Fault diagnosis method, system, equipment and medium for on-load tap changer

Abstract

The present application discloses a fault diagnosis method, system, device and medium for on-load tap changer, which mainly relates to the field of fault diagnosis technology, and is used to solve the problem that the current mainstream adaptive signal decomposition method cannot effectively solve the problem of modal aliasing. It includes: obtaining a training vibration signal, and then using differential mode decomposition to obtain the fault component spectrum and the normal component spectrum; calculating the sample entropy, fuzzy entropy, and permutation entropy of the fault component spectrum and the normal component spectrum; inputting the fault component spectrum and the normal component spectrum into a preset convolutional network to extract the local features of the corresponding sequence data; then the local features and the sample entropy, fuzzy entropy, and permutation entropy together constitute a feature vector; through the feature vector, training the gated cyclic unit to obtain a trained gated cyclic unit; after obtaining the vibration signal to be trained, calculating the feature vector of the vibration signal to be trained; inputting the feature vector into the trained gated cyclic unit to obtain a prediction result.

Description

Technical Field

[0001] The present application relates to the technical field of on-load tap changers, and in particular to a fault diagnosis method, system, device and medium for an on-load tap changer. Background Art

[0002] The capacity of power equipment and the scale of power grids are constantly expanding, and the structure and operation characteristics of power systems are becoming more and more complex. Therefore, the failure of power equipment has a serious impact on the normal operation of the power system. On - LoadTapChanger , OLTC ) As a key device for adjusting the output voltage of the transformer, its normal operation is crucial to maintaining the stability of the power grid and improving the quality of power. OLTC Frequent operation in the power system makes the mechanical parts inside it very easy to malfunction due to long-term wear and aging, such as contact wear, spring weakness, transmission mechanism jamming, etc. According to statistics, OLTC Mechanical failures account for 70% of all failures. If these failures are not discovered and handled in a timely manner, they may cause abnormal operation of the transformer and even cause failure of the entire power system.

[0003] Currently, for OLTC Fault diagnosis mainly relies on vibration signal analysis technology. Adaptive signal decomposition method is currently OLTC The commonly used signal decomposition method for vibration signal analysis is widely used because it does not need to set a priori transformation basis and can adaptively decompose the target signal into several mode components. At present, the mainstream adaptive signal decomposition methods such as empirical mode decomposition and variational mode decomposition cannot solve the problem of modal aliasing well. Therefore, a fault diagnosis method, system, equipment and medium for on-load tap changers are urgently needed to solve the above problems. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present application provides a fault diagnosis system, method and medium for an on-load tap changer to solve the problem that the current mainstream adaptive signal decomposition method cannot effectively solve the modal aliasing.

[0005] In a first aspect, the present application provides a fault diagnosis method for an on-load tap changer, the method comprising:

[0006] A training vibration signal is obtained, and then the fault component spectrum and the normal component spectrum are obtained by differential mode decomposition; wherein the training vibration signal is divided into a fault signal and a normal signal; the sample entropy, fuzzy entropy and permutation entropy of the fault component spectrum and the normal component spectrum are calculated; the fault component spectrum and the normal component spectrum are input into a preset convolutional network to extract the local features of the corresponding sequence data; then the local features and the sample entropy, fuzzy entropy and permutation entropy together constitute a feature vector; through the feature vector, a gated recurrent unit is trained to obtain a trained gated recurrent unit; after obtaining the vibration signal to be trained, the feature vector of the vibration signal to be trained is calculated; the feature vector is input into the trained gated recurrent unit to obtain a prediction result.

[0007] The fault diagnosis method of the on-load tap changer provided in the embodiment of the present application can separate the fault component spectrum and the normal component spectrum by introducing differential mode decomposition, thereby solving the problem of modal aliasing. In addition, the present application not only uses differential mode decomposition to obtain the component spectrum of the signal, but also calculates the sample entropy, fuzzy entropy and permutation entropy, and these entropy values ​​reflect the different complexity and uncertainty characteristics of the signal. By forming a feature vector with these entropy values ​​and local features, the characteristics of the signal can be more comprehensively described. By training the gated recurrent unit and using the feature vector for input, the prediction result of the vibration signal can be obtained. Since the feature vector contains multiple features of the signal, the neural network can more accurately identify the fault component and the normal component of the signal, thereby improving the accuracy and efficiency of fault diagnosis. By training the gated recurrent unit (a recurrent neural network suitable for sequence data) through the feature vector, the time dependency in the sequence data can be fully utilized to improve the accuracy and robustness of fault prediction.

[0008] In one implementation of the present application, a training vibration signal is obtained, and then differential mode decomposition is used to obtain a fault component spectrum line and a normal component spectrum line, specifically including:

[0009] Calculate the normalized Fourier spectra corresponding to the fault signal and the normal signal respectively;

[0010] By minimizing the objective function :

[0011] ,

[0012] ,

[0013] Compute the optimal difference spectrum ;in, P represents the number of samples of fault signals in the training vibration signal, Q represents the number of samples of normal signals in the training vibration signal, Represents the coefficient of the preset regularization term; represents the preset model parameter vector, Indicates A training vibration signal, Indicates The sample labels of the training vibration signals are When a training vibration signal is a fault signal = 1. When When the training vibration signal is a normal signal = 0, , Indicates The column vector of training vibration signals, Indicates The normalized Fourier spectrum of the training vibration signal;

[0014] The optimal differential spectrum is divided into fault component spectrum lines and normal component spectrum lines.

[0015] In one implementation of the present application, calculating the normalized Fourier spectra corresponding to the fault signal and the normal signal respectively includes:

[0016] By formula:

[0017] ,

[0018] Calculate the normalized Fourier spectrum corresponding to the fault signal and the normal signal; where, represents the sum of the absolute values ​​of all elements, , represents the fast Fourier transform, x Represents the original time domain signal corresponding to the training vibration signal.

[0019] In one implementation of the present application, the optimal differential spectrum is divided into a fault component spectrum line and a normal component spectrum line, specifically including:

[0020] Get the point sequence corresponding to the optimal difference spectrum:

[0021] ;in, Indicates the number of points in the point sequence; P Indicates the spectral line value,

[0022] ,

[0023] in, , and , are the preset coefficients of the two straight lines respectively; represents the optimal change point;

[0024] formula:

[0025] ,

[0026] By adjusting the formula value, calculate the minimum and minimum ;in, Indicates that in the point sequence n Sequence T value, Indicates that in the point sequence n Sequence P value,

[0027] The optimal difference spectrum is higher than and The spectrum line with the maximum value is divided into the fault component spectrum line, and the spectrum below and The spectral lines with minimum values ​​are divided into normal component spectral lines.

[0028] In one implementation of the present application, calculating the sample entropy, fuzzy entropy, and permutation entropy of the fault component spectrum and the normal component spectrum respectively includes:

[0029] For each fault component spectrum line and normal component spectrum line, construct the corresponding time series:

[0030] ;in, represents the component spectrum value, N Indicates the sequence length; and sets the embedding dimension m ,tolerance r ;

[0031] Step 11, build length m The embedding vector of:

[0032] ;

[0033] Step 12, calculate the distance d :

[0034] ;

[0035] Step 13, calculate the ratio of matching vectors, where is a step function:

[0036] ;

[0037] Step 14: Increase the embedding dimension to m +1 and repeat steps 12 and 13 to get ;

[0038] Step 15, calculate and obtain the sequence sample entropy:

[0039] ;

[0040] Step 21, repeat step 11;

[0041] Step 22, repeat step 12;

[0042] Step 23, calculate the fuzzy membership function :

[0043] ;

[0044] in, r represents the parameter controlling the similarity threshold;

[0045] Step 24, for each embedding vector , computed with all other embedding vectors The fuzzy similarity of and take the average:

[0046] ;

[0047] For all i Find the mean, and we get m Fuzzy similarity of dimensional embedding vectors:

[0048] ;

[0049] Step 25, increase the embedding dimension to m +1 and repeat steps 21-24 to get ;

[0050] Step 26, calculate and obtain the fuzzy entropy:

[0051] .

[0052] In one implementation of the present application, calculating the sample entropy, fuzzy entropy, and permutation entropy of the fault component spectrum and the normal component spectrum respectively includes:

[0053] For each fault component spectrum line and normal component spectrum line, construct the corresponding time series:

[0054] ;in, represents the component spectrum value, N Indicates the sequence length; and sets the embedding dimension and time delay ;

[0055] Step 31, construct the length Embedded subsequence of:

[0056] ;

[0057] Step 32, arrange each subsequence by size to obtain m ! possible arrangement patterns;

[0058] Step 33: Count the number of occurrences of each arrangement pattern ,in, g =1,2,..., m !;

[0059] Step 34, calculate the permutation entropy :

[0060] .

[0061] In one implementation of the present application, the preset convolution network includes a causal convolution module, a dilated convolution module and a residual connection module; the fault component spectrum and the normal component spectrum are input into the preset convolution network to extract the local features of the corresponding sequence data, specifically including:

[0062] Set the preset dilation rate of the dilated convolution module to adjust the number of input steps skipped by the convolution kernel of the causal convolution module; input the fault component spectrum and the normal component spectrum into the preset convolution network, and pass through the causal convolution module:

[0063] , maintain the sequential characteristics in time series data;

[0064] in, Represents the time step t The output, Indicates the time step corresponding to the fault component spectrum and the normal component spectrum t The input sequence is represents the weight of the convolution kernel, U is the size of the convolution kernel, S represents the expansion rate; through the residual connection and Add together to form the local features of the final output.

[0065] In one implementation of the present application, training a gated recurrent unit through a feature vector to obtain a trained gated recurrent unit specifically includes:

[0066] The feature vector is divided into a training set and a test set; the gated recurrent unit is trained by the training set to obtain a preliminarily trained gated recurrent unit; the accuracy of the preliminarily trained gated recurrent unit is obtained by the test set; when the accuracy exceeds a preset threshold, the trained gated recurrent unit is obtained; otherwise, training is performed again until the accuracy exceeds the preset threshold.

[0067] In a second aspect, the present application provides a fault diagnosis system for an on-load tap changer, the system comprising:

[0068] The acquisition module is used to acquire the training vibration signal, and then use differential mode decomposition to obtain the fault component spectrum line and the normal component spectrum line; wherein the training vibration signal is divided into a fault signal and a normal signal; the calculation module is used to calculate the sample entropy, fuzzy entropy, and permutation entropy of the fault component spectrum line and the normal component spectrum line respectively; the vector module is used to input the fault component spectrum line and the normal component spectrum line into a preset convolutional network to extract the local features of the corresponding sequence data; then the local features and the sample entropy, fuzzy entropy, and permutation entropy together constitute a feature vector; the training module is used to train the gated recurrent unit through the feature vector to obtain a trained gated recurrent unit; the prediction module is used to calculate the feature vector of the vibration signal to be trained after obtaining the vibration signal to be trained; the feature vector is input into the trained gated recurrent unit to obtain a prediction result.

[0069] In one implementation of the present application, the acquisition module includes an acquisition unit, which is used to calculate the normalized Fourier spectra corresponding to the fault signal and the normal signal respectively;

[0070] By minimizing the objective function :

[0071] ,

[0072] ,

[0073] Compute the optimal difference spectrum ;in, P represents the number of samples of fault signals in the training vibration signal, Q represents the number of samples of normal signals in the training vibration signal, Represents the coefficient of the preset regularization term; represents the preset model parameter vector, Indicates A training vibration signal, Indicates The sample labels of the training vibration signals are When a training vibration signal is a fault signal = 1. When When the training vibration signal is a normal signal = 0, , Indicates The column vector of training vibration signals, Indicates The normalized Fourier spectrum of the training vibration signal;

[0074] The optimal differential spectrum is divided into fault component spectrum lines and normal component spectrum lines.

[0075] In one implementation of the present application, the obtaining unit includes a Fourier spectrum calculation subunit, which is used to calculate the Fourier spectrum through the formula:

[0076] ,

[0077] Calculate the normalized Fourier spectrum corresponding to the fault signal and the normal signal; where, represents the sum of the absolute values ​​of all elements, , represents the fast Fourier transform, x Represents the original time domain signal corresponding to the training vibration signal.

[0078] In one implementation of the present application, the acquisition module includes a division subunit,

[0079] The point sequence used to obtain the optimal difference spectrum:

[0080] ;in, Indicates the number of points in the point sequence; P Indicates the spectral line value,

[0081] ,

[0082] in, , and , are the preset coefficients of the two straight lines respectively; represents the optimal change point;

[0083] formula:

[0084] ,

[0085] By adjusting the formula value, calculate the minimum and minimum ;in, Indicates that in the point sequence n Sequence T value, Indicates that in the point sequence nSequence P value,

[0086] The optimal difference spectrum is higher than and The spectrum line with the maximum value is divided into the fault component spectrum line, and the spectrum below and The spectral lines with minimum values ​​are divided into normal component spectral lines.

[0087] In one implementation of the present application, the calculation module includes a calculation unit, which is used to construct a corresponding time series for each fault component spectrum line and normal component spectrum line:

[0088] ;in, represents the component spectrum value, N Indicates the sequence length; and sets the embedding dimension m ,tolerance r ;

[0089] Step 11, build length m The embedding vector of:

[0090] ;

[0091] Step 12, calculate the distance d :

[0092] ;

[0093] Step 13, calculate the ratio of matching vectors, where is a step function:

[0094] ;

[0095] Step 14: Increase the embedding dimension to m +1 and repeat steps 12 and 13 to get ;

[0096] Step 15, calculate and obtain the sequence sample entropy:

[0097] ;

[0098] Step 21, repeat step 11;

[0099] Step 22, repeat step 12;

[0100] Step 23, calculate the fuzzy membership function :

[0101] ;

[0102] in,r represents the parameter controlling the similarity threshold;

[0103] Step 24, for each embedding vector , computed with all other embedding vectors The fuzzy similarity of and take the average:

[0104] ;

[0105] For all i Find the mean, and we get m Fuzzy similarity of dimensional embedding vectors:

[0106] ;

[0107] Step 25, increase the embedding dimension to m +1 and repeat steps 21-24 to get ;

[0108] Step 26, calculate and obtain the fuzzy entropy:

[0109] .

[0110] In a third aspect, the present application provides a fault diagnosis device for an on-load tap changer, the device comprising:

[0111] processor;

[0112] and a memory having executable codes stored thereon, which, when executed, causes a processor to execute any of the above-mentioned methods for diagnosing faults of an on-load tap changer.

[0113] In a fourth aspect, the present application provides a non-volatile computer storage medium having computer instructions stored thereon, which, when executed, implement a fault diagnosis method for an on-load tap changer as described in any one of the above items.

[0114] Those skilled in the art can understand that the present application has at least the following beneficial effects:

[0115] The present application discloses a fault diagnosis method, system, device and medium for an on-load tap changer. By introducing differential mode decomposition, fault component spectrum lines and normal component spectrum lines are separated, thereby solving the problem of modal aliasing. In addition, the present application not only uses differential mode decomposition to obtain component spectrum lines of a signal, but also calculates sample entropy, fuzzy entropy and permutation entropy. These entropy values ​​reflect the different complexity and uncertainty characteristics of the signal. By forming a feature vector with these entropy values ​​and local features, the characteristics of the signal can be described more comprehensively. By training a gated recurrent unit and using the feature vector for input, the prediction result of the vibration signal can be obtained. Since the feature vector contains multiple features of the signal, the neural network can more accurately identify the fault component and normal component of the signal, thereby improving the accuracy and efficiency of fault diagnosis. In addition, by training the gated recurrent unit (a recurrent neural network suitable for sequence data) through the feature vector, the time dependency in the sequence data can be fully utilized, and the accuracy and robustness of fault prediction can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0116] Some embodiments of the present disclosure are described below with reference to the accompanying drawings, in which:

[0117] Figure 1 It is a flow chart of a fault diagnosis method for an on-load tap changer provided in an embodiment of the present application.

[0118] Figure 2 It is a schematic diagram of the internal structure of a fault diagnosis system for an on-load tap changer provided in an embodiment of the present application.

[0119] Figure 3 It is a schematic diagram of the internal structure of a fault diagnosis device for an on-load tap changer provided in an embodiment of the present application. DETAILED DESCRIPTION

[0120] It should be understood by those skilled in the art that the embodiments described below are only preferred embodiments of the present disclosure, and do not mean that the present disclosure can only be implemented through the preferred embodiments. The preferred embodiments are only used to explain the technical principles of the present disclosure, and are not used to limit the protection scope of the present disclosure. Based on the preferred embodiments provided by the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work should still fall within the protection scope of the present disclosure.

[0121] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0122] At present, the fault diagnosis of OLTC mainly relies on vibration signal analysis technology. Vibration signals contain rich mechanical state information, which can reflect the operating status of various components inside OLTC and potential fault hazards. However, vibration signals usually have nonlinear, non-stationary and chaotic characteristics, which makes their feature extraction and fault diagnosis process complicated and challenging. The adaptive signal decomposition method is currently a commonly used signal decomposition method for OLTC vibration signal analysis. It is widely used because it does not need to set a priori transformation basis and can adaptively decompose the target signal into several mode components. The current mainstream adaptive signal decomposition methods such as empirical mode decomposition and variational mode decomposition cannot solve the problem of modal aliasing well. To solve this problem, a fault diagnosis method, system, equipment and medium for on-load tap changers are proposed.

[0123] The technical solution proposed in the embodiments of the present application is described in detail below with reference to the accompanying drawings.

[0124] The embodiment provides a fault diagnosis method for an on-load tap changer, such as Figure 1 As shown, the method provided in the embodiment of the present application mainly includes the following steps:

[0125] Step 110: Obtain a training vibration signal, and then use differential mode decomposition to obtain a fault component spectrum and a normal component spectrum.

[0126] It should be noted that the training vibration signal is divided into fault signals and normal signals. The differential mode decomposition involved in this step processes nonlinear and non-stationary signals in two stages. First, the normalized Fourier spectra of the fault and normal signals are calculated, and an objective function is established to solve the weight vector ω (the weight vector ω of the hyperplane has spectral characteristics, which is called the optimal differential spectrum). Next, the change point analysis model is used to determine two thresholds to divide the spectrum lines into fault components and normal components. Finally, the fault component spectrum lines and the normal component spectrum lines are converted into time domain representations through the inverse fast Fourier transform for subsequent analysis.

[0127] The above-mentioned solution of normalizing Fourier spectrum can be specifically as follows:

[0128] By formula:

[0129] ,

[0130] Calculate the normalized Fourier spectrum corresponding to the fault signal and the normal signal; where, represents the sum of the absolute values ​​of all elements, , represents the fast Fourier transform, x Represents the original time domain signal corresponding to the training vibration signal.

[0131] As an example, the solution for calculating the optimal difference spectrum may be specifically as follows:

[0132] By minimizing the objective function :

[0133] ,

[0134] , calculate the optimal difference spectrum .

[0135] in, P represents the number of samples of fault signals in the training vibration signal, Q represents the number of samples of normal signals in the training vibration signal, Represents the coefficient of the preset regularization term; represents the preset model parameter vector, Indicates A training vibration signal, Indicates The sample labels of the training vibration signals are When a training vibration signal is a fault signal = 1. When When the training vibration signal is a normal signal = 0, , Indicates A column vector of training vibration signals, Indicates Normalized Fourier spectrum of a training vibration signal.

[0136] In addition, the solution of the present application for dividing the optimal differential spectrum into fault component spectrum lines and normal component spectrum lines can be specifically as follows:

[0137] Get the point sequence corresponding to the optimal difference spectrum .

[0138] in, Indicates the number of points in the point sequence; P Indicates the spectral line value.

[0139] ,

[0140] in, , and , are the preset coefficients of the two straight lines respectively; Indicates the optimal change point.

[0141] formula:

[0142] ,

[0143] By adjusting the formula value, calculate the minimum and minimum ;in, Indicates that in the point sequence n Sequence T value, Indicates that in the point sequence n Sequence P value,

[0144] The optimal difference spectrum is higher than and The spectrum line with the maximum value is divided into the fault component spectrum line, and the spectrum below and The spectral lines with minimum values ​​are divided into normal component spectral lines.

[0145] Step 120: Calculate the sample entropy, fuzzy entropy, and permutation entropy of the fault component spectrum and the normal component spectrum.

[0146] It should be noted that this step is to perform inverse fast Fourier transform on the obtained fault component spectrum and normal component spectrum to obtain the fault component time domain representation and the normal component time domain representation. The sample entropy, fuzzy entropy and permutation entropy of the fault component time domain representation and the normal component time domain representation are calculated respectively as eigenvalues ​​to construct feature vectors, sample entropy, fuzzy entropy and permutation entropy.

[0147] Among them, the scheme for calculating sample entropy can be specifically as follows:

[0148] For each fault component spectrum line and normal component spectrum line, construct the corresponding time series:

[0149] ;in, represents the component spectrum value, N Indicates the sequence length; and sets the embedding dimension m ,tolerance r ;

[0150] Step 11, build length mThe embedding vector of:

[0151] ;

[0152] Step 12, calculate the distance d :

[0153] ;

[0154] Step 13, calculate the ratio of matching vectors, where is a step function:

[0155] ;

[0156] Step 14: Increase the embedding dimension to m +1 and repeat steps 12 and 13 to get ;

[0157] Step 15, calculate and obtain the sequence sample entropy:

[0158] .

[0159] Among them, the scheme for calculating fuzzy entropy can be specifically as follows:

[0160] Step 21, repeat step 11;

[0161] Step 22, repeat step 12;

[0162] Step 23, calculate the fuzzy membership function :

[0163] ;

[0164] in, r represents the parameter controlling the similarity threshold;

[0165] Step 24, for each embedding vector , computed with all other embedding vectors The fuzzy similarity of and take the average:

[0166] ;

[0167] For all i Find the mean, and we get m Fuzzy similarity of dimensional embedding vectors:

[0168] ;

[0169] Step 25, increase the embedding dimension to m +1 and repeat steps 21-24 to get ;

[0170] Step 26, calculate and obtain the fuzzy entropy:

[0171] .

[0172] Among them, the scheme for calculating the permutation entropy can be specifically as follows:

[0173] For each fault component spectrum line and normal component spectrum line, construct the corresponding time series: ;in, represents the component spectrum value, N Indicates the sequence length; and sets the embedding dimension and time delay ;

[0174] Step 31, construct the length Embedded subsequence of:

[0175] ;

[0176] Step 32, arrange each subsequence by size to obtain m ! possible arrangement patterns;

[0177] Step 33: Count the number of occurrences of each arrangement pattern ,in, g =1,2,..., m !;

[0178] Step 34, calculate the permutation entropy :

[0179] .

[0180] Step 130: Input the fault component spectrum and the normal component spectrum into a preset convolutional network to extract local features of the corresponding sequence data; and then the local features together with sample entropy, fuzzy entropy and permutation entropy constitute a feature vector.

[0181] It should be noted that the preset convolutional network can be a temporal convolutional network.

[0182] The preset convolutional network includes a causal convolution module, a dilated convolution module, and a residual connection module;

[0183] The fault component spectrum and the normal component spectrum are input into the preset convolutional network to extract the local features of the corresponding sequence data, including:

[0184] Set the preset dilation rate of the dilated convolution module to adjust the number of input steps skipped by the convolution kernel of the causal convolution module;

[0185] The fault component spectrum and the normal component spectrum are input into the preset convolutional network and passed through the causal convolution module:

[0186] , maintain the sequential characteristics in time series data;

[0187] in, Represents the time step t The output, Indicates the time step corresponding to the fault component spectrum and the normal component spectrum t The input sequence is represents the weight of the convolution kernel, U is the size of the convolution kernel, and S represents the expansion rate;

[0188] Through residual connection and Add together to form the local features of the final output.

[0189] Those skilled in the art will appreciate that the use of residual connections can avoid the gradient vanishing problem and maintain the stability of the model, especially when the number of network layers is deep. Residual connections pass the input directly to the subsequent layers by introducing "skip connections", which can help the model not lose information when capturing complex time dependencies, so that the model can be trained more efficiently.

[0190] Step 140: Train the gated recurrent unit using the feature vector to obtain a trained gated recurrent unit.

[0191] It should be noted that the training process can be specifically as follows:

[0192] The feature vector is divided into a training set and a test set; the gated recurrent unit is trained by the training set to obtain a preliminarily trained gated recurrent unit; the accuracy of the preliminarily trained gated recurrent unit is obtained by the test set; when the accuracy exceeds a preset threshold, the trained gated recurrent unit is obtained; otherwise, training is performed again until the accuracy exceeds the preset threshold.

[0193] Step 150: After obtaining the vibration signal to be trained, calculate and obtain a feature vector of the vibration signal to be trained; input the feature vector into a trained gated recurrent unit to obtain a prediction result.

[0194] It should be noted that the scheme for calculating and obtaining the characteristic vector of the vibration signal to be trained is consistent with the above scheme, and this application will not elaborate on it here.

[0195] Based on the foregoing description, it can be seen that in this embodiment, the time domain representation of the fault component and the normal component generated by differential mode decomposition and the calculated center frequency, sample entropy, fuzzy entropy and permutation entropy are input into the model for training, and the preset convolutional network (time series convolutional network) is used to mine the feature information and the detail information of the changes in the time dimension. The feature information output by the time series convolutional network is combined with the center frequency, sample entropy, fuzzy entropy and permutation entropy as a feature vector and input into the gated recurrent unit to realize the diagnosis and identification of typical OLTC faults.

[0196] In addition, this application Figure 2 A fault diagnosis system for an on-load tap changer is provided in an embodiment of the present application. Figure 2 As shown, the system provided in the embodiment of the present application mainly includes:

[0197] The acquisition module 210 is used to acquire the training vibration signal, and then use differential mode decomposition to obtain the fault component spectrum and the normal component spectrum; wherein the training vibration signal is divided into a fault signal and a normal signal.

[0198] The acquisition module 210 includes an acquisition unit,

[0199] Used to calculate the normalized Fourier spectra corresponding to the fault signal and the normal signal;

[0200] By minimizing the objective function :

[0201] ,

[0202] ,

[0203] Compute the optimal difference spectrum ;in, P represents the number of samples of fault signals in the training vibration signal, Q represents the number of samples of normal signals in the training vibration signal, Represents the coefficient of the preset regularization term; represents the preset model parameter vector, Indicates A training vibration signal, Indicates The sample labels of the training vibration signals are When a training vibration signal is a fault signal = 1. When When the training vibration signal is a normal signal = 0, , Indicates A column vector of training vibration signals, Indicates The normalized Fourier spectrum of the training vibration signal;

[0204] The optimal differential spectrum is divided into fault component spectrum lines and normal component spectrum lines.

[0205] Wherein, the acquisition unit includes a Fourier spectrum calculation subunit,

[0206] Used by the formula:

[0207] ,

[0208] Calculate the normalized Fourier spectrum corresponding to the fault signal and the normal signal; where, represents the sum of the absolute values ​​of all elements, , represents the fast Fourier transform, x Represents the original time domain signal corresponding to the training vibration signal.

[0209] Wherein, the acquisition module includes a division subunit,

[0210] The point sequence used to obtain the optimal difference spectrum:

[0211] ;in, Indicates the number of points in the point sequence; P Indicates the spectral line value,

[0212] ,

[0213] in, , and , are the preset coefficients of the two straight lines respectively; represents the optimal change point;

[0214] formula:

[0215] ,

[0216] By adjusting the formula value, calculate the minimum and minimum ;in, Indicates that in the point sequence n Sequence T value, Indicates that in the point sequence n Sequence P value,

[0217] The optimal difference spectrum is higher than and The spectrum line with the maximum value is divided into the fault component spectrum line, and the spectrum below and The spectral lines with minimum values ​​are divided into normal component spectral lines.

[0218] The calculation module 220 is used to calculate the sample entropy, fuzzy entropy and permutation entropy of the fault component spectrum and the normal component spectrum.

[0219] The calculation module 220 includes a calculation unit,

[0220] Used to construct the corresponding time series for each fault component spectrum line and normal component spectrum line:

[0221] ;in, represents the component spectrum value, N Indicates the sequence length; and sets the embedding dimension m ,tolerance r ;

[0222] Step 11, build length m The embedding vector of:

[0223] ;

[0224] Step 12, calculate the distance d :

[0225] ;

[0226] Step 13, calculate the ratio of matching vectors, where is a step function:

[0227] ;

[0228] Step 14: Increase the embedding dimension to m +1 and repeat steps 12 and 13 to get ;

[0229] Step 15, calculate and obtain the sequence sample entropy:

[0230] ;

[0231] Step 21, repeat step 11;

[0232] Step 22, repeat step 12;

[0233] Step 23, calculate the fuzzy membership function :

[0234] ;

[0235] in, r represents the parameter controlling the similarity threshold;

[0236] Step 24, for each embedding vector , computed with all other embedding vectors The fuzzy similarity of and take the average:

[0237] ;

[0238] For all i Find the mean, and we get m Fuzzy similarity of dimensional embedding vectors:

[0239] ;

[0240] Step 25, increase the embedding dimension to m +1 and repeat steps 21-24 to get ;

[0241] Step 26, calculate and obtain the fuzzy entropy:

[0242] .

[0243] The vector module 230 is used to input the fault component spectrum and the normal component spectrum into a preset convolutional network to extract the local features of the corresponding sequence data; and then the local features and the sample entropy, fuzzy entropy, and permutation entropy together constitute a feature vector.

[0244] The training module 240 is used to train the gated recurrent unit through the feature vector to obtain a trained gated recurrent unit.

[0245] The prediction module 250 is used to obtain the vibration signal to be trained, calculate the feature vector of the vibration signal to be trained, and input the feature vector into the trained gated recurrent unit to obtain the prediction result.

[0246] The above is a method embodiment of the present application. Based on the same inventive concept, the present application embodiment also provides a fault diagnosis device for an on-load tap changer. Figure 3 As shown, the device includes: a processor; and a memory, on which executable codes are stored. When the executable codes are executed, the processor executes as in the above embodiment:

[0247] The training vibration signal is obtained, and then the fault component spectrum and the normal component spectrum are obtained by differential mode decomposition; wherein the training vibration signal is divided into a fault signal and a normal signal; the sample entropy, fuzzy entropy, and permutation entropy of the fault component spectrum and the normal component spectrum are calculated; the fault component spectrum and the normal component spectrum are input into a preset convolutional network to extract the local features of the corresponding sequence data; then the local features and the sample entropy, fuzzy entropy, and permutation entropy together constitute a feature vector; the gated recurrent unit is trained through the feature vector to obtain a trained gated recurrent unit; after obtaining the vibration signal to be trained, the feature vector of the vibration signal to be trained is calculated; the feature vector is input into the trained gated recurrent unit to obtain a prediction result

[0248] In addition, the embodiment of the present application further provides a non-volatile computer storage medium, on which executable instructions are stored. When the executable instructions are executed, the following are implemented:

[0249] A training vibration signal is obtained, and then the fault component spectrum and the normal component spectrum are obtained by differential mode decomposition; wherein the training vibration signal is divided into a fault signal and a normal signal; the sample entropy, fuzzy entropy and permutation entropy of the fault component spectrum and the normal component spectrum are calculated; the fault component spectrum and the normal component spectrum are input into a preset convolutional network to extract the local features of the corresponding sequence data; then the local features and the sample entropy, fuzzy entropy and permutation entropy together constitute a feature vector; through the feature vector, a gated recurrent unit is trained to obtain a trained gated recurrent unit; after obtaining the vibration signal to be trained, the feature vector of the vibration signal to be trained is calculated; the feature vector is input into the trained gated recurrent unit to obtain a prediction result.

[0250] So far, the technical solutions of the present disclosure have been described in combination with the above multiple embodiments, but it is easy for those skilled in the art to understand that the protection scope of the present disclosure is not limited to these specific embodiments. Without departing from the technical principles of the present disclosure, those skilled in the art can split and combine the technical solutions in the above-mentioned various embodiments, and can also make equivalent changes or replacements to the relevant technical features. Any changes, equivalent replacements, improvements, etc. made within the technical concept and / or technical principle of the present disclosure will fall within the protection scope of the present disclosure.

Claims

1. A fault diagnosis method for an on-load tap changer, characterized in that: The method comprises: A training vibration signal is obtained, and then a fault component spectrum line and a normal component spectrum line are obtained by using differential mode decomposition; wherein the training vibration signal is divided into a fault signal and a normal signal; Specifically include: Calculate the normalized Fourier spectra corresponding to the fault signal and the normal signal respectively; By minimizing the objective function : ， ， Compute the optimal difference spectrum ;in, A represents the number of samples of fault signals in the training vibration signal, Q represents the number of samples of normal signals in the training vibration signal, Represents the coefficient of the preset regularization term; represents the preset model parameter vector, c Indicates c A training vibration signal, Indicates c The sample labels of the training vibration signals are c When a training vibration signal is a fault signal = 1. When c When the training vibration signal is a normal signal = 0, , Indicates c The column vector of training vibration signals, Indicates c The normalized Fourier spectrum of the training vibration signal; The optimal difference spectrum is divided into a fault component spectrum line and a normal component spectrum line; Calculate the sample entropy, fuzzy entropy and permutation entropy of the fault component spectrum and the normal component spectrum respectively; The fault component spectrum and the normal component spectrum are input into the preset convolutional network to extract the local features of the corresponding sequence data; then the local features, sample entropy, fuzzy entropy and permutation entropy together form a feature vector; Through the feature vector, the gated recurrent unit is trained to obtain a trained gated recurrent unit; After obtaining the vibration signal to be trained, a feature vector of the vibration signal to be trained is calculated; the feature vector is input into the trained gated recurrent unit to obtain a prediction result.

2. The fault diagnosis method of the on-load tap changer according to claim 1, characterized in that: Calculate the normalized Fourier spectra corresponding to the fault signal and the normal signal, including: By formula: ， Calculate the normalized Fourier spectrum corresponding to the fault signal and the normal signal; where, represents the sum of the absolute values ​​of all elements, , represents the fast Fourier transform, x Represents the original time domain signal corresponding to the training vibration signal.

3. The fault diagnosis method of the on-load tap changer according to claim 1, characterized in that: The optimal differential spectrum is divided into fault component spectrum and normal component spectrum, including: Get the point sequence corresponding to the optimal difference spectrum: ;in, Indicates the number of points in the point sequence; P Indicates the spectral line value, ， in, , and , are the preset coefficients of the two straight lines respectively; represents the optimal change point; formula: ， By adjusting the formula value, calculate the minimum and minimum ;in, Indicates that in the point sequence n Sequence T value, Indicates that in the point sequence n Sequence P value, The optimal difference spectrum is higher than and The spectrum line with the maximum value is divided into the fault component spectrum line, and the spectrum below and The spectral lines with minimum values ​​are divided into normal component spectral lines.

4. The fault diagnosis method of the on-load tap changer according to claim 1, characterized in that: Calculate the sample entropy, fuzzy entropy, and permutation entropy of the fault component spectrum and the normal component spectrum, including: For each fault component spectrum line and normal component spectrum line, construct the corresponding time series: ;in, represents the component spectrum value, N Indicates the sequence length; and sets the embedding dimension m ,tolerance r ; Step 11, build length m The embedding vector of : ； Step 12, calculate the distance d : ； Step 13, calculate the ratio of matching vectors, where is a step function: ； Step 14: Increase the embedding dimension to m +1 and repeat steps 12 and 13 to get ; Step 15, calculate and obtain the sequence sample entropy: ； Step 21, repeat step 11; Step 22, repeat step 12; Step 23, calculate the fuzzy membership function : ； in, r represents the parameter controlling the similarity threshold; Step 24, for each embedding vector , computed with all other embedding vectors The fuzzy similarity of and take the average: ； For all i Find the mean, and we get m Fuzzy similarity of dimensional embedding vectors: ； Step 25, increase the embedding dimension to m +1 and repeat steps 21-24 to get ; Step 26, calculate and obtain the fuzzy entropy: 。 5. The fault diagnosis method of the on-load tap changer according to claim 1, characterized in that: Calculate the sample entropy, fuzzy entropy, and permutation entropy of the fault component spectrum and the normal component spectrum, including: For each fault component spectrum line and normal component spectrum line, construct the corresponding time series: ;in, represents the component spectrum value, N Indicates the sequence length; and sets the embedding dimension and time delay ; Step 31, construct the length Embedded subsequence of: ； Step 32, arrange each subsequence by size to obtain m ! possible arrangement patterns; Step 33: Count the number of occurrences of each arrangement pattern ,in, g =1,2,..., m !; Step 34, calculate the permutation entropy : 。 6. The fault diagnosis method of the on-load tap changer according to claim 1, characterized in that: The preset convolutional network includes a causal convolution module, a dilated convolution module, and a residual connection module; The fault component spectrum and the normal component spectrum are input into the preset convolutional network to extract the local features of the corresponding sequence data, including: Set the preset dilation rate of the dilated convolution module to adjust the number of input steps skipped by the convolution kernel of the causal convolution module; The fault component spectrum and the normal component spectrum are input into the preset convolutional network and passed through the causal convolution module: , maintain the sequential characteristics in time series data; in, Represents the time step t The output, Indicates the time step corresponding to the fault component spectrum and the normal component spectrum t The input sequence is represents the weight of the convolution kernel, U is the size of the convolution kernel, and S represents the expansion rate; Through residual connection and Add together to form the local features of the final output.

7. The fault diagnosis method of the on-load tap changer according to claim 1, characterized in that: Through the feature vector, the gated recurrent unit is trained to obtain the trained gated recurrent unit, which specifically includes: Split the feature vectors into training and testing sets; Train the gated recurrent unit through the training set to obtain a preliminarily trained gated recurrent unit; The accuracy of the initially trained gated recurrent unit is obtained through the test set; When the accuracy exceeds a preset threshold, a trained gated recurrent unit is obtained; Otherwise, train again until the accuracy exceeds the preset threshold.

8. A fault diagnosis system for an on-load tap changer, characterized in that: The system comprises: An acquisition module is used to acquire a training vibration signal, and then use differential mode decomposition to obtain a fault component spectrum line and a normal component spectrum line; wherein the training vibration signal is divided into a fault signal and a normal signal; The acquisition module includes an acquisition unit, Used to calculate the normalized Fourier spectra corresponding to the fault signal and the normal signal; By minimizing the objective function : ， ， Compute the optimal difference spectrum ;in, A represents the number of samples of fault signals in the training vibration signal, Q represents the number of samples of normal signals in the training vibration signal, Represents the coefficient of the preset regularization term; represents the preset model parameter vector, c Indicates c A training vibration signal, Indicates c The sample labels of the training vibration signals are c When a training vibration signal is a fault signal = 1. When c When the training vibration signal is a normal signal = 0, , Indicates c The column vector of training vibration signals, Indicates c The normalized Fourier spectrum of the training vibration signal; The optimal difference spectrum is divided into a fault component spectrum line and a normal component spectrum line; A calculation module, used to calculate the sample entropy, fuzzy entropy and permutation entropy of the fault component spectrum and the normal component spectrum respectively; The vector module is used to input the fault component spectrum and the normal component spectrum into the preset convolution network to extract the local features of the corresponding sequence data; then the local features and the sample entropy, fuzzy entropy, and permutation entropy together constitute a feature vector; A training module, used for training the gated recurrent unit through the feature vector to obtain a trained gated recurrent unit; The prediction module is used to obtain the vibration signal to be trained, calculate the feature vector of the vibration signal to be trained, and input the feature vector into the trained gated recurrent unit to obtain the prediction result.

9. The fault diagnosis system for the on-load tap changer according to claim 8, characterized in that: The acquisition unit includes a Fourier spectrum calculation subunit, Used by the formula: ， Calculate the normalized Fourier spectrum corresponding to the fault signal and the normal signal; where, represents the sum of the absolute values ​​of all elements, , represents the fast Fourier transform, x Represents the original time domain signal corresponding to the training vibration signal.

10. The fault diagnosis system for the on-load tap changer according to claim 8, characterized in that: The acquisition module includes a division subunit, The point sequence used to obtain the optimal difference spectrum: ;in, Indicates the number of points in the point sequence; P Indicates the spectral line value, ， in, , and , are the preset coefficients of the two straight lines respectively; represents the optimal change point; formula: ， By adjusting the formula value, calculate the minimum and minimum ;in, Indicates that in the point sequence n Sequence T value, Indicates that in the point sequence n Sequence P value, The optimal difference spectrum is higher than and The spectrum line with the maximum value is divided into the fault component spectrum line, and the spectrum below and The spectral lines with minimum values ​​are divided into normal component spectral lines.

11. The fault diagnosis system for the on-load tap changer according to claim 8, characterized in that: The computing module includes a computing unit, Used to construct the corresponding time series for each fault component spectrum line and normal component spectrum line: ;in, represents the component spectrum value, N Indicates the sequence length; and sets the embedding dimension m ,tolerance r ; Step 11, build length m The embedding vector of : ； Step 12, calculate the distance d : ； Step 13, calculate the ratio of matching vectors, where is a step function: ； Step 14: Increase the embedding dimension to m +1 and repeat steps 12 and 13 to get ; Step 15, calculate and obtain the sequence sample entropy: ； Step 21, repeat step 11; Step 22, repeat step 12; Step 23, calculate the fuzzy membership function : ； in, r represents the parameter controlling the similarity threshold; Step 24, for each embedding vector , computed with all other embedding vectors The fuzzy similarity of and take the average: ； For all i Find the mean, and we get m Fuzzy similarity of dimensional embedding vectors: ； Step 25, increase the embedding dimension to m +1 and repeat steps 21-24 to get ; Step 26, calculate and obtain the fuzzy entropy: 。 12. A fault diagnosis device for an on-load tap changer, characterized in that: The device comprises: processor; and a memory storing executable codes thereon, wherein when the executable codes are executed, the processor executes a fault diagnosis method for an on-load tap changer according to any one of claims 1 to 7.

13. A non-volatile computer storage medium, characterized in that: Computer instructions are stored thereon, and when the computer instructions are executed, the fault diagnosis method of an on-load tap changer according to any one of claims 1 to 7 is implemented.

Images