GIS equipment fault diagnosis method and device, medium and equipment

Abstract

The invention discloses a GIS equipment fault diagnosis method and device, a medium and equipment. According to the method, the preset acoustic feature vectors and the vibration feature vectors are obtained, the feature vectors are converted into the acoustic atlas and the vibration atlas with the time-frequency characteristics through the specific signal reconstruction technology, and then the acoustic atlas and the vibration atlas are input into the pre-trained deep learning model. According to the model, a multi-head self-attention mechanism is used for constructing an incidence matrix between features, key information can be extracted from an acoustic atlas and a vibration atlas, and then the fault type of GIS equipment is accurately output. The training process of the deep learning model relates to transfer learning, and an initial model is trained by using a historical fault feature database, so that the adaptability of the model to new data and the diagnosis accuracy are improved.

Description

Technical Field

[0001] The present invention relates to the field of fault diagnosis of GIS equipment, and particularly to a fault diagnosis method, device, medium and equipment for GIS equipment. Background Art

[0002] Gas Insulated Switchgear (GIS equipment) has been widely used in high-voltage power systems and plays a crucial role due to its advantages such as easy installation, small floor area, overall modularization, safe and reliable operation, etc. However, during the transportation, production, installation and long-term operation of GIS equipment, defects may be formed inside the equipment due to various uncontrollable factors (such as transportation collision, mechanical wear, non-standard installation, equipment aging, etc.), which may further lead to faults such as equipment breakdown. If these hidden dangers cannot be discovered and eliminated in time, it will not only endanger the stable operation of the power system, but also may cause huge losses to people's lives and property.

[0003] At present, the fault detection means of GIS equipment mainly rely on vibration and noise acquisition instruments, which can collect the acoustic vibration signal waveforms of GIS equipment and calculate their time-frequency characteristic parameters and sound pressure signals. However, the existing detection technologies have the following deficiencies:

[0004] High misjudgment rate: The existing technology only relies on the characteristic values of sound signals for fault judgment, and the selection of characteristic values is relatively single, which is prone to misjudgment. For example, when analyzing only through sound pressure or time-frequency characteristics, it is difficult to comprehensively reflect the actual operating state of the equipment, especially in complex working conditions, and the misjudgment rate is relatively high.

[0005] Lack of effective signal processing and analysis methods: A large amount of detection data collected on site lacks effective sorting and mining means. The existing analysis methods usually can only perform simple waveform acquisition and basic characteristic calculation on acoustic vibration signals, and cannot deeply analyze the complex characteristics and potential fault information in the signals.

[0006] Unable to accurately locate the fault position: The existing technology processes sound signals and vibration signals separately, and it is difficult to determine the specific position where the fault occurs. This not only slows down the work progress of operation and maintenance and repair, but also increases the repair cost and workload.

[0007] High requirement for sensor accuracy: In the existing technology, in order to obtain sufficient fault characteristic information, high-precision sensors are usually required to collect signals in the ultrasonic range. This not only increases the equipment cost, but also poses higher requirements for the installation and maintenance of sensors, which is not conducive to the actual operation of front-line operation and maintenance personnel.

[0008] These deficiencies result in the inability of the existing technology to accurately diagnose faults in GIS equipment. Summary of the Invention

[0009] The present invention provides a fault diagnosis method, device, medium and equipment for GIS equipment, so as to solve the problem that the fault diagnosis of GIS equipment cannot be accurately carried out in the prior art.

[0010] In a first aspect, the present application provides a fault diagnosis method for GIS equipment, including:

[0011] Obtain a preset acoustic feature vector and vibration feature vector;

[0012] According to a preset signal reconstruction method, convert the acoustic feature vector into an acoustic spectrogram with time-frequency characteristics, and convert the vibration feature vector into a vibration spectrogram with time-frequency characteristics;

[0013] Obtain the fault type of the GIS equipment according to the acoustic spectrogram, vibration spectrogram and a preset deep learning model;

[0014] Wherein, the preset deep learning model constructs a feature correlation matrix based on a multi-head self-attention mechanism, and performs transfer learning training on the initial deep learning model according to a historical fault feature database.

[0015] The present application obtains an acoustic feature vector and a vibration feature vector, and uses a signal reconstruction method to convert them into an acoustic spectrogram and a vibration spectrogram with time-frequency characteristics, providing rich multi-modal input data for the deep learning model. This multi-modal data fusion method can more comprehensively reflect the operating state of the GIS equipment, thereby improving the accuracy of fault diagnosis. Further, the preset deep learning model constructs a feature correlation matrix based on a multi-head self-attention mechanism, which can automatically learn the internal correlation relationship between the acoustic mode and the vibration mode, and strengthen the robustness of feature extraction. Combining transfer learning training with a historical fault feature database enables the model to quickly adapt to new fault modes and significantly improve the recognition ability of unknown fault types. The present application effectively solves the problem that the fault diagnosis of GIS equipment cannot be accurately carried out in the prior art.

[0016] As a preferred embodiment of the first aspect, the obtaining of the preset acoustic feature vector and vibration feature vector is specifically:

[0017] Collect acoustic and vibration signals according to a preset acoustic sensor array and a preset vibration sensor;

[0018] Extract features from the acoustic and vibration signals according to a preset blind source separation method to obtain each independent component of the acoustic and vibration signals;

[0019] Calculate the fuzzy entropy values of each independent component, and select the minimum entropy feature from the fuzzy entropy values as the preliminary feature vector;

[0020] Perform multi-scale time window segmentation on the preliminary feature vectors to obtain acoustic feature vectors and vibration feature vectors with noise resistance ability.

[0021] In this preferred embodiment, the present application collects acoustic and vibration signals through a preset acoustic sensor array and vibration sensors, and uses the blind source separation method to extract features from the signals to obtain multiple independent components. By calculating the fuzzy entropy values of these components and screening out the minimum entropy features, preliminary feature vectors are formed. Further, multi-scale time window segmentation is performed on the preliminary feature vectors to generate acoustic feature vectors and vibration feature vectors with noise resistance ability. This process not only effectively extracts the key features in the acoustic and vibration signals, but also enhances the noise resistance and robustness of the features through multi-scale analysis. Therefore, the present invention can significantly improve the accuracy and reliability of feature extraction, provide high-quality input data for subsequent fault diagnosis, and thus improve the performance and accuracy of the overall diagnosis system.

[0022] As a preferred embodiment of the first aspect, after collecting the acoustic and vibration signals according to the preset acoustic sensor array and preset vibration sensors, it further includes:

[0023] According to the beamforming principle, obtain the respective beamforming outputs received by multiple acoustic sensor arrays;

[0024] According to the respective beamforming outputs and a preset formula, calculate the cross-power spectra of each beam;

[0025] Calculate the maximum value points of the cross-power spectra;

[0026] According to the maximum value points and the spatial coordinates of the multiple acoustic sensor arrays, determine the specific location where the fault occurs.

[0027] The calculating the cross-power spectra of each beam according to the respective beamforming outputs and a preset formula is specifically:

[0028] The preset formula is:

[0029]

[0030] Wherein, t is the relative delay of the acoustic signal transmitted to any sensor and the acoustic signal transmitted to the standard sensor; f is the angular frequency of the acoustic signal; N represents the total number of sensors; C ab is the cross-correlation matrix between sensor a and sensor b; j is the imaginary unit;

[0031] According to the respective beamforming outputs and a preset formula, calculate the cross-power spectra of each beam.

[0032] In this preferred embodiment, after collecting the acoustic vibration signal, the present application further utilizes the beamforming principle to process the output signal of the acoustic sensor array, calculates the mutual power spectrum of each beam, and accurately determines the specific location of the fault by analyzing the maximum points of the mutual power spectrum and combining the spatial coordinates of the sensor array. This process can not only effectively capture the propagation direction and time delay information of the acoustic signal, but also accurately locate the source of the fault through the maximum points of the mutual power spectrum. This method significantly improves the accuracy and reliability of fault location, reduces the misjudgment rate, and provides clear fault location information to operation and maintenance personnel, greatly improving maintenance efficiency and the safety of equipment operation.

[0033] As a preferred embodiment of the first aspect, the fault type of the GIS equipment is obtained according to the acoustic spectrum, the vibration spectrum and the preset deep learning model, specifically:

[0034] Inputting the acoustic spectrum and the vibration spectrum into a preset deep learning model, so that the deep learning model can calculate the soundprint features and vibration pattern features of typical faults of GIS equipment according to the extreme learning machine algorithm;

[0035] Compare the voiceprint features and vibration print features of the typical fault of the GIS equipment with the preset fuzzy entropy value;

[0036] If the absolute value of the difference between the preset fuzzy entropy and the voiceprint feature and the vibration print feature is less than or equal to the preset threshold, the fault type of the inspected GIS device is output.

[0037] In this preferred embodiment, the present application inputs the acoustic spectrum and vibration spectrum into a preset deep learning model, and uses the extreme learning machine (ELM) algorithm to calculate the soundprint features and vibration pattern features of typical faults of GIS equipment. By comparing with the preset fuzzy entropy value, when the absolute value of the difference is less than or equal to the preset threshold, the model outputs the fault type. This process not only makes full use of the multimodal information of acoustic and vibration signals, but also improves the accuracy of fault diagnosis through the precise matching of fuzzy entropy values. The setting of the preset threshold further ensures the reliability of the diagnostic results and reduces the misjudgment rate. Therefore, the present invention significantly improves the efficiency and accuracy of GIS equipment fault diagnosis and provides a strong guarantee for the stable operation of the power system.

[0038] As a preferred embodiment of the first aspect, the preset deep learning model is obtained by constructing a feature association matrix based on a multi-head self-attention mechanism and performing transfer learning training on the initial deep learning model according to a historical fault feature database, specifically:

[0039] Obtain a historical fault feature database;

[0040] Input the acoustic spectrogram and vibration spectrogram in the historical fault feature database into the initial deep learning model, so that the initial deep learning model extracts features from the acoustic spectrogram and vibration spectrogram to obtain a feature matrix;

[0041] Divide the feature matrix into image patches, and convert each image patch into serialized data;

[0042] Perform a non-linear transformation on the serialized data and introduce position information to generate a position embedding vector;

[0043] According to the multi-head self-attention mechanism, perform transfer training on the position embedding vector;

[0044] When the parameters of the transfer training reach a preset threshold, stop the training to obtain a preset deep learning model.

[0045] In this preferred embodiment, the present application constructs a preset deep learning model by using the historical fault feature database to perform transfer learning training on the initial deep learning model. The specific process includes: inputting the acoustic spectrogram and vibration spectrogram in the historical fault feature database into the initial model, extracting the feature matrix and dividing it into image patches, converting it into serialized data and then performing non-linear transformation and introducing position information to generate a position embedding vector. Subsequently, based on the multi-head self-attention mechanism, perform transfer training on the position embedding vector until the training parameters reach the preset threshold to complete the model training. This process not only makes full use of historical fault data, enhances the model's learning ability for different fault modes, but also improves the efficiency and accuracy of feature extraction through the multi-head self-attention mechanism. The finally obtained deep learning model can more accurately identify and diagnose the fault types of GIS devices, significantly improving the accuracy and reliability of fault diagnosis, and providing strong support for the efficient operation and maintenance of power systems.

[0046] In a second aspect, the present application provides a fault diagnosis device for a GIS device. The fault diagnosis device for the GIS device includes: an acquisition module, a reconstruction module, and a diagnosis module;

[0047] The acquisition module is used to acquire a preset acoustic feature vector and vibration feature vector;

[0048] The reconstruction module is used to convert the acoustic feature vector into an acoustic spectrogram with time-frequency characteristics and convert the vibration feature vector into a vibration spectrogram with time-frequency characteristics according to a preset signal reconstruction method;

[0049] The diagnosis module is used to obtain the fault type of the GIS device according to the acoustic spectrogram, vibration spectrogram, and a preset deep learning model;

[0050] Among them, the preset deep learning model constructs a feature correlation matrix based on the multi-head self-attention mechanism and is obtained by performing transfer learning training on the initial deep learning model according to the historical fault feature database.

[0051] This device uses three modules to divide the work and coordinate with each other to more accurately diagnose the faults of GIS equipment. This application obtains acoustic feature vectors and vibration feature vectors, and uses the signal reconstruction method to convert them into acoustic spectrograms and vibration spectrograms with time-frequency characteristics, providing rich multi-modal input data for the deep learning model. This multi-modal data fusion method can more comprehensively reflect the operating state of GIS equipment, thereby improving the accuracy of fault diagnosis. Further, the preset deep learning model constructs a feature correlation matrix based on the multi-head self-attention mechanism, which can automatically learn the internal correlation relationship between the acoustic mode and the vibration mode, and strengthen the robustness of feature extraction. Combining with the historical fault feature database for transfer learning training enables the model to quickly adapt to new fault modes and significantly improve the recognition ability for unknown fault types. This application effectively solves the problem that the prior art cannot accurately diagnose the faults of GIS equipment.

[0052] As a preferred embodiment of the second aspect, the acquisition module is used to acquire preset acoustic feature vectors and vibration feature vectors, specifically:

[0053] The acquisition module collects acoustic and vibration signals according to the preset acoustic sensor array and the preset vibration sensor;

[0054] According to the preset blind source separation method, perform feature extraction on the acoustic and vibration signals to obtain each independent component of the acoustic and vibration signals;

[0055] Calculate the fuzzy entropy values of each independent component, and select the minimum entropy feature from the fuzzy entropy values as the preliminary feature vector;

[0056] Perform multi-scale time window segmentation on the preliminary feature vector to obtain acoustic feature vectors and vibration feature vectors with anti-noise ability.

[0057] In this preferred embodiment, the present application collects acoustic and vibration signals through a preset acoustic sensor array and vibration sensors, and uses the blind source separation method to extract the features of the signals to obtain multiple independent components. By calculating the fuzzy entropy values of these components and screening out the minimum entropy features, a preliminary feature vector is formed. Further, the preliminary feature vector is segmented by a multi-scale time window to generate an acoustic feature vector and a vibration feature vector with noise resistance. This process not only effectively extracts the key features in the acoustic and vibration signals, but also enhances the noise resistance and robustness of the features through multi-scale analysis. Therefore, the present invention can significantly improve the accuracy and reliability of feature extraction, provide high-quality input data for subsequent fault diagnosis, and thus improve the performance and accuracy of the overall diagnosis system.

[0058] In a third aspect, the present application provides a computer-readable storage medium, which includes a stored computer program. When the computer program runs, it controls the device where the computer-readable storage medium is located to execute a fault diagnosis method for a GIS device as described above. Its beneficial effects are the same as those of the fault diagnosis method for a GIS device provided in the first aspect of the present application.

[0059] In a fourth aspect, the present application provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements any one of the fault diagnosis methods for a GIS device as described in the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 : A flowchart of an embodiment of the fault diagnosis method for a GIS device provided by the present application;

[0061] Figure 2 : A structural diagram of an embodiment of signal processing and fault diagnosis provided by the present application;

[0062] Figure 3 : A structural diagram of an embodiment of an acoustic feature spectrogram formed at the moment of mechanical fault occurrence of a GIS device provided by the present application;

[0063] Figure 4 : A structural diagram of an embodiment of a vibration feature spectrogram formed at the moment of mechanical fault occurrence of a GIS device provided by the present application;

[0064] Figure 5 : A structural diagram of an embodiment of the fault diagnosis device for a GIS device provided by the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0065] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0066] Embodiment 1

[0067] Please refer to Figure 1 , which is a fault diagnosis method for GIS equipment provided by an embodiment of the present invention.

[0068] In this embodiment, the process of the fault diagnosis method for GIS equipment in this application is described in detail through steps S01-S03.

[0069] The entire process of signal acquisition, signal processing and fault diagnosis in this application is as follows Figure 2 As shown in the figure; first, the multi-dimensional original signal is obtained from the acoustic and vibration sensor array through the ICA blind source separation technology. Then, the fuzzy entropy of these signals is calculated, and the component with the smallest fuzzy entropy is selected as the characteristic signal of the acoustic and vibration signals. Then, the multi-scale fuzzy entropy is calculated for these characteristic signals. Finally, the calculated multi-scale fuzzy entropy is compared with the soundprint and vibration print characteristics of the typical fault of the GIS equipment obtained by the ELM algorithm, and the absolute value of the difference is calculated. If one of these differences is less than or equal to the preset threshold of 0.02, it is determined that this type of fault has occurred in the GIS equipment. This process realizes the accurate diagnosis of GIS equipment faults by comprehensively analyzing the characteristics of acoustic and vibration signals.

[0070] S01: Obtain preset acoustic eigenvectors and vibration eigenvectors.

[0071] As a preferred embodiment of the first embodiment, the obtaining of the preset acoustic feature vector and the vibration feature vector is specifically: collecting acoustic and vibration signals according to a preset acoustic sensor array and a preset vibration sensor;

[0072] According to a preset blind source separation method, feature extraction is performed on the acoustic vibration signal to obtain each independent component of the acoustic vibration signal;

[0073] Calculating each fuzzy entropy value of the independent components, and selecting a minimum entropy feature from each fuzzy entropy value as a preliminary feature vector;

[0074] The preliminary feature vector is segmented into multi-scale time windows to obtain an acoustic feature vector and a vibration feature vector with noise resistance.

[0075] More specifically, the ICA blind source separation technology is adopted to perform blind source separation on the acoustic signal to obtain a multi-dimensional source signal x q (q = 1, 2,..., n, where n is a positive integer);

[0076] The ICA blind source separation technology is adopted to perform blind source separation on the vibration signal to obtain a multi-dimensional source signal y q (q = 1, 2,..., n, where n is a positive integer);

[0077] Calculate the fuzzy entropy of x q and y q respectively, and select the components with the minimum values as the characteristic signals respectively;

[0078] Based on the obtained characteristic signals, combined with Equation (1) and Equation (2), calculate the multi-scale fuzzy entropy of y q :

[0079]

[0080] where L is the time scale factor, that is, the multi-dimensional original signal is divided into L segments, each segment is n / L in length; 1 ≤ r ≤ n, and r is a positive integer; μ is the given embedding dimension, σ is the given similarity tolerance, and ξ is the given fuzzy function gradient;

[0081] Similarly, the multi-scale fuzzy entropy of x q can be obtained.

[0082] In this preferred embodiment, the present application collects acoustic and vibration signals through a preset acoustic sensor array and vibration sensors, and uses the blind source separation method to extract features of the signals to obtain multiple independent components. By calculating the fuzzy entropy values of these components and screening out the minimum entropy features, a preliminary feature vector is formed. Further, the preliminary feature vector is segmented by a multi-scale time window to generate an acoustic feature vector and a vibration feature vector with noise resistance. This process not only effectively extracts the key features in the acoustic and vibration signals, but also enhances the noise resistance and robustness of the features through multi-scale analysis. Therefore, the present invention can significantly improve the accuracy and reliability of feature extraction, provide high-quality input data for subsequent fault diagnosis, and thus improve the performance and accuracy of the overall diagnosis system.

[0083] As a preferred embodiment of Embodiment 1, after collecting the acoustic and vibration signals according to the preset acoustic sensor array and the preset vibration sensors, it further includes:

[0084] According to the beamforming principle, obtain the beamforming outputs received by multiple acoustic sensor arrays;

[0085] According to the beamforming outputs and the preset formula, calculate the cross-power spectra of each beam;

[0086] Calculate the maximum points of the cross-power spectrum;

[0087] Determine the specific location where the fault occurs based on the maximum points and the spatial coordinates of the multiple acoustic sensor arrays.

[0088] More specifically, obtain the outputs of beamforming received at multiple acoustic sensors through the beamforming principle;

[0089] Combining Equation (3) with the output of beamforming, the cross-power spectrum of the beam can be obtained:

[0090]

[0091] where t is the relative delay time of the acoustic signal transmitted to any sensor and the acoustic signal transmitted to the reference sensor; f is the angular frequency of the acoustic signal; N represents the total number of sensors; C ab is the cross-correlation matrix between sensor a and sensor b; j is the imaginary unit;

[0092] From the cross-power spectrum of the beam, it can be obtained that when t a equals the time when the beam actually propagates to the sensor in the propagation direction, it reaches the maximum value. Therefore, the specific location of the faulty GIS device can be deduced based on the propagation direction of the wave and the delay time.

[0093] In this preferred embodiment, after collecting the acoustic vibration signals, the present application further processes the output signals of the acoustic sensor array using the beamforming principle, calculates the cross-power spectra of the beams, and by analyzing the maximum points of the cross-power spectra and combining the spatial coordinates of the sensor array, accurately determines the specific location where the fault occurs. This process can not only effectively capture the propagation direction and time delay information of the acoustic signals, but also accurately locate the fault source through the maximum points of the cross-power spectra. This method significantly improves the accuracy and reliability of fault location, reduces the misjudgment rate, and at the same time provides clear fault location information for maintenance personnel, greatly improving the maintenance efficiency and the safety of equipment operation.

[0094] S02: According to the preset signal reconstruction method, convert the acoustic feature vector into an acoustic spectrogram with time-frequency characteristics, and convert the vibration feature vector into a vibration spectrogram with time-frequency characteristics.

[0095] S03: Obtain the fault type of the GIS device according to the acoustic spectrogram, vibration spectrogram and the preset deep learning model;

[0096] where the preset deep learning model constructs a feature correlation matrix based on the multi-head self-attention mechanism and is obtained by transfer learning training on the initial deep learning model according to the historical fault feature database.

[0097] As a preferred embodiment of the first embodiment, the fault type of the GIS equipment is obtained according to the acoustic spectrum, the vibration spectrum and the preset deep learning model, specifically:

[0098] Inputting the acoustic spectrum and the vibration spectrum into a preset deep learning model, so that the deep learning model can calculate the soundprint features and vibration pattern features of typical faults of GIS equipment according to the extreme learning machine algorithm;

[0099] Among them, the acoustic characteristic spectrum formed when the mechanical failure of the GIS equipment occurs is as follows: Figure 3 As shown in the figure, the vibration characteristic spectrum formed when the mechanical failure of the GIS equipment occurs is as follows Figure 4 As shown;

[0100] Compare the voiceprint features and vibration print features of the typical fault of the GIS equipment with the preset fuzzy entropy value;

[0101] If the absolute value of the difference between the preset fuzzy entropy and the voiceprint feature and the vibration print feature is less than or equal to the preset threshold, the fault type of the inspected GIS device is output.

[0102] More specifically, the ELM algorithm is used to calculate the soundprint features and vibration print features of typical faults of GIS equipment, and then they are compared with the fuzzy entropies obtained in the above steps respectively. When the absolute value of the difference between one of the multi-scale fuzzy entropies of the sound signal and the vibration signal and the corresponding eigenvalue obtained by the ELM algorithm is within 0.02, it can be determined that this type of fault has occurred in the inspected GIS equipment.

[0103] The frequency range of the acoustic vibration signal collected in this application is 0 to 2000 Hz.

[0104] In this preferred embodiment, the present application inputs the acoustic spectrum and vibration spectrum into a preset deep learning model, and uses the extreme learning machine (ELM) algorithm to calculate the soundprint features and vibration pattern features of typical faults of GIS equipment. By comparing with the preset fuzzy entropy value, when the absolute value of the difference is less than or equal to the preset threshold, the model outputs the fault type. This process not only makes full use of the multimodal information of acoustic and vibration signals, but also improves the accuracy of fault diagnosis through the precise matching of fuzzy entropy values. The setting of the preset threshold further ensures the reliability of the diagnostic results and reduces the misjudgment rate. Therefore, the present invention significantly improves the efficiency and accuracy of GIS equipment fault diagnosis and provides a strong guarantee for the stable operation of the power system.

[0105] As a preferred embodiment of the first embodiment, the preset deep learning model is obtained by constructing a feature association matrix based on a multi-head self-attention mechanism and performing transfer learning training on the initial deep learning model according to a historical fault feature database, specifically:

[0106] Obtain the historical fault feature database;

[0107] Input the acoustic spectrogram and vibration spectrogram in the historical fault feature database into the initial deep learning model, so that the initial deep learning model extracts features from the acoustic spectrogram and vibration spectrogram to obtain a feature matrix;

[0108] Divide the feature matrix into image patches and convert each image patch into serialized data;

[0109] Perform a non-linear transformation on the serialized data and introduce position information to generate a position embedding vector;

[0110] According to the multi-head self-attention mechanism, perform transfer training on the position embedding vector;

[0111] When the parameters of the transfer training reach the preset threshold, stop training to obtain a preset deep learning model.

[0112] More specifically, extract the feature spectrogram generated by the spectrogram generation module to obtain a feature matrix;

[0113] Divide the feature matrix and convert the image patches into serialized data, and its conversion method is shown in Equation (4):

[0114] O g,h = Ω0·vec(F(u g,h )) + d g,h (4)

[0115] where O g,h represents the h-th position in the g-th image patch, F represents a specific non-linear transformation function, vec represents flattening the matrix into a vector, and Ω0 and d g,h are learnable parameters;

[0116] Introduce the position information of the image patch in the way of adding position embedding and Patch embedding and convert it into a vector k;

[0117] Perform query, key, and value linear transformations on the input vector k to obtain corresponding multi-head transformation vectors and perform self-attention calculations respectively. The specific calculation processes are shown in Equation (5), Equation (6), and Equation (7):

[0118]

[0119] where, W1 represents the weight matrix obtained after performing a query transformation on the input vector k, W2 represents the weight matrix obtained after performing a key transformation on the input vector k, W3 represents the weight matrix obtained after performing a linear transformation on the input vector k, and E, R, I represent the corresponding vectors obtained after the weight matrix transformation;

[0120] Perform splitting operations on E, R, and I respectively, splitting each into h heads, resulting in For each head v ∈ [1, h], calculate the corresponding attention weight γ:

[0121]

[0122] where ω represents the scaling factor, used to reduce the gradient vanishing error caused by dot product calculations;

[0123] For the attention weight γ of each head v multiply it with its corresponding value matrix I V and concatenate them to obtain:

[0124]

[0125] where Concat represents the concatenation operation, MA represents the multi - head self - attention mechanism, and finally, through linear transformation of the obtained MA(E, R, I), the corresponding query matrix, key matrix, and value matrix of MA(E, R, I) are obtained.

[0126] In this preferred embodiment, the present application constructs a preset deep - learning model by performing transfer learning training on the initial deep - learning model using the historical fault feature database. The specific process includes: inputting the acoustic spectrogram and vibration spectrogram in the historical fault feature database into the initial model, extracting the feature matrix and dividing it into image patches, converting it into serialized data, then performing non - linear transformation and introducing position information to generate position embedding vectors. Subsequently, based on the multi - head self - attention mechanism, transfer training is performed on the position embedding vectors until the training parameters reach the preset threshold, completing the model training. This process not only makes full use of historical fault data, enhancing the model's learning ability for different fault modes, but also improves the efficiency and accuracy of feature extraction through the multi - head self - attention mechanism. The finally obtained deep - learning model can more accurately identify and diagnose the fault types of GIS devices, significantly improving the accuracy and reliability of fault diagnosis, and providing strong support for the efficient operation and maintenance of the power system.

[0127] This application obtains acoustic eigenvectors and vibration eigenvectors, and uses signal reconstruction methods to convert them into acoustic spectra and vibration spectra with time-frequency characteristics, thereby providing rich multimodal input data for the deep learning model. This multimodal data fusion method can more comprehensively reflect the operating status of GIS equipment, thereby improving the accuracy of fault diagnosis. Furthermore, the preset deep learning model constructs a feature association matrix based on a multi-head self-attention mechanism, which can automatically learn the intrinsic correlation between acoustic modes and vibration modes, and enhance the robustness of feature extraction. Transfer learning training is carried out in combination with a historical fault feature database, so that the model can quickly adapt to new fault modes and significantly improve the ability to identify unknown fault types. This application effectively solves the problem that the existing technology cannot accurately diagnose faults in GIS equipment.

[0128] Embodiment 2

[0129] Please refer to Figure 5 , which is a fault diagnosis device for GIS equipment provided in an embodiment of the present application.

[0130] In this embodiment, the fault diagnosis device for GIS equipment includes an acquisition module 10 , a reconstruction module 20 and a diagnosis module 30 .

[0131] The entire process of signal acquisition, signal processing and fault diagnosis in this application is as follows Figure 2 As shown in the figure; first, the multi-dimensional original signal is obtained from the acoustic and vibration sensor array through the ICA blind source separation technology. Then, the fuzzy entropy of these signals is calculated, and the component with the smallest fuzzy entropy is selected as the characteristic signal of the acoustic and vibration signals. Then, the multi-scale fuzzy entropy is calculated for these characteristic signals. Finally, the calculated multi-scale fuzzy entropy is compared with the soundprint and vibration print characteristics of the typical fault of the GIS equipment obtained by the ELM algorithm, and the absolute value of the difference is calculated. If one of these differences is less than or equal to the preset threshold of 0.02, it is determined that this type of fault has occurred in the GIS equipment. This process realizes the accurate diagnosis of GIS equipment faults by comprehensively analyzing the characteristics of acoustic and vibration signals.

[0132] The acquisition module 10 is used to acquire preset acoustic eigenvectors and vibration eigenvectors.

[0133] As a preferred embodiment of the second embodiment, the obtaining of the preset acoustic feature vector and the vibration feature vector is specifically: collecting acoustic and vibration signals according to a preset acoustic sensor array and a preset vibration sensor;

[0134] According to a preset blind source separation method, feature extraction is performed on the acoustic vibration signal to obtain each independent component of the acoustic vibration signal;

[0135] Calculate the fuzzy entropy values of each of the independent components, and select the minimum entropy feature from the fuzzy entropy values as the preliminary feature vector;

[0136] Perform multi-scale time window segmentation on the preliminary feature vector to obtain an acoustic feature vector and a vibration feature vector with noise resistance.

[0137] More specifically, use the ICA blind source separation technology to perform blind source separation on the acoustic signal to obtain a multi-dimensional source signal x q (q = 1, 2,..., n, where n is a positive integer);

[0138] Use the ICA blind source separation technology to perform blind source separation on the vibration signal to obtain a multi-dimensional source signal y q (q = 1, 2,..., n, where n is a positive integer);

[0139] Calculate the fuzzy entropy of x q and y q respectively, and select the component with the minimum value as the feature signal;

[0140] Based on the obtained feature signal, combined with Equation (1) and Equation (2), calculate the multi-scale fuzzy entropy of y q :

[0141]

[0142] where L is the time scale factor, that is, the multi-dimensional original signal is divided into L segments, each segment is n / L in length; 1 ≤ r ≤ n, and r is a positive integer; μ is the given embedding dimension, σ is the given similarity tolerance, and ξ is the given fuzzy function gradient;

[0143] Similarly, the multi-scale fuzzy entropy of x q can be obtained.

[0144] In this preferred embodiment, the present application collects acoustic and vibration signals through a preset acoustic sensor array and vibration sensors, and uses the blind source separation method to extract features of the signals to obtain multiple independent components. By calculating the fuzzy entropy values of these components and selecting the minimum entropy feature, a preliminary feature vector is formed. Further, multi-scale time window segmentation is performed on the preliminary feature vector to generate an acoustic feature vector and a vibration feature vector with noise resistance. This process not only effectively extracts the key features in the acoustic and vibration signals, but also enhances the noise resistance and robustness of the features through multi-scale analysis. Therefore, the present invention can significantly improve the accuracy and reliability of feature extraction, provide high-quality input data for subsequent fault diagnosis, and thus improve the performance and accuracy of the overall diagnosis system.

[0145] As a preferred embodiment of the second embodiment, after collecting the acoustic and vibration signals according to the preset acoustic sensor array and the preset vibration sensors, it further includes:

[0146] According to the beamforming principle, obtain the beamforming outputs received by multiple acoustic sensor arrays;

[0147] According to the respective beamforming outputs and a preset formula, calculate the cross-power spectra of the respective beams;

[0148] Calculate the maximum points of the cross-power spectra;

[0149] According to the maximum points and the spatial coordinates of the multiple acoustic sensor arrays, determine the specific location where the fault occurs.

[0150] More specifically, obtain the beamforming outputs received at multiple acoustic sensors through the beamforming principle;

[0151] Combining Equation (3) with the beamforming output, the cross-power spectrum of the beam can be obtained:

[0152]

[0153] where t is the relative delay between the acoustic signal reaching any sensor and the acoustic signal reaching the reference sensor; f is the angular frequency of the acoustic signal; N represents the total number of sensors; C ab is the cross-correlation matrix between sensor a and sensor b; j is the imaginary unit;

[0154] From the cross-power spectrum of the beam, it can be obtained that when t a equals the time when the beam actually propagates to reach the sensor, it reaches the maximum value. Therefore, the specific location of the faulty GIS device can be deduced based on the propagation direction of the wave and the delay time.

[0155] In this preferred embodiment, after collecting the acoustic vibration signals, the present application further processes the output signals of the acoustic sensor array using the beamforming principle, calculates the cross-power spectra of the respective beams, and by analyzing the maximum points of the cross-power spectra and combining the spatial coordinates of the sensor array, accurately determines the specific location where the fault occurs. This process can not only effectively capture the propagation direction and time delay information of the acoustic signal, but also accurately locate the fault source through the maximum points of the cross-power spectra. This method significantly improves the accuracy and reliability of fault location, reduces the misjudgment rate, and at the same time provides clear fault location information for maintenance personnel, greatly improving the maintenance efficiency and the safety of equipment operation.

[0156] The reconstruction module 20 is configured to convert the acoustic feature vector into an acoustic spectrogram with time-frequency characteristics and convert the vibration feature vector into a vibration spectrogram with time-frequency characteristics according to a preset signal reconstruction method.

[0157] The diagnostic module 30 is used to obtain the fault type of the GIS device according to the acoustic spectrogram, vibration spectrogram and a preset deep learning model;

[0158] Among them, the preset deep learning model constructs a feature correlation matrix based on the multi-head self-attention mechanism and performs transfer learning training on the initial deep learning model according to the historical fault feature database.

[0159] As a preferred embodiment of the second embodiment, the obtaining of the fault type of the GIS device according to the acoustic spectrogram, vibration spectrogram and the preset deep learning model is specifically as follows:

[0160] Input the acoustic spectrogram and vibration spectrogram into the preset deep learning model, so that the deep learning model calculates the acoustic feature and vibration feature of the typical fault of the GIS device according to the extreme learning machine algorithm;

[0161] Among them, the acoustic feature spectrogram formed at the moment of the mechanical fault of the GIS device is as Figure 3 shown, and the vibration feature spectrogram formed at the moment of the mechanical fault of the GIS device is as Figure 4 shown;

[0162] Compare the acoustic feature and vibration feature of the typical fault of the GIS device with the preset fuzzy entropy value;

[0163] If the absolute value of the difference between the preset fuzzy entropy and the acoustic feature and vibration feature is less than or equal to the preset threshold, output the fault type of the inspected GIS device.

[0164] More specifically, the ELM algorithm is used to calculate the acoustic feature and vibration feature of the typical fault of the GIS device, and then compare them with the fuzzy entropy obtained in the above steps respectively. When the absolute value of the difference between one of the multi-scale fuzzy entropies of the acoustic signal and the vibration signal and the corresponding eigenvalue obtained by the ELM algorithm is within 0.02, it can be determined that the inspected GIS device has this type of fault.

[0165] Among them, the frequency range of the acoustic and vibration signals collected in this application is from 0 to 2000 Hz.

[0166] In this preferred embodiment, the present application inputs the acoustic spectrogram and vibration spectrogram into a preset deep learning model, and uses the Extreme Learning Machine (ELM) algorithm to calculate the acoustic fingerprint features and vibration fingerprint features of typical faults of GIS equipment. By comparing with the preset fuzzy entropy value, when the absolute value of the difference is less than or equal to the preset threshold, the model outputs the fault type. This process not only makes full use of the multimodal information of acoustic and vibration signals, but also improves the accuracy of fault diagnosis through the precise matching of fuzzy entropy values. The setting of the preset threshold further ensures the reliability of the diagnosis result and reduces the misjudgment rate. Therefore, the present invention significantly improves the efficiency and accuracy of fault diagnosis of GIS equipment, providing a strong guarantee for the stable operation of the power system.

[0167] As a preferred embodiment of Embodiment 2, the preset deep learning model is constructed by building a feature correlation matrix based on the multi-head self-attention mechanism and performing transfer learning training on the initial deep learning model according to the historical fault feature database. Specifically:

[0168] Obtain the historical fault feature database;

[0169] Input the acoustic spectrogram and vibration spectrogram in the historical fault feature database into the initial deep learning model, so that the initial deep learning model extracts features from the acoustic spectrogram and vibration spectrogram to obtain a feature matrix;

[0170] Divide the feature matrix into image patches and convert each image patch into serialized data;

[0171] Perform a non-linear transformation on the serialized data and introduce position information to generate a position embedding vector;

[0172] According to the multi-head self-attention mechanism, perform transfer training on the position embedding vector;

[0173] When the parameters of the transfer training reach the preset threshold, stop training to obtain the preset deep learning model.

[0174] More specifically, extract the feature spectrogram generated by the spectrogram generation module to obtain a feature matrix;

[0175] Divide the feature matrix and convert the image patches into serialized data, and the conversion method is as shown in Equation (4):

[0176] O g,h =Ω0·vec(F(u g,h ))+d g,h (4)

[0177] where O g,hDenote the h-th position in the g-th image patch, F represents a specific non-linear transformation function, vec represents flattening a matrix into a vector, and Ω0 and d g,h are learnable parameters;

[0178] Introduce the position information of the image patch by adding position embedding and Patch embedding, and convert it into a vector k;

[0179] Perform query, key, and value linear transformations on the input vector k to obtain corresponding multi-head transformation vectors, and perform self-attention calculations respectively. The specific calculation processes are as shown in Equations (5), (6), and (7):

[0180]

[0181] Among them, W1 represents the weight matrix obtained after performing a query transformation on the input vector k, W2 represents the weight matrix obtained after performing a key transformation on the input vector k, W3 represents the weight matrix obtained after performing a linear transformation on the input vector k, and E, R, I represent the corresponding vectors obtained after the weight matrix transformation;

[0182] Perform splitting operations on E, R, and I respectively, and split them into u heads, then there are Calculate the corresponding attention weight γ for each head v ∈ [1, h]:

[0183]

[0184] where ω represents a scaling factor used to reduce the gradient vanishing error caused by dot product calculations;

[0185] For the attention weight γ of each head v and its corresponding value matrix I V perform multiplication and concatenation to obtain:

[0186]

[0187] where Concat represents the concatenation operation, MA represents the multi-head self-attention mechanism, and finally, by performing a linear transformation on the obtained MA(E, R, I), the corresponding query matrix, key matrix, and value matrix of MA(E, R, I) are obtained.

[0188] In this preferred embodiment, the present application constructs a preset deep learning model by performing transfer learning training on an initial deep learning model using a historical fault feature database. The specific process includes: inputting the acoustic spectrogram and vibration spectrogram in the historical fault feature database into the initial model, extracting the feature matrix and dividing it into image patches, converting it into serialized data, performing non-linear transformation and introducing position information to generate a position embedding vector. Subsequently, transfer training is performed on the position embedding vector based on the multi-head self-attention mechanism until the training parameters reach the preset threshold, completing the model training. This process not only makes full use of historical fault data, enhances the model's learning ability for different fault modes, but also improves the efficiency and accuracy of feature extraction through the multi-head self-attention mechanism. The finally obtained deep learning model can more accurately identify and diagnose the fault types of GIS devices, significantly improving the accuracy and reliability of fault diagnosis, and providing strong support for the efficient operation and maintenance of the power system.

[0189] The present application provides rich multi-modal input data for the deep learning model by obtaining acoustic feature vectors and vibration feature vectors and converting them into acoustic spectrograms and vibration spectrograms with time-frequency characteristics using a signal reconstruction method. This multi-modal data fusion method can more comprehensively reflect the operating state of GIS devices, thereby improving the accuracy of fault diagnosis. Further, the preset deep learning model constructs a feature correlation matrix based on the multi-head self-attention mechanism, which can automatically learn the internal correlation relationship between the acoustic modality and the vibration modality, strengthening the robustness of feature extraction. Combining transfer learning training with the historical fault feature database enables the model to quickly adapt to new fault modes and significantly improves the recognition ability for unknown fault types. The present application effectively solves the problem that the prior art cannot accurately diagnose the faults of GIS devices.

[0190] Embodiment Three:

[0191] The embodiment of the present application provides a computer-readable storage medium, which includes a stored computer program. When the computer program runs, it controls the device where the computer-readable storage medium is located to execute the fault diagnosis method of a GIS device as described above;

[0192] Among them, for the fault diagnosis method of the GIS device, when it is implemented in the form of a software functional unit and used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, to implement all or part of the processes in the above-mentioned embodiment methods of the present invention, it can also be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-mentioned various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0193] Embodiment 4

[0194] This application provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements any one of the fault diagnosis methods of the GIS device as described in Embodiment 1.

[0195] The above-mentioned specific embodiments have further elaborated on the purpose, technical solutions, and beneficial effects of the present invention. It should be understood that the above is only the specific embodiments of the present invention and is not used to limit the protection scope of the present invention. It is particularly pointed out that for those skilled in the art, any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A fault diagnosis method for a GIS device, characterized in that, include: Obtaining preset acoustic eigenvectors and vibration eigenvectors; According to a preset signal reconstruction method, the acoustic feature vector is converted into an acoustic spectrum with time-frequency characteristics, and the vibration feature vector is converted into a vibration spectrum with time-frequency characteristics; Obtaining the fault type of the GIS equipment according to the acoustic spectrum, the vibration spectrum and the preset deep learning model; Among them, the preset deep learning model is obtained by constructing a feature association matrix based on a multi-head self-attention mechanism, and performing transfer learning training on the initial deep learning model according to a historical fault feature database.

2. The fault diagnosis method of the GIS device according to claim 1, characterized in that, The step of obtaining the preset acoustic eigenvector and vibration eigenvector is specifically as follows: Acquiring acoustic and vibration signals according to a preset acoustic sensor array and a preset vibration sensor; According to a preset blind source separation method, feature extraction is performed on the acoustic vibration signal to obtain each independent component of the acoustic vibration signal; Calculating each fuzzy entropy value of the independent components, and selecting a minimum entropy feature from each fuzzy entropy value as a preliminary feature vector; The preliminary feature vector is segmented into multi-scale time windows to obtain an acoustic feature vector and a vibration feature vector with noise resistance.

3. The fault diagnosis method of the GIS device according to claim 1, characterized in that After collecting the acoustic and vibration signals according to the preset acoustic sensor array and the preset vibration sensor, the method further includes: According to the beamforming principle, obtaining respective beamforming outputs received by the plurality of acoustic sensor arrays; Calculate the cross power spectrum of each beam according to the output of each beam forming and a preset formula; Calculating the maximum point of the cross power spectrum; The specific location where the fault occurs is determined according to the maximum point and the spatial coordinates of the multiple acoustic sensor arrays.

4. The fault diagnosis method of the GIS device according to claim 3, characterized in that, The cross power spectrum of each beam is calculated based on the output of each beam forming and a preset formula, specifically: The preset formula is: where t is the relative delay of the acoustic signal reaching any sensor and the acoustic signal reaching the standard sensor; f is the angular frequency of the acoustic signal; N represents the total number of sensors; C ab is the cross-correlation matrix between sensor a and sensor b; j is the imaginary unit; The cross power spectrum of each beam is calculated based on the output of each beam forming and a preset formula.

5. The fault diagnosis method of the GIS device according to claim 1, characterized in that The fault type of the GIS equipment is obtained according to the acoustic spectrum, vibration spectrum and the preset deep learning model, specifically: Inputting the acoustic spectrum and the vibration spectrum into a preset deep learning model, so that the deep learning model can calculate the soundprint features and vibration pattern features of typical faults of GIS equipment according to the extreme learning machine algorithm; Compare the voiceprint features and vibration print features of the typical fault of the GIS equipment with the preset fuzzy entropy value; If the absolute value of the difference between the preset fuzzy entropy and the voiceprint feature and the vibration print feature is less than or equal to the preset threshold, the fault type of the inspected GIS device is output.

6. The fault diagnosis method of the GIS device according to claim 1, wherein The preset deep learning model is obtained by constructing a feature association matrix based on a multi-head self-attention mechanism and performing transfer learning training on the initial deep learning model according to a historical fault feature database, specifically: Obtain historical fault feature database; Inputting the acoustic spectrum and the vibration spectrum in the historical fault feature database into the initial deep learning model, so that the initial deep learning model performs feature extraction on the acoustic spectrum and the vibration spectrum to obtain a feature matrix; Dividing the feature matrix into image blocks, and converting each image block into serialized data; Perform a non - linear transformation on the serialized data and introduce location information to generate a location embedding vector; According to the multi - head self - attention mechanism, perform transfer training on the location embedding vector; When the parameters of the transfer training reach a preset threshold, stop the training to obtain a preset deep learning model.

7. A fault diagnosis device for a GIS device, characterized in that, It includes: An acquisition module, a reconstruction module, and a diagnosis module; The acquisition module is used to acquire preset acoustic feature vectors and vibration feature vectors; The reconstruction module is used to convert the acoustic feature vector into an acoustic spectrogram with time - frequency characteristics and convert the vibration feature vector into a vibration spectrogram with time - frequency characteristics according to a preset signal reconstruction method; The diagnosis module is used to obtain the fault type of the GIS device according to the acoustic spectrogram, vibration spectrogram, and a preset deep learning model; Among them, the preset deep learning model is constructed by building a feature correlation matrix based on the multi - head self - attention mechanism and performing transfer learning training on the initial deep learning model according to the historical fault feature database.

8. The fault diagnosis device of the GIS device according to claim 7, characterized in that, The acquisition module is used to acquire preset acoustic feature vectors and vibration feature vectors, specifically: The acquisition module collects acoustic - vibration signals according to a preset acoustic sensor array and a preset vibration sensor; According to a preset blind source separation method, perform feature extraction on the acoustic - vibration signals to obtain each independent component of the acoustic - vibration signals; Calculate the fuzzy entropy values of each independent component, and select the minimum entropy feature from the fuzzy entropy values as a preliminary feature vector; Perform multi - scale time - window segmentation on the preliminary feature vector to obtain an acoustic feature vector and a vibration feature vector with noise - resistant ability.

9. A computer-readable storage medium, characterized in that, The computer - readable storage medium includes a stored computer program, wherein when the computer program runs, it controls the device where the computer - readable storage medium is located to execute the fault diagnosis method of the GIS device according to any one of claims 1 to 6.

10. A terminal device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the fault diagnosis method of the GIS device according to any one of claims 1 to 6.

Images