Spectral Signal Feature Extraction Method for GIS Equipment and Discharge Fault Diagnosis Method

Abstract

The present invention relates to the technical field of discharge detection, and specifically provides a method for extracting spectral signal features of a GIS device and a method for diagnosing discharge faults. The method includes: acquiring spectral data of gas generated by the discharge of a gas-insulated switchgear; performing enhancement processing on the spectral data to obtain one-dimensional spectral data; performing transformation on the one-dimensional spectral data through an improved S-transform to obtain a two-dimensional time-frequency spectrogram, where the improved S-transform is an S-transform function that dynamically adjusts the frequency-domain window width through a Gaussian error function; extracting spectral time-frequency features according to the two-dimensional time-frequency spectrogram through a convolutional neural network, and determining the corresponding fault type according to the spectral time-frequency features. To solve the problems in the detection of gas-insulated switchgears in the related art that rely on chemical sensors or chromatographic analysis, have slow response speed, high cost, and poor real-time performance.

Description

Technical Field

[0001] The present invention relates to the technical field of discharge detection, and particularly to a method for extracting spectral signal characteristics of a GIS device and a method for diagnosing discharge faults. Background Art

[0002] In the power system, gas-insulated switchgear (GIS) is widely used due to its excellent insulation performance and compact structure. However, during operation, GIS devices may experience partial discharge (PD) due to manufacturing defects, material aging, or improper operation, which can cause the decomposition of the mixed gas of sulfur hexafluoride SF6 and nitrogen N2, generating chemical decomposition products such as sulfur dioxide SO2, nitric oxide NO, and carbon disulfide CS2. The presence of these decomposition products not only affects the insulation performance of GIS devices but may also trigger equipment failures. Therefore, the monitoring and analysis of these decomposition products are of great significance for evaluating the health status of GIS and preventing failures.

[0003] Traditional PD decomposition product monitoring methods rely on chemical sensors or chromatographic analysis techniques. Although these methods are accurate, they often suffer from limitations such as slow response speed, high cost, and inability to perform real-time monitoring.

[0004] Regarding the detection of gas-insulated switchgear in related technologies, which rely on chemical sensors or chromatographic analysis, there are problems of slow response speed, high cost, and poor real-time performance, and no effective solution has been proposed yet. Summary of the Invention

[0005] The present invention provides a method for extracting spectral signal characteristics of a GIS device and a method for diagnosing discharge faults, which at least solve the problems of slow response speed, high cost, and poor real-time performance in the detection of gas-insulated switchgear in related technologies that rely on chemical sensors or chromatographic analysis.

[0006] According to one aspect of the embodiments of the present invention, a method for diagnosing discharge faults of a GIS device is provided, including: obtaining spectral data of the gas generated by the discharge of the gas-insulated switchgear; performing enhancement processing on the spectral data to obtain one-dimensional spectral data; performing transformation on the one-dimensional spectral data through an improved S transform to obtain a two-dimensional time-frequency spectrogram, where the improved S transform is an S transform function that dynamically adjusts the frequency-domain window width through a Gaussian error function; extracting spectral time-frequency characteristics according to the two-dimensional time-frequency spectrogram through a convolutional neural network, and determining the corresponding fault type according to the spectral time-frequency characteristics.

[0007] As an alternative, the spectral data is enhanced to obtain one-dimensional spectral data, including: subtracting the background spectral data of the background gas from the spectral data to obtain spectral data with background interference removed; using a target wavelet function as the mother wavelet to perform N-layer wavelet decomposition on the spectral data with background interference removed to obtain the approximation coefficient of the Nth layer and the detail coefficients of the first N layers; performing adaptive threshold estimation on the detail coefficients to obtain the denoising threshold corresponding to the detail coefficients; performing hard threshold denoising processing on the corresponding detail coefficients through the denoising threshold to obtain denoised detail coefficients; and performing wavelet reconstruction on the N-layer decomposition signal according to the denoised detail coefficients and the approximation coefficient to obtain the enhanced one-dimensional spectral data.

[0008] As an alternative, performing adaptive threshold estimation on the detail coefficients to obtain the denoising threshold corresponding to the detail coefficients includes: performing adaptive estimation on the detail coefficients through the following formula to obtain the corresponding denoising threshold:

[0009]

[0010] where ε j is the denoising threshold of the detail coefficient d j at the jth layer, j = 1, 2, 3... N, σ j is the standard deviation estimate of d j N j is the number of sampling points in d j and MAD|d j | represents the calculation of the absolute median difference of d j .

[0011] As an alternative, the one-dimensional spectral data is transformed through an improved S transform to obtain a two-dimensional time-frequency spectrogram, including: calculating the frequency domain window width through the Gaussian error function; calculating the expression of the window function of the improved S transform at the corresponding frequency according to the frequency domain window width; and calculating the two-dimensional time-frequency spectrogram corresponding to the one-dimensional spectral data according to the expression and the calculation formula of the improved S transform.

[0012] As an alternative, calculating the frequency domain window width through the Gaussian error function includes: calculating through the calculation formula of the frequency domain window width width f , and the calculation formula is as follows:

[0013]

[0014] where f maxis the maximum value of the signal frequency, f is the current frequency, a and b are the amplitude and slope adjustment parameters respectively, and GEF() is the Gaussian error function; calculating the expression of the window function of the improved S transform at the corresponding frequency according to the frequency domain window width, including: calculating through the expression, and the expression is as follows:

[0015]

[0016] In the formula, W(α, f) is the expression of the Fourier transform of the window function w() at the frequency f.

[0017] As an optional solution, calculating the two-dimensional time-frequency spectrogram corresponding to the one-dimensional spectral data according to the expression and the calculation formula of the improved S transform, including: the calculation formula of the improved S transform is as follows:

[0018]

[0019] In the formula, X(α) is the Fourier transform of the continuous signal x(t) of the one-dimensional spectral data, α is the frequency variable, with the unit of radian per second (rad / s), t is the time variable, with the unit of second (s), and S(τ, f) is the time-frequency representation at time τ and frequency f, that is, the two-dimensional time-frequency spectrogram corresponding to the one-dimensional spectral data.

[0020] As an optional solution, through a convolutional neural network, extracting spectral time-frequency features according to the two-dimensional time-frequency spectrogram, and determining the corresponding fault type according to the spectral time-frequency features, including: extracting features from the two-dimensional time-frequency spectrogram through a feature extraction network to obtain a feature vector corresponding to the spectral time-frequency features, where the feature extraction network includes a convolutional neural network with multiple convolutional layers and pooling layers, and is trained through training data; determining the corresponding fault type according to the spectral time-frequency features through a fault diagnosis network, where the fault diagnosis network includes a feedforward neural network, which is trained through a preset number of training cycles on a training set and verified through a validation set, and the training set and validation set are shuffled and re-partitioned in each training cycle; where the feedforward neural network includes an input layer, a hidden layer, and an output layer, the input layer is used to receive the one-dimensional feature vector of the spectral time-frequency features, the hidden layer is at least one fully connected layer, which is used to extract features and make decisions on the output data of the input layer, and an activation function is also set after the fully connected layer of the hidden layer, the output layer is a fully connected layer, which is used to output multiple nodes, and each node corresponds to the probability of a fault type.

[0021] As an alternative solution, before enhancing the spectral data to obtain one-dimensional spectral data, the method further includes: using a dynamic gas mixer to mix various gases stored in a gas cylinder according to required ratios to obtain a standard mixed gas, where the various gases include a background gas and a discharge gas, the background gas includes a mixed gas of sulfur hexafluoride (SF6) and nitrogen (N2), and the discharge gas includes at least one of the following: sulfur dioxide (SO2), nitric oxide (NO), and carbon disulfide (CS2); detecting corresponding spectral data of the standard mixed gas using an infrared light source and an infrared spectrometer; and using spectral data corresponding to various different required ratios as training data to train the convolutional neural network.

[0022] According to another aspect of the embodiments of the present invention, there is also provided a method for extracting spectral signal features of a GIS device, including: enhancing spectral data of a mixed gas to obtain one-dimensional spectral data; performing a transformation on the one-dimensional spectral data using an improved S-transform to obtain a two-dimensional time-frequency spectrogram, where the improved S-transform is an S-transform function that dynamically adjusts the width of the frequency-domain window using a Gaussian error function; and extracting spectral time-frequency features from the two-dimensional time-frequency spectrogram using a convolutional neural network as the spectral signal features.

[0023] According to another aspect of the embodiments of the present invention, there is also provided an electronic device, including: a processor, and a memory storing a program, where the program includes instructions that, when executed by the processor, cause the processor to execute the above-mentioned method for extracting spectral signal features of a GIS device and the discharge fault diagnosis method.

[0024] The discharge fault diagnosis method provided by the embodiments of the present invention creates, by enhancing the spectral data of the gas generated by the discharge of a gas-insulated switchgear, the obtained one-dimensional spectral data eliminates background interference and random noise, having a more accurate technical effect; and performs a transformation on the one-dimensional spectral data using an improved S-transform to obtain a two-dimensional time-frequency spectrogram. The improved S-transform sets a Gaussian error function to dynamically adjust the width of the frequency-domain window of the S-transform to optimize the time-frequency resolution, which is particularly suitable for analyzing signals containing noise, can more accurately capture high-frequency components in the signal, and reduce false components in the finally obtained two-dimensional time-frequency spectrogram. The spectral time-frequency features are extracted from the two-dimensional time-frequency spectrogram using a convolutional neural network, and the corresponding fault type is determined based on the spectral time-frequency features, thereby quickly and accurately performing fault diagnosis, improving the efficiency and accuracy of fault diagnosis. Furthermore, it solves the problems in the related art that the detection of gas-insulated switchgear depends on chemical sensors or chromatographic analysis, with slow response speed, high cost, and poor real-time performance, achieving the technical effects of improving the response speed, not requiring professional sensors and analysis devices, reducing costs, and ensuring the real-time performance of fault detection with a high detection speed. Description of the Drawings

[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other embodiments can be obtained based on these drawings.

[0026] Figure 1 It is a flowchart of a method for diagnosing discharge faults of a GIS device according to an embodiment of the present invention.

[0027] Figure 2 It is a schematic diagram of the discharge fault diagnosis process of a GIS device according to an implementation manner of the present invention.

[0028] Figure 3 It is a schematic diagram of the processing result of the standard S transform for multi-frequency signals without improvement according to an implementation manner of the present invention.

[0029] Figure 4 It is a schematic diagram of the processing result of the improved S transform for multi-frequency signals according to an implementation manner of the present invention.

[0030] Figure 5 It is a schematic diagram of the feature extraction network architecture according to an implementation manner of the present invention.

[0031] Figure 6 It is a schematic diagram of the data processing flow of the fault diagnosis network according to an implementation manner of the present invention.

[0032] Figure 7 It is a flowchart of a method for extracting spectral signal features of a GIS device according to an embodiment of the present invention.

[0033] Figure 8 It is a schematic diagram of the structure of the electronic device of the present invention. Detailed implementation manners

[0034] The following will describe the embodiments of the present invention in more detail with reference to the drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments described herein. On the contrary, these embodiments are provided to more thoroughly and completely understand the present invention. It should be understood that the drawings and embodiments of the present invention are only for exemplary purposes and are not used to limit the protection scope of the present invention.

[0035] In view of the fact that traditional PD decomposition product monitoring methods rely on chemical sensors or chromatographic analysis techniques, although these methods are accurate, they often have limitations such as slow response speed, high cost, and inability to monitor in real time. With the development of spectral analysis technology, especially the application of Fourier transform infrared spectroscopy (FTIR), it has become possible to rapidly and real-time monitor PD decomposition products.

[0036] FTIR technology can provide rich molecular vibration information, but how to extract useful features from complex spectral data, especially for the processing of non-linear and high-dimensional data, remains a challenge.

[0037] In recent years, deep learning technology, especially convolutional neural networks (CNNs), has achieved remarkable results in the fields of image recognition and feature extraction. CNNs can automatically learn hierarchical feature representations from images, but their applications are usually limited to two-dimensional data.

[0038] In order to apply CNNs to the feature extraction of spectral signals, it is necessary to convert one-dimensional spectral signals into two-dimensional forms so that the powerful capabilities of CNNs can be utilized. As an effective time-frequency analysis method, the S transform can convert one-dimensional signals into two-dimensional time-frequency spectrograms, but traditional S transforms have problems with insufficient resolution when processing high-frequency components.

[0039] To solve the above technical problems, embodiments of the present invention provide a discharge fault diagnosis method for GIS equipment based on an improved S transform and a convolutional neural network. Figure 1 It is a flowchart of a discharge fault diagnosis method for a GIS device according to an embodiment of the present invention. As Figure 1 shown, the discharge fault diagnosis method for a GIS device provided by the embodiments of the present invention includes the following steps:

[0040] Step S101, obtaining spectral data of the gas generated by the discharge of a gas-insulated switchgear;

[0041] Step S102, performing enhancement processing on the spectral data to obtain one-dimensional spectral data;

[0042] Step S103, converting the one-dimensional spectral data through an improved S transform to obtain a two-dimensional time-frequency spectrogram, where the improved S transform is an S transform function that dynamically adjusts the frequency-domain window width through a Gaussian error function;

[0043] Step S104, extracting spectral time-frequency features from the two-dimensional time-frequency spectrogram through a convolutional neural network, and determining the corresponding fault type according to the spectral time-frequency features.

[0044] The above-mentioned discharge fault diagnosis method provided by the embodiments of the present invention creates enhanced processing on the spectral data of the gas generated by the discharge of gas-insulated switchgear, enabling the obtained one-dimensional spectral data to eliminate background interference and random noise, achieving a more accurate technical effect; and the one-dimensional spectral data is transformed by improving the S transform to obtain a two-dimensional time-frequency spectrogram. The improved S transform sets a Gaussian error function to dynamically adjust the width of the frequency domain window of the S transform to optimize the time-frequency resolution, which is particularly suitable for analyzing signals containing noise, can more accurately capture the high-frequency components in the signal, and reduce the false components in the finally obtained two-dimensional time-frequency spectrogram. The convolutional neural network extracts spectral time-frequency features based on the two-dimensional time-frequency spectrogram and determines the corresponding fault type according to the spectral time-frequency features, thereby quickly and accurately performing fault diagnosis, improving the efficiency and accuracy of fault diagnosis.

[0045] The execution subject of the above steps can be a controller for GIS phase communication of gas-insulated switchgear, or a processor of a detection system that is not connected to the GIS device and independently operates to detect the working faults of the GIS device, or a server. The gas during the operation of the GIS device can be obtained through the devices of the system or third-party devices, and the spectral data of the gas is extracted, and through the above-mentioned execution subject.

[0046] The specific method for obtaining the spectral data of the gas generated by the discharge of the GIS device can be to first obtain the gas generated by its discharge from the GIS device through a gas collection device, and transmit the gas to a spectral acquisition device to collect the corresponding spectral data.

[0047] The above-mentioned spectral acquisition device can include an infrared light source and a corresponding infrared spectrometer. The gas collected is irradiated by the infrared light source to excite the vibrational energy level transition of gas molecules, and then the spectral signal generated after the gas molecules absorb infrared light is received in the infrared spectrometer to form the above-mentioned spectral data.

[0048] There is background interference and random noise in the above-mentioned spectral data. To avoid interference caused by multiple detection results, it is necessary to first perform enhanced processing on the spectral data to remove the random noise therein to obtain one-dimensional spectral data.

[0049] The way of background interference can be achieved by obtaining the spectral data of the background gas and subtracting it from the above-mentioned spectral data to remove the background interference. The ways of denoising can be Fourier transform, wavelet transform, mean filtering and other methods. Since the noise in the spectral data is random noise, the specific way of denoising is preferably wavelet threshold denoising. The specific method will be described later.

[0050] Enhanced processing can weaken background interference and random noise, improve the representation of spectral data for the actual discharge gas, enhance the signal quality, and improve the accuracy of subsequent processing and detection results.

[0051] The improved S transform described above is an S transform function that dynamically adjusts the width of the frequency-domain window through the Gaussian error function. In the standard S transform in the related art, the S transform, also known as the Stockwell transform, is a time-frequency analysis method. In the standard S transform, the window function is related to both time and frequency. That is, the width of the frequency-domain window is related to frequency.

[0052] When the frequency increases, the width of the window function becomes narrower; when the frequency decreases, the width of the window function becomes wider. This characteristic makes the standard S transform have a high frequency resolution and a low time resolution in the low-frequency part, and a high time resolution and a low frequency resolution in the high-frequency part. That is, it is impossible to balance the time-domain resolution and the frequency-domain resolution.

[0053] This embodiment provides an improved S transform based on the Gaussian error function. Through the Gaussian error function, the width of the frequency-domain window is dynamically calculated according to the maximum frequency and the current frequency to form a frequency-domain Gaussian window independent of frequency. Then, using the frequency-domain Gaussian window, a two-dimensional time-frequency spectrogram that balances the time-domain resolution and the frequency-domain resolution is obtained through the S transform. Furthermore, it can better reflect the effective components in the spectral data, which is beneficial for subsequent feature extraction based on the two-dimensional time-frequency spectrogram and fault detection.

[0054] The detailed steps of converting one-dimensional spectral data into a two-dimensional time-frequency spectrogram through the improved S transform will be described later.

[0055] Based on the two-dimensional time-frequency spectrogram that balances the time-domain resolution and the frequency-domain resolution, spectral time-frequency features are extracted from the two-dimensional time-frequency spectrogram through a convolutional neural network, and the corresponding fault types are determined according to the spectral time-frequency features, so as to achieve fast and efficient feature extraction and fault determination of the two-dimensional time-frequency spectrogram. Not only is the efficiency high, but the accuracy can also be guaranteed.

[0056] As an alternative solution, the spectral data is enhanced to obtain one-dimensional spectral data, including: subtracting the background spectral data of the background gas from the spectral data to obtain spectral data with background interference removed; using the target wavelet function as the mother wavelet to perform N-layer wavelet decomposition on the spectral data with background interference removed to obtain the approximation coefficient of the Nth layer and the detail coefficients of the first N layers; performing adaptive threshold estimation on the detail coefficients to obtain the denoising threshold corresponding to the detail coefficients; performing hard threshold denoising processing on the corresponding detail coefficients through the denoising threshold to obtain the denoised detail coefficients; and performing wavelet reconstruction based on the denoised detail coefficients and the approximation coefficient to reconstruct the N-layer decomposition signal to obtain the enhanced one-dimensional spectral data.

[0057] Subtracting the background spectral data of the background gas from the spectral data to obtain spectral data with background interference removed, thereby improving the accuracy of the spectral data.

[0058] The subsequent steps after obtaining the spectral data with background interference removed are all detailed steps for removing random noise through wavelet threshold denoising. The above wavelet decomposition is to decompose the spectral data to obtain multiple wavelet signals.

[0059] The above wavelet signals can be understood as the wavelet signal components of the spectral data in different wavelet functions. The wavelet function is a function with finite duration and oscillation characteristics defined in the related technology, and has good localization properties in both the time domain and the frequency domain.

[0060] Determining the denoising threshold through adaptive threshold estimation can improve the denoising effect and accuracy. The above wavelet reconstruction can be understood as the inverse operation of wavelet decomposition, and reconstructs the multiple wavelet signals obtained by decomposing the spectral data into the processed one-dimensional spectral data.

[0061] Through the above steps, accurate and effective random noise removal can be performed on the spectral data, improving the accuracy of the spectral data and better reflecting the accuracy of the components in the northern gas.

[0062] As an alternative solution, performing adaptive threshold estimation on the detail coefficients to obtain the denoising threshold corresponding to the detail coefficients, including: performing adaptive estimation on the detail coefficients through the following formula to obtain the corresponding denoising threshold:

[0063]

[0064] In the formula, ε j is the denoising threshold of the detail coefficient d j , j = 1, 2, 3... N, σ j is the standard deviation estimate of d j , N j is d jThe number of sampling points in, MAD|d j | represents calculating d j 's median absolute deviation.

[0065] Through the above formula, accurately calculate the denoising threshold of the wavelet signals at each layer, provide the accuracy rate of wavelet signal denoising, and further improve the accuracy rate of spectral data denoising.

[0066] As an alternative solution, transform the one-dimensional spectral data through an improved S transform to obtain a two-dimensional time-frequency spectrogram, including: calculating the frequency-domain window width through the Gaussian error function; calculating the expression of the window function of the improved S transform at the corresponding frequency according to the frequency-domain window width; calculating the two-dimensional time-frequency spectrogram corresponding to the one-dimensional spectral data according to the expression and the calculation formula of the improved S transform.

[0067] When using the improved S transform to transform the one-dimensional spectral data, first calculate the frequency-domain window width through the Gaussian error function. As an alternative solution, calculating the frequency-domain window width through the Gaussian error function includes: calculating through the calculation formula of the frequency-domain window width width f as follows:

[0068]

[0069] In the formula, f max is the maximum value of the signal frequency, f is the current frequency, a and b are the amplitude and slope adjustment parameters respectively, and GEF() is the Gaussian error function.

[0070] Through the formula containing the Gaussian error function, the correlation between the current frequency and the maximum value of the signal frequency can be reflected. Moreover, when the frequency increases, the frequency-domain window width increases, and when the frequency decreases, the frequency-domain window width decreases. Thus, both the frequency resolution and the time resolution are at a relatively high level, taking into account both the frequency resolution and the time resolution, and being able to better reflect the effective components in the spectral data.

[0071] Specifically, the Gaussian error function is:

[0072]

[0073] Then, calculate the expression of the window function of the improved S transform at the corresponding frequency according to the frequency-domain window width, including: calculating through the expression as follows:

[0074]

[0075] In the formula, W(α, f) is the expression of the Fourier transform of the window function w() at the frequency f.

[0076] What the above expression finally calculates is also the Gaussian window in the frequency domain. By using the frequency-domain Gaussian window and performing the S transform, a two-dimensional time-frequency spectrogram that takes into account both time-domain resolution and frequency-domain resolution can be obtained, which can better reflect the effective components in the spectral data and is conducive to subsequent feature extraction and fault detection based on the two-dimensional time-frequency spectrogram.

[0077] As an alternative solution, according to the expression and the calculation formula of the improved S transform, calculate the two-dimensional time-frequency spectrogram corresponding to the one-dimensional spectral data, including: The calculation formula of the improved S transform is as follows:

[0078]

[0079] In the formula, X(α) is the Fourier transform of the continuous signal x(t) of the one-dimensional spectral data, α is the frequency variable, with the unit of radian per second (rad / s), t is the time variable, with the unit of second (s), and S(τ, f) is the time-frequency representation at time τ and frequency f, which is also the two-dimensional time-frequency spectrogram corresponding to the one-dimensional spectral data.

[0080] The above calculation formula can accurately calculate the corresponding two-dimensional time-frequency spectrogram that takes into account both time-domain resolution and frequency-domain resolution based on the one-dimensional spectral data.

[0081] The above calculation formula can be the S transform formula for continuous signals and can be obtained by transforming the discrete formula. Let x(kT), k = 0, 1,..., N - 1 be the discrete sequence obtained by sampling the continuous signal x(t), T be the sampling time interval, and N be the number of sampling points. Then the discrete form of the S transform can be expressed as follows:

[0082]

[0083] Among them, the time index is represented by i: i = 0, 1,..., N - 1; the frequency index is represented by n: n = 0, 1,..., N - 1, and m is the intermediate index variable, m = 0, 1,..., N - 1.

[0084] As an alternative solution, through a convolutional neural network, spectral time-frequency features are extracted based on a two-dimensional time-frequency spectrogram, and the corresponding fault type is determined based on the spectral time-frequency features, including: extracting features from the two-dimensional time-frequency spectrogram through a feature extraction network to obtain a feature vector corresponding to the spectral time-frequency features, where the feature extraction network includes a convolutional neural network with multiple convolutional layers and pooling layers and is trained with training data; determining the corresponding fault type according to the spectral time-frequency features through a fault diagnosis network, where the fault diagnosis network includes a feedforward neural network, is trained with a training set for a preset number of training cycles, and is verified through a validation set, and the training set and the validation set are shuffled and re-partitioned in each training cycle; where the feedforward neural network includes an input layer, a hidden layer, and an output layer, the input layer is used to receive a one-dimensional feature vector of the spectral time-frequency features, the hidden layer is at least one fully connected layer for feature extraction and decision-making on the output data of the input layer, an activation function is also set after the fully connected layer of the hidden layer, and the output layer is a fully connected layer for outputting multiple nodes, and each node corresponds to the probability of a fault type.

[0085] The above-mentioned feature extraction network includes multiple convolutional layers and pooling layers, and these layers can extract features from local to global. As Figure 5 shown, the feature extraction network of this embodiment may include two convolutional layers and two pooling layers.

[0086] The convolutional layer captures local features through filters, and the above filters are also the Figure 5 general name of the rectified linear unit layer in, and its main part is the ReLU activation function, that is, the rectified linear unit activation function. The pooling layer is used to reduce the spatial dimension of the features and at the same time increase the invariance to input changes.

[0087] Through the combination of the above two convolutional layers and two pooling layers, the feature extraction network adopting the convolutional neural network CNN architecture can be trained with training data to learn feature representations from simple to complex. That is, after the feature extraction network is trained, it can directly calculate the feature vector corresponding to the corresponding spectral time-frequency features according to the two-dimensional time-frequency spectrogram quickly and accurately.

[0088] The above-mentioned fault diagnosis network includes a feedforward neural network, is trained with a training set for a preset number of training cycles, and is verified through a validation set, and the training set and the validation set are shuffled and re-partitioned in each training cycle. This helps to reduce overfitting and accelerate convergence.

[0089] As Figure 6As shown in the figure, the feedforward neural network includes an input layer, a hidden layer, and an output layer. The input layer is used to receive the one-dimensional feature vector of the spectral time-frequency features. The hidden layer is at least one fully connected layer, which is used to extract features and make decisions on the output data of the input layer. An activation function, such as the above-mentioned ReLU activation function, is also set after the fully connected layer of the hidden layer. The output layer is a fully connected layer, which is used to output multiple nodes, and each node corresponds to the probability of a fault type.

[0090] The input of the fault diagnosis network is the one-dimensional feature vector output by the feature extraction network, and the output is a 1×5 vector, and each element represents the probability of a fault category. If the values of all elements are 0 or less than the preset threshold, it means that there is no fault in the current state.

[0091] Thus, the above-mentioned feature extraction network and fault judgment network are pre-trained, and then used in actual use after training is completed. In actual use, the corresponding fault type can be quickly and accurately calculated according to the two-dimensional time-frequency spectrogram. Greatly improve the speed and efficiency of fault diagnosis and ensure the accuracy of fault determination.

[0092] As an optional solution, before obtaining the one-dimensional spectral data by enhancing the spectral data, the method further includes: using a dynamic gas mixer to mix a variety of gases stored in a gas tank according to the required ratio to obtain a standard mixed gas, where the variety of gases includes a background gas and a discharge gas, the background gas includes a mixed gas of sulfur hexafluoride SF6 and nitrogen N2, and the discharge gas includes at least one of the following: sulfur dioxide SO2, nitric oxide NO, carbon disulfide CS2; detecting the corresponding spectral data of the standard mixed gas through an infrared light source and an infrared spectrometer; using the spectral data corresponding to a variety of different required ratios as training data to train the convolutional neural network.

[0093] The above-mentioned required ratio may include the ratio of the background gas, the ratio of the discharge gas, and the ratio of the background gas to the discharge gas. When mixing the gases, a dynamic gas mixer is used to configure a series of background gases with different ratios of sulfur hexafluoride SF6 and nitrogen N2, and the discharge gases that may be generated under simulated discharge conditions.

[0094] The reason for choosing sulfur hexafluoride SF6 and nitrogen N2 as the background gas is that sulfur hexafluoride SF6 and nitrogen N2 do not have significant absorption of infrared light and will not be reflected in the spectral data, ensuring a high signal-to-noise ratio of the spectral data.

[0095] When using the spectral data corresponding to a variety of different required ratios as training data, it is necessary to record the corresponding discharge fault type of each mixed gas while collecting the spectral data to form the training data.

[0096] It should be noted that this embodiment also provides an alternative implementation manner, which will be described in detail below.

[0097] The present invention proposes a method for two-dimensional feature extraction of spectral signals based on improved S-transform and CNN. This method first converts one-dimensional spectral signals into two-dimensional spectrograms through improved S-transform, and then uses CNN to extract features from the two-dimensional spectrograms to achieve rapid and accurate diagnosis of discharge faults in GIS equipment. This method not only improves the efficiency and accuracy of feature extraction, but also provides a new technical means for fault diagnosis and preventive maintenance of power systems.

[0098] The method proposed in this embodiment aims to diagnose discharge faults in GIS equipment through improved S-transform and convolutional neural network (CNN). The overall process includes acquisition, preprocessing, two-dimensional processing, feature extraction, and fault diagnosis of spectral data. The overall flow chart is as Figure 2 shown, and the specific implementation processes of each part will be described in detail below.

[0099] 1. Acquisition of spectral data: Obtaining the spectral signals of the mixed gas generated by the discharge of GIS equipment is the primary step of this embodiment. The following are the detailed implementation steps:

[0100] Experimental system design: In order to collect the infrared spectral data of sulfur hexafluoride SF6 / nitrogen N2 mixed gas generated by GIS equipment under different discharge conditions, an experimental device was designed. This device includes the following main components:

[0101] Gas cylinder: Contains sulfur hexafluoride SF6 / nitrogen N2 mixed gas and possible decomposition product gases, such as sulfur dioxide SO2, nitric oxide NO, carbon disulfide CS2, etc.

[0102] Dynamic gas mixer: Used to accurately prepare mixed gases with different ratios to simulate the gas environment under different discharge conditions.

[0103] Infrared light source: Provides stable infrared light for exciting the vibrational energy level transitions of gas molecules.

[0104] Gas cell: Used to hold the mixed gas to ensure sufficient interaction length between the gas and infrared light.

[0105] Infrared spectrometer: Used to capture the spectral signals generated after gas molecules absorb infrared light.

[0106] Data storage / display PC: Used to record and display spectral data in real time for subsequent analysis.

[0107] Spectral data collection: During the experiment, first, a series of sulfur hexafluoride (SF6) / nitrogen (N2) mixed gases with different concentration ratios, as well as decomposition product gases that may be generated under simulated discharge conditions, were prepared using a dynamic gas mixer.

[0108] Then, the prepared gas was introduced into the gas cell, and infrared spectrometer was used to collect spectral data in the wavenumber range from 1700 cm -1 to 2700 cm -1 . In this wavenumber range, decomposition product gases such as sulfur dioxide (SO2), nitric oxide (NO), and carbon disulfide (CS2) have obvious absorption peaks, while the background gases of sulfur hexafluoride (SF6) and nitrogen (N2) have no significant absorption, thus ensuring a high signal-to-noise ratio of the spectral data.

[0109] Data marking: While collecting spectral data, the discharge fault type corresponding to each mixed gas was recorded. These marked information is crucial for subsequent training of the fault diagnosis model because they provide a direct association between the spectral signal and the fault type.

[0110] 2. Data preprocessing: Since there are inevitably background interference and random noise in the actual process of configuring mixed gases, it is necessary to further enhance the spectral data.

[0111] To reduce background interference, before conducting the experiment, the optical path was purged with a mixed gas of 30% sulfur hexafluoride (SF6) and 70% nitrogen (N2) to make the gas cell contain only the sulfur hexafluoride (SF6) / nitrogen (N2) mixed gas as the background. Then, the spectral data at this time was measured and recorded as Data0. On the premise of keeping the on-site experimental environment and equipment unchanged, all subsequent measured spectral data was subtracted from Data0, and the background spectrum could be better removed. At the same time, to suppress the random noise in the spectral data, wavelet threshold denoising technology was introduced. The specific process is as follows:

[0112] Step 1: Using db6 as the mother wavelet, the one-dimensional spectral data was decomposed into 5 layers to obtain the 5th layer approximation coefficient a5 and the first 5 layer detail coefficients d j , j = 1, 2,..., 5;

[0113] Step 2: Adaptive threshold estimation was performed on the detail coefficient d j , and the formula is as follows:

[0114]

[0115] In the formula, ε j is the denoising threshold of the jth layer detail coefficient d j , σ j is the standard deviation estimate of d j , N j is dj The number of sampling points in it. In addition, in the formula, MAD|d j | represents calculating the j median absolute deviation of d.

[0116] Step 3: For d j , use the estimated ε j to perform hard threshold denoising, that is:

[0117]

[0118] In the formula, d ji represents the i-th point of the detail coefficient of the j-th layer.

[0119] Step 4: Add the denoised wavelet decomposition detail coefficients and the approximation coefficients of the fifth layer to obtain the denoised one-dimensional spectral data. Specifically, use the inverse wavelet transform to reconstruct the denoised wavelet decomposition coefficients and the approximation coefficients of the fifth layer respectively, and add the reconstructed signals to obtain the denoised one-dimensional spectral data.

[0120] 3. Feature extraction network: After completing the acquisition and preprocessing of spectral data, the next step is feature extraction, which is a key step in realizing the discharge fault diagnosis of GIS equipment. The feature extraction part includes the following two main links:

[0121] Generating a two-dimensional spectral map by improving the S transform:

[0122] In this embodiment, the one-dimensional spectral signal is converted into a two-dimensional spectral map through an improved S transform. This step enables the original time-frequency analysis method to be utilized by subsequent image processing techniques. The improved S transform uses the Gaussian error function (GEF) to dynamically adjust the width of the frequency domain window to optimize the time-frequency resolution, and is particularly suitable for analyzing signals containing noise. Through this improvement, the high-frequency components in the signal can be captured more accurately, and the false components in the time-frequency map can be reduced.

[0123] As an effective time-frequency analysis method, the S transform (Stockwell Transform) has received the attention of researchers in various fields. This method is a reversible time-frequency localization technique that combines the respective advantages of the short-time Fourier transform (STFT) and the wavelet transform, mainly manifested as: 1) The S transform uses a variable-width window function, enabling it to obtain high-frequency domain resolution at the low-frequency end and high-time domain resolution at the high-frequency end; 2) The S transform can perform multi-resolution analysis while retaining the absolute phase of each frequency component.

[0124] Let \(x(kT)\), where \(k = 0,1,\cdots,N - 1\), be the discrete sequence obtained by sampling the continuous signal \(x(t)\) of one-dimensional spectral data. \(T\) is the sampling time interval and \(N\) is the number of sampling points. Then the discrete form of the S-transform can be expressed as follows:

[0125]

[0126] In the formula, the time index is represented by \(i\): \(i = 0,1,\cdots,N - 1\); the frequency index is represented by \(n\): \(n = 0,1,\cdots,N - 1\), and \(m\) is an intermediate index variable, \(m = 0,1,\cdots,N - 1\).

[0127] Correspondingly, the continuous S-transform can be written in the following form:

[0128]

[0129] In the formula, \(X(\alpha)\) is the Fourier transform of the continuous signal \(x(t)\), α is the frequency variable, with the unit of radians per second (rad / s), t is the time variable, with the unit of seconds (s), and \(W(\alpha,f)\) is the expression of the Fourier transform of the window function \(w(t)\) at frequency \(f\): .

[0130] The above formula is the Gaussian window in the frequency domain. The \(f\) in the above formula corresponds to the width \(width\) of the frequency domain window. In the actual test calculation, first calculate the width \(width\) of the frequency domain window f , and then substitute it into \(f\) in the above formula for calculation. That is to say, the above formula can be equivalent to f . Through this improvement, the calculated two-dimensional spectrogram takes into account both the time domain resolution and the frequency domain resolution, and can thus better reflect the effective components in the spectral data, which is beneficial to the subsequent feature extraction based on the spectrogram. .

[0131] Obviously, according to the above formula, it can be determined that when the signal frequency \(f\) increases, the width of the frequency domain window will increase, resulting in a poor frequency domain resolution of the S-transform, which is not conducive to analyzing the high-frequency components in the signal. Further, if there is noise interference in the signal, it may also cause a large number of false components to appear in the high-frequency part of the time-frequency diagram, seriously affecting the extraction of useful features.

[0132] In actual detection, the improved formula of the continuous S-transform is directly used to calculate the one-dimensional spectral data and convert it into the corresponding two-dimensional time-frequency spectrogram,

[0133] Therefore, in this embodiment, a Gaussian error function (GEF) is designed. The Gaussian error function here is used both in training and actual detection because it is a tool for converting one-dimensional spectral signals into two-dimensional spectrograms. Whether it is training or actual detection, it can only be carried out after obtaining the two-dimensional spectrogram.

[0134] The frequency domain window width width as shown below is obtained f Calculation formula:

[0135]

[0136] In the formula, f max is the maximum value of the signal frequency, f is the current frequency, and a and b are the amplitude and slope adjustment parameters respectively. GEF() is the Gaussian error function, and the expression is:

[0137]

[0138] By appropriately adjusting the parameters a and b in the above formula, the constructed frequency domain window width can meet the optimal design requirements, which can not only ensure the frequency domain resolution of the high-frequency part of the signal but also take into account the time domain resolution of the low-frequency part.

[0139] The amplitude and slope adjustment parameters a and b are hyperparameters, not trainable parameters, and are artificially specified according to different application scenarios. Of course, an algorithm for automatically optimizing the optimal a and b parameters can also be considered, such as an iterative optimization algorithm, etc.

[0140] Figure 3 is a schematic diagram of the processing result of the unimproved standard S transform of the embodiment of the present invention for multi-frequency signals, Figure 4 is a schematic diagram of the processing result of the improved S transform of the embodiment of the present invention for multi-frequency signals, as Figure 3 and Figure 4 shown, which intuitively shows the results of time-frequency transformation of a group of one-dimensional signals containing different frequency components using the standard S transform and the improved S transform proposed in this embodiment. Specifically, Figure 4 in this embodiment, the amplitude and slope adjustment parameters a and b of the improved S transform based on the Gaussian error function are respectively taken as a = 0.41 and b = 2.2. Obviously, from Figure 3 and Figure 4 the comparison shows that the improved S transform of this embodiment can achieve a good balance between frequency domain resolution and time domain resolution.

[0141] 4. Feature Extraction Network: The improved S-transform converts one-dimensional spectral data into a two-dimensional time-frequency spectrogram, where the x-axis represents time, the y-axis represents frequency, and the color intensity represents the signal intensity at that time-frequency point. In this way, the original one-dimensional signal is converted into an image, and the effective CNN in the image field can be used for spectral time-frequency feature extraction.

[0142] In this embodiment, the convolutional neural network CNN is used to identify and extract features related to the discharge fault of GIS equipment in the two-dimensional spectrogram.

[0143] As Figure 5 shown, the structure of the convolutional neural network CNN includes multiple two-dimensional convolutional layers and two-dimensional average pooling layers, which can extract local to global features from one-dimensional spectral data. The two-dimensional convolutional layer captures local features through filters, while the two-dimensional average pooling layer is used to reduce the spatial dimension of features and increase the invariance to input changes. Through the combination of these two-dimensional convolutional layers and two-dimensional average pooling layers, CNN can learn feature representations from simple to complex, providing strong feature support for subsequent fault diagnosis.

[0144] In the design of the feature extraction network, parameter settings of different layers are considered, including the size and number of convolutional kernels in the two-dimensional convolutional layer, and the type and size of the two-dimensional average pooling layer, etc., to ensure that the network can effectively capture the key information in the two-dimensional spectrogram. These parameter settings are based on experiments and the experience of pre-trained models to achieve the best performance.

[0145] Through the above feature extraction steps, a set of feature vectors representing the discharge fault of GIS equipment can be obtained, and these feature vectors will be used for subsequent network training and fault diagnosis.

[0146] 5. Fault Diagnosis Network: After feature extraction, the core of network design is to build a fault diagnosis model that can convert the output of the feature extraction network into a prediction of fault categories.

[0147] The input of the fault diagnosis network is the one-dimensional discharge fault feature vector output by the feature extraction network, and the output is a 1×5 vector, where each element represents the probability of a fault category. If all element values are 0 or less than a preset threshold, it means that no fault occurs in the current state.

[0148] As Figure 6 shown, the fault diagnosis network can be designed as a simple feedforward neural network, including:

[0149] 1. Input layer: Accepts the one-dimensional discharge fault feature vector output by the feature extraction network.

[0150] 2. Hidden layer: One or more fully connected layers, which are used to further extract features and make decisions. After each hidden layer, a rectified linear unit layer (ReLU layer) and an activation function, such as ReLU (rectified linear unit), can be set to introduce non-linear mapping properties.

[0151] 3. Output layer: A fully connected layer that outputs N nodes, and each node corresponds to the probability of a fault category.

[0152] Loss function: Since this is a multi-classification problem, the categorical cross-entropy loss can be selected as the loss function, which measures the difference between the probability distribution predicted by the model and the true labels.

[0153] For each node in the output layer, a threshold (e.g., 0.5) is set. If the output value of the node is greater than this threshold, it is considered that the fault category has occurred; if the output values of all nodes are less than or equal to the threshold, it is considered that there is no fault in the current state.

[0154] In the training parameter settings, the maximum number of iterations is 50, that is, the training process runs for at most 50 epochs; the mini-batch size is set to 32, that is, the number of one-dimensional spectral data and two-dimensional spectral data in each batch are both 32; the initial learning rate is set to 0.001; the validation frequency is set to 10, that is, after 10 training epochs, the performance is evaluated on the validation set; and it is set that at the beginning of each training epoch, the training data is shuffled, which helps to reduce overfitting and accelerate convergence.

[0155] The trained model can be deployed in the monitoring system of GIS equipment to achieve real-time discharge fault diagnosis. When new spectral signals are collected and feature extracted, the model can quickly identify the corresponding fault types and provide timely fault information for operators.

[0156] In summary, the fault diagnosis network of this embodiment can accurately convert the extracted features into predictions of fault categories, providing an efficient and accurate technical means for the fault diagnosis of GIS equipment.

[0157] Based on the above-mentioned discharge fault diagnosis method based on improved S-transform and convolutional neural network provided by the embodiments of the present invention, the embodiments of the present invention also provide a discharge fault diagnosis device based on improved S-transform and convolutional neural network, which is applied to the discharge fault diagnosis of GIS equipment. The device includes: a data acquisition module, a data enhancement module, a data transformation module, and a fault detection module. The device will be described in detail below.

[0158] An acquisition data module for acquiring spectral data of gas generated by the discharge of a gas-insulated switchgear;

[0159] A data enhancement module connected to the above acquisition data module for enhancing the spectral data to obtain one-dimensional spectral data;

[0160] A data transformation module connected to the above data enhancement module for converting the one-dimensional spectral data through an improved S-transform to obtain a two-dimensional time-frequency spectrogram, where the improved S-transform is an S-transform function that dynamically adjusts the width of the frequency-domain window through a Gaussian error function;

[0161] A fault detection module connected to the above data transformation module for extracting spectral time-frequency features according to the two-dimensional time-frequency spectrogram through a convolutional neural network and determining the corresponding fault type according to the spectral time-frequency features.

[0162] The above discharge fault diagnosis device based on the improved S-transform and convolutional neural network provided by the embodiments of the present invention creates, through enhancing the spectral data of the gas generated by the discharge of the gas-insulated switchgear, the obtained one-dimensional spectral data eliminates background interference and random noise, having a more accurate technical effect; and converts the one-dimensional spectral data through the improved S-transform to obtain a two-dimensional time-frequency spectrogram. The improved S-transform sets a Gaussian error function to dynamically adjust the width of the frequency-domain window of the S-transform to optimize the time-frequency resolution, which is particularly suitable for analyzing signals containing noise, can more accurately capture the high-frequency components in the signal, and reduce the false components in the finally obtained two-dimensional time-frequency spectrogram. Extract spectral time-frequency features according to the two-dimensional time-frequency spectrogram through a convolutional neural network and determine the corresponding fault type according to the spectral time-frequency features, thereby quickly and accurately performing fault diagnosis and improving the efficiency and accuracy of fault diagnosis.

[0163] Figure 7 It is a flowchart of a method for extracting spectral signal features of a GIS device according to an embodiment of the present invention. As Figure 7 shown, according to another aspect of the embodiments of the present invention, a method for extracting spectral signal features of a GIS device is also provided, including the following steps:

[0164] Step S701, enhancing the spectral data of the mixed gas to obtain one-dimensional spectral data;

[0165] Step S702, converting the one-dimensional spectral data through the improved S-transform to obtain a two-dimensional time-frequency spectrogram, where the improved S-transform is an S-transform function that dynamically adjusts the width of the frequency-domain window through a Gaussian error function;

[0166] Step S703, extracting spectral time-frequency features according to the two-dimensional time-frequency spectrogram through a convolutional neural network as spectral signal features.

[0167] The spectral signal feature extraction method provided by the embodiment of the present invention creates enhanced processing through spectral data, so that the obtained one-dimensional spectral data eliminates background interference and random noise, and has a more accurate technical effect; and the improved S transform is used to transform the one-dimensional spectral data to obtain a two-dimensional time-frequency spectrogram. The improved S transform sets a Gaussian error function to dynamically adjust the width of the frequency domain window of the S transform to optimize the time-frequency resolution. It is particularly suitable for analyzing signals containing noise, can more accurately capture the high-frequency components in the signal, and reduce the false components in the finally obtained two-dimensional time-frequency spectrogram. The spectral time-frequency features are extracted according to the two-dimensional time-frequency spectrogram through a convolutional neural network. It achieves the technical effect of quickly and accurately extracting spectral signal features to accurately reflect the components of the mixed gas, so as to improve the accuracy of mixed gas component identification and detection.

[0168] In the scenario of GIS equipment detection, the above spectral signal feature extraction method is used to determine the corresponding fault type in combination with spectral signal features, so as to quickly and accurately perform fault diagnosis, improving the efficiency and accuracy of fault diagnosis. Furthermore, it solves the problems in the detection of gas-insulated switchgear in the related art that rely on chemical sensors or chromatographic analysis, with slow response speed, high cost, and poor real-time performance, and achieves the technical effects of improving the response speed, not requiring professional sensors and analysis devices, reducing costs, and ensuring the real-time performance of fault detection with a high detection speed.

[0169] The embodiment of the present invention also provides a non-transitory machine-readable medium storing a computer program, wherein the above computer program, when executed by a processor of a computer, is used to cause the computer to execute the method of the embodiment of the present invention.

[0170] The embodiment of the present invention also provides a computer program product, including a computer program, wherein the computer program, when executed by a processor of a computer, is used to cause the computer to execute the method of the embodiment of the present invention.

[0171] The embodiment of the present invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor. The above memory stores a computer program that can be executed by the at least one processor, and the above computer program, when executed by the at least one processor, is used to cause the electronic device to execute the method of the embodiment of the present invention.

[0172] Reference Figure 8, the structural block diagram of an electronic device such as a server or a client that can be an embodiment of the present invention will now be described. It is an example of a hardware device that can be applied to various aspects of the present invention. The electronic device is intended to represent various forms of digital electronic computer devices, such as, laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as, personal digital processors, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present invention described and / or claimed herein.

[0173] As Figure 8 shown, the electronic device includes a computing unit 801, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 802 or a computer program loaded from a storage unit 808 into a random access memory (RAM) 803. In the RAM 803, various programs and data required for the operation of the electronic device can also be stored. The computing unit 801, the ROM 802, and the RAM 803 are connected to each other via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0174] A plurality of components in the electronic device are connected to the I / O interface 805, including: an input unit 806, an output unit 807, a storage unit 808, and a communication unit 809. The input unit 806 can be any type of device that can input information into the electronic device. The input unit 806 can receive input digital or character information, and generate key signal inputs related to the user settings and / or function controls of the electronic device. The output unit 807 can be any type of device that can present information, and can include, but is not limited to, a display, a speaker, a video / audio output terminal, a vibrator, and / or a printer. The storage unit 808 can include, but is not limited to, magnetic disks, optical disks. The communication unit 809 allows the electronic device to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks, and can include, but is not limited to, a modem, a network card, an infrared communication device, and / or a wireless communication transceiver, such as a Bluetooth device, a WiFi device, a WiMax device, a cellular communication device, and / or the like.

[0175] The computing unit 801 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a CPU, a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing units, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 801 executes the various methods and processes described above. For example, in some embodiments, the method embodiments of the present invention can be implemented as a computer program, which is tangibly contained in a machine-readable medium, such as the storage unit 808. In some embodiments, part or all of the computer program can be loaded and / or installed onto the electronic device via the ROM 802 and / or the communication unit 809. In some embodiments, the computing unit 801 can be configured to execute the above-described method in any other suitable manner (e.g., by means of firmware).

[0176] The computer program for implementing the method of the embodiments of the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when the computer programs are executed by the processor or controller, the functions / operations specified in the flowchart and / or block diagram are implemented. The computer program can be executed entirely on the machine, partially on the machine, as an independent software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0177] In the context of the embodiments of the present invention, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable signal medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, or infrared system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0178] It should be noted that the term "including" and its variants used in the embodiments of the present invention are open-ended, that is, "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The modifications of "one" and "a plurality" mentioned in the embodiments of the present invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise clearly specified in the context, it should be understood as "one or more".

[0179] In the method embodiments provided by the embodiments of the present invention, the steps described can be executed in different orders and / or executed in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The protection scope of the present invention is not limited in this regard.

[0180] The term "embodiment" in this specification means that the specific features, structures or characteristics described in connection with the embodiment may be included in at least one embodiment of the present invention. The phrase appears in various positions in the specification does not necessarily mean the same embodiment, nor does it mean being independent or alternative to other embodiments and mutually exclusive. The various embodiments in this specification are described in a related manner, and the same or similar parts among the various embodiments are cross-referred to. In particular, for the device, equipment, and system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts refer to the partial description of the method embodiments.

[0181] The above-described embodiments only represent several implementation manners of the present invention, and the description is relatively specific and detailed, but it should not be construed as a limitation on the protection scope. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the appended claims.

Claims

1. A method for diagnosing discharge faults of GIS equipment, characterized in that, Including: Obtaining spectral data of the gas generated by the discharge of a gas-insulated switchgear; Subtracting the background spectral data of the background gas from the spectral data to obtain spectral data with background interference removed; Using a target wavelet function as the mother wavelet to perform N-layer wavelet decomposition on the spectral data with background interference removed, obtaining the approximation coefficient of the Nth layer and the detail coefficients of the first N layers; Performing adaptive threshold estimation on the detail coefficients to obtain the denoising threshold corresponding to the detail coefficients; Performing hard threshold denoising processing on the corresponding detail coefficients through the denoising threshold to obtain denoised detail coefficients; Performing wavelet reconstruction, and reconstructing the N-layer decomposition signal according to the denoised detail coefficients and the approximation coefficient to obtain enhanced one-dimensional spectral data; Calculating the frequency domain window width through a Gaussian error function; Calculating the expression of the window function of the improved S-transform at the corresponding frequency according to the frequency domain window width; Calculating the two-dimensional time-frequency spectrogram corresponding to the one-dimensional spectral data according to the expression and the calculation formula of the improved S-transform, where the improved S-transform is an S-transform function that dynamically adjusts the frequency domain window width through a Gaussian error function; Extracting spectral time-frequency features through a convolutional neural network according to the two-dimensional time-frequency spectrogram, and determining the corresponding fault type according to the spectral time-frequency features.

2. The method according to claim 1, characterized in that, Performing adaptive threshold estimation on the detail coefficients to obtain the denoising threshold corresponding to the detail coefficients, including: Performing adaptive estimation on the detail coefficients through the following formula to obtain the corresponding denoising threshold: ； In the formula, ε j is the j detail coefficient of the d j th layer, where j = 1, 2, 3…N, σ j is the d j estimated standard deviation of N j is the d j number of sampling points in d j |MAD| d j represents the calculation of the absolute median difference of 3. The method according to claim 1, characterized in that, Calculating the frequency domain window width through a Gaussian error function, including: Through the frequency domain window width width f perform the calculation according to the following calculation formula: ； wherein, f max is the maximum value of the signal frequency, f is the current frequency, a , b are the amplitude and slope adjustment parameters respectively, GEF () is the Gaussian error function; Calculating the expression of the window function of the improved S-transform at the corresponding frequency according to the frequency domain window width, including: Calculating through the expression, and the expression is as follows: ； In the formula, W(α, f) is the expression of the Fourier transform of the window function w() at frequency f, α is the frequency variable, with the unit of radian per second (rad / s).

4. The method according to claim 3, wherein Calculating the two-dimensional time-frequency spectrogram corresponding to the one-dimensional spectral data according to the expression and the calculation formula of the improved S-transform, including: The calculation formula of the improved S-transform is as follows: ； In the formula, X(α) is the continuous signal of the one-dimensional spectral data x(t) after Fourier transform, t is the time variable, with the unit of second (s), S(τ, f) is the time τ and f is the time-frequency representation at the frequency, that is, the two-dimensional time-frequency spectrogram corresponding to the one-dimensional spectral data.

5. The method according to any one of claims 1 to 4, characterized in that Extracting spectral time-frequency features through a convolutional neural network according to the two-dimensional time-frequency spectrogram, and determining the corresponding fault type according to the spectral time-frequency features, including: Performing feature extraction on the two-dimensional time-frequency spectrogram through a feature extraction network to obtain a feature vector corresponding to the spectral time-frequency features, where the feature extraction network includes a convolutional neural network with multiple convolutional layers and pooling layers, and is trained through training data; Determining the corresponding fault type through a fault diagnosis network according to the spectral time-frequency features, where the fault diagnosis network includes a feedforward neural network, is trained through a preset number of training cycles of a training set, and is verified through a validation set. In each training cycle, the training set and the validation set are shuffled and re-partitioned. Among them, the feedforward neural network includes an input layer, a hidden layer, and an output layer. The input layer is used to receive the one-dimensional feature vector of the spectral time-frequency features. The hidden layer is at least one fully connected layer, which is used to extract features and make decisions on the output data of the input layer. An activation function is also set after the fully connected layer of the hidden layer. The output layer is a fully connected layer, which is used to output multiple nodes, and each node corresponds to the probability of a fault type.

6. The method according to any one of claims 1 to 4, characterized in that, Before performing enhancement processing on the spectral data to obtain one-dimensional spectral data, the method further includes: Using a dynamic gas mixer to mix various gases stored in a gas tank in accordance with required ratios to obtain a standard mixed gas. Among them, the various gases include a background gas and a discharge gas. The background gas includes a mixed gas of sulfur hexafluoride (SF6) and nitrogen (N2). The discharge gas includes at least one of the following: sulfur dioxide (SO2), nitric oxide (NO), and carbon disulfide (CS2); Detecting the corresponding spectral data of the standard mixed gas through an infrared light source and an infrared spectrometer; Using the spectral data corresponding to various different required ratios as training data to train the convolutional neural network.

7. A method for extracting spectral signal features of a GIS device, characterized in that, Including: Subtracting the background spectral data of the background gas from the spectral data to obtain spectral data with background interference removed; Using a target wavelet function as the mother wavelet to perform N-layer wavelet decomposition on the spectral data with background interference removed to obtain the approximation coefficient of the Nth layer and the detail coefficients of the first N layers; Performing adaptive threshold estimation on the detail coefficients to obtain the denoising threshold corresponding to the detail coefficients; Performing hard threshold denoising processing on the corresponding detail coefficients through the denoising threshold to obtain denoised detail coefficients; Calculating the frequency domain window width through a Gaussian error function; Calculating the expression of the window function of the improved S transform at the corresponding frequency according to the frequency domain window width; Calculating the two-dimensional time-frequency spectrogram corresponding to the one-dimensional spectral data according to the expression and the calculation formula of the improved S transform, where the improved S transform is an S transform function that dynamically adjusts the frequency domain window width through a Gaussian error function; Extracting spectral time-frequency features from the two-dimensional time-frequency spectrogram through a convolutional neural network as the spectral signal features.

8. An electronic device, comprising: A processor and a memory storing a program, characterized in that the program includes instructions that, when executed by the processor, cause the processor to execute the discharge fault diagnosis method of the GIS device according to any one of claims 1 to 6, or the spectral signal feature extraction method of the GIS device according to claim 7.

Images