Environmental noise filtering method for acquiring abnormal sound signals of GIS (Gas Insulated Switchgear) equipment

Abstract

The invention discloses an environment noise filtering method for GIS equipment abnormal sound signal acquisition, and relates to the field of power switch equipment environment sound noise filtering, and the method comprises the steps: S1, obtaining sound sensor data, and analyzing the sound sensor data to obtain a sound observation signal x (t); s2, performing blind source separation on the sound observation signal x (t), and decomposing a blind source separation result by using VMD to obtain a plurality of estimation signals; s3, singular spectrum entropy, energy entropy and kurtosis entropy features of the plurality of estimation signals are extracted respectively, and the singular spectrum entropy, energy entropy and kurtosis entropy features form a feature vector G; s4, an NSVDD algorithm is applied to complete selection of circuit breaker closing sound signals; and S5 in the third chapter is used to complete the identification of the mechanical state of the circuit breaker based on the human auditory characteristics and an improved sparse representation classification method. After the blind source separation and the NSVDD algorithm are used for processing the signals, the recognition accuracy is integrally improved.

Description

Technical Field

[0001] The invention relates to the field of environmental sound noise filtering of power switch equipment, in particular to an environmental noise filtering method for collecting abnormal sound signals of gas insulated switch GIS equipment. Background Art

[0002] In actual operation, the environment where the circuit breaker is located is relatively complex. The environmental background may contain noise such as human voices and partial discharge of other equipment. The overlap of these environmental noises with the circuit breaker closing sound signal may affect the judgment of the mechanical state of the circuit breaker. In order to accurately identify the mechanical state of the circuit breaker in a noisy environment, it is necessary to perform noise reduction processing on the sound signal collected by the sound sensor.

[0003] In order to accurately select the circuit breaker closing sound signal from the separation results of blind source separation, appropriate feature extraction is the key. The feature vector extracted by the conventional method has a higher data dimension in order to be able to reflect the difference of the closing sound signal of the circuit breaker in each mechanical state in more detail. For the blind source separation results, the circuit breaker closing sound signal and the noise signal obviously have a large difference. At this time, the accuracy of the feature vector is no longer the main contradiction in feature extraction. It can extract the sound signal feature vector in a lower dimensional presentation while reflecting the difference information of each sound signal, which can effectively improve the calculation speed and generalization of the subsequent recognition and diagnosis algorithm. In order to successfully select the circuit breaker closing sound signal from several separated signals, the present invention adopts the characteristics of entropy values ​​such as singular spectrum entropy, energy entropy and kurtosis entropy in information theory to meet new needs. Summary of the invention

[0004] To solve the above problems, the present application provides an environmental noise filtering method for abnormal sound signal collection of GIS equipment to improve the accuracy of signal recognition.

[0005] The environmental noise filtering method for collecting abnormal sound signals of GIS equipment provided in this application includes the following steps: S1. Acquire sound sensor data, and analyze the sound sensor data to obtain a sound observation signal x(t); S2, using the FastICA algorithm based on the negentropy maximization objective to perform blind source separation on the sound observation signal x(t), and using VMD to decompose the blind source separation result to obtain several BLIMF components; S3, respectively extracting the singular spectrum entropy, energy entropy and kurtosis entropy features of several BLIMF components, and combining the singular spectrum entropy, energy entropy and kurtosis entropy features into a feature vector G; S4. Using the NSVDD algorithm to select the circuit breaker closing sound signal, obtain the target circuit breaker closing sound signal; S5. Based on the joint cepstrum features and improved sparse representation classification method combined with the closing sound signal of the target circuit breaker, the mechanical state of the circuit breaker is identified.

[0006] Preferably, the specific content of using the FastICA algorithm based on the negentropy maximization objective to perform blind source separation on the sound observation signal x(t) in S2 is: ICA is represented by blind source separation. It is assumed that N sound sources emit sound simultaneously, and the signal of each sound source is recorded as a matrix s(t)=[s1(t); s2(t); ...; s N (t)]; Where t represents the time point, and the row vector s i (t) is the i-th sound source signal sequence; Blind source separation is to successfully separate M (M ≥ N) observation signals x(t) = [x1(t); x2(t); ...; x M (t)](x(t)=As(t); Among them, A is the process of separating the source signal from the signal mixing matrix; The mathematical model of blind source separation is as follows: ; In the formula is the blind separation result, which is an estimate of the source signal and should ideally be consistent with the source signal. is the unmixing matrix of the blind separation process.

[0007] Preferably, the premise for the successful application of the blind source separation algorithm is: Each source signal i (t)(i=1, 2, ..., N) are independent of each other; There cannot be more than one Gaussian distribution variable in each source signal; The number of observed signals M is not less than the number of source signals N, that is, (M≥N).

[0008] Preferably, the unmixing matrix based on the maximization of negative entropy The process of obtaining is as follows: Step 1: Normalize and whiten the observed signal x. The signal matrix after processing is recorded as Y, and its covariance matrix is ​​the unit matrix. Step 2: Set up the unmixing matrix The initial weight vector W of the i-th (i=1,2,…,N) row i , constrained by ||W i ||2=1; Step 3: Use Newton's iteration method to complete W i estimates; Step 4: For the k+1th iteration result W i (k+1) is normalized; Step 5: Repeat steps 3 to 5 until W i Convergence, return to step 2, take i=i+1, until all N row vectors of W are found.

[0009] Preferably, the features of singular spectrum entropy, energy entropy and kurtosis entropy of several BLIMF components are extracted respectively, and the singular spectrum entropy, energy entropy and kurtosis entropy are used to form a feature vector G, and the specific content is: The extraction of singular spectral entropy includes: Will The BLIMF components are arranged into a K×L dimensional matrix, where L is the signal length. The matrix is ​​subjected to singular value decomposition, and the singular value vector of the matrix is ​​obtained as , is the i-th singular value λ i The proportion of the whole singular spectrum; The singular spectral entropy is defined as: ; The extraction of energy entropy includes: The energy arrangement vector of each BLIMF is preset as , For the The proportion of BLIMF energy to the sound signal energy; Energy entropy is defined as follows: ; The extraction of kurtosis entropy includes: First, find the kurtosis value of each BLIMF component, and set the kurtosis arrangement of each BLIMF component as , For the The proportion of the kurtosis of a BLIMF component to the sum of the kurtosis of all components; The corresponding kurtosis entropy is defined as: ; The characteristic vector composed of singular spectrum entropy, energy entropy and kurtosis entropy is G, .

[0010] Preferably, the objective function of the NSVDD algorithm is: ; In the formula , Corresponding to the positive sample feature vector subscript, , Corresponding to the negative sample feature vector subscript, K represents the kernel matrix. Solving the optimization problem of the above formula can obtain the support vector of the hypersphere. Combining the support vector, the radius R of the hypersphere and the center a of the hypersphere can be obtained. For the feature vector to be tested , the eigenvector Euclidean distance from the center of the hypersphere The expression is as follows: ; like , T p is a positive sample, otherwise it is a negative sample.

[0011] Preferably, during the training process of the NSVDD model, a particle swarm optimization (PSO) is used to implement the variance of the Gaussian kernel function. Target calculation of penalty coefficients C1 and C2 of NSVDD model.

[0012] In summary, the environmental noise filtering method for abnormal sound signal collection of GIS equipment of the present invention, compared with the traditional technology, uses the FastICA algorithm based on negative entropy maximization to complete the blind separation of each sound source signal, and aims at the signal disorder and amplitude uncertainty problems of the FastICA algorithm, firstly completes the decomposition of each separated signal by using variational mode decomposition (VMD), extracts the entropy value features of the decomposition results, and then uses the support vector description with negative samples (NSVDD) to complete the selection of the circuit breaker closing sound signal, and then uses the joint cepstrum feature and improved sparse representation classification to complete the recognition of the mechanical state of the circuit breaker, thereby improving the recognition accuracy and generalization.

[0013] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a flow chart of an environmental noise filtering method for collecting abnormal sound signals of GIS equipment according to the present invention; Figure 2 A blind source separation process diagram of an environmental noise filtering method for collecting abnormal sound signals of GIS equipment according to the present invention; Figure 3 These are the BLIMF component spectra corresponding to the closing sound signals of the circuit breaker in four mechanical states and a frame of 1s long partial discharge and human speech signals. DETAILED DESCRIPTION

[0015] The technical solution of the present invention is further described below by means of the accompanying drawings and embodiments. It should be noted that unless otherwise specifically stated, the relative arrangement of components and steps, numerical expressions and numerical values ​​described in these embodiments do not limit the scope of the present application.

[0016] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present application, its application, or uses.

[0017] Technologies, systems, and devices known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, systems, and devices should be considered part of the specification.

[0018] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0019] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0020] like Figure 1 As shown, the present invention provides an environmental noise filtering method for collecting abnormal sound signals of GIS equipment, S1, obtaining sound sensor data, analyzing the sound sensor data to obtain a sound observation signal x(t); S2, using the FastICA algorithm based on the negentropy maximization objective to perform blind source separation on the sound observation signal x(t), and using VMD to decompose the blind source separation result to obtain several BLIMF components; Preferably, Figure 2 As shown in FIG. 1 , the specific contents of blind source separation of the sound observation signal x(t) using the FastICA algorithm based on the negentropy maximization objective in S2 are as follows: ICA is represented by blind source separation. It is assumed that N sound sources emit sound simultaneously, and the signal of each sound source is recorded as a matrix s(t)=[s1(t); s2(t); ...; s N (t)]; Where t represents the time point, and the row vector s i (t) is the i-th sound source signal sequence; Blind source separation is to successfully separate M (M ≥ N) observation signals x(t) = [x1(t); x2(t); ...; x M (t)](x(t)=As(t); Among them, A is the process of separating the source signal from the signal mixing matrix; The mathematical model of blind source separation is as follows: ; In the formula is the blind separation result, which is an estimate of the source signal and should ideally be consistent with the source signal. is the unmixing matrix of the blind separation process.

[0021] Preferably, the premise for the successful application of the blind source separation algorithm is: Each source signal si(t) (i=1, 2, ..., N) is independent of each other; There cannot be more than one Gaussian distribution variable in each source signal; The number of observed signals M is not less than the number of source signals N, that is, (M≥N).

[0022] Preferably, the process of obtaining the unmixing matrix W based on the maximization of negative entropy is as follows: Step 1: Normalize and whiten the observed signal x. The signal matrix after processing is recorded as Y, and its covariance matrix is ​​the unit matrix. Step 2: Set up the unmixing matrix The initial weight vector W of the i-th (i=1,2,…,N) row i , constrained by ||W i ||2=1; Step 3: Use Newton's iteration method to complete W i estimates; Step 4: For the k+1th iteration result W i (k+1) is normalized; Step 5: Repeat steps 3 to 5 until W i Convergence, return to step 2, take i=i+1, until The N row vectors of are all found.

[0023] When performing VMD decomposition on the signal, select the appropriate decomposition layer number K. If K is too large, the modes may overlap; if K is too small, there will be a problem of under-decomposition. This paper determines the decomposition layer number by observing the center frequency, that is, gradually increasing the preset decomposition layer number K, and the corresponding maximum center frequency w in each BLIMF component K will also increase accordingly, when w KThe number of decomposition layers corresponding to the stability is the optimal number of decomposition layers. Since the purpose of extracting the signal entropy value feature is to select the circuit breaker closing sound signal from the blind source separation result, the number of decomposition layers of the circuit breaker closing sound signal is used as the benchmark. The corresponding relationship between the center frequency of the BLIMF component obtained by decomposing the closing sound signal of each mechanical state and the number of decomposition layers K is shown in Table 1, Table 2, Table 3 and Table 4. Table 1 shows the center frequency of each mode of the closing sound signal in the normal state / KHz, Table 2 shows the center frequency of each mode of the closing sound signal when the transmission guide rod is stuck / KHz, Table 3 shows the center frequency of each mode of the closing sound signal when the gap between the iron core and the closing spring is too large / KHz, and Table 4 shows the center frequency of each mode of the closing sound signal when the opening spring is loose / KHz.

[0024] Table 1

[0025] Table 2

[0026] Table 3

[0027] Table 4

[0028] It can be seen from the table above that when the decomposition level K is set to 6, the maximum and minimum values ​​of the center frequency of the BLIMF component have become stable. When the decomposition level is set to 6, the BLIMF component spectra corresponding to the closing sound signals of the circuit breaker in four mechanical states and a frame of 1s long partial discharge and human speech signals are as follows: Figure 3 ,in, Figure 3 a in the figure is the BLIMF component spectrum of the normal closing sound. Figure 3 b in the middle is the BLIMF component spectrum of the transmission guide rod jamming and closing sound. Figure 3 c in the figure is the BLIMF component spectrum of the closing sound signal when the gap between the iron core and the closing spring is too large. Figure 3 d in the middle is the BLIMF component spectrum of the closing sound when the opening spring is released and loosened Figure 3 e is the BLIMF component spectrum of the local discharge sound, Figure 3 The f in the figure is the BLIMF component spectrum of human speaking sound.

[0029] S3, respectively extracting the singular spectrum entropy, energy entropy and kurtosis entropy features of several BLIMF components, and combining the singular spectrum entropy, energy entropy and kurtosis entropy features into a feature vector G; Through VMD, the signal is decomposed into several time-frequency subsequences. The singular spectrum entropy can express the complexity and uncertainty of each subsequence signal, the energy entropy can reflect the energy distribution of the signal, and the kurtosis entropy mainly extracts the probability distribution characteristics of the signal. The above three characteristic entropy values ​​can be used to realize the feature extraction of each signal of the blind source separation result, and the dimension is only 3, which meets the requirement of low-dimensional features of the feature vector.

[0030] Preferably, the features of singular spectrum entropy, energy entropy and kurtosis entropy of several BLIMF components are extracted respectively, and the singular spectrum entropy, energy entropy and kurtosis entropy are used to form a feature vector G, and the specific content is: The extraction of singular spectral entropy includes: Will The BLIMF components are arranged into a K×L dimensional matrix, where L is the signal length. The matrix is ​​subjected to singular value decomposition, and the singular value vector of the matrix is ​​obtained as , is the i-th singular value λ i The proportion of the whole singular spectrum; The singular spectral entropy is defined as: ; The extraction of energy entropy includes: The energy arrangement vector of each BLIMF is preset as , For the The proportion of BLIMF energy to the sound signal energy; Energy entropy is defined as follows: ; The extraction of kurtosis entropy includes: First, find the kurtosis value of each BLIMF component, and set the kurtosis arrangement of each BLIMF component as , For the The proportion of the kurtosis of a BLIMF component to the sum of the kurtosis of all components; The corresponding kurtosis entropy is defined as: ; The characteristic vector composed of singular spectrum entropy, energy entropy and kurtosis entropy is G, .

[0031] The closing sound signals under four mechanical states are taken as positive samples, and the partial discharge sound and voice signals are taken as negative samples. First, 120 frames of partial discharge sound and voice signals are collected in a quiet environment, and each frame signal is 1s long. Half of them are taken together with the training set of the closing sound signals of the four mechanical states collected in Chapter 3 to form a training set, and the remaining half is taken together with the test set of the closing sound signals in Chapter 3 to form a test set. The entire training set consists of 360 frames of signals. Taking one frame of each sound signal as an example, the entropy feature vector G of each sound signal is shown in Table 5. Table 5 shows the entropy feature of each sound signal. Table 5

[0032] S4. Using the NSVDD algorithm to select the circuit breaker closing sound signal, obtain the target circuit breaker closing sound signal; Preferably, the objective function of the NSVDD algorithm is: ; In the formula , Corresponding to the positive sample feature vector subscript, , Corresponding to the negative sample feature vector subscript, K represents the kernel matrix. Solving the optimization problem of the above formula can obtain the support vector of the hypersphere. Combining the support vector, we can get the radius R of the hypersphere and the center a of the hypersphere. For the feature vector T to be tested p , the eigenvector T p Euclidean distance from the center of the hypersphere The expression is as follows: ; like , T p is a positive sample, otherwise it is a negative sample.

[0033] Preferably, during the training process of the NSVDD model, a particle swarm optimization (PSO) is used to implement the variance of the Gaussian kernel function. Target calculation with the penalty coefficients C1 and C2 of the model.

[0034] S5. Based on the joint cepstrum features and improved sparse representation classification method combined with the closing sound signal of the target circuit breaker, the mechanical state of the circuit breaker is identified.

[0035] The recognition accuracy of various mechanical states under noise background by using blind source separation and NSVDD algorithm to preprocess the sound signal and the recognition method without preprocessing is shown in Table 6. As can be seen from the table, the recognition accuracy of the recognition method using blind source separation and NSVDD algorithm to preprocess the sound signal is improved compared with the recognition method without preprocessing, but the improvement result is limited. This is mainly because the energy of the partial discharge gun and human voice noise in the experiment of this paper is small, the overall signal-to-noise ratio of the signal is high, and the influence of noise on the target signal is small. Even if the noise is not preprocessed, the original recognition method can also achieve good recognition results. After the signal is preprocessed using blind source separation and NSVDD algorithm, the overall recognition accuracy is improved by 10%, which can prove the effectiveness of the preprocessing link.

[0036] Table 6

[0037] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for filtering environmental noise for collecting abnormal sound signals of GIS equipment, characterized in that: The following steps are involved: S1. Acquire sound sensor data, and analyze the sound sensor data to obtain a sound observation signal x(t); S2, using the FastICA algorithm based on the negentropy maximization objective to perform blind source separation on the sound observation signal x(t), and using VMD to decompose the blind source separation result to obtain several BLIMF components; S3, respectively extract the singular spectrum entropy, energy entropy and kurtosis entropy characteristics of several BLIMF components, and convert the singular spectrum entropy, energy entropy and kurtosis entropy characteristics into Composed feature vector G; S4. Using the NSVDD algorithm to select the circuit breaker closing sound signal, obtain the target circuit breaker closing sound signal; S5. Based on the joint cepstrum features and improved sparse representation classification method combined with the closing sound signal of the target circuit breaker, the mechanical state of the circuit breaker is identified.

2. The environmental noise filtering method for collecting abnormal sound signals of GIS equipment according to claim 1 is characterized in that: The specific content of blind source separation of the sound observation signal x(t) using the FastICA algorithm based on the negative entropy maximization objective in S2 is: ICA is represented by blind source separation. It is assumed that N sound sources emit sound simultaneously, and the signal of each sound source is recorded as a matrix s(t)=[s1(t); s2(t); ...; s N (t)]; Where t represents the time point, and the row vector s i (t) is the i-th sound source signal sequence; Blind source separation is to successfully separate M (M ≥ N) observation signals x(t) = [x1(t); x2(t); ...; x M (t)], x(t)=As(t); Among them, A is the process of separating the source signal from the signal mixing matrix; The mathematical model of blind source separation is as follows: ； In the formula is the blind separation result, which is an estimate of the source signal and should ideally be consistent with the source signal. is the unmixing matrix of the blind separation process.

3. The environmental noise filtering method for collecting abnormal sound signals of GIS equipment according to claim 2 is characterized in that: The premise for the successful application of the blind source separation algorithm is: Each source signal i (t)(i=1, 2, ..., N) are independent of each other; There cannot be more than one Gaussian distribution variable in each source signal; The number of observed signals M is not less than the number of source signals N, that is, M≥N.

4. The environmental noise filtering method for collecting abnormal sound signals of GIS equipment according to claim 3 is characterized in that: Unmixing Matrix Based on Maximization of Negative Entropy The process of obtaining is as follows: Step 1: Normalize and whiten the observed signal x. The signal matrix after processing is recorded as Y, and its covariance matrix is ​​the unit matrix. Step 2: Set up the unmixing matrix The initial weight vector W of the i-th row i , constrained by ||W i ||2=1, i=1,2,…,N; Step 3: Use Newton's iteration method to complete W i estimates; Step 4: For the k+1th iteration result W i (k+1) is normalized; Step 5: Repeat steps 3 to 5 until W i Convergence, return to step 2, take i=i+1, until all N row vectors of W are found.

5. The environmental noise filtering method for collecting abnormal sound signals of GIS equipment according to claim 4 is characterized in that: The characteristics of singular spectrum entropy, energy entropy and kurtosis entropy of several BLIMF components are extracted respectively, and the specific content of the singular spectrum entropy, energy entropy and kurtosis entropy forming the feature vector G is as follows: The extraction of singular spectral entropy includes: Will The BLIMF components are arranged into a K×L dimensional matrix, where L is the signal length. The matrix is ​​subjected to singular value decomposition, and the singular value vector of the matrix is ​​obtained as , is the i-th singular value λ i The proportion of the whole singular spectrum; The singular spectral entropy is defined as: ； The extraction of energy entropy includes: The energy arrangement vector of each BLIMF is preset as , For the The proportion of BLIMF energy to the sound signal energy; Energy entropy is defined as follows: ； The extraction of kurtosis entropy includes: First, find the kurtosis value of each BLIMF component, and set the kurtosis arrangement of each BLIMF component as , For the The proportion of the kurtosis of a BLIMF component to the sum of the kurtosis of all components; The corresponding kurtosis entropy is defined as: ； The characteristic vector composed of singular spectrum entropy, energy entropy and kurtosis entropy is G, .

6. The environmental noise filtering method for collecting abnormal sound signals of GIS equipment according to claim 5 is characterized in that: The objective function of the NSVDD algorithm is: ； In the formula , Corresponding to the positive sample feature vector subscript, , Corresponding to the negative sample feature vector subscript, K represents the kernel matrix. Solving the optimization problem of the above formula can obtain the support vector of the hypersphere. Combining the support vector, the radius R of the hypersphere and the center a of the hypersphere can be obtained. For the feature vector to be tested , the eigenvector Euclidean distance from the center of the hypersphere The expression is as follows: ； like , T p is a positive sample, otherwise it is a negative sample.

7. The environmental noise filtering method for collecting abnormal sound signals of GIS equipment according to claim 6 is characterized in that: In the training process of the NSVDD model, the particle swarm algorithm PSO is used to implement the variance of the Gaussian kernel function. Target calculation of penalty coefficients C1 and C2 of NSVDD model.

Images