer, oil level of the transformer, water-soluble acid value of the transformer, volume resistivity of the transformer, hydrogen content in the oil of the transformer, methane content in the oil of the transformer, ethane content in the oil of the transformer, ethylene content in the oil of the transformer, acetylene content in the oil of the transformer, carbon dioxide content in the oil of the transformer, carbon monoxide content in the oil of the transformer, total hydrocarbon content in the oil of the transformer, gas content in the oil of the transformer;

[0115] The body appearance and accessory performance of the transformer include: color grade of the breather silica gel of the transformer, completeness of the body appearance of the transformer, dirtiness of the transformer, completeness of the identification of the transformer, sealing inspection index of the oil tank leakage of the transformer, rust degree of the transformer, vibration degree of the transformer, noise degree of the transformer, damage degree of the bushing appearance of the transformer, damage degree of the oil tank appearance of the transformer, damage degree of the grounding down lead appearance of the transformer, winding deformation grade of the transformer, evaluation grade of the on-load voltage regulator of the transformer, evaluation grade of the volume relay of the transformer, evaluation grade of the cooler of the transformer, evaluation grade of the temperature measuring device of the transformer;

[0116] The natural factor indexes of the transformer include: lightning and gale weather index of the transformer.

[0117] Therefore, the transformer characteristic quantity of the present application is set, including: the winding DC resistance phase difference of the transformer, the winding insulation resistance of the transformer, the winding absorption ratio of the transformer, the winding polarization index of the transformer, the winding dielectric loss of the transformer, the winding capacitance variation level of the transformer, the capacitive bushing dielectric loss of the transformer, the capacitive bushing insulation resistance of the transformer, the capacitive bushing capacitance value variation level of the transformer, the capacitive bushing end screen insulation resistance variation level of the transformer, the core grounding current of the transformer, the core insulation resistance of the transformer, the no-load loss variation level of the transformer, the no-load current variation level of the transformer, the load loss variation level of the transformer, the impedance voltage variation level of the transformer, the oil breakdown voltage of the transformer, the load rate level of the transformer, the low voltage level of the transformer, the three-phase imbalance rate of the transformer, the breather silica gel color level of the transformer, the body appearance integrity degree of the transformer, the pollution degree of the transformer, the identification completeness degree of the transformer, the oil tank leakage sealing inspection index of the transformer, the rust degree of the transformer, the vibration degree of the transformer, the noise degree of the transformer, the bushing appearance damage degree of the transformer, the oil tank appearance damage degree of the transformer, the grounding down lead appearance damage degree of the transformer, the oil appearance level of the transformer, the insulation oil color level of the transformer, the oil furfural content level of the transformer, the oil sludge and sediment mass fraction level of the transformer, the bushing lead joint temperature of the transformer, the oil micro-water content of the transformer, the flash point of the transformer, the oil temperature of the transformer, the oil level of the transformer, the water-soluble acid value of the transformer, the volume resistivity of the transformer, the winding deformation level of the transformer, the on-load voltage regulator evaluation level of the transformer, the volume relay evaluation level of the transformer, the cooler evaluation level of the transformer, the temperature measuring device evaluation level of the transformer, the oil hydrogen content of the transformer, the oil methane content of the transformer, the oil ethane content of the transformer, the oil ethylene content of the transformer, the oil acetylene content of the transformer, the oil carbon dioxide content of the transformer, the oil carbon monoxide content of the transformer, the oil total hydrocarbon content of the transformer, the oil gas content of the transformer, and the lightning and gale weather index of the transformer.

[0118] In the specific embodiments of the present application, the feature quantity correlation is too strong, the redundancy is too high, and the index system is relatively complex for the 110kV oil-immersed transformer health state index system. The present application extracts the features of the selected 57-dimensional 110kV oil-immersed transformer health state feature quantities, such as the winding direct current resistance phase difference, winding insulation resistance, capacitive bushing dielectric loss, oil temperature, flash point, oil breakdown voltage, transformer accessory evaluation level, and oil gas content of the 110kV oil-immersed transformer, to achieve the purpose of removing redundant indicators and simplifying the index system. The project team collected and sorted the routine maintenance records and fault data of 17 transformers in 8 substations in a city of a province, including substations put into use in the past five years and substations put into operation for ten, fifteen and twenty years, and calculated the health index values of each transformer at the corresponding recording time through the health index calculation platform. Finally, the collected 437 samples were subjected to feature extraction comparison experiment in the Matlab 2014a simulation environment. The influence of different algorithms on the health index calculation error was compared by gradually increasing the feature components. The simulation results obtained are shown in Figure 6 The simulation results in Figure 6 show that, with the increase of the number of feature components, the estimation error of the health index of the three feature extraction algorithms overall presents an initial obvious decrease and then gradually tends to be stable. However, the method based on extreme learning machine (ELM) has a larger fluctuation range than the other two methods. The reason is that the parameters of the ELM network are randomly selected and have no constraints, so the fitting effect of the network will have a lot of randomness and the network fitting performance cannot be guaranteed. The method based on stochastic configuration network (SCN) is an improvement based on ELM, and the random parameters are determined based on the fitting error of the previous network, so the network fitting performance can be better guaranteed and the error curve will be more stable.

[0119] However, from the aspect of transformer health index prediction error, the nonlinear characteristics of the hidden layer node activation function of the feedforward neural network mentioned in the foregoing, whether based on extreme learning machine (ELM) or based on stochastic configuration network (SCN), the prediction error of the transformer health index will be obviously better than that of the traditional principal component analysis method (PCA). The PCA and SCN estimation results are relatively stable, and Table 1 shows the specific values of the PCA and SCN fitting errors with the increase of the feature components. For the error value results of the 110kV oil-immersed transformer health index calculation, the superiority of the SCN algorithm in the estimation error can be obviously seen.

[0120] Table 1

[0121] Characteristic component number 1 2 3 4 5 6 PCA 1.848 1.368 1.294 1.334 1.201 1.191 SCN 1.186 1.127 1.059 0.687 0.670 0.669 7 8 9 10 … 56 57 1.140 1.114 1.182 1.167 … 1.127 1.115 0.669 0.696 0.734 0.741 … 0.710- 0.700

[0122] Combining the relatively smooth PCA and SCN graphics and Table 1 data can be seen directly, when the characteristic component is 8, the health index estimation error is basically stable. It can be concluded that through the foregoing feature extraction algorithm, the minimum number of simplified key features is 8 dimensions.

[0123] But the key features obtained by the feature extraction algorithm at present are the results obtained by transforming the 57-dimensional actual physical characteristic quantities, and have no actual physical meaning. In order to achieve the purpose of simplifying the actual physical characteristics in this paper, in the autoregressive neural network formed in the feature extraction process, we can obtain the output transformation matrix corresponding to the extracted key feature quantity and the actual feature quantity, and the numerical value in the output transformation matrix represents the correlation degree between the key feature quantity and the actual index, and also reflects the contribution degree of the actual physical characteristic quantity in the transformation to form the simplified key feature. Therefore, if the numerical value in the obtained transformation matrix is normalized to the interval [0, 1], and the correlation between the key feature quantity and the actual index is represented as shown in Figure 7 , the lightness of the color in the figure can represent the correlation degree between the key feature component without physical meaning and the actual physical characteristic quantity, and also reflect the contribution degree of the actual physical characteristic quantity to the extracted key feature component. Figure 7 The horizontal axis from left to right in the figure represents the characteristic component increasing in turn, and the vertical axis is the number of 57-dimensional transformer health characteristic quantities with actual physical meaning.

[0124] From the trend of the lightness of the color of the color block in Figure 7 , when the number of key feature components is small, the contribution degree of the key feature components extracted by the feature extraction algorithm to the characteristic quantities such as No. 37 (micro water content in oil), No. 39 (oil temperature), No. 36 (sleeve lead joint temperature), No. 53 (carbon dioxide), No. 54 (carbon monoxide), No. 55 (total hydrocarbon), No. 56 (gas content in oil), and No. 57 (lightning and gale) is higher, which is an important characteristic quantity for measuring the health state of the transformer. With the increase of the number of extracted key state components, the color blocks with deeper color also gradually increase, and the appearance of the color blocks also presents a certain trend. Taking the gas content in oil and the total hydrocarbon content as examples, with the increase of the characteristic component, it will no longer appear as a key feature quantity in the color block, which is because the appearance of hydrogen, methane and acetylene, the total hydrocarbon content and the gas content in oil can be obtained by other gases, and the feature extraction algorithm removes it as redundant quantity. In order to more intuitively reflect the trend of the color block, Figure 8 the trend of the color block is compared with the error trend of the feature extraction algorithm.

[0125] It can be seen that the color block will increase significantly with the feature component number 7 and 19 as the demarcation point. Combined with the error change diagram, it can be divided into under-fitting area, standard fitting area and over-fitting area. When the feature component number is small, the calculation error of the health index is large; when the feature component is too much, the key feature quantity of the transformer health state selected by the corresponding ordinate of the color block distribution diagram is too much, the redundancy and correlation degree between the indicators are high, and the purpose of simplifying the transformer health state index system cannot be achieved. Therefore, the transformer feature quantity corresponding to the vertical axis in the standard fitting area will be selected as the simplified key feature finally selected in the present application, as shown in Table 2.

[0126] Table 2

[0127] Serial number Category Characteristic quantity name Serial number Category Characteristic quantity name 1 Electrical performance Winding DC resistance phase difference 37 Physical and chemical performance Micro water content in oil 2 Electrical performance Winding insulation resistance 39 Physical and chemical performance Oil temperature 7 Electrical performance Capacitive bushing dielectric loss 40 Physical and chemical performance Oil level 11 Electrical performance Core grounding current 41 Physical and chemical performance Water-soluble acid value 12 Electrical performance Core insulation resistance 42 Physical and chemical performance Volume resistivity 13 Electrical performance No-load loss change level 43 Body appearance and accessory performance Winding deformation level 14 Electrical performance No-load current change level 45 Body appearance and accessory performance Gas relay evaluation level 15 Electrical performance Load loss change level 48 Physical and chemical performance Hydrogen 16 Electrical performance Impedance voltage change level 49 Physical and chemical performance Methane 17 Electrical performance Breakdown voltage in oil 50 Physical and chemical performance Ethane 18 Electrical performance Load rate level 51 Physical and chemical performance Ethylene 19 Electrical performance Low voltage level 52 Physical and chemical performance Acetylene 20 Electrical performance Three-phase imbalance rate 53 Physical and chemical performance Carbon dioxide 21 Body appearance and accessory performance Respirator silica gel color level 54 Physical and chemical performance Carbon monoxide 25 Body appearance and accessory performance Sealing inspection 56 Physical and chemical performance Gas content in oil 32 Physical and chemical performance Oil appearance level 57 Natural factors Thunder and lightning gale 36 Physical and chemical performance Bush lead joint temperature

[0128] The application provides a key feature quantity determination system for characterizing the health state of a transformer, as shown in Figure 9 , the system comprises:

[0129] A first determination module is configured to determine a feature component analysis matrix of a to-be-tested transformer according to a feature quantity normalization matrix of the to-be-tested transformer.

[0130] A second determination module is configured to determine a feature component correlation analysis matrix between feature quantities of the to-be-tested transformer according to the feature component analysis matrix of the to-be-tested transformer.

[0131] A third determination module is configured to determine a key feature quantity for characterizing the health state of the to-be-tested transformer according to the feature component correlation analysis matrix between the feature quantities of the to-be-tested transformer.

[0132] Specifically, the first determination module comprises:

[0133] A first acquisition unit is configured to acquire a key feature extraction model with an output layer node number of 1 to S τ ;

[0134] A second acquisition unit is configured to acquire S τ transformer health state calculation models respectively corresponding to the key feature extraction model with the output layer node number of 1 to S τ ;

[0135] A third acquisition unit is configured to input the feature quantity normalization matrix of the to-be-tested transformer into the input layer of S τ transformer health state calculation models respectively, and acquire HI values of the transformer output by the S τ transformer health state calculation models;

[0136] The generating unit is configured to generate, as an element at the λth position in a HI error sequence corresponding to the feature quantity normalization matrix of the transformer to be tested, a difference between an HI value of the transformer output by the transformer health state calculation model corresponding to the key feature extraction model with the output layer node number λ and a standard HI value of the transformer corresponding to the feature quantity normalization matrix of the transformer to be tested.

[0137] The fourth obtaining unit is configured to obtain a HI standard deviation value sequence corresponding to the feature quantity normalization matrix of the transformer to be tested by subtracting a first preset value from each element in the HI error sequence corresponding to the feature quantity normalization matrix of the transformer to be tested.

[0138] The fifth obtaining unit is configured to find, from the first element in the HI standard deviation value sequence corresponding to the feature quantity normalization matrix of the transformer to be tested, a sequence segment with consecutive element values less than 0, and obtain a key feature extraction model corresponding to any element value in the sequence segment.

[0139] The sixth obtaining unit is configured to obtain a feature component analysis matrix of the transformer to be tested by substituting the feature quantity normalization matrix of the transformer to be tested into the key feature extraction model.

[0140] The input layer structure of the transformer health state calculation model corresponding to the key feature extraction model with the output layer node number λ is a key feature extraction model with the output layer node number λ that is constructed in advance, and λ is an element in the range of (1, S τ ). S τ is the number of feature quantity types of the transformer.

[0141] Specifically, the system further includes a first model construction module configured to construct the key feature extraction model with the output layer node number λ in advance, and the first model construction module includes:

[0142] The first initialization unit is configured to initialize θ = 1, and set the number of hidden layer nodes of the initial randomly configured network model to λ.

[0143] The first random generating unit is configured to randomly generate an input weight matrix ω λθ and a hidden layer node bias matrix b λθ of the initial randomly configured network model that satisfies a first preset constraint condition during the θth training.

[0144] The first setting unit is configured to set the feature quantity normalization matrix X of the sample transformer as input layer sample data of the initial randomly configured network model, and set a feature component analysis matrix X of the sample transformer output by the hidden layer of the initial randomly configured network model as output layer sample data of the initial randomly configured network model. Train the initial randomized network model to obtain the randomized network model with λ hidden layer nodes constructed during the θ-th training period and its output matrix Y of the output layer. θ ;

[0145] The first error calculation unit is used to calculate the errors of matrix X and matrix Y. θ Error between

[0146] The first model acquisition unit is used when θ = S ε When, then the sequence will be... minimum value The model structure between the input layer and the hidden layer of the randomly configured network model with λ hidden layer nodes built during the xth training period is used as the model structure between the input layer and the output layer of the key feature extraction model with λ output layer nodes, and the key feature extraction model with λ output layer nodes is obtained; otherwise, let θ = θ + 1, and return to step B;

[0147] Among them, b' λθ Let b' be a λ×N matrix. λθ The values ​​of each element in each row of the matrix are the same as those in matrix b. λθ The values ​​of each row of elements in the middle, Let θ be the bias of the λth hidden layer node randomly generated in the initial randomized configuration of the network model during the θth training iteration. For matrix H λθ The least-squares generalized inverse of the transpose of a matrix, where T is the transpose sign.

[0148]

[0149] x τi Let be the value of the τth type of feature quantity of the i-th sample transformer. Let be the value of the τ-th class feature of the i-th sample transformer output by a randomly configured network model with λ hidden layer nodes built during the θ-th training period. Let be the weight between the λth hidden layer node randomly generated in the initial random configuration of the network model during the θth training period and the τth type of feature of the sample transformer. Let X and Y be matrices θ The characteristic difference value of the i-th sample transformer. i∈(1~N), N is the number of sample transformers, S ε This is the first preset number of training iterations.

[0150] Specifically, the first preset constraint condition is determined by the following formula:

[0151]

[0152] wherein, the difference between the output matrix of the key feature extraction model with the output layer node number λ-1 and the characteristic quantity normalized matrix X of the sample transformer, and the characteristic quantity difference of the i-th sample transformer between the output matrix of the key feature extraction model with the output layer node number λ-1 and the characteristic quantity normalized matrix X of the sample transformer, r is any value in the interval (0, 1);

[0153] H λθ,λi is the matrix H λθ is the value of the element in the i-th column of the λ-th row, b g is any positive real number, μ p,L is a non-negative real number sequence is the L-th element in the sequence, and μ p,L satisfies the constraint condition lim λ→+∞ μ p,L = 0 and μ p,L ≤ 1-r, L ∈ (1~S L ), S L is the number of elements in the non-negative real number sequence.

[0154] Specifically, the system further comprises a second model construction unit for pre-constructing a transformer health state calculation model corresponding to the key feature extraction model with the output layer node number λ, comprising:

[0155] The second initialization unit is configured to initialize δ = 1 and set the number of hidden layer nodes of the initial extreme learning machine network model as P.

[0156] The second random generation unit is configured to randomly generate the input weight matrix ω λ,Pδ and the hidden layer node bias matrix b λ,Pδ of the initial extreme learning machine network model during the δ-th training.

[0157] The second setting unit is configured to set the key feature extraction model with the output layer node number λ as the input layer structure of the initial extreme learning machine network model, and set the characteristic quantity normalized matrix X of the sample transformer as the input layer sample data of the initial extreme learning machine network model.

[0158] The second setting unit is configured to set the hidden layer output matrix of the initial extreme learning machine network model, and set the output weight matrix of the initial extreme learning machine network model. λ,δ ;

[0159] The second error calculation unit is configured to calculate the matrix HIλ With matrix HI λ,δ Error between

[0160] The second model acquisition unit is used when δ≠S δ If the condition is met, let δ = δ + 1 and return to step 2; otherwise, change the sequence. minimum value The extreme learning machine network model constructed during the corresponding v-th training period serves as the transformer health state calculation model corresponding to the key feature extraction model with λ output layer nodes.

[0161] in, Let H be the P-th hidden layer node and matrix H in the initial extreme learning machine network model during the δ-th training period. λx The weights between the feature components in the λth row, H λx To substitute the normalized feature matrix X of the sample transformer into the output matrix of the key feature extraction model with λ output layer nodes, HI λ =[HI1…HI i …HI N ] T , Let HI be the HI value corresponding to the feature matrix of the i-th sample transformer output by the Extreme Learning Machine Network model built during the δ-th training period. i Let HI be the eigenvalue corresponding to the feature matrix of the i-th sample transformer. T is the transpose symbol, b' λ,Pδ Let b' be a P×N matrix. λ,Pδ The values ​​of each element in each row are equivalent to those in matrix b. λ,Pδ The values ​​of each row of elements in the middle, Let be the bias of the Pth hidden layer node randomly generated in the initial extreme learning machine network model during the δth training period. For matrix H λ,Pδ The least-squares generalized inverse of the transpose of the matrix, i∈(1~N), where N is the number of sample transformers, S δ This is the second preset number of training iterations.

[0162] Specifically, the second determining module includes:

[0163] The computational unit is used to calculate the least-squares generalized inverse matrix of the transpose of the eigencomponent analysis matrix of the transformer under test.

[0164] The first determining unit is used to determine the matrix. Determine the transpose matrix of the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be measured

[0165] Wherein, X T The transpose matrix of the characteristic quantity normalization matrix of the transformer to be measured, and T is the transpose symbol.

[0166] Specifically, the third determination module is used to:

[0167] If the number of elements with values greater than the second preset threshold in the τth row elements of the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be measured exceeds the third preset threshold, the characteristic quantity corresponding to the τth row elements of the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be measured is the key characteristic quantity representing the health state of the transformer to be measured.

[0168] Wherein, τ∈(1~S τ ), S τ The number of characteristic quantities of the transformer.

[0169] Specifically, the transformer characteristic quantity includes: a winding direct current resistance phase difference of the transformer, a winding insulation resistance of the transformer, a winding absorption ratio of the transformer, a winding polarization index of the transformer, a winding dielectric loss of the transformer, a winding capacitance variation level of the transformer, a capacitive bushing dielectric loss of the transformer, a capacitive bushing insulation resistance of the transformer, a capacitive bushing capacitance value variation level of the transformer, a capacitive bushing end screen insulation resistance variation level of the transformer, a core grounding current of the transformer, a core insulation resistance of the transformer, a no-load loss variation level of the transformer, a no-load current variation level of the transformer, a load loss variation level of the transformer, an impedance voltage variation level of the transformer, a breakdown voltage in oil of the transformer, a load rate level of the transformer, a low voltage level of the transformer, a three-phase imbalance rate of the transformer, a breather silica gel color level of the transformer, a body appearance integrity degree of the transformer, a pollution degree of the transformer, an identification completeness degree of the transformer, an oil tank leakage sealing inspection index of the transformer, a rust degree of the transformer, a vibration degree of the transformer, a noise degree of the transformer, a bushing appearance damage degree of the transformer, an oil tank appearance damage degree of the transformer, a grounding down lead appearance damage degree of the transformer, an oil appearance level of the transformer, an insulation oil color level of the transformer, a furfural content level in oil of the transformer, an oil sludge and sediment mass fraction level of the transformer, a bushing lead joint temperature of the transformer, a micro-water content in oil of the transformer, a flash point of the transformer, an oil temperature of the transformer, an oil level of the transformer, a water-soluble acid value of the transformer, a volume resistivity of the transformer, a winding deformation level of the transformer, an on-load voltage regulator evaluation level of the transformer, a volume relay evaluation level of the transformer, a cooler evaluation level of the transformer, a temperature measuring device evaluation level of the transformer, a hydrogen content in oil of the transformer, a methane content in oil of the transformer, an ethane content in oil of the transformer, an ethylene content in oil of the transformer, an acetylene content in oil of the transformer, a carbon dioxide content in oil of the transformer, a carbon monoxide content in oil of the transformer, a total hydrocarbon content in oil of the transformer, a gas content in oil of the transformer, and a thunderstorm gale weather index of the transformer.

[0170] Those skilled in the art will appreciate that embodiments of the application can be supplied as methods, systems, or computer program products. Accordingly, the application can be embodied in the form of complete hardware embodiments, complete software embodiments, or embodiments combining software and hardware aspects. Furthermore, the application can be embodied in the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage media, etc.) having computer usable program code embodied thereon.

[0171] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 one or more flow or blocks

[0172] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 one or more flow or blocks

[0173] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 one or more flow or blocks

[0174] Finally, it should be noted that the above-mentioned embodiments are merely intended to illustrate the technical solutions of the present application, but not to limit the same. Although the present application has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalent replaced without departing from the spirit and scope of the present application, and any modification or equivalent replacement should be covered within the scope of protection of the claims of the present application.

Claims

1. A method of determining a key feature quantity characterizing a health state of a transformer, characterized by, The method comprises: determining a characteristic component analysis matrix of the transformer to be tested according to a characteristic quantity normalization matrix of the transformer to be tested; determining a characteristic component correlation analysis matrix between characteristic quantities of the transformer to be tested according to the characteristic component analysis matrix of the transformer to be tested; determining a key characteristic quantity representing the health state of the transformer to be tested according to the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be tested; the determining of the characteristic component analysis matrix of the transformer to be tested according to the characteristic quantity normalization matrix of the transformer to be tested comprises: The number of output layer nodes is 1 to S τ key feature extraction model is constructed in advance; obtain a pre-constructed S τ key feature extraction model corresponding to the 1 to S τ transformer health state calculation model respectively; The characteristic quantity normalization matrix of the transformer to be tested is respectively taken as the input quantity of the input layer of the S τ transformer health state calculation model, and the HI value of the transformer output by the S τ transformer health state calculation model is obtained. taking a difference between a HI value of the transformer output by a transformer health state calculation model corresponding to a key characteristic extraction model with an output layer node number of λ and a standard HI value of the transformer corresponding to the characteristic quantity normalization matrix of the transformer to be tested as a λth element in an HI error sequence corresponding to the characteristic quantity normalization matrix of the transformer to be tested, and generating the HI error sequence corresponding to the characteristic quantity normalization matrix of the transformer to be tested; taking a difference between each element in the HI error sequence corresponding to the characteristic quantity normalization matrix of the transformer to be tested and a first preset value to obtain an HI standard deviation sequence corresponding to the characteristic quantity normalization matrix of the transformer to be tested; starting from a first element in the HI standard deviation sequence corresponding to the characteristic quantity normalization matrix of the transformer to be tested, finding a sequence segment with first continuous element values less than 0, and obtaining a key characteristic extraction model corresponding to any element value in the sequence segment; substituting the characteristic quantity normalization matrix of the transformer to be tested into the key characteristic extraction model to obtain the characteristic component analysis matrix of the transformer to be tested; The input layer structure of the transformer health state calculation model corresponding to the key feature extraction model with the output layer node number of λ is a key feature extraction model with an output layer node number of λ pre-constructed, λ∈(1~S τ ), S τ is the feature quantity type number of the transformer.

2. The method of claim 1, wherein, a training process of the key characteristic extraction model with the output layer node number of λ pre-constructed comprises: step A: initializing θ=1, and setting an initial random configuration network model hidden layer node number as λ; Step B: during the θth training, an input weight matrix ω of the initial random configuration network model satisfying the first preset constraint condition is randomly generated λθ and an implicit layer node bias matrix b λθ ; Step C: taking the characteristic quantity normalization matrix X of the sample transformer as the input layer sample data of the initial random configuration network model, and setting the characteristic component analysis matrix of the sample transformer output by the hidden layer of the initial random configuration network model the output weight matrix of the initial random configuration network model training the initial random configuration network model to obtain the output matrix Y of the random configuration network model with the number of hidden layer nodes being λ and the output layer thereof constructed during the θth training θ ; Step D: calculating the error between matrix X and matrix Y θ ​ Step E: If θ = S ε When, then the sequence will be... minimum value The model structure between the input layer and the hidden layer of the randomly configured network model with λ hidden layer nodes built during the xth training period is used as the model structure between the input layer and the output layer of the key feature extraction model with λ output layer nodes, and the key feature extraction model with λ output layer nodes is obtained; otherwise, let θ = θ + 1, and return to step B; wherein b' λθ is a matrix of order λ x N, and b' λθ each row element of the matrix b λθ each row element of the matrix b is the bias of the λth hidden layer node randomly generated in the initial randomly configured network model during the θth training, is the least squares generalized inverse matrix of the transpose matrix of the matrix H λθ , T is the transpose symbol, x τi is the value of the τth feature quantity of the ith sample transformer, is the value of the τth feature quantity of the ith sample transformer output by the random configuration network model with λ hidden layer nodes constructed during the θth training, is the weight between the λth hidden layer node randomly generated in the initial random configuration network model and the τth feature quantity of the sample transformer during the θth training, is the difference between the ith sample transformer and the (i+1)th sample transformer, θ λ, τ ∈ (1~S τ ), i ∈ (1~N), N is the number of sample transformers, S ε is the first preset number of training.​ 3. The method of claim 2, wherein, the first preset constraint condition is determined as follows: In the formula, The difference between the output matrix of the key feature extraction model with the input / output layer node number of λ-1 and the characteristic quantity normalized matrix X of the sample transformer and the characteristic quantity of the i-th sample transformer in the characteristic quantity normalized matrix X is r, which is any value in the interval (0, 1). H λθ,λi is a matrix H λθ is the value of the element in the λth row and the ith column of matrix H g is any positive real number, μ p,L is a non-negative real number sequence is the Lth element in the sequence, and μ p,L satisfies the constraint condition lim λ→+∞ μ p,L = 0 and μ p,L ≤ 1-r, L ∈ (1~S L ), S L is the number of elements in the non-negative real number sequence.

4. The method of claim 1, wherein, a training process of a transformer health state calculation model corresponding to the key characteristic extraction model with the output layer node number of λ comprises: step 1: initializing δ=1, and setting an initial extreme learning machine network model hidden layer node number as P; Step 2: During the δth training, input weight matrix ω of the initial extreme learning machine network model is randomly generated λ,Pδ and hidden layer node bias matrix b λ,Pδ ; step 3: taking the key characteristic extraction model with the output layer node number of λ as an input layer structure of the initial extreme learning machine network model, and taking the characteristic quantity normalization matrix X of the sample transformer as input layer sample data of the initial extreme learning machine network model; Setting an output matrix of a hidden layer of an initial extreme learning machine network model Setting an output weight matrix of the initial extreme learning machine network model Training the initial extreme learning machine network model to obtain an output matrix H1 of an output layer of an extreme learning machine network model constructed during a δth training λ,δ ; Step 4: Compute matrix HI λ Error between matrix HI λ,δ and matrix H Step 5: When δ≠ S δ , let δ = δ + 1 and return to Step 2; otherwise, the limit learning machine network model constructed during the vth training is taken as the transformer health state calculation model corresponding to the key feature extraction model with the output layer node number λ. Step 5: When δ≠ S δ , let δ = δ + 1 and return to Step 2; otherwise, the limit learning machine network model constructed during the vth training is taken as the transformer health state calculation model corresponding to the key feature extraction model with the output layer node number λ. wherein, is the weight between the Pth hidden layer node in the initial extreme learning machine network model during the δth training and the λth feature component in the matrix H λx λx is the output matrix of the key feature extraction model after the characteristic quantity normalization matrix X of the sample transformer is substituted into the key feature extraction model with the output layer node number λ, HI λ = [HI1…HI i …HI N ] T , is the HI value corresponding to the characteristic quantity matrix of the ith sample transformer output by the extreme learning machine network model constructed during the δth training, i is the HI value corresponding to the characteristic quantity matrix of the ith sample transformer, T is a transpose symbol, b' λ,Pδ is a P×N order matrix, and b' λ,Pδ each row element value in the matrix b λ,Pδ is equal to each row element value in the matrix b is the bias of the Pth hidden layer node randomly generated in the initial extreme learning machine network model during the δth training, is the least square generalized inverse matrix of the transpose matrix of the matrix H λ,Pδ , i∈(1~N), N is the number of sample transformers, S δ is the second preset training number.​ 5. The method of claim 1, wherein, the determining of the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be tested according to the characteristic component analysis matrix of the transformer to be tested comprises: Computing a least squares generalized inverse of a transpose of a feature component analysis matrix of a transformer under test According to the matrix determining a transpose matrix of a characteristic component correlation analysis matrix between characteristic quantities of the transformer to be measured wherein X T is the transpose of the characteristic quantity normalization matrix of the transformer to be tested, T is the transpose symbol.

6. The method of claim 1, wherein, the determining of the key characteristic quantity representing the health state of the transformer to be tested according to the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be tested comprises: if the number of elements with values greater than a second preset threshold in the τth row elements of the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be tested exceeds a third preset threshold, the characteristic quantity corresponding to the τth row elements of the characteristic component correlation analysis matrix between the characteristic quantities of the transformer to be tested is the key characteristic quantity representing the health state of the transformer to be tested; where τ∈(1 ~ S τ ), S τ is the number of characteristic quantities of the transformer.

7. The method of claim 1, wherein, The transformer feature quantity includes: a winding direct current resistance phase difference of the transformer, a winding insulation resistance of the transformer, a winding absorption ratio of the transformer, a winding polarization index of the transformer, a winding dielectric loss of the transformer, a winding capacitance variation level of the transformer, a capacitive bushing dielectric loss of the transformer, a capacitive bushing insulation resistance of the transformer, a capacitive bushing capacitance value variation level of the transformer, a capacitive bushing end screen insulation resistance variation level of the transformer, a core grounding current of the transformer, a core insulation resistance of the transformer, a no-load loss variation level of the transformer, a no-load current variation level of the transformer, a load loss variation level of the transformer, an impedance voltage variation level of the transformer, a breakdown voltage in oil of the transformer, a load rate level of the transformer, a low voltage level of the transformer, a three-phase imbalance rate of the transformer, a breather silica gel color level of the transformer, a body appearance integrity degree of the transformer, a pollution degree of the transformer, a complete identification degree of the transformer, an oil tank leakage sealing inspection index of the transformer, a rust degree of the transformer, a vibration degree of the transformer, a noise degree of the transformer, a bushing appearance damage degree of the transformer, an oil tank appearance damage degree of the transformer, a grounding down lead appearance damage degree of the transformer, an oil appearance level of the transformer, an insulation oil color level of the transformer, a furfural content level in oil of the transformer, an oil sludge and sediment mass fraction level of the transformer, a bushing lead joint temperature of the transformer, a micro-water content in oil of the transformer, a flash point of the transformer, an oil temperature of the transformer, an oil level of the transformer, a water-soluble acid value of the transformer, a volume resistivity of the transformer, a winding deformation level of the transformer, an on-load voltage regulator evaluation level of the transformer, a volume relay evaluation level of the transformer, a cooler evaluation level of the transformer, a temperature measuring device evaluation level of the transformer, a hydrogen content in oil of the transformer, a methane content in oil of the transformer, an ethane content in oil of the transformer, an ethylene content in oil of the transformer, an acetylene content in oil of the transformer, a carbon dioxide content in oil of the transformer, a carbon monoxide content in oil of the transformer, a total hydrocarbon content in oil of the transformer, a gas content in oil of the transformer, and a lightning stroke and strong wind weather index of the transformer.

8. A system for determining key characteristic quantities representing a health state of a transformer, characterized in that The system comprises: A first determination module configured to determine a feature component analysis matrix of the transformer to be measured according to the feature quantity normalization matrix of the transformer to be measured; A second determination module configured to determine a feature component correlation analysis matrix between the feature quantities of the transformer to be measured according to the feature component analysis matrix of the transformer to be measured; A third determination module configured to determine a key feature quantity representing the health state of the transformer to be measured according to the feature component correlation analysis matrix between the feature quantities of the transformer to be measured; The first determination module comprises: The first obtaining unit is configured to obtain a pre-constructed key feature extraction model with an output layer node number of 1 to S τ . The second acquisition unit is configured to acquire S τ transformer health state calculation models respectively corresponding to the S τ key feature extraction models constructed in advance. The third acquisition unit is configured to acquire the characteristic quantity normalized matrix of the transformer to be tested as the input quantity of an input layer of S τ transformer health state calculation models, and acquire the HI value of the transformer output by the S τ transformer health state calculation models. A generation unit configured to generate an HI error sequence corresponding to the feature quantity normalization matrix of the transformer to be measured by taking, as the λth element in the HI error sequence corresponding to the feature quantity normalization matrix of the transformer to be measured, a difference between an HI value of the transformer output by a transformer health state calculation model corresponding to a key feature extraction model with an output layer node number λ and a standard HI value of the transformer corresponding to the feature quantity normalization matrix of the transformer to be measured. The fourth obtaining unit is configured to subtract each element in the HI error sequence corresponding to the characteristic quantity normalization matrix of the transformer to be tested from a first preset value to obtain an HI standard deviation value sequence corresponding to the characteristic quantity normalization matrix of the transformer to be tested; The fifth obtaining unit is configured to find a sequence segment with the first continuous element value less than 0 from the first element in the HI standard deviation value sequence corresponding to the characteristic quantity normalization matrix of the transformer to be tested, and obtain a key feature extraction model corresponding to any element value in the sequence segment; The sixth obtaining unit is configured to substitute the characteristic quantity normalization matrix of the transformer to be tested into the key feature extraction model to obtain a characteristic component analysis matrix of the transformer to be tested; The input layer structure of the transformer health state calculation model corresponding to the key feature extraction model with the output layer node number of λ is a key feature extraction model with an output layer node number of λ pre-constructed, λ∈(1~S τ ), S τ is the feature quantity type number of the transformer.

9. The system of claim 8, wherein, The second determining module comprises: A computing unit is configured to calculate a least square generalized inverse matrix of a transposed matrix of a characteristic component analysis matrix of a transformer to be measured The first determination unit is configured to determine a transposed matrix of a characteristic component correlation analysis matrix between characteristic quantities of the transformer under test according to the matrix The first determination unit is configured to determine a transposed matrix of a characteristic component correlation analysis matrix between characteristic quantities of the transformer under test according to the matrix wherein X T is the transpose of the characteristic quantity normalization matrix of the transformer under test, T being the transpose symbol.

10. The system of claim 8, wherein, The third determining module is configured to: If the number of elements with values greater than the second preset threshold in the τth row of the characteristic component correlation analysis matrix of the characteristic quantities of the transformer to be tested exceeds a third preset threshold, the characteristic quantity corresponding to the τth row of the characteristic component correlation analysis matrix of the characteristic quantities of the transformer to be tested is a key characteristic quantity representing the health state of the transformer to be tested; where τ∈(1 ~ S τ ), S τ is the number of characteristic quantities of the transformer.

Images