GIS partial discharge soundprint detection method based on energy operator improved wavelet packet

Abstract

The present invention discloses a GIS partial discharge voiceprint detection method based on energy operator improved wavelet packet, selects wavelet packet basis function to perform wavelet packet decomposition and noise reduction on GIS voiceprint signal, adopts improved Teager energy operator wavelet packet coefficient to perform instantaneous energy calculation, obtains instantaneous energy sequence, combines sliding window function to perform background threshold judgment, records energy abnormal points; then combines kurtosis entropy, fuzzy entropy, instantaneous energy and energy abnormal points to construct joint features, and uses RBF neural network algorithm to perform fault judgment. The present invention can quickly and accurately discover GIS operation hidden dangers.

Description

Technical Field

[0001] The invention belongs to the technical field of electric power detection equipment, and specifically designs a GIS partial discharge soundprint detection method based on energy operator improved wavelet packet. Background Art

[0002] Gas Insulated Switchgear (GIS) has been widely used in the field of power systems due to its small footprint, high reliability, large transmission capacity, simple maintenance, and long maintenance cycle. However, in the actual production, shipment, and actual operation of GIS, it will inevitably produce some safety hazards such as dust, air gaps, and metal conductive particles due to problems such as processing technology, collision, and complex operating environment, which will cause GIS to produce various forms of partial discharge. Partial discharge is a defect that does not immediately cause the insulation of the equipment to break down, but once it occurs, the surrounding medium will be continuously eroded, leading to penetrating insulation failure. There are many types of partial discharge in GIS, and different partial discharges have different hazards to the equipment. Therefore, in the case of large-scale use of GIS equipment, identifying the discharge type is a prerequisite for ensuring the safe operation of GIS and the stability of the power system.

[0003] Partial discharge is closely related to the type of applied voltage, the physical properties of the insulating material itself, the space charge generated during the discharge process, and the electric field distribution at the defect location. As for the existing partial discharge detection methods, most of them are to capture and decompose the abnormal physical quantities that appear during the discharge. When partial discharge occurs, some waves, light pulses and electric pulses will be radiated, local overheating will occur, some new gases will be generated, and rich sound signals will be generated, which contain a lot of equipment operation information. As a result, optical detection methods, chemical detection methods, pulse current detection methods, infrared thermal imaging methods and detection methods based on voiceprints have emerged.

[0004] However, in actual application of the above methods, the optical detection method is easily affected by the external environment, has a limited detection range, and is difficult to operate and maintain; the chemical detection method is based on the fact that the decomposition products of the gas inside the entire GIS are small in proportion and the concentration is low, making it difficult to detect components, and it is impossible to perform online detection; it is easily interfered by electromagnetic signals, and due to its narrow frequency band, the signal contains less information, and it is difficult to distinguish pulse interference currents; infrared thermal imaging has blind spots in detecting thermal faults in large generators, transformers and GIS.

[0005] The method of detecting GIS partial discharge based on voiceprint has the following advantages: there is no need to disassemble the GIS equipment, and the detection can be ensured when the GIS is operating normally, reducing the use of manpower and material resources caused by maintenance; non-invasive detection will not cause any impact on the GIS equipment itself, and will not have any impact on the operation of the power grid; the detection range is wide, and each component of the GIS can be tested to fully understand the overall operation status of the GIS. Summary of the invention

[0006] In order to solve the problem of abnormal discharge fault detection and identification in GIS equipment, a GIS partial discharge soundprint detection method based on energy operator improved wavelet packet is proposed. The soundprint data of abnormal discharge signal inside GIS is analyzed in real time to monitor the internal abnormal discharge phenomenon. Under the premise of ensuring the normal operation of GIS, it faces abnormal discharge faults, discovers operation hazards in advance, reduces economic losses and ensures the safe operation of GIS, which has important engineering application value.

[0007] The present invention is implemented by the following technical solutions: A GIS partial discharge voiceprint detection method based on energy operator improved wavelet packet, comprising the following steps:

[0008] Step 1: Select GIS voiceprint signals and establish a sample data set;

[0009] Step 2: Select wavelet packet basis function to perform wavelet packet decomposition and noise reduction on GIS voiceprint signal;

[0010] Step 3: Use the improved Teager energy operator wavelet packet coefficient to calculate the instantaneous energy and obtain the instantaneous energy sequence. Calculate the average value and standard deviation of the instantaneous energy within the step length, and compare the energy value in the area with the reference threshold. If it is higher than the threshold, it is recorded as an energy anomaly point, otherwise it is judged as background;

[0011] Step 4: For energy outliers, the fuzzy entropy, kurtosis entropy, instantaneous energy sequence and energy outlier mark position information are combined to form a joint feature;

[0012] Step 5: Based on the joint feature data set, the fault diagnosis model is constructed using the RBF neural network, and the fault diagnosis model is used to diagnose and judge the GIS voiceprint signal monitored online.

[0013] Further preferably, the sample data set includes partial discharge fault data, and the fault types include: metal tip defect, floating potential defect, free metal particle defect, insulation surface or hole discharge defect;

[0014] Further preferably, the process of step 2 is as follows:

[0015] Step 2.1: Use the 4-layer Morse basis function to perform wavelet packet decomposition on the GIS voiceprint signal to obtain the wavelet packet coefficients;

[0016] Step 2.2: De-noise the wavelet packet coefficients obtained by wavelet packet decomposition:

[0017]

[0018]

[0019] Where σ is the noise standard deviation of the signal; N is the signal length; η(d j ,λ) is the wavelet packet coefficient d j The processing result after threshold processing, d j is the wavelet packet coefficient, λ is a threshold parameter used to control the amplitude of threshold processing;

[0020] Step 2.3 reconstructs the processed wavelet packet coefficients to obtain a signal with a higher signal-to-noise ratio.

[0021] Further preferably, the process of step 3 is as follows:

[0022] Step 3.1 uses the improved Teager energy operator to calculate the instantaneous energy of the wavelet packet coefficients based on step 2, where the improved Teager energy operator is defined as:

[0023]

[0024] Where r(t) is the value of the processed signal, ψ[r(t)] represents the Teager energy value at the sample point at time t;

[0025] The instantaneous energy expression of the Teager energy operator is:

[0026] E (i,j) =ψ[r(t)] 2

[0027] In the formula, E (i,j) is the instantaneous energy of the i-th node in the j-th layer, then the instantaneous energy sequence is:

[0028] E N ={E (1,1) ,……,E (i,j)}

[0029] In the formula, E N is the instantaneous energy sequence, E (1,1) is the instantaneous energy of the first node in layer 1, E (i,j) is the instantaneous energy of the i-th node in the j-th layer;

[0030] Step 3.2: Calculate the energy background threshold. According to the instantaneous energy sequence of the GIS voiceprint signal in normal operation, dynamically determine the background threshold of the instantaneous energy sequence of the GIS voiceprint signal that is similar in time domain to see if there is an energy anomaly point, and record the location of the energy anomaly point in S location .

[0031] Further preferably, step 3.2 is specifically as follows:

[0032] Step 3.2.1: Define the starting window size Window size and step length Step size , respectively represent the number of data in the initial window and the initial sliding step size;

[0033]

[0034]

[0035] F s is the sampling rate of GIS voiceprint signal, Window size (i) is the i-th window size, Step size (i) is the i-th step length in seconds, T(i) is the length of the i-th frame in seconds, Step len is the number of sampling points in the i-th window, and i is the sequence number of the i-th step length;

[0036] Step 3.2.2: Calculate the standard deviation of the initial window and determine the threshold E based on the standard deviation of the sample data bg_threshold , and E bg_threshold As a criterion, the points exceeding the threshold are determined as energy abnormal points, and the energy abnormal points are marked;

[0037]

[0038] E bg_threshold =E (i,avg) +2*E (i,std)

[0039] In the formula, E (i,std) is the instantaneous energy standard deviation of each node in the initial window, E (i,avg) is the average instantaneous energy of each node;

[0040] Then, it is determined whether the signal energy of the current frame exceeds the background threshold, and the energy abnormal point is recorded in the energy abnormal point mark position information S location

[0041] Step 3.2.3: Repeat steps 3.2.1 and 3.2.2 for the next step interval until the data ends.

[0042] Further preferably, the process of using the training set to carry out learning and training of the RBF neural network and constructing the fault diagnosis model is as follows:

[0043] Step 5.2.1 Determine the input vector X = [x1, x2, ... x n ] T , where n represents the number of input units in the input layer, x n represents the input signal of the nth input unit; determines the output vector Y=[y1,y2…y q ] T , q represents the number of output units in the output layer, y q represents the output signal of the qth output unit; determine the expected output O = [o1, o2…o m ] T , m is the number of output units;

[0044] Furthermore, the connection weight W from the hidden layer to the output layer is initialized v =[w v1 w v2 ...w vp ] T , w vp is the connection weight from the vth neuron in the hidden layer to the pth neuron in the output layer, T represents transposition, and the weight initialization formula is:

[0045] w vp =rand(min v ,max v )

[0046] In the formula, w vp is the connection weight from the vth neuron in the hidden layer to the pth neuron in the output layer. Random initialization is used here to increase the diversity of the network and avoid falling into the local optimal solution. v is the minimum value of all expected outputs of the vth output neuron in the training set; max v is the maximum value of all expected outputs of the vth output neuron in the training set;

[0047] Initialize the central parameters c of each neuron in the hidden layer j =[c j1 ,c j2 ……c ji ] T ;

[0048]

[0049] In the formula, c ji is the central parameter of the i-th input of the j-th neuron in the hidden layer, g is the total number of neurons in the hidden layer, and p is the number of output units in the output layer; mini is the minimum value of all input information of the i-th feature in the training set, max i It is the maximum value of all input information of the i-th feature in the training set;

[0050] Initialize the width vector D j =[d j1 ,d j2 ……d ji ],

[0051]

[0052] Where, d ji is the width of the ith input of the jth neuron in the hidden layer, represents the i-th input of the v-th sample in the training set, N represents the number of samples in the training set, and d f is the width adjustment coefficient, the value is less than 1;

[0053] Step 5.2.2 Calculate the output value z of the jth neuron in the hidden layer j , X represents the input vector of the input layer, j = 1, 2...g, C j is the center vector of the jth neuron in the hidden layer, which is composed of the center components of all neurons in the input layer corresponding to the jth neuron in the hidden layer;

[0054] C j =[c j1 ,c j2 ,……c jn ] T ;D j is the width vector of the jth neuron in the hidden layer, and C j Correspondingly,

[0055] D j =[d j1 ,d j2 ……d ji ] T , D j The larger it is, the greater the influence of the hidden layer on the input layer, and the smoothness between neurons is better. ||.|| is the Euclidean norm;

[0056] Step 5.2.3: Calculate the output of the output layer neurons:

[0057] Y=[y1,y2……y q ] T

[0058]

[0059] In the formula, v = 1, 2, ... q, where wvj is the adjustment weight between the vth neuron in the output layer and the jth neuron in the hidden layer;

[0060] Step 5.2.4: Use the improved Adam algorithm to adjust the learning rate:

[0061]

[0062] Where α(u) is the learning rate of the uth iteration, α0 is the initial learning rate, β1 is the constant factor, and t is the number of iterations;

[0063] The iterative calculation is as follows:

[0064]

[0065]

[0066]

[0067] In the formula, w vj (u) The adjustment weight between the vth output neuron and the jth hidden layer neuron in the uth iteration calculation, v = 1, 2, ... q, j = 1, 2 ... g; c ji (u) is the central component of the jth hidden layer neuron for the ith input neuron at the uth iteration; d ji (u) for learning center c ji (u) corresponding width; Indicates w vj The first moment estimate of (u), Indicates w vj The second moment estimate of (u), Indicates c ji The first moment estimate of (u), Indicates c ji The second moment estimate of (u), Indicates d ji The first moment estimate of (u), Indicates d ji The second-order moment estimates of (u) are represented respectively, their initial values ​​are 0, ∈ is a minimum value to avoid the denominator being 0, and η is the learning factor;

[0068] E is the RBF neural network evaluation function:

[0069]

[0070] In the formula, O lv is the expected output value of the vth output neuron at the lth input sample; y lvis the network output value of the vth output neuron at the lth input sample.

[0071] Further optimization, use the test set to verify the feasibility of the algorithm:

[0072] A1: Initialize the neural network parameters according to step 5.2.1, and give the values ​​of η and α and the value of iteration accuracy ε;

[0073] A2: Calculate the root mean square error (RMS) of the network output. If RMS ≤ ε, the training ends. Otherwise, go to step A3.

[0074] A3: According to the weight iteration calculation in step 5.2.4, iteratively calculate the adjustment weight, center and width parameters;

[0075] A4: Return to step A2.

[0076] The present invention utilizes the characteristic that the improved Teager energy operator is sensitive to pulse signals, first uses wavelet packet analysis to decompose and reduce noise on GIS voiceprint signals; after obtaining the wavelet packet coefficients of each sub-band, the improved Teager energy operator is used to calculate the instantaneous energy to obtain the instantaneous energy series, and the background threshold is judged in combination with the sliding window function, and the energy anomalies are recorded; then, the kurtosis entropy, fuzzy entropy, instantaneous energy and energy anomalies are combined to construct joint features, and the RBF neural network algorithm is used to make fault judgments, so as to discover operation hazards in advance and reduce economic losses to ensure the safe operation of GIS, which has important engineering application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 This is a flow chart of the GIS partial discharge voiceprint detection method based on energy operator improved wavelet packet.

[0078] Figure 2 It is the RBF neural network training graph. DETAILED DESCRIPTION

[0079] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0080] Reference Figure 1 , a GIS partial discharge soundprint detection method based on energy operator improved wavelet packet includes the following steps:

[0081] Step 1: Select GIS voiceprint signals and establish a sample data set. The sample data set contains partial discharge fault data. The fault types include: metal tip defects, floating potential defects, free metal particle defects, insulation surface or hole discharge defects;

[0082] Step 2: Select wavelet packet basis function to perform wavelet packet decomposition and noise reduction on GIS voiceprint signal;

[0083] Step 2.1: According to the characteristics of GIS partial discharge signal, a 4-layer Morse basis function is used to perform wavelet packet decomposition on the GIS voiceprint signal, where the function expression of Morse in the frequency domain is:

[0084]

[0085] In the formula, ψ β,γ represents the Morse wavelet basis function, K β,γ =2(eγ / β) β / γ is a standard constant, e is an Euler number, H(w) is a unit step function, β and γ are two variable parameters, and both are greater than 0;

[0086]

[0087]

[0088] Where W 2n (t) represents the wavelet coefficient when the decomposition level is 2n, W 2n+1 (t) represents the wavelet coefficient when the decomposition level is 2n+1, h0(k) and h1(k) represent the conjugate filter coefficients, t represents the time variable, that is, the position on the time axis, k represents the translation of the wavelet function, W(2t-k) represents the scaling transformation on the time axis, and n represents the number of levels of wavelet function decomposition; when n=0, represents the scaling function, W1(t)=ψ(t) represents the wavelet function, and the wavelet packet corresponding to W0(t) is {W n (t)}, where n∈Z; Z represents a positive integer, It represents a scaling function, which is a low-pass filter used to smooth the signal and extract the low-frequency part of the signal. ψ(t) is a band-pass filter used to analyze the high-frequency part of the signal.

[0089] Since the wavelet packet decomposition results in a binary tree structure, the binary tree node is recorded as (j, i), where j is the wavelet packet decomposition layer number and i is the number of nodes. Therefore, the wavelet packet coefficient of the i-th node in the j-th layer is:

[0090]

[0091]

[0092] In the formula, is the low-frequency wavelet packet coefficient of the i-th node in the j-th layer is the high-frequency wavelet packet coefficient of the i-th node in the j-th layer; It means that the signal with offset k is filtered by low-pass filter at the i-th node of the j-1 layer to obtain the low-frequency coefficient of the j-th layer. The i-th node of the representation layer uses a high-pass filter to filter the signal with an offset of k to obtain the high-frequency coefficient of the j-th layer, where k is the offset;

[0093] Step 2.2: De-noise the wavelet packet coefficients obtained by wavelet packet decomposition. Based on the defects and shortcomings of the traditional threshold function, the present invention adopts an improved wavelet packet threshold method:

[0094]

[0095]

[0096] Where σ is the noise standard deviation of the signal; N is the signal length; η(d j ,λ) is the wavelet packet coefficient d j The processing result after threshold processing, d j is the wavelet packet coefficient, λ is a threshold parameter used to control the amplitude of threshold processing;

[0097] Step 2.3 reconstructs the processed wavelet packet coefficients to obtain a signal with a higher signal-to-noise ratio.

[0098] Step 3: Use the improved Teager energy operator wavelet packet coefficient to calculate the instantaneous energy and obtain the instantaneous energy sequence. Calculate the mean value and standard deviation of the instantaneous energy within the step length, and compare the energy value in the area with the reference threshold. If it is higher than the threshold, it is recorded as an energy anomaly point, otherwise it is judged as the background (i.e., the normal signal area).

[0099] Step 3.1 uses the improved Teager energy operator to calculate the instantaneous energy of the wavelet packet coefficients based on step 2, where the improved Teager energy operator is defined as:

[0100]

[0101] Among them, r(t) is the value of the processed signal, ψ[r(t)] represents the Teager energy value at t sample points in time. On the basis of the original Teager energy operator, adding half of the scores of two adjacent points can improve the resolution and accuracy of the processed data;

[0102] The instantaneous energy expression of the Teager energy operator is:

[0103] E (i,j) =ψ[r(t)] 2

[0104] In the formula, E (i,j)is the instantaneous energy of the i-th node in the j-th layer, t is the position on the time axis at this time, then the instantaneous energy sequence is:

[0105] E N ={E (1,1) ,……,E (i,j)}

[0106] In the formula, E N is the instantaneous energy sequence, E (1,1) is the instantaneous energy of the first node in layer 1, E (i,j) is the instantaneous energy of the i-th node in the j-th layer.

[0107] Step 3.2: Calculate the energy background threshold. According to the instantaneous energy sequence of the GIS voiceprint signal in normal operation, dynamically determine the background threshold of the instantaneous energy sequence of the GIS voiceprint signal that is similar in time domain to see if there is an energy anomaly point, and record the location of the energy anomaly point in S location ;

[0108] Step 3.2.1 Define the starting window size Window size and step length Step size , respectively represent the number of data in the initial window and the initial sliding step size;

[0109]

[0110]

[0111] F s is the sampling rate of GIS voiceprint signal, Window size (i) is the i-th window size, Step size (i) is the i-th step length in seconds, T(i) is the length of the i-th frame in seconds, Step len is the number of sampling points in the i-th window, and i is the sequence number of the i-th step length.

[0112] Step 3.2.2 Calculate the standard deviation of the initial window and determine the threshold E based on the standard deviation of the sample data bg_threshold , and E bg_threshold As a criterion, the points exceeding the threshold are determined as energy abnormal points, and the energy abnormal points are marked;

[0113]

[0114] E bg_threshold =E (i,avg) +2*E (i,std)

[0115] In the formula, E (i,std) is the instantaneous energy standard deviation of each node in the initial window, E(i,avg) is the average instantaneous energy of each node.

[0116] Then, it is determined whether the signal energy of the current frame exceeds the background threshold, and the energy abnormal point is recorded in the energy abnormal point mark position information S location

[0117] Step 3.2.3 repeats steps 3.2.1 and 3.2.2 in the next step interval until the data ends.

[0118] Step 4: For energy outliers, the fuzzy entropy, kurtosis entropy, instantaneous energy sequence and energy outlier mark position information are combined to form a joint feature;

[0119] Fuzzy entropy: S FE (X,m,r,N)=lnΨ m (r)-lnΨ m+1 (r)

[0120] Where X = {x(i), i = 1, 2, ..., N} is a time series with a length of N, x(i) is a time series data element, m is the dimension, r is the similarity tolerance, Ψ m for in

[0121]

[0122]

[0123] i,j=1,2,…,Nm,i≠j;

[0124] x0(i) is the mean of m consecutive data in the time series data X;

[0125] is the m-dimensional vector after the phase space reconstruction of the time series X = {x(i), i = 1, 2, ..., N}, for and The maximum absolute value of the corresponding element difference, yes and The similarity.

[0126] Kurtosis entropy:

[0127] Where: represents the ratio of the kurtosis of the GIS voiceprint signal frequency distribution in the i-th step to the sum of the kurtosis of all GIS voiceprint signal frequency distributions; η iis the kurtosis of the frequency distribution of the GIS voiceprint signal within the i-th step; N2 represents the number of distinguishable frequency distribution characteristics of the GIS voiceprint signal

[0128] Joint feature T E =[S FE ,S QE ,E N ,S location ].

[0129] Step 5: Based on the joint feature data set, the fault diagnosis model is constructed using the RBF neural network, and the fault diagnosis model is used to diagnose and judge the GIS voiceprint signal monitored online. Figure 2 , the specific process is as follows:

[0130] Step 5.1 divide the joint feature dataset into a training set and a test set;

[0131] Step 5.2: Use the training set to train the RBF neural network;

[0132] Step 5.2.1 Determine the input vector X = [x1, x2, ... x n ] T , where n represents the number of input units in the input layer, x n represents the input signal of the nth input unit; determines the output vector Y=[y1,y2…y q ] T , q represents the number of output units in the output layer, y q represents the output signal of the qth output unit; determine the expected output O = [o1, o2…o m ] T , m is the number of output units.

[0133] Furthermore, the connection weight W from the hidden layer to the output layer is initialized v =[w v1 w v2 ...w vp ] T , w vp is the connection weight from the vth neuron in the hidden layer to the pth neuron in the output layer, T represents transposition, and the weight initialization formula is:

[0134] w vp =rand(min v ,max v )

[0135] In the formula, w vp is the connection weight from the vth neuron in the hidden layer to the pth neuron in the output layer. Random initialization is used here to increase the diversity of the network and avoid falling into the local optimal solution. vis the minimum value of all expected outputs of the vth output neuron in the training set; max v is the maximum value of all expected outputs of the vth output neuron in the training set.

[0136] Furthermore, the central parameters c of each neuron in the hidden layer are initialized j =[c j1 ,c j2 ……c ji ] T

[0137]

[0138] In the formula, c ji is the central parameter of the i-th input of the j-th neuron in the hidden layer, g is the total number of neurons in the hidden layer, and p is the number of output units in the output layer; min i is the minimum value of all input information of the i-th feature in the training set, max i It is the maximum value of all input information of the i-th feature in the training set;

[0139] Furthermore, initialize the width vector D j =[d j1 ,d j2 ……d ji ],

[0140]

[0141] Where, d ji is the width of the ith input of the jth neuron in the hidden layer, represents the i-th input of the v-th sample in the training set, N represents the number of samples in the training set, and d f is the width adjustment coefficient, which is less than 1. Its function is to make each hidden layer neuron more sensitive to local information, which is beneficial to improve the local response ability of the RBF neural network.

[0142] Step 5.2.2 Calculate the output value z of the jth neuron in the hidden layer j , X represents the input vector of the input layer, j = 1, 2...g, C j is the center vector of the jth neuron in the hidden layer, which is composed of the center components of all neurons in the input layer corresponding to the jth neuron in the hidden layer.

[0143] C j =[c j1 ,c j2 ,……c jn ] T ;D jis the width vector of the jth neuron in the hidden layer, and C j Correspondingly,

[0144] D j =[d j1 ,d j2 ……d ji ] T , D j The larger it is, the greater the influence of the hidden layer on the input layer, and the smoothness between neurons is better. ||.|| is the Euclidean norm;

[0145] Step 5.2.3 calculates the output of the output layer neurons:

[0146] Y=[y1,y2……y q ] T

[0147]

[0148] In the formula, v = 1, 2, ... q, where w kj is the adjustment weight between the vth neuron in the output layer and the jth neuron in the hidden layer;

[0149] Step 5.2.4 In view of the fact that traditional RBF neural network weight parameter training methods, such as the gradient descent method, will fall into local optimal solutions and slow convergence, the improved Adam algorithm is used to adjust the learning rate to better control the model training process and improve the performance and robustness of the model.

[0150] Optimize the Adam algorithm and dynamically adjust the learning rate based on the learning rate decay method:

[0151]

[0152] Among them, α(u) is the learning rate of the uth iteration, α0 is the initial learning rate, β1 is the constant factor, and u is the number of iterations;

[0153] The iterative calculation is as follows:

[0154]

[0155]

[0156]

[0157] In the formula, w vj (u) The adjustment weight between the vth output neuron and the jth hidden layer neuron in the uth iteration calculation, v = 1, 2, ... q, j = 1, 2 ... g; c ji(u) is the central component of the jth hidden layer neuron for the ith input neuron at the uth iteration; d ji (u) for learning center c ji (u) corresponding width; Indicates w vj The first moment estimate of (u), Indicates w vj The second moment estimate of (ut), Indicates c ji The first moment estimate of (u), Indicates c ji The second moment estimate of (u), Indicates d ji The first moment estimate of (u), Indicates d ji The second-order moment estimates of (u) are respectively expressed as follows: their initial values ​​are 0, ∈ is a minimum value to avoid the denominator being 0, and η is the learning factor

[0158] E is the RBF neural network evaluation function:

[0159]

[0160] In the formula, O lv is the expected output value of the vth output neuron at the lth input sample; y lv is the network output value of the vth output neuron at the lth input sample;

[0161] In summary, the learning algorithm of RBF neural network is given, and the feasibility of the algorithm is verified using the test set:

[0162] A1: Initialize the neural network parameters according to step 5.2.1, and give the values ​​of η and α and the value of iteration accuracy ε;

[0163] A2: Calculate the root mean square error (RMS) of the network output according to the following formula. If RMS ≤ ε, the training ends. Otherwise, go to step A3.

[0164] A3: According to the weight iteration calculation in step 5.2.4, iteratively calculate the adjustment weight, center and width parameters;

[0165] A4: Return to step A2.

[0166] The above-described embodiments are merely descriptions of preferred implementations of the present invention, and are not intended to limit the concept and scope of the present invention. Without departing from the design concept of the present invention, various modifications and improvements made by ordinary persons in the art to the technical solution of the present invention should all fall within the scope of protection of the present invention.

Claims

1. A GIS partial discharge soundprint detection method based on energy operator improved wavelet packet, characterized in that: The following steps are involved: Step 1: Select GIS voiceprint signals and establish a sample data set; Step 2: Select wavelet packet basis function to perform wavelet packet decomposition and noise reduction on GIS voiceprint signal; Step 3: Use the improved Teager energy operator wavelet packet coefficient to calculate the instantaneous energy and obtain the instantaneous energy sequence. Calculate the average value and standard deviation of the instantaneous energy within the step length, and compare the energy value in the area with the reference threshold. If it is higher than the threshold, it is recorded as an energy anomaly point, otherwise it is judged as background. The improved Teager energy operator is defined as: ； in, is the value of the processed signal, Represents the Teager energy value at the sample point at time t; The instantaneous energy expression of the Teager energy operator is: ； In the formula, is the instantaneous energy of the i-th node in the j-th layer, then the instantaneous energy sequence is: ={ ……, }； In the formula, is the instantaneous energy sequence, is the instantaneous energy of the first node in layer 1, is the instantaneous energy of the i-th node in the j-th layer; Step 4: For energy outliers, the fuzzy entropy, kurtosis entropy, instantaneous energy sequence and energy outlier mark position information are combined to form a joint feature; Step 5: Based on the joint feature data set, the fault diagnosis model is constructed using the RBF neural network, and the fault diagnosis model is used to diagnose and judge the GIS voiceprint signal monitored online.

2. According to claim 1, the GIS partial discharge soundprint detection method based on energy operator improved wavelet packet is characterized in that: The sample data set contains partial discharge fault data, and the fault types include: metal tip defects, floating potential defects, free metal particle defects, insulation surface or hole discharge defects.

3. The GIS partial discharge soundprint detection method based on energy operator improved wavelet packet according to claim 1 is characterized in that: The process for step 2 is as follows: Step 2.1: Use the 4-layer Morse basis function to perform wavelet packet decomposition on the GIS voiceprint signal to obtain the wavelet packet coefficients; Step 2.2: De-noise the wavelet packet coefficients obtained by wavelet packet decomposition: ； ； In the formula, is the noise standard deviation of the signal; N is the signal length; is the wavelet packet coefficient The processing result after threshold processing is is the wavelet packet coefficient, is a threshold parameter used to control the amplitude of threshold processing; Step 2.3 reconstructs the processed wavelet packet coefficients to obtain a signal with a higher signal-to-noise ratio.

4. The GIS partial discharge soundprint detection method based on energy operator improved wavelet packet according to claim 1 is characterized in that: The process for step 3 is as follows: Step 3.1 Use the improved Teager energy operator to calculate the instantaneous energy of the wavelet packet coefficients; Step 3.2: Calculate the energy background threshold. According to the instantaneous energy sequence of the GIS voiceprint signal in normal operation, dynamically determine the background threshold of the instantaneous energy sequence of the GIS voiceprint signal that is similar in time domain to see if there is an energy anomaly point, and record the location of the energy anomaly point in the .

5. The GIS partial discharge soundprint detection method based on energy operator improved wavelet packet according to claim 4 is characterized in that: Step 3.2 is as follows: Step 3.2.1: Define the starting window size With step length , respectively represent the number of data in the initial window and the initial sliding step size; ； ； is the GIS voiceprint signal sampling rate, is the i-th window size, is the i-th step length, in seconds, is the length of the i-th frame in seconds, is the number of sampling points in the i-th window, and i is the sequence number of the i-th step length; Step 3.2.2: Calculate the standard deviation of the initial window and determine the threshold based on the sample data standard deviation , and As a criterion, the points exceeding the threshold are determined as energy abnormal points, and the energy abnormal points are marked; ； ； In the formula, is the instantaneous energy standard deviation of each node in the initial window, is the average instantaneous energy of each node; Then determine whether the signal energy of the current frame exceeds the background threshold, and record the energy abnormal point in the energy abnormal point mark position information ; Step 3.2.3: Repeat steps 3.2.1 and 3.2.2 for the next step interval until the data ends.

6. The GIS partial discharge soundprint detection method based on energy operator improved wavelet packet according to claim 1 is characterized in that: The process of using the training set to train the RBF neural network and build a fault diagnosis model is as follows: Step 5.2.1 Determine the input vector , where n represents the number of input units in the input layer, Represents the input signal of the nth input unit; determines the output vector , q represents the number of output units in the output layer, represents the output signal of the qth output unit; determines the expected output , m is the number of output units; Initialize the connection weights from the hidden layer to the output layer , is the connection weight from the vth neuron in the hidden layer to the pth neuron in the output layer, T represents transposition, and the weight initialization formula is: ； In the formula, is the connection weight from the vth neuron in the hidden layer to the pth neuron in the output layer. Random initialization is used here to increase the diversity of the network and avoid falling into the local optimal solution. is the minimum value of all expected outputs of the vth output neuron in the training set; is the maximum value of all expected outputs of the vth output neuron in the training set; Initialize the central parameters of each neuron in the hidden layer ; ； In the formula, is the central parameter of the i-th input of the j-th neuron in the hidden layer, g is the total number of neurons in the hidden layer, and p represents the number of output units in the output layer; is the minimum value of all input information of the i-th feature in the training set, is the maximum value of all input information of the i-th feature in the training set; Initialize the width vector ], ； In the formula, is the width of the ith input of the jth neuron in the hidden layer, represents the i-th input of the v-th sample in the training set, N represents the number of samples in the training set, is the width adjustment coefficient, the value is less than 1; Step 5.2.2 Calculate the output value of the jth neuron in the hidden layer , ,X represents the input vector of the input layer, j=1,2……g, is the center vector of the jth neuron in the hidden layer, which is composed of the center components of all neurons in the input layer corresponding to the jth neuron in the hidden layer; ; is the width vector of the jth neuron in the hidden layer, and Correspondingly, , The larger it is, the greater the influence of the hidden layer on the input layer, and the smoothness between neurons is also better. is the Euclidean norm; Step 5.2.3: Calculate the output of the output layer neurons: ； ； In the formula, v = 1, 2, ... q, where is the adjustment weight between the vth neuron in the output layer and the jth neuron in the hidden layer; Step 5.2.4: Use the improved Adam algorithm to adjust the learning rate: ； in, is the learning rate of the u-th iteration, is the initial learning rate, is a constant factor, is the number of iterations; The iterative calculation is as follows: ； ； ； In the formula, The adjustment weight between the vth output neuron and the jth hidden layer neuron at the uth iteration calculation, v=1,2,...q, j=1,2...g; is the central component of the j-th hidden layer neuron for the i-th input neuron at the u-th iteration; For and Learning Center The corresponding width; express The first moment estimate of express The second moment estimate of express The first moment estimate of express The second moment estimate of express The first moment estimate of express The second-order moment estimates of are respectively expressed as follows: their initial values ​​are 0, is a minimum value, avoiding the denominator to be 0, is the learning factor; E is the RBF neural network evaluation function: ； In the formula, is the expected output value of the vth output neuron at the lth input sample; is the network output value of the vth output neuron at the lth input sample.

7. The GIS partial discharge soundprint detection method based on energy operator improved wavelet packet according to claim 6 is characterized in that: Use the test set to verify the feasibility of the algorithm: A1: Initialize the neural network parameters according to step 5.2.1 and give and The value and iteration accuracy of The value of A2: Calculate the root mean square error RMS value of the network output. , the training ends, otherwise go to step A3; A3: According to the weight iteration calculation in step 5.2.4, iteratively calculate the adjustment weight, center and width parameters; A4: Return to step A2.

Images