A device fault fusion judgment method based on voiceprint model evolution for power grid monitoring

Abstract

The present invention provides a method for fusion judgment of device faults of voiceprint model evolution for power grid monitoring, including: real-time audio acquisition, interference noise removal preprocessing by dual classification denoising method based on wavelet transform; extracting MFCC feature vectors of device sound signal data under various normal states after preprocessing, respectively obtaining sample data sets under various normal states, establishing and updating Gaussian mixture models of target devices under various normal states; extracting audio sample data to be tested and establishing a multi-Gaussian mixture model evolution method based on voting to match the characteristic law of voiceprint of target devices that changes over time; based on the multi-Gaussian mixture model, performing fusion judgment of voiceprint faults of target devices. The method of the present invention can eliminate the influence of weak abnormal interference and strong abnormal interference of the environment, can match the changing law of sample data in real time, effectively monitor power grid monitoring scenes, realize abnormal monitoring and timely early warning.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid monitoring and voiceprint fault monitoring and early warning, and specifically but not limited to a method for fault fusion judgment of voiceprint model evolution equipment for power grid monitoring. Background Art

[0002] With the continuous development of science and technology, more and more power distribution industrial equipment has appeared in people's vision. Grid monitoring is to supervise and control the operation of the grid, and to promptly discover and solve faults or abnormal operation conditions in the operation of the grid. In the process of the development of smart grids, the intelligence of grid monitoring has become an important aspect, which is related to the stability and safety of the entire power system. At present, the traditional monitoring method generally adopts large sample machine learning algorithms, which not only has high maintenance costs and low efficiency, but also has a serious shortage of abnormal samples in grid monitoring scenarios, affecting the accuracy of discrimination. However, the existing maintenance methods rely on prior knowledge, and the identification capabilities of different personnel are different. There is no effective method that can replace the monitoring of maintenance personnel.

[0003] In response to the above problems, monitoring methods based on adaptive algorithms have received more and more attention. Matching the algorithm to the main equipment in the power grid monitoring scenario for one-to-one monitoring can not only reduce the dispatching cost, but also greatly improve the maintenance efficiency. Existing fault monitoring methods usually focus on sensing aspects such as humidity, pressure and temperature, but they are not as direct and effective as visual monitoring or auditory monitoring. Although dispatching professionals for regular maintenance can reduce the occurrence of faults in the power grid monitoring center to a certain extent, the labor cost is high and the monitoring efficiency is low. Therefore, the audio monitoring method has gradually become known to people because of its high efficiency and low cost. Fault monitoring and early warning based on voiceprint can detect faults in time and issue early warnings in time.

[0004] In view of this, it is necessary to provide a new method to solve at least part of the above problems. Summary of the invention

[0005] In response to one or more problems in the prior art, the present invention proposes a voiceprint model evolution device fault fusion judgment method for power grid monitoring, which can monitor the power grid monitoring scene in real time and effectively, realize abnormal monitoring and timely warning.

[0006] The technical solution to achieve the purpose of the present invention is:

[0007] A method for fusion judgment of device faults based on voiceprint model evolution for power grid monitoring, comprising:

[0008] Step 1: Collect the real-time sound signals of the power grid equipment, and use double classification denoising based on wavelet transform to pre-process the real-time sound signals to remove the environmental interference noise;

[0009] Step 2: Extract the MFCC feature vectors of the preprocessed sound signals of the power grid equipment in each normal state, construct a sample data set in each normal state, and establish and update the Gaussian mixture model of the target equipment in each normal state according to the sample data set;

[0010] Step 3: Extract the sound sample data to be tested and establish a voting-based multi-Gaussian mixture model evolution method to match the voiceprint feature law of the target device that changes over time;

[0011] Step 4: Perform voiceprint fault fusion judgment on the target device based on the multi-Gaussian mixture model.

[0012] Furthermore, the voiceprint model evolution device fault fusion judgment method for power grid monitoring of the present invention, step 1 specifically includes:

[0013] Step 1-1: The environmental interference noise of the target power grid equipment includes the superposition of background noise and abnormal interference. The real-time sound signal X(t) is subjected to discrete wavelet transform to obtain the wavelet coefficient W i,k :

[0014]

[0015] Among them, a i =2 -i is the scaling factor, which is discretized according to the power series; i is the number of wavelet decomposition layers, 1≤i≤I, I represents the total number of wavelet decomposition layers, k is the translation factor, representing the wavelet coefficient number; t is the discrete acquisition time of the signal; is the wavelet function;

[0016] Step 1-2: Perform the first level of background noise removal: If the real-time sound signal amplitude is less than the background noise discrimination threshold λ′ i , then remove the corresponding wavelet coefficient label, otherwise keep it;

[0017] The background noise discrimination threshold λ′ i for:

[0018]

[0019] Among them, σ 1 Represents the background noise compensation coefficient, which is proportional to the background noise intensity, σ 1 >1; r represents the average distance from the surrounding equipment to the target power grid equipment; r 0 is the distance normalization constant used to overcome the near-far effect; γ is the signal length; α represents the enhancement factor of the abnormal factor. If there is no abnormal factor, α = 1. If there is an abnormal factor, then:

[0020] α=1+LΔ

[0021] Among them, Δ is the enhancement step length, and L is the number of abnormal interference factors;

[0022] Step 1-3: Perform the second level of abnormal interference removal: If the real-time sound signal amplitude is greater than the abnormal interference judgment threshold Then remove the corresponding wavelet coefficient label, otherwise keep it;

[0023] The abnormal interference discrimination threshold for:

[0024]

[0025] Among them, σ 2 represents the abnormal interference compensation coefficient under abnormal conditions, which is proportional to the abnormal interference intensity, and σ 2 >>σ 1 ;

[0026] The abnormal interference factors are obtained by using the video monitoring system and the sensor data collection system. The current abnormal interference factors are L in total. The sound center frequency point generated by the lth abnormal interference factor is obtained by prior statistics as f l , and according to X(t)>th 1 Get the corresponding The central point t of the time period τ , get the expansion factor a corresponding to the abnormal interference factor i And the wavelet coefficient index k:

[0027]

[0028] k=v(t τ )

[0029] Among them, 1 is the first threshold, u(·) represents the exponential approximation process, v(·) represents the rounding process, 1≤l≤L,

[0030] Then the wavelet decomposition layer number i and wavelet coefficient number k corresponding to the abnormal interference factor are obtained;

[0031] Step 1-4: De-noising by double classification to obtain the reconstructed wavelet coefficients W′ i,k :

[0032]

[0033] Among them, W i,k ′ is the wavelet coefficient reconstructed after threshold processing;

[0034] Step 1-5: Reconstruct the wavelet coefficients W′i,k The corresponding inverse discrete wavelet transform is performed to obtain the sound signal X′(t) of the target power grid equipment after removing the interference noise.

[0035] Furthermore, in the voiceprint model evolution device fault fusion judgment method for power grid monitoring of the present invention, the step of extracting the MFCC feature vector in step 2 includes:

[0036] Step 2-1: Divide the sound signal X′(t) of the target power grid equipment into I frames, each frame contains N=512 samples, and adjacent frames have 256 samples overlapped. Then the audio signal x′(t) of the lth frame is l (n) is:

[0037] x l (n) = x(l*N+n)w(n)

[0038] Among them, x is the audio signal after framing, l is the audio signal frame number, N is the window length, n is the number of window movement steps, x(.) represents framing processing, and the Hanning window w(n) is:

[0039]

[0040] Step 2-2: Design a filter with 36 frequency bands, the number of filters is L = 36, the frequency range is 0 Hz to 5512 Hz, and the output of each triangular filter is m(q), where q = 1, 2…, L;

[0041] Step 2-3: Perform logarithmic operation on the output m(q) of all filters, and then perform discrete cosine transform to obtain a set of MFCC parameters:

[0042]

[0043] Among them, β=0,1,…,R,1≤R≤L, β is the MFCC feature vector label index.

[0044] Furthermore, in the voiceprint model evolution device fault fusion judgment method for power grid monitoring of the present invention, in step 2, N Gaussian mixture models are established according to the sample data set, each Gaussian mixture model is a mixture of M Gaussian distributions, wherein the nth normal state Gaussian mixture model is expressed as:

[0045]

[0046] Where, 1≤n≤N, 1≤m≤M, n represents the index of the Gaussian mixture model, m represents the index of each sub-model of the Gaussian mixture model, and x q represents the input of the Gaussian mixture model, q represents the input sample number, 1≤q≤Q, Q is the cumulative number of input samples, P(xq |u nm ,δ nm ) is the mth Gaussian distribution at x q The probability density value, w nm is the weight of the mth Gaussian distribution of the nth normal Gaussian mixture model, u nm is the mean of the mth Gaussian distribution of the nth normal Gaussian mixture model, δ nm is the variance of the mth Gaussian distribution of the nth normal Gaussian mixture model.

[0047] Furthermore, in the voiceprint model evolution device fault fusion judgment method for power grid monitoring of the present invention, the parameter update rule of the Gaussian mixture model is:

[0048] 1) Initialize the parameters u of the nth normal state Gaussian mixture model nm (0) ,δ nm (0) ;

[0049] 2) Calculate the probability P based on the current Gaussian mixture model nm (x q ):

[0050]

[0051] Among them, j represents the traversal label, P nm (x q ) represents the sample x q The probability generated by the mth sub-model of the nth normal Gaussian mixture model;

[0052] 3) Calculate the model parameters for the next iteration:

[0053]

[0054]

[0055]

[0056] Among them, w nm ′、μ nm ′、δ nm ′ respectively represent the weight, mean, and variance of the mth Gaussian distribution of the nth normal state Gaussian mixture model after iteration;

[0057] 4) Repeat steps 2) and 3) until the Gaussian mixture model of the nth normal state converges.

[0058] Furthermore, in the voiceprint model evolution device fault fusion judgment method for power grid monitoring of the present invention, step 3 specifically includes:

[0059] Step 3-1: Calculate the audio sample data x to be tested c The probability P of belonging to the Gaussian mixture model of the nth normal state n (x c ):

[0060]

[0061] Among them, m represents the index of each sub-model of the Gaussian mixture model, 1≤m≤M, w nm is the weight of the mth Gaussian distribution of the nth normal Gaussian mixture model, u nm is the mean of the mth Gaussian distribution of the nth normal Gaussian mixture model, δ nm is the variance of the mth Gaussian distribution of the nth normal Gaussian mixture model, P(x c |u nm ,δ nm ) is the mth Gaussian distribution at x c The probability density value of ;

[0062] According to the probability P n (x c ) Get the current audio sample data x to be tested c The Gaussian mixture model of the most likely normal state n * for:

[0063]

[0064] Step 3-2: According to the fault fusion decision rule, if the current audio sample data to be tested x c If the current audio sample data x is in normal state, c Belongs to the nth v The normal Gaussian mixture model votes to determine whether to evolve it. If it evolves, go to step 3-3;

[0065] Step 3-3: For the nth * The normal Gaussian mixture model is evolved as follows:

[0066]

[0067]

[0068]

[0069]

[0070] Among them, h and j represent the corresponding traversal labels respectively. Indicates the value of x cThe average value of , H represents the voting threshold, For each x in the historical records c value, ρ is the evolution rate, [·] c+1 ,[·] c Respectively represent the values ​​of the corresponding variables after evolution and before evolution, [·] T Represents vector transpose;

[0071] Step 3-4: Let n * Voting variables for Gaussian mixture models And clear the history of the model.

[0072] Furthermore, in the voiceprint model evolution device fault fusion judgment method for power grid monitoring of the present invention, the voting rule in step 3-2 is:

[0073] 1) Initialize the nth * Voting variables for Gaussian mixture models

[0074] 2) If

[0075]

[0076] Then

[0077]

[0078] Among them, 2 is the second threshold, |·| represents the norm, For nth * The weight of the mth Gaussian distribution of a normal Gaussian mixture model, For nth * The mean of the mth Gaussian distribution of a normal Gaussian mixture model;

[0079] like

[0080]

[0081] Then

[0082]

[0083] Among them, 3 is the third threshold, and th 3 >th 2 ;

[0084] like

[0085]

[0086] Then

[0087]

[0088] 3) Judgment Is it true? If true, then for the nth * The Gaussian mixture model with a normal state is evolved, otherwise it does not evolve, and H represents the voting threshold.

[0089] Furthermore, in the voiceprint model evolution device fault fusion judgment method for power grid monitoring of the present invention, the rule for performing voiceprint fault fusion judgment of the target device according to the multi-Gaussian mixture model in step 4 is:

[0090] Rule 1: The probability of belonging to the Gaussian mixture model of the most likely normal state at present

[0091]

[0092] Rule 2: Entropy E of the current system c :

[0093] E c >th 5

[0094]

[0095] Rule 3: There exists a Gaussian mixture model n:

[0096]

[0097] Among them, 4 is the failure probability threshold, th 5 is the entropy threshold, th 6 is the evolution threshold, w nm The initial value of for u nm The initial value of

[0098] If any one of the above three rules is met, it is judged as a fault; if none of them are met, it is a normal state.

[0099] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0100] 1. The voiceprint model evolution device fault fusion judgment method for power grid monitoring of the present invention can eliminate the influence of weak abnormal interference and strong abnormal interference in the environment, and effectively monitor the power grid monitoring scene.

[0101] 2. The detection phase of the voiceprint model evolution device fault fusion judgment method for power grid monitoring of the present invention uses a voting-based multi-Gaussian mixture model evolution method, which can match the change law of sample data in real time, realize abnormal monitoring and timely warning.

[0102] 3. The voiceprint model evolution device fault fusion judgment method for power grid monitoring of the present invention has high detection accuracy, which can effectively improve maintenance efficiency and reduce dispatching costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0103] The accompanying drawings are used to provide a further understanding of the present invention and are used together with the description to explain the embodiments of the present invention, but do not constitute a limitation of the present invention. In the accompanying drawings:

[0104] Figure 1 The flowchart of the voiceprint model evolution device fault fusion judgment method for power grid monitoring of the present invention is shown. DETAILED DESCRIPTION

[0105] In order to further understand the present invention, preferred embodiments of the present invention are described below in conjunction with examples. However, it should be understood that these descriptions are only for further illustrating the features and advantages of the present invention, rather than limiting the claims of the present invention.

[0106] The description in this section is only for typical embodiments, and the present invention is not limited to the scope of the embodiments. The combination of different embodiments, the mutual replacement of some technical features in different embodiments, and the mutual replacement of the same or similar prior art means with some technical features in the embodiments are also within the scope of the description and protection of the present invention.

[0107] A voiceprint model evolution device fault fusion judgment method for power grid monitoring, such as Figure 1 As shown, the specific steps include:

[0108] Step 1: Collect sound signals in real time and use a double classification denoising method based on wavelet transform to remove interference noise preprocessing. Specifically include:

[0109] The target power grid equipment environmental interference noise is regarded as the superposition of background noise and abnormal interference. First, the signal X(t) is subjected to discrete wavelet transform:

[0110]

[0111] Among them, a i =2 -i is the scaling factor, which is discretized according to the power series; i is the number of wavelet decomposition layers, 1≤i≤I, I represents the total number of wavelet decomposition layers, k is the translation factor, representing the wavelet coefficient number; t is the discrete acquisition time of the signal; is the wavelet function;

[0112] Perform the first level of background noise removal, and the background noise discrimination threshold is:

[0113]

[0114] Among them, i is the number of wavelet decomposition layers, γ is the signal length, σ 1 represents the background noise compensation coefficient under normal circumstances, which is proportional to the background noise intensity, σ 1 >1, α represents the reinforcement factor of abnormal factors. If there is no abnormal factor, α=1. If there is an abnormal factor, then:

[0115] α=1+LΔ (1-3)

[0116] Among them, Δ is the enhancement step length, L is the number of factors, such as the number of staff entering the monitoring environment, the number of natural phenomena such as thunderstorms, etc.; r represents the average distance from the surrounding equipment to the target equipment, r 0 is the distance normalization constant used to overcome the near-far effect.

[0117] If the real-time sound signal amplitude is less than the above threshold λ′ i , then remove the corresponding wavelet coefficient label, otherwise keep it.

[0118] The second level of abnormal interference removal is performed, and external auxiliary methods such as video monitoring systems and sensor data collection systems are used to obtain abnormal interference factors. Abnormal interference factors include: knowing the staff entering the environment through face recognition, and knowing the current weather conditions by accessing the meteorological system, including whether it is thundering, whether it is raining, etc.; the current abnormal interference factors are L in total, and the sound center frequency generated by the lth factor is f l , which can be obtained by prior statistics and by X(t)>th 1 , th 1 is the threshold, and the corresponding 1 of The central point t of the time period τ , thus we get the corresponding wavelet layer number i, namely:

[0119]

[0120] b i =t τ (1-5)

[0121] Among them, 1≤l≤L,

[0122] The second level of abnormal interference discrimination threshold for:

[0123]

[0124] Among them, σ 2 It represents the abnormal interference compensation coefficient under abnormal conditions, which is proportional to the abnormal interference intensity, σ2 >>σ 1 If the amplitude of the real-time sound signal is greater than the above threshold, the corresponding wavelet coefficient label is removed, otherwise it is retained.

[0125] Through the double classification denoising method, the reconstructed wavelet coefficients are obtained:

[0126]

[0127] Among them, W i,k ' is the wavelet coefficient after threshold processing. The wavelet coefficient after threshold processing is subjected to corresponding discrete wavelet inverse transform to obtain the target device sound signal X'(t) with interference removed.

[0128] Step 2: Extract the MFCC feature vectors of the pre-processed device sound signal data under various normal states, obtain the sample data sets under various normal states, and establish and update the Gaussian mixture model of the target device under various normal states. Specifically:

[0129] The sound signal data of various power grid equipment under normal conditions are collected, and the MFCC feature vectors are extracted after preprocessing to obtain sample data sets under normal conditions. N Gaussian mixture models are established and updated. Each generated Gaussian mixture model is a mixture of M Gaussian distributions, where the weight of the mth Gaussian distribution of the nth normal state Gaussian mixture model is w nm , the mean value is u nm , the variance is δ nm .

[0130] Among them, the Gaussian mixture model of the nth normal state can be expressed as:

[0131]

[0132] Where, 1≤n≤N, 1≤m≤M, n represents the index of each type of normal Gaussian mixture model, and m represents the index of each sub-model of the Gaussian mixture model; x q represents the input of the Gaussian mixture model, q represents the input sample label, 1≤q≤Q, P(x q |u nm ,δ nm ) is the mth Gaussian distribution at x q The probability density value of .

[0133] The parameter update rule of the Gaussian mixture model is:

[0134] 1) Initialize the parameters of the nth Gaussian mixture model;

[0135] 2) Calculate P based on the current model parameters nm (x q ):

[0136]

[0137] Among them, j represents the traversal label, P nm (x q ) represents the sample x q The probability generated by the mth sub-model of the nth normal Gaussian mixture model;

[0138] 3) Calculate the model parameters for the next iteration:

[0139]

[0140]

[0141]

[0142] Among them, w nm ′、μ nm ′、δ nm ′ respectively represent the weight, mean, and variance of the mth Gaussian model of the nth normal state Gaussian mixture model after iteration;

[0143] 4) Repeat steps 2 and 3 until the nth normal state Gaussian mixture model converges.

[0144] The audio signal X′(t) to be tested is divided into I frames, and a Hanning window is added. Each frame contains N=512 samples, and adjacent frames overlap by 256 samples.

[0145] x l (n) = x(l*N+n)w(n) (2-6)

[0146] Where N is the window length, w(n) is the Hanning window, and x l (n) represents the lth frame audio signal;

[0147]

[0148] Design a filter with 36 bands, L = 36, frequency range 0 to 5512HZ, using overlapping triangular bands, the output of each triangular filter is m(q),

[0149] Where, q = 1, 2…, L, L is the number of filters;

[0150] Perform logarithmic operations on all filter outputs and further perform discrete cosine transform (DCT) to obtain a set of MFCC parameters.

[0151]

[0152] Among them, β=0,1,…,R,1≤R≤L.

[0153] Step 3: Extract the audio sample data to be tested and establish a voting-based multi-Gaussian mixture model evolution method to match the characteristic law of the target device voiceprint that changes over time. Specifically:

[0154] The audio sample data x to be tested c Substituting into each Gaussian mixture model respectively, the probability of belonging to the nth Gaussian distribution is,

[0155]

[0156] Among them, P(x c |u nm ,δ nm ) is the mth Gaussian distribution at x c The probability density value of ;

[0157] Then the Gaussian mixture model of the normal state to which the current audio sample data to be tested is most likely to belong is:

[0158]

[0159] Among them, n * For x c The Gaussian mixture model of the most likely normal state, and then according to the fault fusion decision rule, if the current sample x c is a normal state, then for the current nth * A normal state Gaussian mixture model votes to decide whether to evolve.

[0160] The voting rules are as follows:

[0161] 1) Initialize model n * Voting variables

[0162] 2) If

[0163]

[0164] th 2 is a very small threshold, |·| represents the norm, then let

[0165]

[0166] like

[0167]

[0168] th 3 is a large threshold, then

[0169]

[0170] like

[0171]

[0172] Then

[0173]

[0174] Record x at the same time c The results and further judgment Is it true? If true, then for the nth * Gaussian mixture models are evolved if the model is not used, otherwise no evolution occurs; where H represents the voting threshold.

[0175] The specific evolution method of the Gaussian mixture model is as follows:

[0176]

[0177]

[0178]

[0179]

[0180] Among them, h represents the traversal label, Indicates the value of x c The average value of For each x in the historical records c value, ρ is the evolution rate, [·] c+1 ,[·] c Respectively represent the values ​​of the corresponding variables after evolution and before evolution, [·] T Represents the transpose of a vector.

[0181] Final Order And clear the history of the model.

[0182] Step 4: Based on the multi-Gaussian mixture model, the target device voiceprint fault fusion judgment is performed. The judgment rules are as follows:

[0183] Rule 1: The probability of belonging to the Gaussian mixture model of the most likely normal state at present

[0184]

[0185] Rule 2: Entropy E of the current system c ,

[0186] E c >th 5 (4-2)

[0187]

[0188] Rule 3: There exists a Gaussian mixture model n,

[0189]

[0190] Among them, 4 is the failure probability threshold, th 5 is the entropy threshold, th 6 is the evolution threshold, w nm The initial value of for u nm The initial value of

[0191] If any one of the above three rules is met, it is judged as a fault; if none of them are met, it is a normal state.

[0192] The description and application of the present invention here are illustrative, and it is not intended to limit the scope of the present invention to the above-mentioned embodiments. The effects or advantages involved in the specification may not be reflected in the actual experimental examples due to the uncertainty of specific condition parameters or other factors, and the effects or advantages are not used to limit the scope of the invention. The deformation and change of the embodiments disclosed here are possible, and the replacement of the embodiments and the various equivalent parts are well known to those of ordinary skill in the art. It should be clear to those skilled in the art that the present invention can be implemented in other forms, structures, arrangements, proportions, and with other components, materials and parts without departing from the spirit or essential features of the present invention. Other deformations and changes can be made to the embodiments disclosed here without departing from the scope and spirit of the present invention.

Claims

1. A method for fusion judgment of device faults based on voiceprint model evolution for power grid monitoring, characterized in that: include: Step 1: Collect the real-time sound signals of the power grid equipment, and use the double classification denoising based on wavelet transform to pre-process the real-time sound signals to remove the environmental interference noise; specifically include: Step 1-1: The environmental interference noise of the target power grid equipment includes the superposition of background noise and abnormal interference. The real-time sound signal X(t) is subjected to discrete wavelet transform to obtain the wavelet coefficient W i,k : Among them, a i =2 -i is the scaling factor, which is discretized according to the power series; i is the number of wavelet decomposition layers, 1≤i≤I, I represents the total number of wavelet decomposition layers, k is the translation factor, representing the wavelet coefficient number; t is the discrete acquisition time of the signal; is the wavelet function; Step 1-2: Perform the first level of background noise removal: If the real-time sound signal amplitude is less than the background noise discrimination threshold λ′ i , then remove the corresponding wavelet coefficient label, otherwise keep it; The background noise discrimination threshold λ′ i for: Among them, σ1 represents the background noise compensation coefficient, which is proportional to the background noise intensity, σ1>1; r represents the average distance from the peripheral equipment to the target power grid equipment; r0 is the distance normalization constant, which is used to overcome the near-far effect; γ is the signal length; α represents the enhancement factor of the abnormal factor. If there is no abnormal factor, α=1. If there is an abnormal factor, then: α=1+LΔ Among them, Δ is the enhancement step length, and L is the number of abnormal interference factors; Step 1-3: Perform the second level of abnormal interference removal: If the real-time sound signal amplitude is greater than the abnormal interference judgment threshold Then remove the corresponding wavelet coefficient label, otherwise keep it; The abnormal interference discrimination threshold for: Among them, σ2 represents the abnormal interference compensation coefficient under abnormal conditions, which is proportional to the abnormal interference intensity, and σ2>>σ1; The abnormal interference factors are obtained by using the video monitoring system and the sensor data collection system. The current abnormal interference factors are L in total. The sound center frequency point generated by the lth abnormal interference factor is obtained by prior statistics as f l , and according to X(t)>th1, we can get the central time point t of the corresponding ┏ time period τ , get the expansion factor a corresponding to the abnormal interference factor i And the wavelet coefficient index k: k=v(t τ ) Wherein, th1 is the first threshold, u(·) represents the exponential approximation process, v(·) represents the rounding process, 1≤l≤L, 1≤τ≤┏; Then the wavelet decomposition layer number i and wavelet coefficient number k corresponding to the abnormal interference factor are obtained; Step 1-4: De-noising by double classification to obtain the reconstructed wavelet coefficients W′ i,k : Among them, W i,k ′ is the wavelet coefficient reconstructed after threshold processing; Step 1-5: Reconstruct the wavelet coefficients W′ i,k Perform the corresponding inverse discrete wavelet transform to obtain the sound signal X′(t) of the target power grid equipment after removing the interference noise; Step 2: Extract the MFCC feature vectors of the preprocessed sound signals of the power grid equipment in each normal state, construct a sample data set in each normal state, and establish and update the Gaussian mixture model of the target equipment in each normal state according to the sample data set; Step 3: Extract the sound sample data to be tested and establish a voting-based multi-Gaussian mixture model evolution method to match the voiceprint feature law of the target device that changes over time; specifically including: Step 3-1: Calculate the audio sample data x to be tested c The probability P of belonging to the Gaussian mixture model of the nth normal state n (x c ): Where m represents the index of each sub-model of the Gaussian mixture model, 1≤m≤M, M is the number of Gaussian distributions included in each Gaussian mixture model, and w nm is the weight of the mth Gaussian distribution of the nth normal Gaussian mixture model, u nm is the mean of the mth Gaussian distribution of the nth normal Gaussian mixture model, δ nm is the variance of the mth Gaussian distribution of the nth normal Gaussian mixture model, P(x c |u nm ,δ nm ) is the mth Gaussian distribution at x c The probability density value of ; According to the probability P n (x c ) Get the current audio sample data x to be tested c The Gaussian mixture model of the most likely normal state n * for: Step 3-2: According to the fault fusion decision rule, if the current audio sample data to be tested x c If the current audio sample data x is in normal state, c Belongs to the nth * The normal Gaussian mixture model is voted to determine whether to evolve it. If it is evolved, go to step 3-3; the voting rules are: 1) Initialize the nth * Voting variables for Gaussian mixture models 2) If Then Among them, th2 is the second threshold, |·| represents the norm, For nth * The weight of the mth Gaussian distribution of a normal Gaussian mixture model, For nth * The mean of the mth Gaussian distribution of a normal Gaussian mixture model; like Then Wherein, th3 is the third threshold, and th3>th2; like Then 3) Judgment Is it true? If true, then for the nth * The Gaussian mixture model of the normal state is evolved, otherwise it does not evolve, and H represents the voting threshold; Step 3-3: For the mth * The normal Gaussian mixture model is evolved as follows: Among them, h and j represent the corresponding traversal labels respectively. Indicates the value of x c The average value of , H represents the voting threshold, For each x in the historical records c value, ρ is the evolution rate, [·] c+1 ,[·] c Respectively represent the values ​​of the corresponding variables after evolution and before evolution, [·] T Represents vector transpose; Step 3-4: Let n * Voting variables for Gaussian mixture models And clear the history of the model; Step 4: Perform voiceprint fault fusion judgment on the target device based on the multi-Gaussian mixture model.

2. According to claim 1, the method for determining fault fusion of voiceprint model evolution equipment for power grid monitoring is characterized in that: The steps of extracting MFCC feature vectors in step 2 include: Step 2-1: Divide the sound signal X′(t) of the target power grid equipment into I frames, each frame contains N=512 samples, and adjacent frames have 256 samples overlapped. Then the audio signal x′(t) of the lth frame is l (n) is: x l (n)=x(l*N+n)w(n) Among them, x is the audio signal after framing, l is the audio signal frame number, N is the window length, n is the number of window movement steps, x(.) represents framing processing, and the Hanning window w(n) is: Step 2-2: Design a filter with 36 frequency bands, the number of filters is L = 36, the frequency range is 0 Hz to 5512 Hz, and the output of each triangular filter is m(q), where q = 1, 2…, L; Step 2-3: Perform logarithmic operation on the output m(q) of all filters, and then perform discrete cosine transform to obtain a set of MFCC parameters: Among them, β=0,1,…,R,1≤R≤L, β is the MFCC feature vector label index.

3. According to claim 1, the method for determining fault fusion of voiceprint model evolution equipment for power grid monitoring is characterized in that: In step 2, N Gaussian mixture models are established according to the sample data set. Each Gaussian mixture model is a mixture of M Gaussian distributions. The nth normal Gaussian mixture model is expressed as: Where, 1≤n≤N, 1≤m≤M, n represents the index of the Gaussian mixture model, m represents the index of each sub-model of the Gaussian mixture model, and x q represents the input of the Gaussian mixture model, q represents the input sample number, 1≤q≤Q, Q is the cumulative number of input samples, P(x q |u nm ,δ nm ) is the mth Gaussian distribution at x q The probability density value, w nm is the weight of the mth Gaussian distribution of the nth normal Gaussian mixture model, u nm is the mean of the mth Gaussian distribution of the nth normal Gaussian mixture model, δ nm is the variance of the mth Gaussian distribution of the nth normal Gaussian mixture model.

4. The method for fusion judgment of device faults based on voiceprint model evolution for power grid monitoring according to claim 1 or 3, characterized in that: The parameter update rule of the Gaussian mixture model is: 1) Initialize the parameters u of the nth normal state Gaussian mixture model nm (0) , δ nm (0) ; 2) Calculate the probability P based on the current Gaussian mixture model nm (x q ): Among them, j represents the traversal label, P nm (x q ) represents the sample x q The probability generated by the mth sub-model of the nth normal Gaussian mixture model; 3) Calculate the model parameters for the next iteration: Among them, w nm ′、μ nm ′、δ nm ′ respectively represent the weight, mean, and variance of the mth Gaussian distribution of the nth normal state Gaussian mixture model after iteration; 4) Repeat steps 2) and 3) until the Gaussian mixture model of the nth normal state converges.

5. The method for determining fault fusion of voiceprint model evolution equipment for power grid monitoring according to claim 1 is characterized in that: In step 4, the rule for making the voiceprint fault fusion judgment of the target device based on the multi-Gaussian mixture model is: Rule 1: The probability of belonging to the Gaussian mixture model of the most likely normal state at present Rule 2: Entropy E of the current system c : AND c >th5 Rule 3: There exists a Gaussian mixture model n: Among them, th4 is the fault probability threshold, th5 is the entropy threshold, th6 is the evolution threshold, w nm The initial value of for u nm The initial value of If any one of the above three rules is met, it is judged as a fault; if none of them are met, it is a normal state.

Images