A Partial Discharge Detection Method for an Airborne Acoustic Camera

Abstract

The present invention discloses a partial discharge detection method for an airborne acoustic camera, which includes the following steps: deploying an acoustic camera on a drone to collect and locate ultrasonic waves generated by partial discharge of electrical equipment; converting the received ultrasonic wave signals into electrical signals by a microphone array, then separating the noise from the electrical signals to reduce noise, and transmitting the noise-reduced signals to a signal enhancement module; the signal enhancement module enhances the signal intensity in the direction of the ultrasonic wave signal source and reduces the noise of the acoustic wave signals incident from other angles; the feature recognition module receives the signals processed by the signal enhancement module and outputs a two-dimensional spectrogram image through the MFCC algorithm; and identifying different operating states of the electrical equipment by passing the MFCC features through a trained convolutional neural network according to the voiceprint and vibration characteristics of the ultrasonic wave signals. The present invention uses ultrasonic sensors to detect partial discharge of power equipment, avoiding electromagnetic interference during the operation of power equipment, and can achieve better detection effects.

Description

Technical Field

[0001] The invention belongs to the technical field of power system detection, and in particular relates to a partial discharge detection method of an airborne acoustic camera. Background Art

[0002] With the development of industry and social progress, the power system is developing towards large capacity, ultra-high voltage and ultra-high voltage, and the requirements for system operation reliability are getting higher and higher. Power equipment is the basic component of the power system, and its working condition is directly related to the safe and economical operation of the power system. The insulating materials of electrical equipment are mostly organic materials, such as mineral oil, insulating paper or various organic synthetic materials. The electric field borne by each area of ​​the insulator is generally uneven, and the dielectric itself is usually uneven. Some are composite insulators composed of different materials, such as gas-solid composite insulation, liquid-solid composite insulation and solid-solid composite insulation. Although some are single materials, some bubbles or other impurities will remain during the manufacturing or use process, so the electric field strength in some areas inside or on the surface of the insulator will be higher than the average electric field strength, or the breakdown field strength in some areas will be lower than the average breakdown field strength. Therefore, discharge will occur first in some areas, while other areas still maintain insulation characteristics, which forms partial discharge (abbreviated as partial discharge). Practice has proved that partial discharge is the main reason for the final insulation breakdown of high-voltage electrical equipment, so it is particularly important to monitor partial discharge of electrical equipment.

[0003] When high-voltage equipment discharges electricity, the air around it will be ionized, and during the ionization process, ultrasonic waves will be generated. Using this principle, the ultrasonic signals generated when high-voltage electrical equipment discharges can be detected, providing a more reliable basis for evaluating the operating status of the equipment. Practice has proved that ultrasonic detection technology can effectively and intuitively observe the discharge of high-voltage equipment, providing a new and powerful diagnostic method for live detection. Ultraviolet imaging technology and ultrasonic imaging technology are complementary, with completely different principles. Each has its own unique advantages, and the detection purpose and application method also have their own characteristics. It can be foreseen that the combined application of these two technologies will greatly enhance the comprehensive detection capability of high-voltage equipment fault points, and also provide a comprehensive means for the reliability research of high-voltage products.

[0004] The main difficulty of airborne acoustic partial discharge detection is that the huge wind noise when the drone is flying will cause significant interference to the acoustic equipment. The microphone is a highly sensitive measuring instrument. Under strong interference, it is difficult to accurately obtain other signals from the sound field. The vibration generated by the rotation of the drone rotor will be transmitted to the voiceprint collection unit through the gimbal, which will reduce the stability of the voiceprint camera. The above problems need to be solved to improve the availability of airborne acoustic equipment. Summary of the invention

[0005] The present invention provides a partial discharge detection method for an airborne acoustic camera, aiming to solve the problems of insufficient sensitivity and stability in ultrasonic partial discharge detection in the prior art.

[0006] To solve the above technical problems, the present invention proposes a partial discharge detection method for an airborne acoustic camera, including the following steps:

[0007] S1: Deploy an acoustic camera on a drone, control the drone to fly towards the electrical equipment to be measured, collect and locate the ultrasonic waves generated by the partial discharge of the electrical equipment. The acoustic camera includes a microphone array, a noise separation module, a signal enhancement module, and a feature recognition module for locating the ultrasonic wave source.

[0008] S2: The microphone array converts the received ultrasonic wave signal into an electrical signal, and then transmits the electrical signal to the signal processing module. The signal processing module separates the noise from the electrical signal, decomposes the separated signal subspace and noise subspace, judges the direction of the noise source through the orthogonal vector of the noise subspace and suppresses it to reduce the noise, and transmits the noise-reduced signal to the signal enhancement module.

[0009] S3: The signal enhancement module uses the signal superposition technology based on time delay, and through the array electronic steering technology based on the steering vector, enhances the signal intensity in the direction of the ultrasonic wave signal source. For the acoustic wave signals incident at other angles, noise reduction is performed through the wavelet packet noise reduction and reconstruction algorithm. The wavelet packet noise reduction can decompose the input signal into multiple levels, and all levels retain all the information of the input signal. The wavelet packet coefficients of each node after decomposition can be noise-reduced through dynamic threshold calculation. After noise reduction, the optimal wavelet packet basis is selected for signal reconstruction by examining the Shannon entropy of the parent node and the child node through the optimal wavelet packet basis algorithm. The selection criterion for the optimal wavelet packet basis is the entropy criterion.

[0010] S4: The feature recognition module receives the signal processed by the signal enhancement module and outputs a two-dimensional time-frequency spectrum image through the MFCC algorithm. The specific method for feature recognition by the MFCC algorithm includes:

[0011] S4-1: Perform pre-emphasis, framing, and windowing operations on the input signal to obtain a short-time signal.

[0012] S4-2: Perform a fast Fourier transform on each short-time signal to obtain the corresponding linear spectrum.

[0013] S4-3: Take the square of the modulus of the linear spectrum to obtain a discrete power spectrum.

[0014] S4-4: Filter the obtained discrete power spectrum through a Mel filter bank, and then calculate the logarithmic energy of the output of the filter bank.

[0015] S4-5: Perform discrete cosine transform on the logarithmic energy to obtain MFCC features.

[0016] S5: Pass the MFCC features through a trained convolutional neural network to identify different operating states of the electrical equipment according to the voiceprint and vibration features of the ultrasonic signal.

[0017] Preferably, the convolutional neural network adopts a Chunk-based Attention module, which divides the input audio frames into Chunks of equal size, and only performs Attention operations within each Chunk for the frames within the Chunk.

[0018] Preferably, the convolutional neural network adopts an encoder-decoder-like structure to solve the problem of gradient disappearance during algorithm training through residual connections.

[0019] Preferably, the encoder-decoder-like structure is provided with a first downsampling module, a second downsampling module, a third downsampling module, a fourth downsampling module, a first one-dimensional convolutional module, a first upsampling and downsampling module, a second upsampling and downsampling module, a third upsampling and downsampling module, a fourth upsampling and downsampling module, and a second one-dimensional convolutional module in the processing order.

[0020] Preferably, the pre-emphasis adopts time-domain technology or frequency-domain technology to enhance the high-frequency components of the signal to compensate for the excessive attenuation of the high-frequency components during transmission.

[0021] Preferably, the time-domain technology makes corresponding changes to the amplitude of the signal to be transmitted. When a bit signal is different from the previous bit signal transmitted, the amplitude of the current bit signal is increased by a set multiple. When a bit signal is the same as the previous bit signal, no processing is performed.

[0022] Preferably, the frequency-domain technology sets an active high-pass filter, and the active high-pass filter increases the energy of the high-frequency components of the signal to be transmitted to pre-compensate for the attenuation of the high-frequency components of the signal by the transmission line.

[0023] Preferably, the wavelet packet denoising divides the frequency band of the input signal into multiple levels, and each decomposed frequency band is further halved. Then, the corresponding frequency band is selected according to the characteristics of the signal to be analyzed to match the signal spectrum. The wavelet packet denoising adopts a four-layer structure.

[0024] Preferably, the windowing method is preferably the Hamming window, and its mathematical formula is as follows:

[0025]

[0026] Wherein, N is the window length of the Hamming window, and n is the step size; the window length is preferably 25 ms, and the step size is preferably 10 ms.

[0027] Compared with the prior art, the present invention has the following technical effects:

[0028] 1. The detection method proposed by the present invention uses an ultrasonic sensor to detect partial discharge on the outer shell part of the power equipment, avoiding strong electromagnetic interference during the operation of the power equipment, and better detection effects can be obtained.

[0029] 2. The ultrasonic partial discharge detection proposed by the present invention can be widely applied to various primary equipment in the power system. According to different ultrasonic signal propagation paths, the unmanned aerial vehicle equipped with an acoustic camera can fly close to the power equipment that is inconvenient for personnel to approach for detection, expanding the applicable range of acoustic detection.

[0030] 3. The detection method proposed by the present invention denoises the signal through wavelet packet denoising and reconstruction algorithms, retains the characteristics of the signal in each frequency band, and has great flexibility during signal reconstruction, which can effectively improve the signal-to-noise ratio of the signal and highlight the main features.

[0031] 4. The detection method proposed by the present invention realizes the effect of enhancing the sound source in any direction through the array electronic steering technology of the steering vector, separates the noise sound source, so as to better locate the sound source under high interference conditions; aiming at the characteristic that the partial discharge signal is strongly correlated with the power frequency period, the signal-to-noise ratio of the partial discharge signal is improved, thereby improving the sensitivity of partial discharge detection.

[0032] 5. The neural network model adopted by the detection method proposed by the present invention introduces background noise during training for sample enhancement, expands the samples by introducing white noise within the range of 10 dB of the partial discharge noise magnitude, and improves the generalization ability of the model. Description of the Drawings

[0033] Figure 1 is the flowchart of a partial discharge detection method for an airborne acoustic camera according to the present invention;

[0034] Figure 2 is the processing flowchart of the feature recognition module of a partial discharge detection method for an airborne acoustic camera according to the present invention;

[0035] Figure 3 is the structure and processing flowchart of a partial discharge detection method for an airborne acoustic camera according to the present invention. Detailed Embodiments

[0036] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present application and with reference to the accompanying drawings.

[0037] Please refer to Figure 1 , a partial discharge detection method for an airborne acoustic camera, comprising the following steps:

[0038] S1: Deploy an acoustic camera on a drone, control the drone to fly towards the electrical equipment to be tested, collect and locate the ultrasonic waves generated by the partial discharge of the electrical equipment. The acoustic camera includes a microphone array, a noise separation module, a signal enhancement module, and a feature recognition module for locating the ultrasonic wave source.

[0039] S2: The microphone array converts the received ultrasonic wave signal into an electrical signal, and then transmits the electrical signal to the signal processing module. The signal processing module separates the noise from the electrical signal, decomposes the separated signal subspace and noise subspace, determines the direction of the noise source through the orthogonal vector of the noise subspace and suppresses it to reduce the noise, and transmits the noise-reduced signal to the signal enhancement module.

[0040] S3: The signal enhancement module uses the signal superposition technology based on time delay and enhances the signal intensity in the direction of the ultrasonic wave signal source through the array electronic steering technology based on the steering vector. For the acoustic wave signals incident at other angles, noise reduction is performed through the wavelet packet noise reduction and reconstruction algorithm. The wavelet packet noise reduction can decompose the input signal at multiple levels, and all levels retain all the information of the input signal. The wavelet packet coefficients of each node after decomposition can be noise-reduced through dynamic threshold calculation. After noise reduction, the optimal wavelet packet basis is selected for signal reconstruction by examining the Shannon entropy of the parent node and the child node through the optimal wavelet packet basis algorithm. The selection criterion for the optimal wavelet packet basis is the entropy criterion.

[0041] S4: The feature recognition module receives the signal processed by the signal enhancement module and outputs a two-dimensional spectrogram through the MFCC algorithm. The specific method for feature recognition by the MFCC algorithm includes:

[0042] S4-1: Perform pre-emphasis, framing, and windowing operations on the input signal to obtain a short-time signal.

[0043] S4-2: Perform a fast Fourier transform on each short-time signal to obtain the corresponding linear spectrum.

[0044] S4-3: Take the square of the modulus of the linear spectrum to obtain a discrete power spectrum.

[0045] S4-4: Filter the obtained discrete power spectrum through a Mel filter bank, and then calculate the logarithmic energy of the output of the filter bank.

[0046] S4-5: Perform a discrete cosine transform on the logarithmic energy to obtain MFCC features.

[0047] S5: Identify different operating states of the electrical equipment based on the MFCC features through the trained convolutional neural network according to the voiceprint and vibration features of the ultrasonic signal.

[0048] During the training of the convolutional neural network, background noise is introduced for sample enhancement. White noise within the range of 10 dB of the partial discharge noise magnitude is introduced to augment the samples, improving the generalization ability of the model. The convolutional neural network adopts a Chunk-based Attention module, which divides the input audio frames into equal-sized Chunks, and the frames within each Chunk only perform Attention operations within the Chunk. Traditional Attention algorithms rely on all future data and cannot be applied to streaming audio decoding. Therefore, the Chunk-based Attention module is adopted in this embodiment. Additionally, it is also allowed to perform Attention with a certain length of frames on the left side.

[0049] The noise separation technology is a very effective signal processing means in various scenarios. In the calculation process of sound source localization, a large number of matrix operations are involved. Utilizing the properties of matrices, the signal subspace and the noise subspace can be decomposed, and the sound source direction can be estimated through the orthogonal vectors of the noise subspace, ultimately achieving the effect of reducing irrelevant noise.

[0050] In a microphone array including M microphones, the sound source wave equation at the microphone is as follows:

[0051]

[0052] where d is the distance between every two adjacent microphones in the microphone array, m is the serial number of each microphone, and θ k is the angle between the sound source and the vertical direction of the microphone array. Therefore, at the m-th microphone, the superposition of the waves of all sound sources can be expressed as:

[0053]

[0054] From the geometric relationship, the theoretical time delay can be obtained by the following formula:

[0055]

[0056] After introducing an artificial noise term into the equation, the wave equation of the microphone array can be expressed in matrix form, where the last term is the introduced noise term:

[0057]

[0058] Here, it can be expressed in a concise matrix form:

[0059] X(n) = S(n) + V(n)

[0060] Where X is the vibration data received by the microphone, A is the steering vector, S is the sound source fluctuation model, and V is the noise.

[0061] X is the observed data, which contains other noises in the scene. It should be assumed that the noises of each microphone element are independent of each other and satisfy a Gaussian distribution with a variance of Introduce the covariance matrix R

[0062] R = E[X(n)X H (n)

[0063] Through theoretical derivation:

[0064]

[0065] Where:

[0066] R s = [S(n)S H (n)

[0067] Perform eigenvalue decomposition on R and calculate according to the properties of eigenvalue decomposition:

[0068] R = UΛU H

[0069] UU H = H U = I

[0070] Here, U is a unitary matrix composed of eigenvectors and has orthogonal properties. Λ is a diagonal matrix composed of matrix eigenvalues. According to the properties of the eigenvalue diagonal matrix, substitute the covariance matrix:

[0071]

[0072] Since Rs is a K-dimensional matrix and has at most K eigenvalues, when the number of microphones M exceeds K, the matrix is not a full-rank matrix, and the subsequent diagonal elements must be zero. After introducing the noise term, the expression form of the eigenvalue matrix changes to:

[0073]

[0074] Based on the basic assumption that the intensity of the noise space is less than that of the sound source to be monitored, there must be value jumps in the diagonal elements of the diagonal matrix, from which the signal source components and noise components in the eigenvalue matrix can be determined. Then, by using the properties of the block matrix, the coupled sound sources in the scene are decomposed into the signal subspace and the noise subspace. Through further research on the noise of urban substations, the distribution of signals can be further modeled to further improve the decomposition accuracy of the noise subspace.

[0075] The signal enhancement module enhances the signal intensity in the direction of the ultrasonic signal source through the array electronic steering technology based on the steering vector. For the acoustic wave signals incident at other angles, noise reduction is performed through the wavelet packet noise reduction and reconstruction algorithm. The principle of the microphone array for the sound source gain is based on the signal superposition technology of time delay. Through the array electronic steering technology based on the steering vector, the effect of enhancing the sound source in any direction can be achieved, thereby improving the sensitivity of partial discharge detection.

[0076] The steering vector describes the time delay of the signals incident on the microphone array from various angles. Based on the assumption of the far-field sound field model, when the sound wave arrives at the array plane perpendicular to the direction of the microphone array, theoretically, all microphones receive the sound wave with the same phase. After superimposing the signals received by all microphones, the positions of the wave peaks and wave valleys will obtain a certain degree of gain, while the random noise part will be suppressed, realizing the positive signal gain through the microphone array.

[0077] For the sound waves incident from other angles, the steering vector can be used to calculate the phase delay based on the time delay between the current microphone and the arrival time of the plane wave. The signal is denoised through the wavelet packet noise reduction and reconstruction algorithm. The wavelet packet can decompose the original signal into multiple levels, and all levels retain all the information of the signal. The wavelet packet coefficients of each node after decomposition can be denoised through dynamic threshold calculation. After denoising, through the optimal wavelet packet basis algorithm, that is, by examining the Shannon entropy of the parent node and the child node, the optimal wavelet packet basis is selected for signal reconstruction. This method retains the characteristics of the signal in each frequency band and has great flexibility in signal reconstruction, which can effectively improve the signal-to-noise ratio of the signal and highlight the main features. The wavelet packet noise reduction divides the frequency band of the input signal into multiple levels, and each decomposed frequency band is further halved, and the corresponding frequency band is selected according to the characteristics of the signal to be analyzed to match the signal spectrum. The wavelet packet noise reduction adopts a four-layer structure.

[0078] Although the unenhanced signal can distinguish the partial discharge signals related to the power frequency cycle, there is obviously a large amount of abnormal noise; after the signal is enhanced, the partial discharge characteristics are significantly strengthened.

[0079] Next, the feature recognition module receives the signal processed by the signal enhancement module and outputs a two-dimensional spectrogram image through the MFCC algorithm. The microphone array uses the two-dimensional spectrogram image features of the signals output by the traditional voiceprint and vibration signal analysis MFCC / VQ algorithms as the input of the deep neural network CNN network for deep learning, constructs an MFCC-CNN recognition model, and then realizes the extraction of voiceprint and vibration features and the pattern recognition of different operating states.

[0080] MFCC is the representation of the short-term power spectrum of sound and is the frequency reciprocal coefficient of Mel parameters. It is also widely used in the fields of anomaly detection, speech recognition, and voiceprint recognition. It is one of the most widely used speech features, and its feature extraction process is as Figure 2 shown.

[0081] Perform preprocessing operations such as pre-emphasis, framing, and windowing on the original speech to obtain a short-time signal.

[0082] Pre-emphasis increases the high-frequency energy of the effective speech sequence, improves the signal-to-noise ratio, and obtains a pre-emphasized speech sequence. Speech information is often mixed with various other sound information in the environment. Due to the characteristics of human pronunciation, after frequency conversion, most of the speech information is concentrated in the low-frequency band, resulting in too high low-frequency energy and too low high-frequency energy, making it difficult to effectively extract high-frequency speech information. Pre-emphasis adds high-frequency signals in advance. After superimposing with the original speech signal, the energy of the high-frequency band and the low-frequency band is equivalent, which significantly improves the subsequent recognition efficiency.

[0083] Framing and windowing segment the pre-emphasized speech sequence at a set time interval, and then use a band-pass filter to filter the signal to reduce the signal error and obtain a frame sequence dependent on time. A whole signal is unstable, but locally, the speech signal can be assumed to be short-term stationary (it can be considered that the speech signal is approximately unchanged within 10 - 30 ms for the pronunciation of a phoneme, and generally 25 ms is taken). Therefore, it is necessary to perform framing processing on the entire speech signal. In this embodiment, the Hamming window is used for windowing. Since the Hamming window is added, only the middle data is reflected, and the data information on both sides is lost. Therefore, there is an overlapping part between adjacent windows. The window length of this embodiment is 25 ms, and the step size is 10 ms, that is, the last 15 ms of each window and the first 15 ms of the subsequent adjacent window are the overlapping parts. The mathematical formula of the Hamming window is as follows:

[0084]

[0085] where N is the window length of the Hamming window and n is the step size.

[0086] Perform a fast Fourier transform (FFT) on each short-time signal to obtain the corresponding linear spectrum. The fast Fourier transform converts the frame sequence from the time-domain graph to the spectrum of each frame. The characteristics of the speech signal in the time domain are not obvious, so it is usually converted to the energy distribution in the frequency domain. Perform a fast Fourier transform on the signal processed by the window function for each frame to convert the time-domain graph into the spectrum of each frame, and then we can superimpose the spectra of each window to obtain the linear spectrum.

[0087] Take the square of the modulus of the linear spectrum to obtain the discrete power spectrum; filter the obtained discrete power spectrum through a Mel filter bank, and then calculate the logarithmic energy of the output of the filter bank; perform a discrete cosine transform (DCT) on the logarithmic energy to obtain the MFCC. This transformation formula can be simplified as:

[0088]

[0089] In the formula, C n represents the coefficient of the MFCC; L represents the order of the MFCC.

[0090] After obtaining the MFCC features, the device judges the certain change trends presented by the voiceprint and vibration features under the excitation of different voltage levels, harmonic orders or DC contents, realizes classifying the data under the same working condition into one category and performing unified labeling, and performs learning on the parameter-free feature quantities through a convolutional neural network. At the same time, apply the network structure for extracting the voiceprint and vibration features of the processing device, and construct a typical defect / fault vibration acoustic feature library of the device to realize the adaptive detection of abnormal signals and improve the application ability of the algorithm.

[0091] In this embodiment, the convolutional neural network adopts a class encoder-decoder structure to solve the problem of gradient disappearance during algorithm training through residual connections. The class encoder-decoder structure is as Figure 3 shown. There are a first downsampling module, a second downsampling module, a third downsampling module, a fourth downsampling module, a first one-dimensional convolutional module, a first upsampling and downsampling module, a second upsampling and downsampling module, a third upsampling and downsampling module, a fourth upsampling and downsampling module, and a second one-dimensional convolutional module arranged in the processing order; each module receives the output of the previous-level module. At the same time, the first upsampling and downsampling module receives the output of the fourth downsampling module, the second upsampling and downsampling module receives the output of the third downsampling module, the third upsampling and downsampling module receives the output of the fourth downsampling module, and the fourth upsampling and downsampling module receives the output of the first downsampling module.

[0092] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the inventive concept of the present invention, several modifications and improvements can be made, and these all belong to the protection scope of the present invention.

Claims

1. A method for detecting partial discharge of an airborne acoustic camera, characterized in that, It includes the following steps: S1: Deploy an acoustic camera on the UAV, control the UAV to fly towards the electrical equipment to be measured, collect and locate the ultrasonic waves generated by the partial discharge of the electrical equipment. The acoustic camera includes a microphone array, a noise separation module, a signal enhancement module, and a feature recognition module, and is used to locate the ultrasonic wave source; S2: The microphone array converts the received ultrasonic wave signal into an electrical signal, and then transmits the electrical signal to the signal processing module. The signal processing module separates the noise from the electrical signal, decomposes the separated signal subspace and noise subspace, judges the direction of the noise source through the orthogonal vector of the noise subspace and suppresses it to reduce the noise, and transmits the noise-reduced signal to the signal enhancement module; S3: The signal enhancement module uses the method of signal superposition based on time delay, and through the array electronic steering technology based on the steering vector, enhances the signal intensity in the direction of the ultrasonic wave signal source. For the acoustic wave signals incident at other angles, noise reduction is performed through the wavelet packet noise reduction and reconstruction algorithm. The wavelet packet noise reduction decomposes the input signal into multiple levels, and all levels retain all the information of the input signal. The wavelet packet coefficients of each node after decomposition are denoised through dynamic threshold calculation. After denoising, the Shannon entropy of the parent node and the child node is examined through the optimal wavelet packet basis algorithm to select the optimal wavelet packet basis for signal reconstruction. The selection criterion for the optimal wavelet packet basis is the entropy criterion; S4: The feature recognition module receives the signal processed by the signal enhancement module and outputs a two-dimensional spectrogram image through the MFCC algorithm. The specific method for the MFCC algorithm to perform feature recognition includes: S4-1: Perform pre-emphasis, framing, and windowing operations on the input signal to obtain a short-time signal; S4-2: Perform a fast Fourier transform on each short-time signal to obtain the corresponding linear spectrum; S4-3: Take the square of the modulus of the linear spectrum to obtain a discrete power spectrum; S4-4: Filter the obtained discrete power spectrum through a Mel filter bank, and then calculate the logarithmic energy of the output of the filter bank; S4-5: Perform a discrete cosine transform on the logarithmic energy to obtain MFCC features; S5: Pass the MFCC features through a trained convolutional neural network, and identify different operating states of the electrical equipment according to the voiceprint and vibration characteristics of the ultrasonic wave signal.

2. The partial discharge detection method of an airborne acoustic camera according to claim 1, characterized in that The convolutional neural network adopts a Chunk-based Attention module, which divides the input audio frames into Chunks of equal size, and the frames within each Chunk only perform Attention operations within the Chunk.

3. The partial discharge detection method of an airborne acoustic camera according to claim 1, characterized in that The convolutional neural network adopts a class encoder-decoder structure to solve the problem of gradient disappearance during algorithm training through residual connections.

4. The partial discharge detection method of an airborne acoustic camera according to claim 3, characterized in that, The class encoder-decoder structure is provided with a first downsampling module, a second downsampling module, a third downsampling module, a fourth downsampling module, a first one-dimensional convolutional module, a first upsampling and downsampling module, a second upsampling and downsampling module, a third upsampling and downsampling module, a fourth upsampling and downsampling module, and a second one-dimensional convolutional module in the processing order.

5. The partial discharge detection method of an airborne acoustic camera according to claim 1, characterized in that, The pre-emphasis uses a time-domain method to make corresponding changes to the amplitude of the signal to be transmitted or a frequency-domain method to compensate for the attenuation of high-frequency components during transmission.

6. The partial discharge detection method of an airborne acoustic camera according to claim 5, characterized in that, The time-domain method makes corresponding changes to the amplitude of the signal to be transmitted. Specifically, when a bit signal is different from the previous bit signal transmitted, the amplitude of the current bit signal is increased by a set multiple. When a bit signal is the same as the previous bit signal, no processing is performed.

7. A partial discharge detection method for an airborne acoustic camera according to claim 5, characterized in that, The frequency-domain method pre-compensates for the attenuation of the high-frequency components of the signal by the transmission line by setting an active high-pass filter, which enhances the energy of the high-frequency components of the signal to be transmitted.

8. The partial discharge detection method of an airborne acoustic camera according to claim 1, characterized in that, The wavelet packet denoising method divides the frequency band of the input signal into a four-layer structure according to a setting, divides each divided frequency band in half again, and selects the corresponding frequency band according to the characteristics of the signal to be analyzed to match the signal spectrum.

9. The partial discharge detection method of an airborne acoustic camera according to claim 1, characterized in that, The windowing method is a Hamming window, and its mathematical formula is as follows: Among them, is the window length of the Hamming window, is the step size; the window length is 25 ms and the step size is 10 ms.

Images