Transformer on-load tap-changer fault diagnosis method based on background sound texture

Abstract

The present invention relates to a transformer on-load tapchanger fault diagnosis method based on background sound texture. The method comprises the following steps: preprocessing raw audio data; using a deep neural network to generate a time-frequency mask of equipment sounds, and using the time-frequency mask to separate pure equipment sounds; using a primary analysis model to pre-determine the fault status of the separated sound activity; performing start-end endpoint detection on audio events in audio data areas determined to be in a fault state, extracting audio event segments; and using a reanalysis model to accurately identify the detected audio event segments to determine the fault type. This method is beneficial for improving the accuracy of fault diagnosis.

Description

Technical Field

[0001] The present invention belongs to the technical field of power equipment fault diagnosis, and in particular relates to a transformer on-load tap changer fault diagnosis method based on background sound texture. Background Art

[0002] The working status of electrical equipment is affected by a combination of internal and external factors. Existing automated monitoring methods mainly focus on monitoring electrical and chemical parameters. For some auxiliary information, such as sound, they mainly rely on on-site manual inspections by operation and maintenance personnel, which makes it difficult to achieve real-time and comprehensive coverage. In addition, the background interference of the on-site environment is relatively large, which makes it difficult to detect equipment hidden dangers and faults in the first time, thus laying hidden dangers for the normal operation of the equipment.

[0003] Transformers, reactors, and other critical electrical equipment in power systems perform critical tasks such as voltage conversion within the system and the distribution and transmission of electrical energy in substations. They play a vital role in providing high-quality power services and ensuring the safe, reliable, high-quality, and economical operation of power systems. Due to the high volume of these electrical equipment, their diverse capacity levels and specifications, and their long operating times, their accident rates are also increasing. With technological advancements, monitoring and diagnostic techniques for electrical equipment are also gradually improving. Acoustic anomaly monitoring is gradually becoming a new and effective method for detecting electrical equipment anomalies.

[0004] Sound is a mechanical wave that radiates energy through vibration into a sound-transmitting medium. Acoustic signals contain a wealth of vibration information and are a crucial indicator for analyzing equipment operating conditions. During normal operation, equipment produces sound when the motion between the equipment, components, and the components themselves changes. The sound changes accordingly. Electrical equipment is complex and contains a wide variety of components. For example, a three-phase oil-immersed transformer consists of primary and secondary windings, an iron core, an oil tank, a base, high- and low-voltage bushings, a radiator (cooler), an oil purifier, an oil conservator, a gas relay, a safety airway, a thermometer, a tap changer, and other related components and accessories. Electrical equipment can experience various faults in high voltage and strong electromagnetic environments, resulting in changes in the sound it produces. To determine the type of equipment fault, substation operators often place one end of an insulating rod against the equipment and the other end to their ear to listen carefully. Although this method is simple to operate, it cannot achieve remote and continuous online detection, which is inconsistent with the development trend of unmanned substations. It also requires the detection personnel to have rich practical experience, which brings unstable factors to the accuracy of the judgment.

[0005] Compared with other types of monitoring methods for electrical equipment, acoustic anomaly monitoring does not require equipment shutdown when diagnosing equipment status, and is an online monitoring method. In addition, acoustic anomaly monitoring does not require contact with the electrical measuring equipment being measured and can operate at a location far away from the equipment being measured, making it safer than measurement methods that require contact.

[0006] However, the complex acoustic environment surrounding the transformer itself severely interferes with acoustic anomaly monitoring, resulting in a low fault identification rate. This limits existing acoustic anomaly monitoring technology, and its effectiveness needs to be improved. Summary of the Invention

[0007] The object of the present invention is to provide a transformer on-load tap changer fault diagnosis method based on background sound texture, which is conducive to improving the accuracy of fault diagnosis.

[0008] To achieve the above object, the present invention adopts a technical solution: a transformer on-load tap changer fault diagnosis method based on background sound texture, comprising the following steps:

[0009] Preprocess the raw audio data;

[0010] Use a deep neural network to obtain the time-frequency mask of the device sound, and use the time-frequency mask to separate the pure device sound;

[0011] Use the initial analysis model to pre-judge the fault status of the separated sound activities;

[0012] Performing start-end endpoint detection of an audio event on the audio data area judged to be in a fault state, and extracting an audio event segment;

[0013] The detected audio event segments are accurately identified using a reanalysis model to obtain the type of fault.

[0014] Furthermore, the specific method for preprocessing the original audio data is:

[0015] First, the collected original audio data is divided into frames according to the set frame length and frame shift to obtain the framed data s i (n), where i represents the frame subscript of the signal and n represents the number of discrete time points;

[0016] Perform peak normalization on the framed data so that the data are distributed between (-1, 1);

[0017] The framed data s i (n) Perform short-time Fourier transform. The specific method is as follows:

[0018]

[0019] Where S(n,f) represents the time domain signal s i (m), w(nm) represents the short-time Fourier transform of the window function w(m) after mirroring and shifting it to the right by n discrete time points, M = 1024 is the window length, j represents the imaginary unit, and f is the instantaneous frequency of the signal;

[0020] Take the base 10 logarithm of S(n,f) and normalize it to get the spectrum S D (n,f)=log 10 S(n,f).

[0021] Furthermore, a training dataset was constructed using previously collected high signal-to-noise ratio transformer on-load tapchanger switching phase sound signals and background sound signals. This was then fed into a deep neural network for training. Actual on-site audio was then fed into the deep neural network and separated to obtain the time-frequency mask of the equipment sound. This mask was then used to isolate the pure equipment sound.

[0022] The training dataset is constructed by using the previously collected high signal-to-noise ratio transformer on-load tapchanger switching phase sound signal and background sound signal to calculate the corresponding ideal ratio time-frequency mask (IRM). The time-frequency mask is calculated as follows:

[0023]

[0024] Where N D (n,f) represents the spectrogram of the background sound signal;

[0025] The time-frequency mask of the separated device sound is multiplied by the amplitude spectrum of the short-time Fourier transform S(n,f) of the original signal to obtain the amplitude spectrum of the signal after background sound separation, and the phase of S(n,f) is used to obtain the estimated device sound spectrum S P (n,f), that is:

[0026] S P (n,f)=|S(n,f)|·IRM(n,f)·e j∠S(n,f)

[0027] Where |S(n,f)| represents the amplitude spectrum of the signal, and ∠S(n,f) represents the phase of the signal;

[0028] Device sound spectrum S P (n,f) performs inverse Fourier transform to obtain the pure equipment sound signal s P (n).

[0029] Furthermore, the initial analysis model based on deep convolutional neural network is used to judge the separated device sound spectrum S PWhether (n,f) is in a fault state, the network structure is as follows: two-dimensional input layer, convolution layer, normalization layer, maximum pooling layer, convolution layer, normalization layer, maximum pooling layer, convolution layer, normalization layer, maximum pooling layer, convolution layer, normalization layer, maximum pooling layer, fully connected layer, sigmoid output layer; if the pre-judgment result is a fault, subsequent processing is performed.

[0030] Furthermore, the start-end endpoint detection of the audio event is performed on the audio data area judged to be in a fault state, and the audio event fragment is intercepted, which specifically includes the following steps:

[0031] 1) Use short-time Fourier transform to transform the sound spectrum data S that is judged as the fault type P (n,f) is restored to the purified sound waveform data s P (n);

[0032] 2) Use Hilbert transform to calculate the envelope of the sound waveform data to be recognized

[0033]

[0034] Where * represents convolution operation;

[0035] 3) According to the envelope contour Set a high threshold, T2, and the first intersection point, N1, between the envelope and T2. Determine a low threshold, T1, based on the background noise energy. Start searching from before N1 and find the point N2 that intersects T1. Use N2 as the starting point of the device's sound activity and N3 = N2 + 5.2 * Fs as the end point, where Fs is the sampling rate of the sound signal.

[0036] 4) The purified sound waveform data s P (n) is intercepted with the starting point N2 and the ending point N3 to obtain the reanalysis data s R (n).

[0037] Furthermore, the reanalysis model is used to analyze the detected reanalysis data s R (n) to accurately identify and obtain the type of fault. The reanalysis model uses a deep neural network for fault identification. The network type is a convolutional neural network. The network structure is as follows: two-dimensional input layer, convolution layer, convolution layer, fully connected layer, activation function layer, fully connected layer, activation function layer, fully connected layer, activation function layer, fully connected layer, and Sigmoid output layer. The label type corresponding to the maximum probability of the output layer is selected as the fault type of the equipment.

[0038] Compared with the prior art, the present invention has the following beneficial effects: it provides a transformer on-load tap-changer fault diagnosis method based on background sound texture. The method utilizes a multi-stage processing method of separation-detection-initial identification-re-identification to make up for the shortcomings of traditional methods, improve the accuracy and reliability of transformer on-load tap-changer fault diagnosis and identification, and provide more reliable technical support for the normal operation of the tap-changer. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flowchart of a method implementation according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0041] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present application belongs.

[0042] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0043] like Figure 1 As shown, this embodiment provides a transformer on-load tap changer fault diagnosis method based on background sound texture, comprising the following steps:

[0044] (1) Preprocess the original audio data. The specific method is:

[0045] First, the collected original audio data is divided into frames according to the frame length of 15.6s and the frame shift of 10.4ms, and the data after frame division is obtained. i (n), where i represents the frame index of the signal and n represents the number of discrete time points.

[0046] The peak normalization is performed on the framed data so that the data are distributed between (-1, 1).

[0047] The framed data s i (n) Perform short-time Fourier transform. The specific process is as follows:

[0048]

[0049] Where S(n,f) represents the time domain signal s i (m), w(nm) represents the short-time Fourier transform of the window function w(m) after mirroring and shifting it to the right by n discrete time points, M = 1024 is the window length, j represents the imaginary unit, and f is the instantaneous frequency of the signal.

[0050] Take the base 10 logarithm of S(n,f) and normalize it to get the spectrum S D (n,f)=log 10 S(n,f).

[0051] (2) Use a deep neural network to obtain the time-frequency mask of the device sound, and use the time-frequency mask to separate the pure device sound. The specific method is:

[0052] A training dataset was constructed using the sound signals of the transformer on-load tapchanger switching phase with high signal-to-noise ratio and background sound signals collected in the early stage. The dataset was then input into a deep neural network for training. The actual on-site audio collected was then input into the deep neural network and separated to obtain the time-frequency mask of the equipment sound. The time-frequency mask of the equipment sound was then used to separate the pure equipment sound.

[0053] The training dataset is constructed by using the previously collected high signal-to-noise ratio transformer on-load tapchanger switching phase sound signal and background sound signal to calculate the corresponding ideal ratio time-frequency mask (IRM). The time-frequency mask is calculated as follows:

[0054]

[0055] Where N D (n,f) represents the spectrogram of the background sound signal.

[0056] The time-frequency mask of the separated device sound is multiplied by the amplitude spectrum of the short-time Fourier transform S(n,f) of the original signal to obtain the amplitude spectrum of the signal after background sound separation, and the phase of S(n,f) is used to obtain the estimated device sound spectrum S P (n,f), that is:

[0057] S P (n,f)=|S(n,f)|·IRM(n,f)·e j∠S(n,f)

[0058] Where |S(n,f)| represents the amplitude spectrum of the signal, and ∠S(n,f) represents the phase of the signal.

[0059] Device sound spectrum S P (n,f) performs inverse Fourier transform to obtain the pure equipment sound signal s P (n).

[0060] (3) Use the initial analysis model to pre-judge the fault status of the separated sound activities.

[0061] Specifically, the initial analysis model based on deep convolutional neural network is used to judge the separated device sound spectrum S P Whether (n,f) is in a fault state, the network structure is as follows: two-dimensional input layer, convolution layer, normalization layer, maximum pooling layer, convolution layer, normalization layer, maximum pooling layer, convolution layer, normalization layer, maximum pooling layer, convolution layer, normalization layer, maximum pooling layer, fully connected layer, sigmoid output layer; if the pre-judgment result is a fault, subsequent processing is performed.

[0062] (4) Perform audio event start-end endpoint detection on the audio data area judged to be in a fault state and extract the audio event fragment. Specifically, the following steps are included:

[0063] 1) Use short-time Fourier transform to transform the sound spectrum data S that is judged as the fault type P (n,f) is restored to the purified sound waveform data s P (n).

[0064] 2) Use Hilbert transform to calculate the envelope of the sound waveform data to be recognized

[0065]

[0066] Where * represents the convolution operation.

[0067] 3) According to the envelope contour Set a high threshold T2 and the first intersection point N1 between the envelope and T2. Determine a low threshold T1 based on the energy of the background noise. Start searching from before point N1 and find point N2 that intersects with T1. Use N2 as the starting point of the device's sound activity and use N3 = N2 + 5.2 * Fs as the end point of the sound activity, where Fs is the sampling rate of the sound signal.

[0068] 4) The purified sound waveform data s P (n) is intercepted with the starting point N2 and the ending point N3 to obtain the reanalysis data s R (n).

[0069] (5) Use the reanalysis model to accurately identify the detected audio event fragments and obtain the type of fault.

[0070] Specifically, the reanalysis model is used to analyze the detected reanalysis data s R(n) to accurately identify and obtain the type of fault. The reanalysis model uses a deep neural network for fault identification. The network type is a convolutional neural network. The network structure is as follows: two-dimensional input layer, convolution layer, convolution layer, fully connected layer, activation function layer, fully connected layer, activation function layer, fully connected layer, activation function layer, fully connected layer, and Sigmoid output layer. The label type corresponding to the maximum probability of the output layer is selected as the fault type of the equipment.

[0071] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other manner. Any person skilled in the art may utilize the above-disclosed technical content to modify or modify the present invention into equivalent embodiments. However, any simple modifications, equivalent variations, and modifications to the above embodiments that do not depart from the technical content of the present invention and are based on the technical essence of the present invention remain within the scope of protection of the present invention.

Claims

1. A transformer on-load tap changer fault diagnosis method based on background sound texture, characterized in that: The following steps are involved: Preprocess the raw audio data; Use a deep neural network to obtain the time-frequency mask of the device sound, and use the time-frequency mask to separate the pure device sound; Use the initial analysis model to pre-judge the fault status of the separated sound activities; Performing start-end endpoint detection of an audio event on the audio data area judged to be in a fault state, and extracting an audio event segment; Use the reanalysis model to accurately identify the detected audio event fragments and obtain the type of fault; A training dataset was constructed using previously collected high signal-to-noise ratio transformer on-load tapchanger switching phase sound signals and background sound signals. This was then fed into a deep neural network for training. Actual on-site audio was then fed into the deep neural network and separated to obtain the time-frequency mask of the equipment sound. This mask was then used to isolate the pure equipment sound. The training dataset is constructed by using the previously collected high signal-to-noise ratio transformer on-load tapchanger switching phase sound signal and background sound signal to calculate the corresponding ideal ratio time-frequency mask (IRM). The time-frequency mask is calculated as follows: Where, is the instantaneous frequency of the signal, a spectrogram representing a background sound signal; The time-frequency mask of the separated equipment sound is short-time Fourier transformed with the original signal The amplitude spectrum of the background sound separation signal is obtained by multiplying the amplitude spectrum of the background sound separation signal, and using The phase of the device sound spectrum is estimated ,Right now: Where, represents the amplitude spectrum of the signal, Indicates the phase of the signal; Device sound spectrum Perform inverse Fourier transform to obtain pure sound waveform data ; Performing start-end endpoint detection of an audio event on an audio data area judged to be in a fault state and extracting an audio event segment specifically includes the following steps: 1) Use short-time Fourier transform to analyze the sound spectrum of the device that is judged to be in a fault state Restore to purified sound waveform data ; 2) Use Hilbert transform to calculate the envelope of the sound waveform data to be recognized ; Where, Represents the convolution operation; 3) According to the envelope contour Set a high threshold , the envelope and The first intersection of ; Determine a low threshold based on the energy of the background noise ;from Start searching before clicking and find Intersection point ;Will As the starting point of the device sound activity, As the end point of the sound activity, is the sampling rate of the sound signal; 4) Purified sound waveform data According to the starting point , the end point is Intercept and obtain reanalysis data .

2. The transformer on-load tap changer fault diagnosis method based on background sound texture according to claim 1 is characterized in that: The specific method for preprocessing raw audio data is: First, the collected original audio data is divided into frames according to the set frame length and frame shift to obtain the framed data. ,in Indicates the frame subscript of the signal, Represents the number of discrete time points; Perform peak normalization on the framed data so that the data are distributed between (-1, 1); Framed data Perform short-time Fourier transform. The specific method is: Where, Represents a time domain signal The short-time Fourier transform of Represents the window function After mirror flipping and moving right by n discrete time points, For the window length, represents the imaginary unit, is the instantaneous frequency of the signal; right Take the logarithm with base 10 and normalize it to get the spectrum .

3. The transformer on-load tap-changer fault diagnosis method based on background sound texture according to claim 1 is characterized in that: Use the initial analysis model based on deep convolutional neural network to judge the separated device sound spectrum Whether it is a fault state, the deep convolutional neural network structure is: two-dimensional input layer, convolution layer, normalization layer, maximum pooling layer, convolution layer, normalization layer, maximum pooling layer, convolution layer, normalization layer, maximum pooling layer, convolution layer, normalization layer, maximum pooling layer, fully connected layer, sigmoid output layer; if the pre-judgment result is a fault, subsequent processing is performed.

4. The transformer on-load tap changer fault diagnosis method based on background sound texture according to claim 1 is characterized in that: The reanalysis data detected using the reanalysis model Accurately identify the fault type and obtain it. The reanalysis model uses a deep neural network for fault identification. The network type is a convolutional neural network. The network structure is as follows: two-dimensional input layer, convolution layer, convolution layer, fully connected layer, activation function layer, fully connected layer, activation function layer, fully connected layer, activation function layer, fully connected layer, and Sigmoid output layer. The label type with the maximum probability corresponding to the output layer is selected as the fault type of the equipment.

Images