Transformer soundprint fault diagnosis method for new energy stations based on wave search optimization

Abstract

The present invention belongs to the field of online detection of electric power equipment, and relates to a voiceprint fault diagnosis method for transformers in new energy stations based on wave search optimization. The preprocessed transformer voiceprint signal is processed by a Mel filter and a Gammatone filter respectively, so as to extract the Mel frequency cepstral coefficient and the Gammatone frequency cepstral coefficient, and then linearly combine them, and apply the Fisher value after introducing L2 norm regularization to screen, obtain the optimized mixed cepstral coefficient characteristics, and then form a feature training set; use the improved wave search algorithm to optimize the BP neural network, select the optimal regularization parameter and momentum factor; collect the transformer voiceprint signal of the new energy station transformer in real-time operation to form a test set, import the test set into the trained BP neural network model, judge whether the transformer is faulty and derive the fault type. The present invention can reduce the loss rate of training and improve the accuracy of voiceprint fault detection.

Description

Technical Field

[0001] The present invention belongs to the field of online detection of electric power equipment, and in particular relates to a new energy station transformer voiceprint fault diagnosis method based on wave search optimization. Background Art

[0002] The box-type transformer used in new energy stations integrates the transformer body, switchgear, measuring instruments and protection equipment, and is encapsulated in a sealed steel box. Compared with traditional transformers, it has many advantages such as compact structure, small size, small footprint, low operating cost, convenient installation and easy maintenance. In order to ensure the safe and stable operation of such transformers, it is very important to implement online monitoring and maintenance. However, traditional monitoring methods are often restricted by factors such as electromagnetic interference and temperature changes, resulting in low reliability and durability. In addition, the distribution range of transformers in new energy stations is relatively wide, and relying on traditional manual inspection and diagnosis methods has high requirements for inspection personnel and low inspection efficiency. Therefore, studying the acoustic diagnosis and monitoring method of transformers in new energy stations is of great significance to the reliable and high-quality operation of transformers in new energy stations.

[0003] When the transformer of a new energy station is running, it will produce sound. The sound signal contains a lot of status information. For example, when the transformer of a new energy station is defective, its operating sound is different from that of normal operation. The status can be monitored through its transformer voiceprint signal. However, the sound signal contains other irrelevant information, which will increase the computing cost. Therefore, the feature extraction method can greatly improve the computing efficiency.

[0004] BP neural network is a multi-layer feedforward network with forward signal transmission and reverse error propagation. It continuously corrects the network weights or thresholds according to the error between the actual value and the predicted value until the output error is within the set range. BP neural network has strong self-learning and adaptability, and is particularly suitable for multi-factor, uncertain, and fuzzy information processing. The network consists of input layer, hidden layer, output layer, etc. The disadvantage of BP neural network is that the network structure is not easy to determine. When the input parameters are highly correlated, the generalization ability of the network will be weakened. In addition, BP neural network uses a gradient descent algorithm with a slow convergence speed. When optimizing complex problems, the network is prone to fall into local optimality, affecting the prediction effect.

[0005] Wave Search Algorithm (WSA) is a new meta-heuristic algorithm inspired by radar technology, using new initialization methods and boundary constraints. It is a wireless communication technology that uses electromagnetic waves to detect the position, speed and shape of targets. It detects, locates, tracks and identifies targets by transmitting radio waves, receiving reflected echoes, and processing and analyzing the echoes. Use the wave search algorithm to optimize the BP neural network and find the optimal regularization parameter and momentum factor , so that the BP neural network can jump out of the local optimum, reduce the loss rate of training and improve the accuracy. Summary of the invention

[0006] In order to solve the above technical problems, the present invention proposes a new energy station transformer voiceprint fault detection method based on wave search optimization, which extracts feature parallel linear combinations through Mel filter and Gammatone filter respectively, optimizes BP neural network with wave search algorithm, and finds the optimal regularization parameter and momentum factor , so that the BP neural network jumps out of the local optimum, reduces the training loss rate, and improves the accuracy of voiceprint fault detection.

[0007] The present invention is implemented by the following technical solutions: A new energy station transformer voiceprint fault diagnosis method based on wave search optimization, comprising the following steps:

[0008] Step 1: Process the preprocessed transformer voiceprint signal through a Mel filter and a Gammatone filter respectively, so as to extract the Mel frequency cepstral coefficients and the Gammatone frequency cepstral coefficients;

[0009] Step 2: Linearly combine the Mel frequency cepstral coefficients and the gammatone frequency cepstral coefficients, and apply the Fisher value after L2 norm regularization to filter them to obtain the optimized mixed cepstral coefficient features, and then form a feature training set;

[0010] Step 3: Build a BP neural network, use an improved wave search algorithm to optimize the BP neural network, and select the optimal regularization parameter and momentum factor; the improved wave search algorithm includes an initialization stage, a global development stage, a local exploration stage, and deterministic optimization;

[0011] On the basis of the initialization phase of the wave search algorithm, a variation intensity coefficient is set to control the amplitude of the Gaussian perturbation, and the generated Gaussian random number is multiplied by the variation intensity coefficient and then added to the current position of the electromagnetic wave particle to generate a new position;

[0012] The local exploration stage includes transmitting electromagnetic waves, reflecting electromagnetic waves, and receiving electromagnetic waves. When receiving electromagnetic waves, an inertia weight is introduced that decreases linearly with the number of iterations. , correction factor for receiving electromagnetic waves Make improvements;

[0013] The deterministic optimization is performed by using a fitting gradient descent method based on a central difference method; and the electromagnetic wave particles beyond the boundary are randomly set within the search range;

[0014] Step 4: Collect transformer voiceprint signals during real-time operation of new energy station transformers to form a test set, import the test set into the trained BP neural network model, determine whether the transformer is faulty and derive the fault type.

[0015] Further preferably, the preprocessing refers to performing frame and window processing on the collected transformer voiceprint signal.

[0016] Further preferably, in step 1, a Mel filter is used to transform the transformer voiceprint signal from a linear frequency to a Mel frequency, and then the obtained Mel frequency is logarithmically transformed and discrete cosine transformed to obtain the final Mel frequency cepstrum coefficients.

[0017] Further preferably, in step 1, a gammatone filter is used to obtain the gammatone frequency cepstrum coefficient, and the formula is:

[0018] ;

[0019] in, For the The gammatone frequency cepstrum coefficients obtained by the Gammatone filter at time L, A and k are the gain and center frequency of the Gammatone filter respectively; It is The phase of a Gammatone filter; is the order of the Gammatone filter; is the number of Gammatone filters; It is Gammatone filter attenuation factor.

[0020] Further preferably, in step 2, the Fisher value calculation formula after introducing L2 norm regularization is:

[0021] ;

[0022] in, is the regularization parameter, represents the square of the L2 norm of the projection vector, is the projection vector, is the transpose of the projection vector, is the inter-class discrete matrix, is the intra-class dispersion matrix, is the Fisher value after L2 norm regularization.

[0023] Further preferably, when building the BP neural network, a hyperbolic tangent function is used as an activation function of the BP neural network.

[0024] In the signal acquisition link, real-time acquisition will consume huge storage space, and the use of unprocessed raw data detection will generate a huge amount of calculation. Therefore, the method of retaining only characteristic signals is adopted in the present invention. The principle is: the fluctuation range of the signal is usually fixed in a rough range. When a fault occurs, the signal will fluctuate violently, with sudden increases and decreases. After feature extraction, only the data at the abnormal moment will be retained for fault identification. The present invention uses Mel filter and Gammatone filter processing to extract Mel frequency cepstral coefficients with moderate noise resistance and more robust GFCC features for linear combination, and uses the improved Fisher ratio method to eliminate components with lower contribution, and introduces regularization means to deal with the overfitting problem that may be caused by the Fisher ratio.

[0025] In the present invention, the BP neural network model is selected to perform voiceprint fault diagnosis of transformers in new energy stations. The trained BP neural network can accurately identify various types of transformer faults. In order to make the BP neural network jump out of the local optimum and reduce the training loss rate, the present invention uses a wave search algorithm to optimize the regularization parameter of the BP neural network. and momentum factor , and finally the selected regularization parameter and momentum factor Import BP neural network to realize transformer fault identification in new energy stations.

[0026] Beneficial effects of the present invention: The present invention adopts a feature extraction method to extract Mel frequency cepstral coefficients with moderate noise resistance and more robust GFCC features, forms feature fusion parameters based on Fisher ratio dimensionality reduction, and introduces regularization means to deal with the overfitting problem that may be caused by Fisher ratio. In addition, the present invention uses an improved wave search algorithm to optimize the BP neural network, reduce sample loss, greatly improve the accuracy of monitoring, and provide a new idea for the BP neural network to jump out of the local optimum. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a flow chart of the transformer voiceprint fault diagnosis method for new energy stations based on wave search optimization.

[0028] Figure 2 Convergence diagram of the fitness function of the wave search algorithm before and after improvement. DETAILED DESCRIPTION

[0029] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0030] A new energy station transformer voiceprint fault diagnosis method based on wave search optimization includes four steps from step 1 to step 4.

[0031] Step 1: The preprocessed transformer voiceprint signal is processed by a Mel filter and a Gammatone filter respectively, so as to extract the Mel frequency cepstral coefficients with moderate noise resistance and the more robust Gammatone frequency cepstral coefficients.

[0032] Step 1.1: Perform preprocessing such as framing and windowing on the collected transformer voiceprint signal, use a Mel filter to transform the linear frequency to Mel frequency, and then perform logarithmic transformation and discrete cosine transformation on the obtained Mel frequency to obtain the final Mel frequency cepstral coefficient (MFCC). The corresponding relationship between Mel frequency and linear frequency is:

[0033] ;

[0034] Where: is the linear frequency, is the Mel frequency corresponding to the linear frequency.

[0035] Step 1.1.1: The Mel filter is a set of triangular filters with equal spacing on the Mel frequency. Its formula on the Mel frequency is:

[0036] ;

[0037] In the formula, is the serial number of the triangular filter in the Mel filter; For the A triangular filter; Respectively The center frequency corresponding to the triangular filter.

[0038] Step 1.1.2: After filtering by the Mel filter, the audio signal is subjected to logarithmic transformation and discrete cosine transformation to obtain the Mel frequency cepstral coefficient (MFCC) of the audio signal. The transformation formula is:

[0039] ;

[0040] In the formula, For the Vermeer frequency cepstral coefficient values, is the first The output of a triangular filter, for Length, is the sequence number of the discrete cosine transform component, is the number of triangular filters in the Mel filter, is the order of the Mel-frequency cepstrum coefficients.

[0041] Step 1.2: The gammatone filter can well simulate the sharp filtering characteristics of the basilar membrane of the human ear. The gammatone frequency cepstral coefficient (GFCC) is obtained using the gammatone filter. The formula is:

[0042] ;

[0043] in, For the The gammatone frequency cepstrum coefficients obtained by the Gammatone filter at time L, A and k are the gain and center frequency of the Gammatone filter respectively; It is The phase of a Gammatone filter; is the order of the Gammatone filter; is the number of Gammatone filters; It is The attenuation factor of the first Gammatone filter is The bandwidth of a Gammatone filter is related to the formula:

[0044] ;

[0045] ;

[0046] In the formula, To obtain an equivalent rectangular bandwidth, the center frequency of each gammatone filter is equally divided on the ERB scale and then mapped to a linear scale.

[0047] Step 2: Linearly combine the Mel frequency cepstral coefficients and the gammatone frequency cepstral coefficients, and apply the Fisher value after L2 norm regularization to filter them to obtain the optimized mixed cepstral coefficient features, and then form a feature training set.

[0048] Step 2.1: Since the contribution of each dimension coefficient of the box-type transformer noise to classification and recognition is different, some parameters may contain more redundant information, resulting in increased computational complexity and even affecting recognition performance. Therefore, the Fisher ratio is introduced. The Fisher ratio is defined as:

[0049] ;

[0050] in is the inter-class discrete matrix, representing the mean variance between different feature components; is the intra-class dispersion matrix, which represents the mean variance between different samples of the same feature component; That is the obtained Fisher value, The larger the value, the greater the contribution of the dimension component to the recognition performance;

[0051] ;

[0052] ;

[0053] In the formula, represents the total number of samples; Indicates Class The mean of the dimensional components; Represents all categories The mean of the dimensional components; Indicates The number of samples in the class; Indicates Class Dimensional feature parameters.

[0054] Step 2.2: A small number of samples in a data set or a high feature dimension may lead to overfitting problems in Fisher discriminant analysis. To solve this problem, a regularization term can be introduced to limit the complexity of the model, thereby obtaining a more robust dimensionality reduction result. By introducing a regularization term to limit the complexity of the model, the robustness of the dimensionality reduction result is improved. In regularized Fisher discriminant analysis, a regularization term is usually added to the objective function (i.e., the Fisher criterion). The following is the formula for calculating the Fisher value after introducing L2 norm regularization:

[0055] ;

[0056] in, is the regularization parameter, which is used to control the strength of the regularization term. represents the square of the L2 norm of the projection vector, is the projection vector, is the transpose of the projection vector, The Fisher value after L2 norm regularization is introduced. This application introduces a regularization method to deal with the overfitting problem that may be caused by the Fisher ratio.

[0057] Step 3: Build a BP neural network, use the improved wave search algorithm to optimize the BP neural network, and select the optimal regularization parameter and momentum factor .

[0058] Step 3.1: Use the hyperbolic tangent function (Tanh) as the activation function of the BP neural network. Compared with other activation functions, Tanh has a larger output range and can improve the learning efficiency of the network. The formula is as follows:

[0059] ;

[0060] Among them, x is the input of the hyperbolic tangent function, e is a natural constant, and the output range of the Tanh function is -1 to 1.

[0061] Step 3.2: Initialization phase;

[0062] First, perform initialization preparation and set the number of electromagnetic wave particles to , the dimension of the problem to be optimized is set to , use matrix to simulate the position of electromagnetic wave particle group :

[0063] ;

[0064] in, is the first-dimensional position of the first electromagnetic wave particle, is the second-dimensional position of the first electromagnetic wave particle, is the d-dimensional position of the first electromagnetic wave particle, is the first-dimensional position of the nth electromagnetic wave particle, is the d-dimensional position of the n-th electromagnetic wave particle;

[0065] Use with Indicates The fitness value of the electromagnetic wave particle, the fitness value F of the population can be expressed by the following vector:

[0066] ;

[0067] use Random numbers are used to initialize the positions of electromagnetic wave particles.

[0068] ;

[0069] ;

[0070] in, For the The current position of the battery wave electromagnetic wave particle, Represents a random number between 0 and 1. yes The result of averaging is yes A random value in and are the upper and lower bounds of the search space.

[0071] In order to broaden the upper and lower boundaries of the search space, each electromagnetic wave particle generates a Gaussian random number that obeys the standard normal distribution and sets a variation intensity coefficient , controls the amplitude of the Gaussian disturbance, multiplies the generated Gaussian random number by the variation intensity coefficient, and then adds it to the current position of the electromagnetic wave particle to generate a new position. The formula is as follows:

[0072] ;

[0073] in, is the new position of the ith battery wave electromagnetic wave particle, is a random number from a standard normal distribution.

[0074] Step 3.3: Global development phase;

[0075] ;

[0076] ;

[0077] In the formula, Is A vector consisting of the minimum values ​​in each dimension, Is A vector consisting of the minimum values ​​in each dimension, is the average value of the fitness of all electromagnetic wave particles, is a random number between 0 and 1, is the current fitness value of the electromagnetic wave particle; the advantage of the above formula is that the search range can be gradually narrowed, the newly generated points are generated within the narrowed range, and the position of the population near the optimal position can be controlled, thereby improving the search efficiency.

[0078] Step 3.4: Local exploration stage;

[0079] Step 3.4.1: Emitting electromagnetic waves;

[0080] ;

[0081] ;

[0082] ;

[0083] In the formula, is the waveform size control coefficient, is the current iteration number, is the total number of iterations, is a column vector whose elements follow a normal distribution and are arranged in order, is the current optimal position. In this step, the electromagnetic wave particles are The purpose is to simulate the outward diffusion of electromagnetic waves, reduce the possibility of falling into the local optimum, improve the search efficiency, and introduce an improved greedy mechanism to make the group position no worse than the current group position when the group position fluctuates outward.

[0084] Step 3.4.2: Reflecting electromagnetic waves;

[0085] ;

[0086] ;

[0087] ;

[0088] In the formula, Indicates Electromagnetic wave particles, is the reflection intensity coefficient, is a random value between 0 and 1. To simulate the number of particles reflecting electromagnetic waves, e is a natural constant, It is the position of the electromagnetic wave particles after they are rearranged in the order of fitness value from small to large. It simulates the situation that some electromagnetic wave particles (electromagnetic wave particles with lower fitness value) encounter obstacles and move towards Reflection, while the remaining electromagnetic wave particles (electromagnetic wave particles with higher fitness values) move away Continue to spread outward.

[0089] Step 3.4.3: Receiving electromagnetic waves;

[0090] ;

[0091] ;

[0092] ;

[0093] ;

[0094] ;

[0095] in is the receiving coefficient, is a random number from normal distribution, To simulate the number of particles receiving electromagnetic waves, is the correction factor, is the historical optimal position obtained by convolution; , , , is a random number between 0 and 1; are the historical optimal positions of the 1st, 2nd, 3rd, …, tth iterations respectively, are the weights of the 1st, 2nd, 3rd, …, tth iterations respectively, The above formula simulates that the radar usually receives electromagnetic waves normally, but sometimes it is interfered with and needs to be corrected and processed. Its function is to make the electromagnetic wave particle group search in the current optimal direction, so that it has a certain probability of updating the weight to reduce the possibility of falling into the local optimum.

[0096] Introduce an inertia weight that decreases linearly with the number of iterations , for the correction factor Improvements are made so that as the number of iterations increases, the search is more focused on the vicinity of the local optimal solution, which is used to improve the accuracy of receiving electromagnetic wave particles. The formula is as follows:

[0097] ;

[0098] ;

[0099] in and are the maximum and minimum values ​​of the inertia weight, is the updated inertia weight.

[0100] As the number of iterations increases, Linear decrease enables electromagnetic wave particles to search more accurately near the local optimal solution in the later stage, thereby improving the search accuracy. The moving step size of the electromagnetic wave particles is limited, so that the electromagnetic wave particles can conduct a more detailed search in the local area and avoid skipping the optimal solution due to an excessively large step size.

[0101] Step 3.5: Deterministic optimization: To improve the convergence speed of the algorithm, the fitted gradient descent method based on the central difference method is used for deterministic optimization. The expression is:

[0102] ;

[0103] ;

[0104] In the formula, express The fitness value of electromagnetic wave particles, , express The fitness value of electromagnetic wave particles, ; Indicated in The position of the electromagnetic wave particle at the iteration, Indicated in The position of the electromagnetic wave particle at the iteration, the gradient descent coefficient of the electromagnetic wave particle , is the gradient, Assume that the initial step length is , determine the final step length through the step length test; the step length test method is as follows: if the fitness value of the initial step length is less than or equal to the current fitness value, then ,otherwise ,in is the scaling factor. The central difference method is used to fit the analytical information of the problem to be optimized and to search for the optimal solution to improve the search efficiency and accuracy.

[0105] Since the generalized boundary restriction rule is to place the electromagnetic wave particles that cross the boundary on the boundary, in order to increase the diversity of the electromagnetic wave particle group, this method randomly sets the electromagnetic wave particles that exceed the boundary within the search range, and the expression is:

[0106] ;

[0107] in, and are the upper and lower boundaries of the search range, respectively. is a d-dimensional randomized column vector whose elements range from 0 to 1.

[0108] The convergence curves of the fitness function of the wave search algorithm before and after the improvement are as follows: Figure 2 As shown in the figure, the improved wave search algorithm has smaller fluctuations and is more stable in the later stages of iteration, and the fitness function is closer to 0 as the number of iterations increases.

[0109] Step 4: Collect transformer voiceprint signals during real-time operation of new energy station transformers to form a test set, import the test set into the trained BP neural network model, determine whether the transformer is faulty and derive the fault type.

[0110] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A new energy station transformer voiceprint fault diagnosis method based on wave search optimization, characterized in that: The following steps are involved: Step 1: Process the preprocessed transformer voiceprint signal through a Mel filter and a Gammatone filter respectively, so as to extract the Mel frequency cepstral coefficients and the Gammatone frequency cepstral coefficients; Step 2: Linearly combine the Mel frequency cepstral coefficients and the gammatone frequency cepstral coefficients, and apply the Fisher value after L2 norm regularization to filter them to obtain the optimized mixed cepstral coefficient features, and then form a feature training set; Step 3: Build a BP neural network, use an improved wave search algorithm to optimize the BP neural network, and select the optimal regularization parameter and momentum factor; the improved wave search algorithm includes an initialization stage, a global development stage, a local exploration stage, and deterministic optimization; On the basis of the initialization phase of the wave search algorithm, a variation intensity coefficient is set to control the amplitude of the Gaussian perturbation, and the generated Gaussian random number is multiplied by the variation intensity coefficient and then added to the current position of the electromagnetic wave particle to generate a new position; The local exploration stage includes transmitting electromagnetic waves, reflecting electromagnetic waves, and receiving electromagnetic waves. When receiving electromagnetic waves, an inertia weight is introduced that decreases linearly with the number of iterations. , correction factor for receiving electromagnetic waves Make improvements; The deterministic optimization is performed by using a fitting gradient descent method based on a central difference method; and the electromagnetic wave particles beyond the boundary are randomly set within the search range; Step 4: Collect transformer voiceprint signals during real-time operation of new energy station transformers to form a test set, import the test set into the trained BP neural network model, determine whether the transformer is faulty and derive the fault type.

2. The voiceprint fault diagnosis method for transformers in new energy stations based on wave search optimization according to claim 1 is characterized in that: The preprocessing refers to framing and windowing the collected transformer voiceprint signal.

3. The new energy station transformer voiceprint fault diagnosis method based on wave search optimization according to claim 1 is characterized in that: In step 1, a Mel filter is used to transform the transformer voiceprint signal from linear frequency to Mel frequency, and then the obtained Mel frequency is logarithmically transformed and discrete cosine transformed to obtain the final Mel frequency cepstrum coefficient.

4. The new energy station transformer voiceprint fault diagnosis method based on wave search optimization according to claim 1 is characterized in that: In step 2, the calculation formula of the Fisher value after introducing L2 norm regularization is: ； in, is the regularization parameter, represents the square of the L2 norm of the projection vector, is the projection vector, is the transpose of the projection vector, is the inter-class discrete matrix, is the intra-class dispersion matrix, is the Fisher value after L2 norm regularization.

5. The new energy station transformer voiceprint fault diagnosis method based on wave search optimization according to claim 1 is characterized in that: When building a BP neural network, the hyperbolic tangent function is used as the activation function of the BP neural network.

6. The new energy station transformer voiceprint fault diagnosis method based on wave search optimization according to claim 1 is characterized in that: Introduce an inertia weight that decreases linearly with the number of iterations , correction factor for receiving electromagnetic waves To improve, the formula is as follows: ； ； in and are the maximum and minimum values ​​of the inertia weight, is the updated correction factor, is the current iteration number, is the total number of iterations.

Images