Deep learning based off-grid sparse bayesian partial discharge doa estimation method

Abstract

The present application relates to a kind of off-grid sparse bayesian partial discharge DOA estimation methods based on deep learning, belong to partial discharge detection field.The method includes that sampled discharge sound array signal is operated twice iteration using off-grid sparse bayesian algorithm, the mean and variance of the first two off-grid sparse bayesian algorithm are obtained;Then the mean and variance obtained are converted into normalized real number and input to deep learning network LSTM, to solve the DOA of unknown sound source;Compared with the traditional off-grid sparse bayesian algorithm, the positioning speed of the method is obviously improved under the positioning accuracy close to off-grid sparse bayesian, and the anti-frequency fluctuation ability and noise robustness are obviously enhanced.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of partial discharge detection, and relates to an off-grid sparse Bayesian partial discharge DOA estimation method based on deep learning. BACKGROUND

[0002] Partial discharge detection plays an important role in judging the insulation deterioration degree of power equipment, judging the operation condition and fault warning. At present, the time-frequency analysis is performed on the electromagnetic or acoustic signals collected by sensors by using the ultrahigh frequency method or the ultrasonic method, so as to judge whether partial discharge occurs or the type of partial discharge. Compared with the positioning method based on an acoustic array, this kind of method cannot accurately position the specific position of partial discharge.

[0003] At present, the mature acoustic source positioning method is a positioning method based on a subspace, but the traditional positioning method based on a subspace has poor noise robustness and is sensitive to the size of an array and the layout of array elements, and the operation effect is poor under complex working conditions. In recent years, the algorithm based on the compression sensing theory can be beneficial to the improvement of noise robustness, but this kind of algorithm is not easy to obtain a global optimal solution and often needs a special toolbox for solving, which leads to poor universality of the algorithm. In comparison, the Bayesian algorithm is proved to be able to obtain a global optimal solution, but the algorithm introduces off-grid operation in order to improve the positioning accuracy, further increases the calculation complexity, and is difficult to be applied in practical engineering.

[0004] In summary, at present, there is an urgent need for a positioning method with high positioning accuracy, low calculation complexity and high robustness, which can be applied in the field of real-time partial discharge detection to improve the safety of power transmission equipment operation. SUMMARY

[0005] Therefore, the purpose of the present application is to provide an off-grid sparse Bayesian partial discharge DOA estimation method based on deep learning.

[0006] To achieve the above purpose, the present application provides the following technical scheme:

[0007] The off-grid sparse Bayesian partial discharge DOA estimation method based on deep learning specifically comprises the following steps:

[0008] S1: extracting a target frequency from the collected acoustic array signal;

[0009] S2: performing twice off-grid sparse Bayesian algorithm iteration on the acoustic array signal containing only the target frequency after extraction, to obtain the mean value, variance and position of the discharge acoustic source;

[0010] S3: converting the mean value and variance after twice off-grid sparse Bayesian algorithm iteration into normalized real numbers;

[0011] S4: input the normalized real mean and variance into a deep learning network (Long Short-Term Memory, LSTM) to obtain the amplitude of the partial discharge sound source.

[0012] Further, in S1, the target frequency of the collected acoustic array signal is extracted;

[0013] S11: directly using a band-pass filter to filter the acoustic array signal;

[0014] S12: the filter is set to be centered on the target frequency, and a narrow-band acoustic array signal with a bandwidth of 1 kHz is extracted.

[0015] Further, in S2, the acoustic array signal containing only the target frequency after extraction is subjected to two iterations of off-grid sparse Bayesian algorithm to obtain the mean, variance and direction of the discharge sound source.

[0016] S21: define the signal model:

[0017] A two-stage hierarchical model is adopted, the first layer is a complex Gaussian probability distribution with zero mean for a signal X with length L:

[0018]

[0019] In the above formula, Π represents the multiplication operation, P(A|B) represents the conditional probability, that is, the probability of occurrence of condition A under the condition of condition B, Λ = diag(ζ), diag(·) is a diagonal operation, ζ is the signal accuracy, X .,l represents the lth column of X, represents that the variable u follows a complex Gaussian distribution with mean μ and covariance matrix Σ; the second layer is the gamma distribution of ζ, which is represented as:

[0020]

[0021] In the formula, ζ n is the nth element of ζ, N is the number of grids defined by the off-grid sparse Bayesian, and ρ is a fixed prior parameter; and the gamma distribution of any variable u is represented as Γ(u|a,b) = [Γ(a)] -1 b a u a-1 e -bu , where Γ(·) is the gamma function, and a and b are set close to 0.

[0022] S22: define the noise model:

[0023] The zero-mean complex Gaussian distribution is used for noise E, which is represented as:

[0024]

[0025] where α determines the precision of the noise, and if δ 2 is the variance of the noise, then α = δ -2 ; E .,l denotes the lth column of E, and I M is the M-dimensional identity matrix; α is assumed to follow a gamma distribution, which is denoted as:

[0026] p(α|c,d) = Γ(α|c,d)

[0027] where c and d are constants close to 0;

[0028] S23: Sparse Bayesian algorithm inference:

[0029] According to S21 and S22, the likelihood function of the received data Y is obtained as:

[0030]

[0031] In the above distribution A is the transfer matrix of the microphone array and the search plane, where and are the first-order derivatives of A in the elevation angle θ direction and the azimuth angle direction, respectively; Y .,l denotes the lth column of Y, and I M is the M-dimensional identity matrix, where M is the number of array elements;

[0032] The expectation-maximization (EM) algorithm is used to solve the sparse Bayesian algorithm model, and the posterior distribution of X is obtained as a complex Gaussian distribution with mean μ and covariance Σ:

[0033]

[0034] where μ .,l is the lth column of μ:

[0035]

[0036] In the above formula, Σ t is an intermediate variable, and Σ t = ΦΛΦ H + α -1 I N , where I N is the N-dimensional identity matrix; where:

[0037]

[0038] where is the expectation calculation operation, and ||·||2 represents the two-norm operation, the nth row of X, Σ n,n denotes the nth element of the diagonal of Σ

[0039]

[0040] where κ n is an intermediate variable, and κ n = 1 - [ζ n ] -1 Σ n,n , ||·| F denotes the Frobenius norm

[0041] S24: Off-grid operation

[0042] Assume that the elevation and azimuth deviations of the discharge source in space relative to the preset grid are Δθ and respectively, and satisfy uniform distribution:

[0043]

[0044] where r θ and denote the resolutions of the search grid in the θ and axis directions respectively, C ~ U(A, B) denotes that C is uniformly distributed between A and B, and is obtained by the EM algorithm:

[0045]

[0046] where ⊙ is the Kronecker product, is a real number operation, [·] * is a conjugate operation, and ν are intermediate variables, so that

[0047] Further, in the S3, the mean and variance after twice iteration of the off-grid sparse Bayesian algorithm are converted into normalized real numbers;

[0048] S31: Take the absolute value of the mean and variance after twice iteration of the off-grid sparse Bayesian algorithm:

[0049]

[0050] where |·| is an absolute value operation, μ abs and Σ abs are the mean and variance after taking the absolute value respectively;

[0051] S32: Normalize the mean and variance after taking the absolute value:

[0052]

[0053] wherein and respectively represent the normalized values of μ abs and Σ abs , min(·) and max(·) represent the minimum value and the maximum value respectively.

[0054] Further, in the S4, the normalized real mean value and variance are input into an LSTM deep learning network to obtain the amplitude of the discharge sound source:

[0055] S41: 3 layers of LSTM are set to contain 128, 256 and 128 hidden units respectively, and 4 inputs are set to input the mean value and variance of the two iterations of the off-grid sparse Bayesian algorithm, so as to obtain the amplitude distribution of the discharge sound source;

[0056] S42: the azimuth angle of the discharge sound source calculated by the off-grid sparse Bayesian algorithm is combined and the elevation angle θ = θ G + Δθ, wherein θ G and respectively represent a preset grid and the sound source amplitude predicted by the LSTM network, so as to realize the DOA estimation of the partial discharge sound source.

[0057] The present application has the advantages of ensuring the positioning accuracy, low calculation complexity and high robustness, and can be applied to the online identification and early warning equipment of partial discharge.

[0058] Other advantages, objects and features of the present application will be set forth in part in the following specification, and in part will become apparent to those skilled in the art upon examination of the following specification, or can be learned from practice of the present application. The objects and other advantages of the present application can be realized and attained by the methods and instrumentalities particularly pointed out in the following description. BRIEF DESCRIPTION OF DRAWINGS

[0059] In order to make the objects, technical solutions and advantages of the present application clearer, the preferred detailed description of the present application will be combined with the drawings, wherein:

[0060] Figure 1 is the flow chart of the present application. DETAILED DESCRIPTION

[0061] The advantages and effects of the present application can be easily understood by those skilled in the art from the description of the present application. The present application can also be implemented or applied by different specific embodiments, and various modifications or changes can be made to the details in the description based on different views and applications without departing from the spirit of the present application. It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present application in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0062] The accompanying drawings are only used for illustrative description, and the representation is only a schematic diagram, not a physical diagram, and cannot be understood as a limitation of the present application; in order to better illustrate the embodiments of the present application, some components in the drawings are omitted, enlarged or reduced, and do not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings can be omitted.

[0063] The same or similar reference numerals in the drawings of the embodiments of the present application correspond to the same or similar components; in the description of the present application, it should be understood that if the terms "upper", "lower", "left", "right", "front", "back" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore the terms describing the positional relationship in the drawings are only used for illustrative description, and cannot be understood as a limitation of the present application, for those skilled in the art, the specific meaning of the above terms can be understood according to the specific situation.

[0064] Figure 1 The present application provides an off-grid sparse Bayesian estimation method for DOA of partial discharge based on deep learning, which specifically comprises the following steps:

[0065] S1: extracting the target frequency from the collected acoustic array signal;

[0066] S2: performing twice iteration of off-grid sparse Bayesian algorithm on the acoustic array signal containing only the target frequency after extraction, to obtain the mean value, variance and direction of the discharge sound source;

[0067] S3: converting the mean value and variance after twice iteration of the off-grid sparse Bayesian algorithm into normalized real numbers;

[0068] S4: inputting the normalized real mean value and variance into the LSTM deep learning network to obtain the amplitude of the partial discharge sound source.

[0069] In S1, the target frequency is extracted from the collected acoustic array signal;

[0070] S11: Directly filter the acoustic array signal using a band-pass filter;

[0071] S12: The filter is set to be centered at the target frequency, and a narrow-band acoustic array signal with a bandwidth of 1 kHz is extracted.

[0072] In S2, the acoustic array signal containing only the target frequency after extraction is iterated twice using the off-grid sparse Bayesian algorithm to obtain the mean, variance, and azimuth of the discharge sound source.

[0073] S21: Define the signal model:

[0074] A two-stage hierarchical model is used, and the first layer is a complex Gaussian probability distribution with zero mean for a signal X with length L:

[0075]

[0076] In the above formula, Π represents the multiplication operation, P(A|B) represents the conditional probability, i.e., the probability of event A occurring under the condition B, Λ = diag(ζ), diag(·) is the diagonal operation, ζ is the signal precision, X .,l represents the lth column of X, represents that the variable u follows a complex Gaussian distribution with mean μ and covariance matrix Σ. The second layer is a gamma distribution of ζ, which is represented as:

[0077]

[0078] In the formula, ζ n is the nth element of ζ, N is the number of grids defined by the off-grid sparse Bayesian algorithm, and ρ is a fixed prior parameter. And for any variable u, the gamma distribution is represented as Γ(u|a,b) = [Γ(a)] -1 b a u a-1 e -bu , where Γ(·) is the gamma function, and a and b are usually set close to 0;

[0079] S22: Define the noise model:

[0080] A zero-mean complex Gaussian distribution is used for noise E, which is represented as:

[0081]

[0082] In the formula, α determines the precision of the noise, and if δ 2 is the variance of the noise, then α = δ -2 . E .,l represents the lth column of E, and I M is an M-dimensional identity matrix. α adopts a gamma distribution, which is represented as:

[0083] p(a|c,d) = Γ(a|c,d)

[0084] In the above equation, c and d are usually constants close to 0.

[0085] S23: Sparse Bayesian algorithm inference:

[0086] According to S21 and S22, the likelihood function of the received data Y can be obtained as:

[0087]

[0088] In the above distribution A is the transfer matrix of the microphone array and the search plane, where and are the first-order derivatives of A in the direction of the elevation angle θ and the azimuth angle respectively. Y .,l represents the lth column of Y, I M is an M-dimensional unit matrix, and M is the number of array elements.

[0089] The solution to the sparse Bayesian algorithm model can be obtained by using the expectation-maximization (EM) algorithm, and the posterior distribution X follows a complex Gaussian distribution with mean μ and covariance Σ:

[0090]

[0091] where μ .,l is the lth column of μ:

[0092]

[0093] In the above equation, Σ t is an intermediate variable, Σ t = ΦΛΦ H + α -1 I N , I N is an N-dimensional unit matrix. Where:

[0094]

[0095] where is the expectation calculation operation, ||·||2 represents the two-norm operation, represents the nth row of X, and Σ n,n represents the nth diagonal element of Σ.

[0096]

[0097] In the above equation κn is an intermediate variable, and κ n = 1 - [ζ n ] -1 Σ n,n ,||·| F denotes the Frobenius norm

[0098] S24: Off-grid operation

[0099] Assume that the elevation and azimuth deviations of the discharge source in space relative to the preset grid are Δθ and and satisfy uniform distribution:

[0100]

[0101] where r θ and denote the resolution of the search grid in the θ and axis direction, respectively, C ~ U(A, B) denotes that C is subject to uniform distribution between A and B, and is obtained by EM algorithm:

[0102]

[0103] where ⊙ is the Kronecker product, is the real number operation, [·] * is the conjugate, and v are intermediate variables, so that

[0104] In the S3, the mean and variance after twice iteration of the sparse Bayesian algorithm are converted into normalized real numbers;

[0105] S31: Take the absolute value of the mean and variance after twice iteration of the off-grid sparse Bayesian algorithm:

[0106]

[0107] where |·| is the absolute value operation, μ abs and Σ abs are the mean and variance after taking the absolute value, respectively.

[0108] S32: Normalize the mean and variance after taking the absolute value:

[0109]

[0110] where and denote μ abs and Σ absnormalized value, min(·) and max(·) represent the minimum and maximum values, respectively.

[0111] In S4, the normalized real mean and variance are input into the LSTM deep learning network to obtain the amplitude of the discharge sound source:

[0112] S41: Set 3 layers of LSTM to contain 128, 256, and 128 hidden units, respectively, and set 4 inputs to input the two iteration parameters (mean and variance) of the off-grid sparse Bayesian algorithm to obtain the amplitude distribution of the discharge sound source;

[0113] S42: Combine the azimuth angle θ of the discharge sound source calculated by the off-grid sparse Bayesian algorithm and the elevation angle θ = θ G + Δθ, where θ G and represent the preset grid and the sound source amplitude predicted by the LSTM network, thereby realizing the DOA estimation of the partial discharge sound source.

[0114] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the purpose and scope of the technical solutions, and they should be covered in the scope of the claims of the present application.

Claims

1. A deep learning-based method for estimating the DOA of off-grid sparse Bayesian partial discharge, characterized by: The method specifically includes the following steps: S1: Extract the target frequency from the acquired acoustic array signal; S2: The extracted acoustic array signal containing only the target frequency is subjected to two off-grid sparse Bayes algorithm iterations to obtain the mean, variance, and location of the discharge sound source; In step S2, the extracted acoustic array signal containing only the target frequency is subjected to two off-grid sparse Bayes algorithm iterations to obtain the mean, variance, and location of the discharge sound source. S21: Define the signal model: A two-stage hierarchical model is adopted. The first layer is a signal X of length L that follows a complex Gaussian probability distribution with zero mean: In the above formula, Π represents the multiplication operation, P(A|B) represents the conditional probability, that is, the probability of condition A occurring given condition B, Λ=diag(ζ), diag(·) is the diagonal operation, ζ is the signal precision, and X .,l Represents the l-th column of X. The variable u follows a complex Gaussian distribution with mean μ and covariance matrix Σ; the second layer is a gamma distribution with ζ, expressed as: In the formula ζ n Let Γ be the nth element of ζ, N be the number of grids defined by the off-grid sparse Bayesian method, and ρ be a fixed prior parameter; and the gamma distribution for any variable u is represented as Γ(u|a,b)=[Γ(a)]. -1 b a u a-1 e -bu , where Γ(·) is the gamma function, and a and b are set close to 0; S22: Define the noise model: Applying a zero-mean complex Gaussian distribution to the noise E, we can express it as: In the formula, α determines the accuracy of the noise; if δ 2 Let α be the variance of the noise, then δ -2 E .,l Represents the l-th column of E, I M Let α be an M-dimensional identity matrix; α follows a gamma distribution and is represented as: p(α|c,d)=Γ(α|c,d) In the above formula, c and d are constants close to 0; S23: Sparse Bayesian algorithm inference: Based on S21 and S22, the likelihood function of the received data Y is obtained as follows: In the above distribution A is the transfer matrix between the microphone array and the search plane, where and These represent A in the direction of elevation angle θ and azimuth angle, respectively. The first derivative on the square; Y .,l Represents the l-th column of Y, I M Let M be an M-dimensional identity matrix, where M is the number of matrix elements; The Expectation-Maximization (EM) algorithm is used to solve the sparse Bayesian algorithm model, and the posterior distribution X follows a complex Gaussian distribution with mean μ and covariance Σ. Where μ .,l It is the l-th column of μ: In the above formula Σ t Σ is an intermediate variable. t =ΦΛΦ H +α -1 I N I N Let be an N-dimensional identity matrix; where: in For the expected computation operation, ||·||2 denotes the L2 norm operation. Let Σ represent the nth row of X. n,n This represents the nth element of the Σ diagonal; In the formula κ n It is an intermediate variable, and κ n =1-[ζ n ] -1 Σ n,n ,||·|| F Describing the Frobenius norm S24: Offline Operation Assume the elevation and azimuth deviations of the discharge source in space relative to the preset grid are Δθ and Δθ, respectively. And it satisfies a uniform distribution: Where r θ and The search grid is represented at θ and θ respectively. The resolution along the axis, C~U(A,B), indicates that C follows a uniform distribution from A to B, obtained through the EM algorithm: Where ⊙ represents the Kronecker product. To perform the real number operation, [·] * To obtain conjugate, ν is an intermediate variable, thus obtaining S3: Convert the mean and variance of the off-network sparse Bayes algorithm after two iterations into normalized real numbers; S4: Input the normalized real mean and variance into the LSTM deep learning network to obtain the amplitude of the partial discharge sound source.

2. The method for estimating the Off-grid Sparse Bayesian Partial Discharge (DOA) based on deep learning according to claim 1, characterized in that: In step S1, the target frequency is extracted from the acquired acoustic array signal; S11: Directly use a bandpass filter to filter the acoustic array signal; S12: The filter is set to extract a narrowband acoustic array signal with a bandwidth of 1kHz, centered on the target frequency.

3. The method for estimating the Off-grid Sparse Bayesian Partial Discharge (DOA) based on deep learning according to claim 1, characterized in that: In S3, the mean and variance after two iterations of the sparse Bayes algorithm are converted into normalized real numbers. S31: Take the absolute values ​​of the mean and variance after two iterations of the off-network sparse Bayesian algorithm: In the formula, |·| represents the absolute value operation, μ abs and Σ abs These are the mean and variance after taking their absolute values, respectively; S32: Normalize the mean and variance after taking their absolute values: In the formula and They represent μ respectively abs and Σ abs The normalized values ​​are min(·) and max(·), which represent taking the minimum and maximum values, respectively.

4. The method for estimating the Off-grid Sparse Bayesian Partial Discharge (DOA) based on deep learning according to claim 1, characterized in that: In step S4, the normalized real mean and variance are input into the LSTM deep learning network to obtain the amplitude of the discharge sound source: S41: Set up a 3-layer LSTM with 128, 256 and 128 hidden units respectively, set up 4 inputs, and input the two iteration parameters of the off-grid sparse Bayesian, namely the mean and variance, to obtain the amplitude distribution of the discharge sound source. S42: Azimuth angle of the discharge sound source calculated using the off-grid sparse Bayesian algorithm And the angle of elevation θ = θ G +Δθ, where θ G and These represent the preset grid and the source amplitude predicted by the LSTM network, respectively, thus enabling DOA estimation of the partial discharge source.

Images