Power transformer fault soundprint detection method based on extreme learning machine

Abstract

The application discloses a power transformer fault sound print detection method based on an extreme learning machine, first, audio signals of transformer faults are collected for framing, and preprocessed through fast Fourier transform; then, sound print characteristic parameters of the audio signals are extracted through an improved Mel filter, a fault sound print detection model based on the extreme learning machine is constructed, and model training is conducted; finally, a test set is established, real-time test samples are input into the trained fault sound print detection model, test identification is conducted, sound print identification results of transformer faults are obtained, and real-time fault sound print early warning is realized. The improved Mel filter can improve the spectral resolution of transformer sound print feature extraction, reduce the system complexity, the fault sound print detection model is constructed through the extreme learning machine algorithm, the transformer fault detection accuracy is effectively improved, and the method is easy to implement in engineering.

Description

Technical Field

[0001] This invention relates to a method for detecting the acoustic signature of power transformer faults, and particularly to a method for detecting the acoustic signature of power transformer faults based on an extreme learning machine. Background Technology

[0002] As a critical piece of equipment in the power system, the fault-free operation of power transformers plays a vital role in ensuring a reliable power supply. When a transformer fails, it can cause localized power outages or, in severe cases, disrupt the safe operation of the entire power system, resulting in significant economic losses. With the expansion of power grids and the development of intelligent systems, condition monitoring and fault early warning of substation equipment have become crucial means to ensure the stable operation of the power system. Substation personnel can assess the operating status of transformers by listening to their operating sounds during on-site inspections; however, approaching a transformer emitting abnormal sounds can threaten their personal safety. Utilizing modern digital signal processing technology for transformer acoustic signature fault diagnosis can improve the accuracy of acoustic signature identification while also ensuring the safety of personnel.

[0003] During the operation of substation equipment, acoustic signature information changes with the equipment's operating status. Therefore, by monitoring and analyzing transformer acoustic signatures, it is possible to preliminarily determine whether the transformer's operating status is normal, promptly identify potential faults, and provide a basis for transformer maintenance and repair. Based on the acoustic signature characteristics of transformers under different operating states, mathematical models can be established to describe the normal and fault states of the equipment. For example, algorithms such as BP neural networks, Naive Bayes, SVM, and CNN can be used to achieve transformer fault detection.

[0004] However, the audio signals from substations are characterized by non-steady-state behavior and large fluctuations. This is because the operating status of power equipment and environmental factors can both affect the audio signals, leading to instability and volatility. When the transformer's operating status fluctuates, the extracted feature quantities overlap between normal and fault states. In the short term, these fluctuations can be treated as relatively stable signals. Therefore, appropriate signal processing techniques and algorithms are needed to preprocess the audio signals, extracting feature information reflecting the equipment's operating status to achieve real-time monitoring and fault early warning of the equipment's operating condition. Summary of the Invention

[0005] The problem to be solved by this invention is to provide a power transformer fault acoustic signature detection method based on extreme learning machine. By analyzing the acoustic signature information of the power equipment during operation, key feature information can be extracted quickly and accurately to realize real-time monitoring of the equipment's working status and fault early warning.

[0006] This invention adopts the following technical solution: a power transformer fault acoustic signature detection method based on extreme learning machine, comprising the following steps:

[0007] S1. Audio data preprocessing: Acquire audio signals of transformer faults, use overlapping frame division method to divide the audio signals into frames, preprocess each frame of data after frame division by fast Fourier transform, convert the time domain signal to the frequency domain, and obtain spectrum data.

[0008] S2. Voiceprint feature extraction: Using an improved Mel filter, the voiceprint feature parameters of the audio signals in the training set are extracted to construct the training set;

[0009] S3. Construct a fault voiceprint detection model based on extreme learning machine and train it with samples to obtain a trained fault voiceprint detection model.

[0010] S4. Establish a test set. Select three consecutive frames from the middle position of each audio signal to evaluate the voiceprint features and obtain real-time test samples to form a test set.

[0011] S5. Perform fault detection by inputting real-time test samples into the trained fault voiceprint detection model for test recognition, and obtain the voiceprint recognition results of transformer faults.

[0012] Specifically, in step S1, the audio signal of the transformer fault has a sampling frequency of 48kHz and a length of 10s;

[0013] The audio signal is divided into frames using an overlapping framing method. Each frame has a length of 8192 points and a frame shift of 4096 points. Each audio signal is divided into 116 frames. The data x of the i-th frame of the audio signal... i Represented as: x i =[x i1 ,x i2 ,…,x i8192 ], 1≤i≤116;

[0014] For each frame of data x i Perform a Fast Fourier Transform to remove negative frequencies and retain positive frequencies to obtain the spectral data y. i , represented as: y i =[y i1 ,y i2 ,…,y i4096 ].

[0015] Specifically, in step S2, the Mel filter is improved by replacing the triangular filter bank in the traditional Mel filter with a rectangular filter bank to reduce the impact of amplitude modulation extracted by filtering on the voiceprint features. In addition, according to the principle of low frequency density and high frequency sparseness, the number of low frequency band filters is increased to improve the spectral resolution and reduce complexity.

[0016] Further, in step S2, based on the spectrum data obtained in step S1, the energy spectrum of the audio signal in the training set in the frequency domain of 0-1200Hz and 19000-20400Hz is extracted using an improved Mel filter, and the output of the i-th frame of the audio signal is z. i The transformation formula is as follows:

[0017]

[0018] The improved Mel filter outputs a 1×M dimensional vector, z i =[z i1 ,z i2 ,…,z iM ].

[0019] Specifically, in step S3, the Extreme Learning Machine is a training algorithm based on a single hidden layer feedforward neural network, including an input layer, a hidden layer, and an output layer; the Extreme Learning Machine determines and assigns values ​​to the weights and biases of the input layer of the single hidden layer feedforward neural network by random selection.

[0020] Construct a fault voiceprint detection model based on extreme learning machine for N different audio signal samples (x j ,t j ), x j =[x j1 ,x j2 ,…,x jn ] T ,t j =[t j1 ,t j2 ,…t jm ] T , t j The audio label, represented by a single-hidden-layer feedforward neural network with K hidden nodes, is as follows:

[0021]

[0022] In the formula, g(x) is the activation function, j = 1, ..., N, w i w represents the weights from the input layer to the hidden layer. i =[w i1 ,w i2 ,…,w in ] T w represents the bias connecting the i-th hidden node. i ·x j For w i and x j The inner product, b i For the bias of the hidden layer, β j This represents the weights from the hidden layer to the output layer for the j-th sample;

[0023] The single-hidden-layer feedforward neural network infinitely approximates N samples, meaning there exists a β. i ω i b i Make:

[0024]

[0025] K equations are represented as:

[0026] Hβ=T

[0027] in:

[0028]

[0029] In the formula, H is the output matrix of the hidden layer of a single hidden layer feedforward neural network, the i-th column of H is the output of the i-th node of the hidden layer of the neural network, and β is the weight from the hidden layer to the output layer.

[0030] Furthermore, a fault voiceprint detection model based on Extreme Learning Machine (ELM) is constructed. The input and output dimensions of the ELM are set according to the extracted feature dimension and label dimension. The labels adopt one-hot encoding, and the normal sound is t. normal =[1,0] T The fault sound is t abnormal =[0,1] T The number of hidden nodes in the extreme learning machine is set to 1200, and the activation function chosen is Sigmoid, as shown below:

[0031]

[0032] Among them, e -x This represents the reciprocal of an exponential function with the natural base e. The model is trained using a training set, and the trained fault soundprint detection model is saved for real-time fault soundprint detection of power transformers.

[0033] Specifically, in step S4, a test set is established by selecting three consecutive frames from the middle position of each audio signal and averaging them using the following formula to obtain the real-time test sample P. j :

[0034]

[0035] Where, p j_58 ,p j_59 ,p j_60 This is the data from frames 58-60 of the j-th audio signal after feature extraction.

[0036] Specifically, in step S5, fault detection is performed, and the test sample P is... jInput the trained fault voiceprint detection model to obtain the detection output t of the j-th sample. j :

[0037] t j =[t j1 ,t j2 ] T

[0038] If t j1 =[1,0] T The test results were normal; t j2 =[0,1] T The test result was abnormal.

[0039] On the other hand, the present invention also provides: an electronic device, comprising:

[0040] One or more processors;

[0041] A storage device on which one or more programs are stored;

[0042] When the one or more programs are executed by the one or more processors, the one or more processors implement any of the above-described methods for detecting power transformer faults based on extreme learning machines.

[0043] The present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in any of the above-mentioned methods for detecting power transformer faults based on extreme learning machines.

[0044] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0045] 1. The present invention provides a method for detecting power transformer fault acoustic signatures. It analyzes the acoustic signature information of power equipment during operation and uses the ELM algorithm for transformer fault acoustic signature detection. ELM randomly selects the input weights of SLFNs. During the execution of the ELM algorithm, no manual parameter adjustment is required, avoiding the iterative process of traditional training algorithms. It converges quickly, greatly reducing training time, improving the accuracy and speed of detection, and is easy to implement in engineering.

[0046] 2. The power transformer fault acoustic signature detection method of the present invention improves the spectral resolution of transformer acoustic signature feature extraction by using an improved Mel filter, reduces system complexity, and effectively improves the fault detection accuracy of the ELM algorithm. Attached Figure Description

[0047] Figure 1 This is a flowchart of the steps in the power transformer fault acoustic fingerprint detection method of the present invention;

[0048] Figure 2 This is a graph showing the relationship between Mel frequency and actual frequency in an embodiment of the present invention.

[0049] Figure 3 The Mel filter composition is shown in the embodiment of the present invention;

[0050] Figure 4 The above are acoustic signature frequency domain diagrams of the transformer under four different operating conditions in this embodiment of the invention.

[0051] Figure 5 This is a time-frequency comparison diagram of voiceprint feature extraction before and after the improvement of this invention;

[0052] Figure 6 This is a structural diagram of the SLFN of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the application will be further described in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in this invention. All non-innovative embodiments based on these embodiments by other researchers in the art are within the protection scope of this invention. Furthermore, the step numbers in the embodiments of this invention are only set for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0054] This invention proposes a power transformer fault acoustic signature detection method based on Extreme Learning Machine (ELM). By analyzing the acoustic signature information of power equipment during operation, it achieves real-time monitoring of equipment operating status and fault early warning, such as... Figure 1 As shown, it includes the following steps:

[0055] S1. Audio data preprocessing: Collect audio signals of transformer faults, construct training sets, use overlapping frame division method to divide the audio signals of the training sets into frames, and preprocess each frame of data after frame division by fast Fourier transform to convert the time domain signal to the frequency domain and obtain spectrum data.

[0056] S2. Voiceprint feature extraction: Using an improved Mel filter, the voiceprint feature parameters of the audio signals in the training set are extracted to construct the training set;

[0057] S3. Construct an ELM-based fault voiceprint detection model and train it with samples to obtain a trained fault voiceprint detection model.

[0058] S4. Establish a test set. Select three consecutive frames from the middle position of each audio signal to evaluate the voiceprint features and obtain real-time test samples to form a test set.

[0059] S5. Perform fault detection by inputting real-time test samples into the trained fault voiceprint detection model for test recognition, and obtain the voiceprint recognition results of transformer faults.

[0060] In one embodiment of the present invention, an audio signal of a power transformer fault is collected at a sampling frequency of 48kHz and a length of 10s. The power transformer fault voiceprint detection method based on extreme learning machine is as follows:

[0061] S1, Audio Data Preprocessing

[0062] In signal processing, framing is a common strategy that divides the speech signal into multiple short-time segments, each called a frame. To maintain continuity and correlation, it is necessary to ensure that adjacent frames overlap; this method is called overlapping framing.

[0063] In this embodiment, the acquired audio signal is divided into multiple short-time signal segments, each of which is called a frame. To maintain continuity and correlation, it is necessary to ensure that there is overlap between adjacent frames. Assuming a frame length of 8192 points and a frame shift of 4096 points, a certain audio segment can be divided into 116 frames.

[0064] The number of points in a Fast Fourier Transform (FFT) is typically consistent with the frame length. With a fixed audio sampling rate, a longer frame length results in higher frequency resolution, but also increases the computational complexity of the FFT. Considering all factors, this embodiment selects a frame length of 8192 points and a frame shift of 4096 points. For each 10-second audio clip with a sampling rate of 48kHz, it can be divided into 116 frames. This ensures both frequency resolution and keeps the computational load within an acceptable range.

[0065] Specifically, the data of the i-th frame of audio is represented as x. i =[x i1 ,x i2 ,…,x i8192 For each frame, perform an FFT (Fast Fourier Transform) to remove negative frequencies and retain positive frequencies, resulting in the spectral data y, where 1 ≤ i ≤ 116. i =[y i1 ,y i2 ,…,y i4096 ], 1≤i≤116.

[0066] S2, Voiceprint Feature Extraction

[0067] Commonly used microphone sampling rates include 44.1kHz, 48kHz, and 96kHz. Without feature extraction, the data volume would be too large, causing significant inconvenience to detection and recognition. Therefore, this embodiment uses FFT to convert the time-domain signal to the frequency domain and employs a dimensionality reduction filter to extract feature parameters. This not only reduces the data dimensionality but also improves the category representation capability, thereby enhancing the performance of the classification algorithm.

[0068] Studies have shown that the human ear is more sensitive to low-frequency signals than high-frequency signals, and the design of Mel filter banks is based on the perceptual characteristics of the human ear at different frequencies. Therefore, in a Mel filter bank, the number of filters is denser at low frequencies and sparser at high frequencies. The specific idea is to map the frequency domain signal to the Mel frequency domain and then divide it into equal-width Mel bands. The Mel filter is analogous to the human cochlea, simulating the sensitivity of the human ear to low and high frequencies, reflecting the nonlinear response of the human ear to frequency, such as... Figure 2 As shown.

[0069] Traditional Mel filters consist of several non-uniformly distributed triangular filters. The filters are densely distributed at low frequencies, resulting in higher frequency resolution and making it easier to extract low-frequency features. Figure 3 As shown.

[0070] The frequency response of the filter is:

[0071]

[0072] Where k is the frequency, f(m) is the center frequency of the m-th triangular filter, and the energy output by each triangular filter is the extracted feature.

[0073] Mel filters have been widely used in voiceprint recognition and fault diagnosis, achieving good results. However, for transformer voiceprints that are time-stable and whose frequency domain distribution is limited to high and low frequencies, Mel filters cannot accurately extract the desired voiceprint.

[0074] Acoustic frequency domain diagrams of transformers under four different operating conditions, such as Figure 4 As shown, considering that the transformer's operating frequency is mainly concentrated in the 50Hz octave and does not exceed 1kHz, significant changes in these frequencies are usually associated with abnormal transformer faults. Therefore, maintaining the resolution of this frequency band is particularly important during frequency extraction. Although the Mel filter is approximately linear within 1kHz, calculations show that the peak frequency of the Mel filter bank is difficult to align with the 50Hz octave. This means that using the Mel filter bank may result in the loss of key frequency characteristics below 1kHz. Furthermore, the fault tone of partial discharge has a high power output in the 19000-20400Hz range.

[0075] Given these potential issues, the application of Mel filter banks in transformer acoustic signature feature extraction requires careful consideration. In-depth analysis of numerous audio data time-frequency plots within the State Grid database revealed that the mid-frequency band of transformer fault sounds contains almost no effective feature information. However, for certain types of fault sounds, such as partial discharge, there are high-power frequency components in the high-frequency band (around 20kHz). Therefore, during feature extraction, we should focus on the feature information in the low-frequency and high-frequency bands. Furthermore, the frequency response of Mel filter banks is complex, and increasing the dimensionality can affect operating speed.

[0076] Therefore, to address the aforementioned problems, this embodiment improves the transformer acoustic signature characteristics by optimizing the Mel filter for feature extraction. A time-frequency comparison of acoustic signature feature extraction before and after the improvement is shown below. Figure 5 As shown, the method in this embodiment can effectively reduce the algorithm complexity.

[0077] Specifically, in this embodiment, the Mel filter is improved to extract the energy spectrum in the frequency domain of 0-1200Hz and 19000-20400Hz. The triangular filter bank is replaced with a high-efficiency rectangular filter bank, which can reduce the impact of amplitude modulation of the filtered features on the voiceprint characteristics. According to the principle of dense low-frequency and sparse high-frequency, the number of low-frequency filters is increased to improve the spectral resolution and effectively reduce the algorithm complexity.

[0078] The output of the improved Mel filter can be represented by a 1×M dimensional vector. Preferably, in this embodiment, M is 61.

[0079] After processing with the improved Mel filter, the output of the i-th frame is z. i =[z i1 ,z i2 ,…,z i61 ], 1≤i≤116, the transformation formula is as follows:

[0080]

[0081] S3. Construct an ELM-based fault voiceprint detection model

[0082] Extreme Learning Machine (ELM) is a training algorithm for Single-Hidden Layer Feedforward Network (SLFN). It has extremely fast training speed and can achieve high generalization performance, making it suitable for situations with a large number of training samples.

[0083] Unlike traditional training algorithms, the ELM algorithm in this embodiment randomly assigns weights and biases to the input layer, and then directly solves for the weights from the hidden layer to the output layer by finding the Moore-Penrose generalized inverse matrix.

[0084] A typical SLFN consists of an input layer, hidden layers, and an output layer, such as... Figure 6 As shown, its main characteristic is that the number of neurons in its hidden layer is much larger than the number of neurons in the input layer. This results in SLFN being able to quickly generate random connection weights during training, thus accelerating the training process. In SLFN, the connection weights between the input layer and the super-hidden layer are usually randomly initialized, while the connection weights of the output layer can be directly calculated using analytical solutions or pseudo-inverse methods, without requiring iterative backpropagation.

[0085] In this embodiment, ELM randomly selects the input weights of SLFN and determines the output weights through analysis and judgment. Compared with traditional SLFN, ELM has significant advantages in practical applications: fast training speed, no need for repeated parameter adjustment, excellent generalization performance, and the ability to easily obtain the global optimum.

[0086] The learning mode of the fault voiceprint detection model in this embodiment is as follows: for any N different samples (x j ,t j ), where x j =[x j1 ,x j2 ,…,x jn ] T ,t j =[t j1 ,t j2 ,…t jm ] T , t j The label representing the audio.

[0087] Specifically, in this embodiment, m=2, the normal tone is [1,0], and the fault tone is [0,1].

[0088] The standard SLFN with K hidden nodes is represented as follows:

[0089]

[0090] j = 1, ..., N

[0091] Where g(x) is the activation function, w i =[w i1 ,w i2 ,...,w in ] T w is the bias used to connect the i-th hidden node. i ·x j For w i and x j The inner product of.

[0092] Such an SLFN can approximate these N samples infinitely: In other words, there exists β i ω i b i Make:

[0093]

[0094] These K equations can be written as:

[0095] Hβ=T

[0096] in:

[0097]

[0098] In the formula, H is the output matrix of the hidden layer of the neural network, and the i-th column of H is the output of the i-th node of the hidden layer of the neural network.

[0099] It is important to note that if the activation function g(x) is infinitely differentiable, then the weights w from the input layer to the hidden layer can be chosen randomly. i and the bias b of the hidden layer i SLFN can be treated as a linear system, and the entire network can be determined simply by analytically calculating the weights β from the hidden layer to the output layer.

[0100] Hβ=T, where H is an N×K matrix. If N=K, then H is invertible; if N≠K, then the Moore-Penrose generalized inverse matrix of H is calculated. The network obtained in this way not only has the minimum training error, but also the minimum norm of the weights.

[0101] In this embodiment, the input and output dimensions of the ELM are set to 61 and 2, respectively, based on the extracted feature dimension and label dimension. The labels use one-hot encoding, and the normal tone is t. normal =[1,0] T The fault sound is t abnormal =[0,1] T The number of hidden nodes in the ELM is set to 1200, and the activation function chosen is Sigmoid.

[0102]

[0103] The trained hidden layer model is saved for subsequent fault voiceprint detection.

[0104] In this embodiment, the ELM algorithm does not require manual parameter adjustment during execution, avoiding the iterative process of traditional training algorithms, and converges quickly, greatly reducing training time. With its excellent generalization ability and extremely fast training speed, ELM is superior to SVM in both detection accuracy and speed. The time for voiceprint detection is greatly shortened, and the accuracy is further improved to 98.6%.

[0105] S4. Test Set Setup

[0106] The test set is consistent with the training set in terms of feature extraction. The difference is that the training set needs to be divided into frames, while the test set only needs to take the three middle frames and average the features.

[0107] In this embodiment, the data p from frames 58, 59, and 60 after feature extraction is... j_58 ,p j_59 ,p j_60 Perform the following processing:

[0108]

[0109] Obtain real-time test sample P j .

[0110] S5, Fault Detection

[0111] Inputting real-time recorded audio data into a trained model can output a label t. i =[t i1 ,t i2 ] T As a test result, a normal result is [1,0]. T An exception will be [0,1]. T .

[0112] In summary, the power transformer fault acoustic signature detection method of this invention analyzes the acoustic signature information of power equipment during operation and uses the ELM algorithm for transformer fault acoustic signature detection. ELM randomly selects the input weights of SLFNs, eliminating the need for manual parameter adjustment during execution and avoiding the iterative process of traditional training algorithms. This results in rapid convergence, significantly reducing training time and improving detection accuracy and speed, while also being easy to implement in engineering. Furthermore, the improved Mel filter enhances the spectral resolution of transformer acoustic signature feature extraction, reduces system complexity, and effectively improves the fault detection accuracy of the ELM algorithm.

[0113] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A power transformer fault acoustic print detection method based on extreme learning machine, characterized in that, Comprise the following steps: S1, audio data preprocessing: collecting the audio signal of transformer fault, using overlapping frame method to frame the audio signal, pre-processing each frame data through fast Fourier transform after framing, converting time domain signal to frequency domain, obtaining frequency spectrum data; S2, voiceprint feature extraction: extract voiceprint feature parameters through improved mel filter, and construct training set; The mel filter is improved, the triangular filter bank in the traditional mel filter is replaced by a rectangular filter bank, the influence of amplitude modulation on voiceprint features is reduced, the number of low frequency band filters is increased according to the principle of low frequency density and high frequency sparsity, the spectral resolution is improved, and the complexity is reduced; S3, constructing a fault voiceprint detection model based on extreme learning machine, and training samples to obtain a trained fault voiceprint detection model; The extreme learning machine is a training algorithm based on single hidden layer forward neural network, including input layer, hidden layer and output layer; The extreme learning machine determines the input layer weight and bias of the single hidden layer forward neural network by random selection and assignment; A fault voiceprint detection model based on extreme learning machine is constructed for N different audio signal samples , , , represents the label of the audio, a single-hidden-layer feedforward neural network with implicit nodes is represented as: ； wherein, is an activation function, , is a weight from the input layer to the hidden layer, denotes a bias connected to the i-th hidden node, is and an inner product of is a bias of the hidden layer, denotes a weight from the j-th sample hidden layer to the output layer; The single-hidden-layer feedforward neural network, infinite approximation A sample, i.e. there is , , So that: ； The equation is represented as: ； Wherein: ； wherein is the output matrix of the hidden layer of the single hidden layer feedforward neural network, the first column is the output of the first node of the hidden layer of the neural network, and β is the weight from the hidden layer to the output layer. S4, establishing a test set, selecting 3 continuous frames from the middle position of each audio signal for voiceprint feature evaluation to obtain real-time test samples and form a test set; S5, fault detection, inputting real-time test samples into the trained fault voiceprint detection model for test identification to obtain the voiceprint recognition result of transformer fault.

2. The power transformer fault acoustic print detection method of claim 1, wherein, In step S1, the audio signal of transformer fault has a sampling frequency of 48kHz and a length of 10s; The audio signals are framed by using the overlapping frame method, each frame has a length of 8192 points, and a frame shift of 4096 points, and each audio signal is divided into 116 frames; the data of the i-th frame of the audio signal is represented as: is represented as: , ; For each frame of data Take the fast Fourier transform, remove the negative frequencies, and keep the positive frequencies to get the spectral data , denoted as .

3. The power transformer fault acoustic print detection method of claim 2, wherein, In step S2, based on the obtained spectrum data in step S1, the energy spectrum of the audio signal in the frequency domain of 0-1200 Hz and 19000-20400 Hz in the training set is extracted by the improved Mel filter, and the output of the i-th frame of the audio signal is The transformation formula is as follows: ； wherein the improved mel filter output is a 1 x M dimensional vector, .

4. The power transformer fault acoustic print detection method of claim 3, wherein, In step S3, a fault voiceprint detection model based on an extreme learning machine is constructed. The input and output dimensions of the extreme learning machine are set according to the extracted feature dimension and label dimension. The label is encoded by using a one-hot encoding. The normal voice is , and the fault voice is . The number of hidden nodes of the extreme learning machine is set to 1200, and the selected activation function is Sigmoid, which is expressed as follows: ； wherein, represents the inverse of the exponential function with the natural base e, the model training is performed through the training set, and the trained fault voiceprint detection model is saved for real-time detection of the power transformer fault voiceprint.

5. The power transformer fault acoustic print detection method of claim 4, wherein, In step S4, the test set is established, and 3 frames of continuous audio signals are selected from the middle position of each audio signal. The numerical average is obtained by the following formula to obtain the real-time test sample P j : ； wherein, is the data extracted from frames 58-60 of the jth audio signal.

6. The power transformer fault acoustic print detection method of claim 5, wherein, In step S5, fault detection is performed on the test sample The trained fault voiceprint detection model is input to obtain the jth sample detection output : ； If , the detection result is normal; , the detection result is abnormal.

7. An electronic device, comprising: Comprise: One or more processors; Storage device, one or more programs are stored on the storage device; When the one or more programs are executed by the one or more processors, the one or more processors implement the power transformer fault voiceprint detection method of any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, The computer program is stored on the storage device, and the program is executed by the processor to realize the steps in the power transformer fault voiceprint detection method of any one of claims 1-6.

Images