A method for identifying the operating status of converter transformer based on vibration signals

Through the variational modal decomposition algorithm and the depth limit learning machine, the accuracy of the operation state recognition of the converter transformer is solved, and the state recognition is realized under live operating conditions is improved, which is the safety and stability of the power system.

CN112307963BActive Publication Date: 2024-11-08MAINTENANCE & TEST CENTRE CSG EHV POWER TRANSMISSION CO
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202011190818.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-10-30
Publication Date
2024-11-08
Estimated Expiration
2040-10-30

AI Technical Summary

Technical Problem

It is difficult for the prior art to accurately identify the operating status of the converter transformer, especially when the equipment is live, the identification accuracy of the traditional vibration analysis method is low and cannot meet the safety and stability needs of the DC transmission system.

Method used

Using a vibration signal-based identification method, the acceleration signal is collected through multi-channel vibration sensors, and the signal feature extraction and classification is performed using a variational mode decomposition algorithm and a depth limit learning machine. A variable model with constraints is established to optimize parameters and optimize parameters. Build a deep limit learning machine network model to realize the operational state recognition of the converter transformer.

Benefits of technology

The characteristic vector of the vibration signal of the converter transformer is effectively extracted, which realizes accurate operating status recognition, facilitates maintenance and improves the safety and stability of the power system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112307963B_ABST
    Figure CN112307963B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for identifying the operating state of a converter transformer based on a vibration signal, which is used to identify the operating state of a converter transformer. The method collects the vibration signal of the converter transformer surface box, selects the variational mode decomposition algorithm as the signal feature extraction method, completes the parameter optimization work of the variational mode decomposition algorithm through the beetle whisker search algorithm, determines the optimal parameters of the decomposition, and uses the optimized variational mode decomposition algorithm to perform modal decomposition and reconstruction on the measured converter transformer surface box vibration signal, calculates the energy proportion of each reconstructed signal after decomposition in the original signal and constructs the corresponding state feature vector. On the basis of the extreme learning machine, an autoencoder is constructed and used as a basic unit to establish a deep extreme learning machine network model, the state feature vector is divided into a training set and a test set to import into the deep extreme learning machine network model, and the effective identification of the operating state of the converter transformer is achieved through model training and testing.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of power systems, and in particular to a method for identifying the operating state of a converter transformer based on vibration signals. Background Art

[0002] The converter transformer is the core equipment to ensure the stable operation of the converter station and is an important part of the ultra-high voltage direct current transmission project. Whether the converter transformer can work safely and stably is directly related to the operational reliability of the entire direct current transmission system. When the converter transformer fails, it will cause the converter valve to lock, resulting in partial or even large-scale power outages in the power system, causing significant losses to the development of the social economy. At present, several direct current converter stations have been completed and put into operation at home and abroad. According to the statistics of CIGRE, about 60% of the faults in the direct current transmission system are caused by converter transformer failures, among which mechanical failures caused by windings, cores and tap switches in the converter transformer account for about 50% of the converter transformer failures. According to the statistical results of CIGRE, the failure rate of converter transformers is increasing every year, most of which are caused by abnormalities in the mechanical structure of the converter transformer. Therefore, it is very important to identify the operating status of the mechanical components inside the converter transformer, which is of great significance for improving the status maintenance level of direct current transmission equipment and ensuring the safe and reliable operation of the direct current transmission system.

[0003] Currently, there are few methods for monitoring the status of converter transformers. If the operating status of the mechanical components inside the converter transformer cannot be identified in time, it will be difficult to find potential safety hazards inside the converter transformer, posing a threat to the safe and stable operation of the DC transmission system. Most of the existing converter transformer status identification methods are based on electrical signals, and require the converter transformer to be shut down for maintenance. It is not possible to carry out detection when the equipment is energized. Conventional planned maintenance and dismantling inspections consume a lot of manpower, material and financial resources, causing power outage losses and also affecting the life of the equipment.

[0004] Vibration analysis is a state recognition method that can be realized outside the transformer body. It can use sensors to obtain vibration signals during the operation of the converter transformer at different times and locations, extract the time domain and frequency domain feature information of the vibration signal, identify the operating state of the transformer, and promptly discover the latent defects that cause abnormal vibration of the transformer. However, at this stage, transformer vibration technology research is mainly focused on AC transformers. There are obvious differences between the vibration characteristics of converter transformers and ordinary AC transformers, and the signal characteristics are more complex and changeable. The signal analysis method used by the traditional vibration analysis method has a low recognition accuracy and is difficult to meet the current needs for converter transformer state recognition. It is necessary to introduce a new method to realize converter transformer state recognition based on vibration signals. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a method for identifying the operating state of a converter transformer based on a vibration signal, which is used to identify the operating state of the converter transformer.

[0006] To achieve the above object, the technical solution of the present invention is as follows:

[0007] A method for identifying the operating state of a converter transformer based on vibration signals is implemented by the following steps:

[0008] Collect vibration acceleration signals of converter transformers;

[0009] Dividing the vibration acceleration signal into a training sample set and a test sample set; wherein the training sample set is used to seek the optimal set number and penalty factor of the modal function;

[0010] According to the variational mode decomposition algorithm, using the optimal set number of the modal functions and the penalty factor, a variational model with constraints is established;

[0011] Extracting energy features from the constrained variational model to obtain a vibration sample signal feature set and a test sample signal feature set, and training the vibration sample signal feature set through a deep extreme learning machine to obtain a trained model;

[0012] The test sample signal feature set is input into the trained model to obtain a classification result of the vibration acceleration signal, and the operating state of the converter transformer is identified according to the classification result.

[0013] The method for identifying the operating state of a converter transformer based on a vibration signal according to claim 1 is characterized in that the method for seeking the optimal set number of modal functions and penalty factors using a training sample set is as follows:

[0014] Establish the objective function model;

[0015] Initialize the starting position for optimization in the parameter space;

[0016] Convert the position in parameter space into the set number of pre-set mode functions and penalty factors;

[0017] Using the preset number of sets of modal functions and penalty factors, processing is performed according to the variational modal decomposition algorithm;

[0018] Obtain the objective function value of all training sample sets;

[0019] Determine whether the objective function value reaches the maximum; if so, use the preset number of modal function sets and penalty factors as the optimal number of modal function sets and penalty factors; if not, use the beetle whisker search algorithm to update the position of the next parameter space, convert the updated parameter space position into the preset number of modal function sets and penalty factors, and repeat the above steps until the optimal number of modal function sets and penalty factors are obtained.

[0020] The above-mentioned method for identifying the operating state of a converter transformer based on a vibration signal, further, the acquisition of the vibration acceleration signal of the converter transformer is specifically performed by the following method:

[0021] A multi-channel vibration sensor is installed on the surface of the converter transformer box; the vibration acceleration signal output from the surface of the converter transformer box is transmitted to a computer through an integrated measuring device, and the state of the collected vibration acceleration signal is identified in the computer.

[0022] The above-mentioned method for identifying the operating state of a converter transformer based on vibration signals, further, the variational mode decomposition algorithm comprises the following steps:

[0023] Each sub-signal is defined as a modal function U i (t), the original signal can be expressed as:

[0024]

[0025] Among them, K represents the number of modes, A i (t) represents the instantaneous amplitude, It represents the phase, instantaneous frequency

[0026] Variational model with constraints:

[0027] Among them: i ] represents the set of K modal functions, [ω i ] represents the set of corresponding center frequencies, represents the gradient operation, and R(t) represents the real part of the modal function;

[0028] In order to obtain the optimal solution of the model, the Lagrange multiplication operator λ and the penalty factor α are used for the constrained variation problem.

[0029] Turn it into the unconstrained case and build the augmented Lagrangian function:

[0030]

[0031] The saddle point of the above equation is obtained by the alternating direction multiplier method, and the modal components in the frequency domain can be obtained by converting the time domain signal into the frequency domain signal using Fourier transform. and center frequency Expression of

[0032]

[0033]

[0034] Iterative update during the operation ω i as well as The optimal solution can be obtained. The specific process is as follows:

[0035] 1) Initialization and λ 1 =n=0;

[0036] 2) Iterative update based on formula (4) and ω i ;

[0037] 3) Update the Lagrange multiplication operator based on formula (6) Where τ represents the update parameter of the Lagrange multiplication operator;

[0038]

[0039] 4) Take equation (7) as the constraint condition for stopping the operation, and repeat steps (1) and (3) until it is established. In the equation, ε represents the accuracy of the iteration. When the iteration is completed, the result is the modal component and center frequency of the variational mode decomposition.

[0040]

[0041] The above-mentioned method for identifying the operating state of a converter transformer based on vibration signals, further, the beetle whisker search algorithm comprises the following steps:

[0042] The position of the longicorn can be defined as s n (n=1,2,···,N), where N represents the number of iterations of the search. The food odor is composed of the position-related objective function F(s), whose extreme value is the final position of the food;

[0043] The global optimization of the algorithm in multidimensional data space can be expressed as follows:

[0044]

[0045] The global optimization of the algorithm can be summarized as exploration behavior and action behavior. In the former, the direction of the beetle's whiskers can be represented by a normalized random vector D.

[0046]

[0047] In formula (9), Random is a random function; d is the dimension of the data space. The left and right tentacle positions s l and r It can be defined as

[0048]

[0049] In the formula, a n is the length of the antennae; and the walking behavior is to determine the step length and direction of the beetle according to the size relationship of the objective function values ​​corresponding to the positions of the left and right antennae, thereby updating the position of the beetle; its updated position can be expressed as

[0050] s n =s n-1 +ξ n Dsign[f(x r )-f(x l )] (11)

[0051] In formula (11), ξ n is the step length of each iteration, sign is the sign function; each execution of exploration behavior and action behavior can be defined as an iteration process of the algorithm, and after each iteration a n and n Need to be updated, the update formula is

[0052]

[0053] In formula (12), the update coefficients c1 and c3 are set to 0.95, and c2 is set to 0.01. After the algorithm reaches the preset number of iterations, the data space position corresponding to the best objective function value in all iterations is the optimization result.

[0054] The above-mentioned converter transformer operating state identification method based on vibration signal, further, the parameter optimization objective function model F(s) is implemented by the following steps:

[0055] Assuming the length of a vibration signal y(t) is L, if the proportion of the reconstructed signal energy after variational mode decomposition in the original signal energy is defined as E, then its mathematical model can be expressed as

[0056]

[0057] In formula (13), U i(t) represents the reconstructed signal amplitude of the i-th mode, and K represents the number of modes;

[0058] The center frequency variation range of the modal bandwidth estimation is called b, which can be defined as

[0059]

[0060] Where: max and min are functions for finding the maximum and minimum values ​​respectively; f s is the signal sampling frequency; ω i (n) is the center frequency estimated in the iterative process for the i-th mode; M is the number of center frequencies estimated for this mode;

[0061]

[0062] Where s = [p, q] (p∈K, q∈α) is any parameter combination in the optimization space; N is the number of qualified training samples; E j and b j are E and b corresponding to the j-th training sample vibration signal; 0.5 is the weight coefficient that can control the calculation result to be between [0, 100].

[0063] The above-mentioned converter transformer operating state identification method based on vibration signal, further, the construction process of the deep extreme learning machine is as follows:

[0064] Build an Extreme Learning Machine:

[0065] Get T different labeled samples, Represents the i-th sample, corresponding to label l i =[l iz ,l iz ,···,l iz ] T , where z represents the number of categories; if the sample Belongs to the mth category, then the mth value of the corresponding label vector is set to 1, and the remaining (m-1) values ​​are set to -1; the extreme learning machine with H hidden nodes can be expressed by formula (16):

[0066]

[0067] Among them, Q i =[Q i1 ,Q i2 ,···,Q iz ] T It represents the output weight connecting the i-th hidden layer and the output layer, A i =[A i1 ,A i2 ,···,A iz] T is the input weight connecting the input layer and the i-th hidden layer, B i is the bias of the i-th hidden layer, G(A i ,B i ,E j ) is the output of the i-th hidden layer; assuming that the activation function of the i-th hidden layer is g(x), the output of the hidden layer is:

[0068] G(A i ,B i ,E j )=g(A i ·E j +B i ) (17)

[0069] Formula (16) can be summarized in matrix form as

[0070]

[0071] in,

[0072] In order to meet the expected value of the output after network training, it is necessary to find the parameter A i , B i , Q i The optimal value of So that:

[0073]

[0074] Then the minimization loss function is

[0075]

[0076] In the extreme learning machine, the parameter (A i ,B i ) is randomly initialized, so Q i is uniquely determined; its solution can be transformed into:

[0077]

[0078] The norm of the solution Q is the smallest and unique. In order to make the model have better generalization performance, the L1 regularization term can be added, and the above solution problem can be transformed into formula (21):

[0079]

[0080] Among them, C L is the regularization coefficient; the solution is essentially a ridge regression problem, and its solution is shown in the following formula:

[0081]

[0082] Where I is the identity matrix;

[0083] Construct an autoencoder based on extreme learning machine: Apply the idea of ​​autoencoder to extreme learning machine so that the input data is also used for output, that is, output Y = X;

[0084] The output of the extreme learning machine-based autoencoder can be transformed from equation (16) into:

[0085]

[0086] Among them, A T A=I,B T B = 1;

[0087] Among them, A is A i The matrix composed of B is i For sparse expression and dimension compression, the hidden layer output weight Q can be transformed from formula (23) into:

[0088]

[0089] Where E = [E1, E2, ···, E8] is the feature vector composed of the energy proportions of each mode after variational mode decomposition of the input vibration signal; for equal-dimensional feature mapping, the weight Q can be calculated by formula (26):

[0090] Q=C -1 l (26)

[0091] Among them, Q T Q = I;

[0092] Building a Deep Extreme Learning Machine:

[0093] Based on the feature representation ability of the extreme learning machine-based autoencoder, it is used as the basic unit of the deep extreme learning machine;

[0094] The input data sample E is used as the target output of the first extreme learning machine-based autoencoder, and then the output weight Q1 is obtained; then the output matrix C1 of the first hidden layer of the deep extreme learning machine is used as the input and target output of the next autoencoder based on the extreme learning machine, and so on. The last layer is trained with the extreme learning machine, and the output weight Q of the last hidden layer of the deep extreme learning machine is solved using formula (21): i+1 .

[0095] Compared with the prior art, the present invention has the following beneficial effects: when the converter transformer is in different operating states, its vibration signal will change, but due to the complex internal structure of the converter transformer, it is difficult to determine its significant features using the traditional vibration analysis method, so that accurate classification and identification cannot be performed. After decomposing the vibration signal of the converter transformer using the optimized variational mode decomposition algorithm, the effective feature vector of the vibration signal can be effectively obtained, thereby accurately classifying the signal using a deep extreme learning machine to determine the operating state of the converter transformer, facilitating maintenance and improving the safety and stability of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0097] Figure 1 It is a converter transformer vibration measurement system;

[0098] Figure 2 It is a layout diagram of multi-channel sensors;

[0099] Figure 3 It is the structure diagram of the extreme learning machine;

[0100] Figure 4 It is the ELM-AE structure diagram;

[0101] Figure 5 It is a deep extreme learning machine model diagram;

[0102] Figure 6 is the overall method flow chart;

[0103] Figure 7 This is the result of deep extreme learning machine training;

[0104] Figure 8 This is the test result graph of the deep extreme learning machine. DETAILED DESCRIPTION

[0105] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0106] Example:

[0107] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations of the embodiments of the present invention are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0108] See also Figures 1 to 8 , Figure 1 It is a converter transformer vibration measurement system; Figure 2 It is a layout diagram of multi-channel sensors; Figure 3 It is the structure diagram of the extreme learning machine; Figure 4 It is the ELM-AE structure diagram; Figure 5 It is a deep extreme learning machine model diagram; Figure 6 This is the overall method flow chart Figure 7 This is the result of deep extreme learning machine training. Figure 8 This is the test result graph of the deep extreme learning machine.

[0109] The purpose of the present invention is to propose a method for identifying the operating state of a converter transformer based on a vibration signal, which is used to identify the operating state of a converter transformer. The method adopts a multi-channel vibration acceleration sensor, and obtains the vibration signal of the converter transformer surface box through multi-measurement point layout measurement. The signal feature extraction method selects the variational mode decomposition algorithm, and completes the parameter optimization work of the variational mode decomposition algorithm through the longhorn beard search algorithm, determines the optimal parameters of the decomposition, and uses the optimized variational mode decomposition algorithm to perform modal decomposition and reconstruction on the measured converter transformer surface box vibration signal, calculates the energy proportion of each reconstructed signal after decomposition in the original signal, and constructs the corresponding state feature vector. On the basis of the extreme learning machine, an autoencoder is constructed and used as a basic unit to establish a deep extreme learning machine network model. The state feature vector is divided into a training set and a test set to import the deep extreme learning machine network model, and the effective identification of the operating state of the converter transformer is achieved through model training and testing.

[0110] The present invention proposes a method for identifying the operating state of a converter transformer based on a vibration signal, which is characterized by being implemented by the following steps:

[0111] (1) Measurement and collection of vibration signals

[0112] The multi-channel vibration sensor is installed on the surface of the converter transformer box. The vibration acceleration signal it outputs is transmitted to the computer through the integrated measurement device. The entire vibration measurement system is as follows: Figure 1 The installation position of the vibration sensor on the surface of the converter transformer box is as shown in Figure 2 shown.

[0113] (2) Extraction of vibration signal energy features

[0114] ①Variational Mode Decomposition (VMD) algorithm

[0115] Variational mode decomposition is an adaptive quasi-orthogonal decomposition technique that has emerged in recent years. The algorithm can decompose the original signal X(t) containing multiple signal components and decompose it into a finite number of sub-signals, and these sub-signals have special sparsity and can realize amplitude and frequency modulation. Each sub-signal is defined as a mode function U i (t), whose sparsity is generally determined by the bandwidth of the frequency domain, then the original signal can be expressed as:

[0116]

[0117] Among them, K represents the number of modes, that is, the decomposition scale of the signal, and A i (t) represents the instantaneous amplitude, Indicates the phase. Instantaneous frequency Compared with the phase, the instantaneous frequency and instantaneous amplitude change at a slower rate, so the modal function U i (t) can be regarded as a harmonic signal in the range of [t-δ, t+δ], and its instantaneous amplitude is A i (t), the center frequency is ω i (t).

[0118] The essence of the variational modal decomposition algorithm is to establish a variational model with constraints, and to find the optimal solution of the model under the modal estimation bandwidth and the minimum constraints, so as to determine each modal component and its corresponding center frequency. In order to effectively estimate the modal bandwidth, it is first necessary to perform Hilbert transform on each modal component to obtain its unilateral spectrum, use the exponential correction method to modulate the center frequency obtained by the modal estimation to the baseband, and use the Gaussian smoothing method to obtain the bandwidth of each frequency band, so that the whole process is defined as a variational problem with bandwidth constraints, and its mathematical model is:

[0119]

[0120] Among them: i ] represents the set of K modal functions, [ω i ] represents the set of corresponding center frequencies, represents the gradient operation, and R(t) represents the real part of the modal function.

[0121] In order to obtain the optimal solution of the model, the Lagrangian multiplication operator λ and the penalty factor α are used for the constrained variational problem, which is transformed into an unconstrained case and the augmented Lagrangian function is established.

[0122]

[0123] The saddle point of the above equation is obtained by the alternating direction multiplier method, and the modal components in the frequency domain can be obtained by converting the time domain signal into the frequency domain signal using Fourier transform. and center frequency expression.

[0124]

[0125]

[0126] Iterative update during the operation ω i as well as The optimal solution can be obtained. The specific process is as follows:

[0127] 1) Initialization and λ 1 =n=0;

[0128] 2) Iterative update based on formula (4) and ω i ;

[0129] 3) Update the Lagrange multiplication operator based on formula (6) Where τ represents the update parameter of the Lagrange multiplication operator;

[0130]

[0131] 4) Take equation (7) as the constraint condition for stopping the operation, and repeat steps (1) and (3) until it is established. In the equation, ε represents the accuracy of the iteration. When the iteration is completed, the result is the modal component and center frequency of the variational mode decomposition.

[0132]

[0133] From the principle of variational mode decomposition, it can be seen that multiple parameters need to be preset before performing the decomposition operation. Among them, τ and ε mainly determine the number of iterations and accuracy, while K and α determine the effect of decomposition. When analyzing signals, it is necessary to screen the optimal [K, α]. Therefore, this patent introduces the beetle whisker search algorithm for optimization.

[0134] ② Optimization of decomposition algorithm

[0135] The principle of the longhorn beetle whisker search algorithm imitates the two equidistant whiskers on the left and right of the longhorn beetle's head when foraging, randomly exploring the intensity of the smell of nearby food, and then moving in the direction of the tentacle with a stronger food smell, and determining the location of the food through a cycle until the smell is the strongest point.

[0136] When constructing the mathematical model of the algorithm, the position of the longhorn beetle can be defined as s n (n=1,2,···,N), where N represents the number of search iterations. The food smell is composed of the objective function F(s) related to the position, and its extreme value is the final position of the food. The global optimization of the algorithm in the multidimensional data space can be expressed as follows:

[0137]

[0138] The global optimization of the algorithm can be summarized as exploration behavior and action behavior. In the former, the direction of the beetle's whiskers can be represented by a normalized random vector D

[0139]

[0140] In formula (9), Random is a random function; d is the dimension of the data space. The left and right tentacle positions s l and r It can be defined as

[0141]

[0142] In the formula, a n is the length of the antenna. The walking behavior is to determine the step length and direction of the longhorn beetle based on the size relationship of the objective function values ​​corresponding to the positions of the left and right antennae, and then update the position of the longhorn beetle. Its updated position can be expressed as

[0143] s n =s n-1 +ξ n Dsign[f(x r )-f(x l )] (11)

[0144] In formula (11), ξ n is the step length of each iteration, and sign is the sign function. Each execution of exploration behavior and action behavior can be defined as an iteration of the algorithm. After each iteration, a n and n Need to be updated, the update formula is

[0145]

[0146] In formula (12), the update coefficients c1 and c3 are generally set to 0.95, and c2 is generally set to 0.01. After the algorithm reaches the preset number of iterations, the data space position corresponding to the best objective function value in all iterations is the optimization result.

[0147] ③Construction of feature vector

[0148] The objective function used in the optimization algorithm for the beetle whisker search needs to fully reflect the characteristics of the converter transformer vibration signal under variational modal decomposition, and its extreme value should also represent the best average decomposition effect. According to the theory, variational modal decomposition can divide the signal into several components existing in different bandwidths, which naturally distributes the signal energy to each mode, and each modal bandwidth can determine the estimation effect of the center frequency. Therefore, the change in the center frequency of the modal energy and bandwidth estimation can represent the decomposition performance of the variational modal decomposition, so the objective function is designed based on this principle.

[0149] First, assume that the length of a vibration signal y(t) is L. If the proportion of the reconstructed signal energy after variational mode decomposition in the original signal energy is defined as E, then its mathematical model can be expressed as:

[0150]

[0151] In formula (13), U i (t) represents the reconstructed signal amplitude of the i-th mode, and K represents the number of modes. The larger the E, the closer the decomposed signal is to the original signal and the more sufficient the original signal information is.

[0152] Secondly, each mode has a certain bandwidth, which can determine the center frequency. If the bandwidth span is lower, the fluctuation of the center frequency during the iteration process is smaller, the signal resolution is stronger, and the decomposition is more complete. Therefore, the center frequency variation range of the modal bandwidth estimation is called b, which can be defined as

[0153]

[0154] Where: max and min are functions for finding the maximum and minimum values ​​respectively; f s is the signal sampling frequency; ω i (n) is the estimated center frequency of the i-th mode during the iteration process; M is the number of estimated center frequencies under this mode. The larger the b, the smaller the percentage of the estimated center frequency fluctuation relative to the entire signal frequency range, that is, the narrower the bandwidth and the finer the decomposition.

[0155] Therefore, the sum of E and b of each signal represents the balance between the degree of decomposition and the amount of signal information in the variational mode decomposition processing of the signal. The larger the result, the more complete the decomposition, the more information, and the better the signal processing performance; and the average result of the sum of E and b of all signals in the entire qualified training sample set represents the degree of commonality that the decomposition performance of the overall sample can achieve. Each set of parameters corresponds to an average result, and the maximum value of the average result in the parameter space reflects the highest average processing performance of the variational mode decomposition of the overall sample. The variational mode decomposition unified parameters suitable for each signal are the parameter combination corresponding to the maximum value.

[0156] In summary, the objective function F(s) of parameter optimization can be summarized as follows

[0157]

[0158] Where s = [p, q] (p∈K, q∈α) is any parameter combination in the optimization space; N is the number of qualified training samples; E j and b j are E and b corresponding to the vibration signal of the jth training sample; 0.5 is the weight coefficient that can control the calculation result to be in [0, 100]. According to past historical experience, the frequency spectrum of the vibration signal of the converter transformer is mainly distributed in 0-800Hz, so the K value is set to 8, and the optimal α is calculated according to the objective function, and the characteristic vector E = [E1, E2, ···, E K ](K=8).

[0159] (3) Classification and recognition of vibration signals

[0160] ① Basic principles of extreme learning machine (ELM)

[0161] Extreme learning machine is a supervised learning method, which belongs to a single hidden layer feedforward neural network. The input weights and biases of the extreme learning machine network training are randomly generated, which makes it have a faster learning speed, and can randomly map the input data to the Hilbert space, with strong generalization ability. Its basic network structure is as follows Figure 3 shown.

[0162] (4) obtains T different labeled samples, Represents the i-th sample, corresponding to label l i =[l iz ,l iz ,···,l iz ] T , where z represents the number of categories. If it belongs to the mth class, the mth value of the corresponding label vector is set to 1, and the remaining (m-1) values ​​are set to -1. The extreme learning machine with H hidden nodes can be expressed by formula (16):

[0163]

[0164] Among them, Q i =[Q i1 ,Q i2 ,···,Q iz ] T It represents the output weight connecting the i-th hidden layer and the output layer, A i =[A i1 ,A i2 ,···,A iz ] T is the input weight connecting the input layer and the i-th hidden layer, B i is the bias of the i-th hidden layer, G(A i ,B i ,E j ) is the output of the i-th hidden layer. Assuming that the activation function of the i-th hidden layer is g(x), the output of the hidden layer is:

[0165] G(A i ,B i ,E j )=g(A i ·E j +B i ) (17)

[0166] Formula (16) can be summarized in matrix form as

[0167]

[0168] in,

[0169] In order to meet the expected value of the output after network training, it is necessary to find the parameter A i , B i , Q i The optimal value of So that:

[0170]

[0171] Since the goal of the extreme learning machine is to minimize the error between the actual output and the expected output, the minimization loss function is

[0172]

[0173] In the extreme learning machine, the parameter (A i ,B i ) is randomly initialized, so Q i is uniquely determined. Its solution can be transformed into:

[0174]

[0175] The norm of the solution Q is the smallest and unique. In order to make the model have better generalization performance, the L1 regularization term can be added, and the above solution problem can be transformed into formula (21):

[0176]

[0177] Among them, C L is the regularization coefficient. The solution is essentially a ridge regression problem, and its solution is shown as follows:

[0178]

[0179] Where I is the identity matrix.

[0180] ②Extreme Learning Machine-Based Autoencoder (ELM-AE)

[0181] An autoencoder (AE) can be trained to copy input to output. Since no labeled data is required, training an autoencoder is unsupervised. Therefore, the idea of ​​an autoencoder is applied to an extreme learning machine so that the input data is also used for the output, that is, the output Y = X. The network structure of an autoencoder based on an extreme learning machine is shown in the figure below. Figure 4 Shown

[0182] like Figure 2 In the example, if w>H, ELM-AE realizes dimensional compression, mapping high-dimensional data into low-dimensional feature expression; if w=H, ELM-AE realizes equal-dimensional feature expression; if w<H, ELM-AE realizes sparse expression, that is, high-dimensional feature expression of the original data. In summary, ELM-AE is a universal approximator, which makes the output of the network the same as the input, and the input parameters of the hidden layer (A i ,B i ) are randomly generated and then orthogonalized.

[0183] The output of ELM-AE can be transformed from equation (16) into:

[0184]

[0185] Among them, A T A=I,B T B=1.

[0186] Among them, A is A i The matrix composed of B is i For sparse expression and dimension compression, the hidden layer output weight Q can be transformed from formula (23) into:

[0187]

[0188] Where E = [E1, E2, ···, E8] is the feature vector composed of the energy proportion of each mode after the variational mode decomposition of the input vibration signal. For equal-dimensional feature mapping, the weight Q can be calculated by formula (26):

[0189] Q=C -1 l (26)

[0190] Among them, Q T Q=I.

[0191] ③Deep Extreme Learning Machine (DELM)

[0192] Based on the feature representation ability of ELM-AE, it is used as the basic unit of deep extreme learning machine. Similar to traditional deep learning algorithms, deep extreme learning machine also uses layer-by-layer greedy training method to train the network. The input weight of each hidden layer of deep extreme learning machine is initialized with ELM-AE, and layer-by-layer unsupervised training is performed. However, unlike traditional deep learning algorithms, ELM-AE does not require reverse fine-tuning process. The idea of ​​deep extreme learning machine is to minimize the reconstruction error so that the output can be infinitely close to the original input. After training each layer, the high-level features of the original data can be learned. Figure 5 The training process of the deep extreme learning machine model is described. The input data sample E is used as the target output of the first ELM-AE (E1=E), and then the output weight Q1 is obtained; then the output matrix C1 of the first hidden layer of the deep extreme learning machine is used as the input and target output of the next ELM-AE (E2=C1), and so on. The last layer is trained with ELM, and the output weight Q of the last hidden layer of the deep extreme learning machine is solved using formula (21) i+1 . Figure 5 Middle C i+1 is the output matrix of the last hidden layer, and l is the sample label. The input weight matrix of each hidden layer of the deep extreme learning machine is Therefore, the operating status of the converter transformer can be effectively identified by inputting the feature vector after the vibration signal is decomposed into the deep extreme learning machine and classified. The overall method process is as follows: Figure 6 shown.

[0193] With the support of the above theory, the vibration signal of the converter transformer is measured to obtain a set of vibration signals. The optimized variational mode decomposition algorithm is used for decomposition, and the eigenvectors obtained are shown in Table 1.

[0194] Table 1 Characteristic vectors of vibration signals of converter transformers

[0195]

[0196] It can be seen from the results that the eigenvalues ​​of different components in the eigenvector of the separated vibration signal are significantly different, which is suitable for signal classification. Therefore, a total of 1834 samples of the vibration signal of the converter transformer were collected. The training set sample set was divided into a ratio of 8:2, and the corresponding labels were calibrated according to the operating status of the converter transformer. The training set was input into the DELM model for training, and the training results are shown in Figure 2. Figure 7 As shown in the figure, after 100 iterations, the classification accuracy of the training set reached 99.4548%, indicating that the training results are good. Then the test set is input into the trained model, and the test results are as follows: Figure 8 As shown in the figure, the classification accuracy of the test set reached 99.4536%, indicating that the model can effectively realize the recognition of the operating status of the converter transformer based on the vibration signal.

[0197] In summary, when the converter transformer is in different operating states, its vibration signal will change. However, due to the complex internal structure of the converter transformer, it is difficult to determine its significant features using traditional vibration analysis methods, making it impossible to accurately classify and identify it. After decomposing the converter transformer vibration signal using the optimized variational mode decomposition algorithm, the effective feature vector of the vibration signal can be effectively obtained. Therefore, the deep extreme learning machine can be used to accurately classify the signal to determine the operating state of the converter transformer, which is convenient for maintenance and improves the safety and stability of the power system.

[0198] The method of the present invention is based on a variational mode decomposition optimization algorithm and a deep extreme learning machine, and realizes state recognition of a converter transformer by analyzing vibration signals, thereby reducing the risk of abnormal operation of the converter transformer and improving the safety and reliability of power system operation.

[0199] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable ordinary technicians in the field to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made based on the essence of the content of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for identifying the operating state of a converter transformer based on vibration signals, characterized in that: This is accomplished by the following steps: Collect vibration acceleration signals of converter transformers; Dividing the vibration acceleration signal into a training sample set and a test sample set; wherein the training sample set is used to seek the optimal set number and penalty factor of the modal function; According to the variational mode decomposition algorithm, using the optimal set number of the modal functions and the penalty factor, a variational model with constraints is established; Extracting energy features from the constrained variational model to obtain a vibration sample signal feature set and a test sample signal feature set, and training the vibration sample signal feature set through a deep extreme learning machine to obtain a trained model; The test sample signal feature set is input into the trained model to obtain a classification result of the vibration acceleration signal, and the operating state of the converter transformer is identified according to the classification result, wherein: The method of using the training sample set to find the optimal set number of modal functions and the penalty factor is as follows: Establish the objective function model; Initialize the starting position for optimization in the parameter space; Convert the position in parameter space into the set number of pre-set mode functions and penalty factors; Using the preset number of sets of modal functions and penalty factors, processing is performed according to the variational modal decomposition algorithm; Obtain the objective function value of all training sample sets; Determine whether the objective function value reaches the maximum; if so, use the preset set number of modal functions and penalty factor as the optimal set number of modal functions and penalty factor; if not, use the beetle whisker search algorithm to update the position of the next parameter space, convert the updated position of the parameter space into the preset set number of modal functions and penalty factor, and repeat the above steps until the optimal set number of modal functions and penalty factor are obtained; The variational mode decomposition algorithm comprises the following steps: Each sub-signal is defined as a modal function U i (t), the original signal can be expressed as: Among them, K represents the number of modes, A i (t) represents the instantaneous amplitude, It represents the phase, instantaneous frequency Variational model with constraints: Among them: i ] represents the set of K modal functions, [ω i ] represents the set of corresponding center frequencies, represents the gradient operation, and R(t) represents the real part of the modal function; In order to obtain the optimal solution of the model, the Lagrangian multiplication operator λ and the penalty factor α are used for the constrained variation problem, which is transformed into an unconstrained case and the augmented Lagrangian function is established: The saddle point of the above equation is obtained by the alternating direction multiplier method, and the modal components in the frequency domain can be obtained by converting the time domain signal into the frequency domain signal using Fourier transform. and center frequency Expression of Iterative update during the operation ω i as well as The optimal solution can be obtained. The specific process is as follows: 1) Initialization and λ 1 =n=0; 2) Iterative update based on formula (4) and ω i ; 3) Update the Lagrange multiplication operator based on formula (6) Where τ represents the update parameter of the Lagrange multiplication operator; 4) Take equation (7) as the constraint condition for stopping the operation, and repeat steps (1) and (3) until it is established. In the equation, ε represents the accuracy of the iteration. When the iteration is completed, the result is the modal component and center frequency of the variational mode decomposition. The longicorn beard search algorithm comprises the following steps: The position of the longicorn can be defined as s n (n=1,2,···,N), where N represents the number of iterations of the search. The food odor is composed of the position-related objective function F(s), whose extreme value is the final position of the food; The global optimization of the algorithm in multidimensional data space can be expressed as follows: The global optimization of the algorithm can be summarized as exploration behavior and action behavior. In the former, the direction of the beetle's whiskers can be represented by a normalized random vector D. In formula (9), Random is a random function; d is the dimension of the data space; and the left and right tentacle positions s l and r It can be defined as In the formula, a n is the length of the antennae; and the walking behavior is to determine the step length and direction of the beetle according to the size relationship of the objective function values ​​corresponding to the positions of the left and right antennae, thereby updating the position of the beetle; its updated position can be expressed as s n =s n-1 +ξ n Dsign[f(x r )-f(x l )] (11) In formula (11), ξ n is the step length of each iteration, sign is the sign function; each execution of exploration behavior and progress behavior is defined as an iteration process of the algorithm, and after each iteration a n and n Need to be updated, the update formula is In formula (12), the update coefficients c1 and c3 are set to 0.95, and c2 is set to 0.

01. After the algorithm reaches the preset number of iterations, the data space position corresponding to the best objective function value in all iterations is the optimization result. The objective function model F(s) for parameter optimization is implemented using the following steps: Assuming the length of a vibration signal y(t) is L, if the proportion of the reconstructed signal energy after variational mode decomposition in the original signal energy is defined as E, then its mathematical model can be expressed as In formula (13), U i (t) represents the reconstructed signal amplitude of the i-th mode, and K represents the number of modes; The center frequency variation range of the modal bandwidth estimation is called b, which can be defined as Where: max and min are functions for finding the maximum and minimum values ​​respectively; f s is the signal sampling frequency; ω i (n) is the center frequency estimated in the iterative process for the i-th mode; M is the number of center frequencies estimated for this mode; Where s = [p, q] (p∈K, q∈α) is any parameter combination in the optimization space; N is the number of qualified training samples; E j and b j are E and b corresponding to the j-th training sample vibration signal; 0.5 is the weight coefficient that can control the calculation result to be between [0, 100].

2. The method for identifying the operating state of a converter transformer based on a vibration signal according to claim 1, characterized in that: The acquisition of the vibration acceleration signal of the converter transformer is specifically performed by the following method: A multi-channel vibration sensor is installed on the surface of the converter transformer box; the vibration acceleration signal output from the surface of the converter transformer box is transmitted to a computer through an integrated measuring device, and the state of the collected vibration acceleration signal is identified in the computer.

3. The method for identifying the operating state of a converter transformer based on a vibration signal according to claim 1, characterized in that: The construction process of the deep extreme learning machine is as follows: Build an Extreme Learning Machine: Get T different labeled samples, Represents the i-th sample, corresponding to label l i =[l iz ,l iz ,···,l iz ] T , where z represents the number of categories; if the sample Belongs to the mth category, the mth value of the corresponding label vector is set to 1, and the remaining (m-1) values ​​are set to -1; the extreme learning machine with H hidden nodes is expressed by formula (16): Among them, Q i =[Q i1 ,Q i2 ,···,Q iz ] T It represents the output weight connecting the i-th hidden layer and the output layer, A i =[A i1 ,A i2 ,···,A iz ] T is the input weight connecting the input layer and the i-th hidden layer, B i is the bias of the i-th hidden layer, G(A i ,B i ,E j ) is the output of the i-th hidden layer; assuming that the activation function of the i-th hidden layer is g(x), the output of the hidden layer is: G(A i ,B i ,E j )=g(A i ·E j +B i ) (17) Formula (16) can be summarized in matrix form as in, In order to meet the expected value of the output after network training, it is necessary to find the parameter A i , B i , Q i The optimal value of So that: Then the minimization loss function is In the extreme learning machine, the parameter (A i ,B i ) is randomly initialized, so Q i is uniquely determined; its solution is transformed into: The norm of the solution Q is the smallest and unique. In order to make the model have better generalization performance, the L1 regularization term is added, and the above solution problem is transformed into formula (21): Among them, C L is the regularization coefficient; the solution is essentially a ridge regression problem, and its solution is shown in the following formula: Where I is the identity matrix; Construct an autoencoder based on extreme learning machine: Apply the idea of ​​autoencoder to extreme learning machine so that the input data is also used for output, that is, output Y = X; The output of the extreme learning machine-based autoencoder is transformed from equation (16) into: Among them, A T A = I, B T B = 1; Among them, A is A i The matrix composed of i For sparse expression and dimension compression, the hidden layer output weight Q is transformed from formula (23) into: Where E = [E1, E2, ···, E8], the input vibration signal is a feature vector composed of the energy proportions of each mode after variational mode decomposition. For equal-dimensional feature mapping, the weight Q is calculated by formula (26): Q=C -1 l (26) Among them, Q T Q = I; Building a Deep Extreme Learning Machine: Based on the feature representation ability of the extreme learning machine-based autoencoder, it is used as the basic unit of the deep extreme learning machine; The input data sample E is used as the target output of the first extreme learning machine-based autoencoder, and then the output weight Q1 is obtained; then the output matrix C1 of the first hidden layer of the deep extreme learning machine is used as the input and target output of the next autoencoder based on the extreme learning machine, and so on. The last layer is trained with the extreme learning machine, and the output weight Q of the last hidden layer of the deep extreme learning machine is solved using formula (21): i+1 .