Mixed gas insulation electric power equipment fault diagnosis system and method

Abstract

The invention discloses a mixed gas insulation electric power equipment fault diagnosis system and method, and the system comprises a multi-modal data obtaining module which is used for obtaining the multi-modal data of GIS equipment, and the multi-modal data comprises a voiceprint signal, an ultrasonic signal, and an ultrahigh frequency signal; the abnormity analysis module is used for sampling and detecting characteristic decomposition products of the mixed gas when abnormity occurs in the multi-modal data, and calculating three ratios; and the fault diagnosis module is used for performing fusion diagnosis analysis based on the voiceprint signal, the ultrasonic signal, the ultrahigh frequency signal and the three-ratio to obtain the fault type and the fault severity. According to the method, by combining online monitored multi-modal data and offline detected characteristic decomposition product components and contents, the accidental defect of electric signal capture is made up, the defect that misjudgment may be caused only by means of instantaneous acoustoelectric signals is avoided, and analysis of equipment fault types and severity is achieved.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring technology, specifically to a fault diagnosis system and method for mixed gas insulated power equipment. Background Technology

[0002] With the accelerated construction of new power systems, the number of power equipment in my country has experienced explosive growth, bringing new challenges to the safe and stable operation of the power grid. Accurate sensing of power grid equipment status, reliable transmission of sensing data, precise assessment of equipment status, and early warning of defects / faults are fundamental to ensuring the safe and stable operation of the power grid.

[0003] Current sensing technologies suffer from limitations such as incomplete and singular online monitoring instruments, low accuracy, and inability to provide early warnings of defects. Optical and electrical signal online sensing devices are hampered by power supply constraints and insufficient self-powered output, making it difficult to achieve high acquisition frequencies and long-term online operation. Furthermore, insufficient sensor network scale and incomplete coverage make reliable data transmission challenging. Therefore, traditional methods relying on single-mode data analysis for fault diagnosis, such as ultra-high frequency pulse counting, are ill-suited for complex faults. Since traditional gas-insulated electrical equipment typically uses sulfur hexafluoride (SF6), a gas with a strong greenhouse effect (its global warming potential is approximately 24,300 times that of CO2), the reduction and replacement of SF6 emissions are imperative given the increasingly severe global environmental situation. Currently, the power industry primarily employs two technical solutions: one is the use of SF6 mixed gases (such as SF6 / N2), and the other is the use of new, environmentally friendly mixed gases with low greenhouse effects (such as perfluoroisobutyronitrile / carbon dioxide, C4F7N / CO2). For SF6 / N2 mixed gas or C4F7N / CO2 mixed gas insulation equipment, the type, energy level, and location of equipment faults can be obtained by detecting and analyzing the characteristic decomposition products of the mixed gas.

[0004] Therefore, this paper proposes a method for online monitoring, early warning and diagnosis of defects in mixed gas GIS equipment based on photoelectric signals and multi-dimensional characteristic parameters of gas characteristic decomposition products. Summary of the Invention

[0005] The technical problem to be solved by this invention is to accurately determine the fault type and severity of GIS equipment using multimodal data.

[0006] The present invention solves the above-mentioned technical problems through the following technical means: A fault diagnosis system and method for mixed gas insulated electrical equipment, including: Multimodal data acquisition module: used to acquire multimodal data from GIS equipment, including acoustic signature signals, ultrasonic signals, and ultra-high frequency signals; Anomaly Analysis Module: When anomalies occur in multimodal data, sample and detect characteristic decomposition products of the mixed gas, and calculate the three ratios. Fault diagnosis module: Based on the fusion diagnosis analysis of acoustic signal, ultrasonic signal, ultra-high frequency signal and three ratios, the fault type and fault severity are obtained.

[0007] Furthermore, the fusion diagnostic analysis based on acoustic signature signals, ultrasonic signals, ultra-high frequency signals, and the composition and content of characteristic decomposition products to obtain fault diagnosis results and issue early warnings includes: The Transformer model was used to process the voiceprint signal, and the fault analysis result was obtained. By using a graph neural network to process the UHF signal, fault analysis result two was obtained. The ultrasonic signal was processed using a generative adversarial network to obtain fault analysis result three; The components and contents of the characteristic decomposition products were analyzed using the three-ratio method, and the failure analysis results were obtained. Based on fault analysis results 1, 2, 3, and 4, a multimodal uncertainty fusion diagnostic analysis is performed to obtain fault diagnosis results and issue early warnings.

[0008] Furthermore, before performing the fusion diagnostic analysis based on the acoustic signature signal, ultrasonic signal, ultra-high frequency signal, and the components and content of characteristic decomposition products, the acoustic signature signal, ultrasonic signal, and ultra-high frequency signal are cleaned and aligned to obtain... , in, These are the denoising time sequences for voiceprint signals, ultrasonic signals, and ultra-high frequency signals, respectively. This is a time series of partial discharge acoustic-optical-electrical integrated signals.

[0009] Furthermore, the three-ratio method is used to analyze the components and content of characteristic decomposition products. The three-ratio calculation method is as follows: based on previous research on the content of several types of decomposition products under different fault types and severity conditions, a ratio method is constructed to establish a correspondence with fault type and severity for direct application in practical scenarios; the several types of decomposition products are defined as characteristic decomposition products; the time series sequence of characteristic decomposition products is as follows: ,in, The ratio of the content of characteristic decomposition product gas components. , This represents the peak area ratio of chromatographic data for different characteristic decomposition products.

[0010] Furthermore, the calculation process for the fault type and fault severity is as follows: All modal data are aligned to a uniform time scale to form a synchronous multimodal time series:

[0011] After extracting features using an LSTM network, the evaluation results are output through a fully connected layer of Softmax.

[0012] This invention also provides a fault diagnosis system and method for mixed gas insulated power equipment, comprising: Steps for acquiring multimodal data: This involves acquiring multimodal data from GIS equipment, including acoustic signature signals, ultrasonic signals, and ultra-high frequency signals. The steps of anomaly analysis are as follows: When anomalies occur in multimodal data, sample and detect characteristic decomposition products of the mixed gas, and calculate the three ratios. The steps of fault diagnosis are as follows: Based on the acoustic signal, ultrasonic signal, ultra-high frequency signal and the three ratios, a fusion diagnostic analysis is performed to obtain the fault type and fault severity.

[0013] Furthermore, the fusion diagnostic analysis based on acoustic signature signals, ultrasonic signals, ultra-high frequency signals, and the composition and content of characteristic decomposition products to obtain fault diagnosis results and issue early warnings includes: The Transformer model was used to process the voiceprint signal, and the fault analysis result was obtained. By using a graph neural network to process the UHF signal, fault analysis result two was obtained. The ultrasonic signal was processed using a generative adversarial network to obtain fault analysis result three; The components and contents of the characteristic decomposition products were analyzed using the three-ratio method, and the failure analysis results were obtained. Based on fault analysis results 1, 2, 3, and 4, a multimodal uncertainty fusion diagnostic analysis is performed to obtain fault diagnosis results and issue early warnings.

[0014] Furthermore, before performing the fusion diagnostic analysis based on the acoustic signature signal, ultrasonic signal, ultra-high frequency signal, and the components and content of characteristic decomposition products, the acoustic signature signal, ultrasonic signal, and ultra-high frequency signal are cleaned and aligned to obtain... , in, These are the denoising time sequences for voiceprint signals, ultrasonic signals, and ultra-high frequency signals, respectively. This is a time series of partial discharge acoustic-optical-electrical integrated signals.

[0015] Furthermore, the three-ratio method is used to analyze the components and content of characteristic decomposition products. The three-ratio calculation method is as follows: based on previous research on the content of several types of decomposition products under different fault types and severity conditions, a ratio method is constructed to establish a correspondence with fault type and severity for direct application in practical scenarios; the several types of decomposition products are defined as characteristic decomposition products; the time series sequence of characteristic decomposition products is as follows: ,in, The ratio of the content of characteristic decomposition product gas components. , This represents the peak area ratio of chromatographic data for different characteristic decomposition products.

[0016] Furthermore, the calculation process for the fault type and fault severity is as follows: All modal data are aligned to a uniform time scale to form a synchronous multimodal time series:

[0017] After extracting features using an LSTM network, the evaluation results are output through a fully connected layer of Softmax.

[0018] The advantages of this invention are: When a potential insulation defect begins to produce a weak partial discharge, UHF signals can capture this early warning immediately. However, electrical signals (UHF signals) are usually susceptible to electromagnetic and noise interference in the field, making it impossible to accurately quantify the PD discharge quantity. Ultrasonic signals and vibration fingerprints can roughly determine the physical location of the discharge or mechanical vibration, providing key spatial information for judging the fault type (such as free metal particles or suspended potential bodies). However, acoustic signals (ultrasound and acoustic fingerprints, etc.) are difficult to calibrate the discharge quantity and have extremely poor anti-interference capabilities. Furthermore, the generation of electrical and acoustic signals is an instantaneous process, making them difficult to capture. Therefore, the fusion of UHF signals, ultrasonic signals, and acoustic fingerprint signals cannot provide an objective and accurate assessment of the internal insulation status of GIS equipment. On the other hand, the detection of decomposition products in the mixed gas filled into the GIS equipment indicates the presence of defects in the equipment, such as discharge or overheating faults. The fault type can be determined by the decomposition products, thus compensating for the randomness of electrical signal capture.

[0019] Therefore, this application monitors the multimodal data of GIS equipment online in real time. When anomalies occur in the multimodal data, it can be preliminarily considered that a fault has occurred. Further sampling and detection of the characteristic decomposition product components and contents of the mixed gas, combined with the three ratios of the characteristic decomposition products of the mixed gas, can determine the fault type and severity. Therefore, the intelligent equipment status assessment system and centimeter-level partial discharge location method based on the characteristic parameters of decomposition products and the acoustic-optical-electric partial discharge signal of the mixed gas established by this invention, constructs fault severity classification criteria and diagnostic algorithms, and realizes early and rapid warning of equipment latent defects and judgment of fault type and severity.

[0020] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation thereof. Figure 1 This is a flowchart illustrating a fault diagnosis system and method for mixed gas insulated power equipment according to an embodiment of the present invention. Figure 2 This is a fault diagnosis architecture diagram that integrates multi-source information in one embodiment of the present invention; Figure 3 This is a schematic diagram of the principle of a fault diagnosis system and method for mixed gas insulated power equipment according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the internal component arrangement of a multi-parameter online monitoring device for mixed gas in one embodiment of the present invention; Figure 5 This is a schematic diagram of the appearance of a C4F7N / CO2 gas composition detection device according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the distribution of experimental data on discharge and overheating defect forms in a triangular coordinate system composed of three ratio "characteristic quantity groups" in one embodiment of the present invention; Figure 7 This is a preliminary diagram showing the defect distribution area displayed by the three ratio "feature quantity group" under triangular coordinates in one embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of an early warning and diagnosis system for defects in mixed gas GIS equipment based on multimodal data, proposed in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figures 1 to 2 As shown, the first embodiment of the present invention proposes a fault diagnosis system and method for mixed gas insulated power equipment. The system includes: Multimodal data acquisition module: used to acquire multimodal data from GIS equipment, including acoustic signature signals, ultrasonic signals, and ultra-high frequency signals; Anomaly Analysis Module: When anomalies occur in multimodal data, samples are taken to detect the components and content of decomposition products of the mixed gas. Fault diagnosis module: Based on multimodal data, component and content of characteristic decomposition products, fusion diagnostic analysis is performed to obtain fault type and fault severity.

[0024] The fault diagnosis method using the above system is described in detail below, and the method includes the following steps: S10. Online monitoring of multimodal data of GIS equipment, wherein the multimodal data includes acoustic signature signals, ultrasonic signals, ultra-high frequency signals, and the composition and content of decomposition products of mixed gas; It should be noted that vibration acoustic signals and ultrasonic signals can be collected by vibration sensors and ultrasonic sensors, ultra-high frequency signals can be collected by partial discharge sensors, and the decomposition products of mixed gases can be collected by gas composition detection devices.

[0025] S20. When anomalies are found in multimodal data, sample and detect the components and content of the decomposition products of the mixed gas. It should be noted that when anomalies are found in multimodal data, it can be preliminarily determined that the GIS equipment has malfunctioned. Further sampling and detection of the characteristic decomposition product components of the mixed gas can be performed. By combining the characteristic decomposition product components detected offline, early and refined fault diagnosis can be carried out.

[0026] S30. Based on the fusion diagnostic analysis of acoustic signature signal, ultrasonic signal, ultra-high frequency signal and the composition and content of characteristic decomposition products, fault diagnosis results are obtained and early warning is given.

[0027] It should be noted that this embodiment monitors the multimodal data of GIS equipment online in real time. When anomalies occur in the multimodal data, a fault can be initially considered to have occurred. Further sampling and detection of the characteristic decomposition product components and contents of the mixed gas are performed. By combining the characteristic decomposition product components and contents detected offline, early and refined fault diagnosis is carried out. Through multi-parameter uncertainty fusion fault diagnosis analysis at the probability distribution level, the weakness and intermittency of early electrical and acoustic signals are overcome, and the defects that may be caused by misjudgment by relying solely on instantaneous electrical and acoustic signals are overcome. This enables accurate perception and judgment of early trace signals. Therefore, the intelligent equipment status assessment system and centimeter-level partial discharge location method based on the characteristic parameters of decomposition products and acoustic-optical-electric partial discharge signals established in this invention, along with the construction of fault severity classification criteria and diagnostic algorithms, realize early and rapid warning of latent defects in GIS equipment and fault type diagnosis and location.

[0028] As a further preferred technical solution, such as Figure 3 As shown, step S30, which involves performing a fusion diagnostic analysis based on acoustic signature signals, ultrasonic signals, ultra-high frequency signals, and the components and content of characteristic decomposition products to obtain fault diagnosis results and issue an early warning, specifically includes the following steps: S31. The voiceprint signal is processed using the Transformer model to obtain fault analysis result one; S32. Using a graph neural network to process the UHF signal, we obtain fault analysis result two; S33. Using generative adversarial networks to process ultrasonic signals, we obtain fault analysis result three; S34. The components and contents of the characteristic decomposition products were analyzed using the three-ratio method to obtain the fourth fault analysis result. S35 performs multimodal uncertainty fusion diagnostic analysis based on fault analysis results one, two, three, and four to obtain fault diagnosis results and issue warnings.

[0029] It should be noted that Fault Analysis Result 1, Fault Analysis Result 2, Fault Analysis Result 3, and Fault Analysis Result 4 are fault probability distribution results.

[0030] As a further preferred technical solution, after step S10: online monitoring of multimodal data of GIS equipment, wherein the multimodal data includes acoustic signature signals, ultrasonic signals, ultra-high frequency signals, and the composition and content of decomposition products, or before step S30: performing fusion diagnostic analysis based on acoustic signature signals, ultrasonic signals, ultra-high frequency signals, and the composition and content of characteristic decomposition products, the method further includes the following steps: S11' The wavelet threshold denoising technique is used to denoise the voiceprint signal to obtain the denoised voiceprint signal for diagnostic analysis. S12' The ultrasonic signal is denoised using wavelet-frequency domain joint denoising technology to obtain a denoised ultrasonic signal for diagnostic analysis. S13' Adaptive Kalman filtering denoising technology is used to denoise the UHF signal to obtain the denoised UHF signal for diagnostic analysis.

[0031] It should be noted that the electrical signals of early GIS equipment failures are weak, highly susceptible to electromagnetic interference, and easily confused with noise. Acoustic signals have extremely poor anti-interference capabilities. Therefore, this embodiment performs denoising processing on the acoustic signal, ultrasonic signal, and ultra-high frequency signal respectively, and enhances the denoised data to retain the unique physical properties of each mode, ensuring that the generated enhanced sample is consistent with the original mode characteristics and also has a new failure evolution form.

[0032] As a further preferred technical solution, step S11': using wavelet threshold denoising technology to denoise the voiceprint signal to obtain the denoised voiceprint signal, specifically includes the following steps: The acoustic signature signal is decomposed using a wavelet basis, and the decomposed signal is expressed by the following formula:

[0033]

[0034] In the formula, W j,k These are wavelet coefficients. j To decompose the scale, k These are displacement index parameters used to index wavelet coefficients. Voiceprint signal, For wavelet basis functions, This is the form of the mother wavelet function after scaling and translation transformations. t It is a time variable; Based on the wavelet coefficients, the voiceprint signal is reconstructed, and the denoised voiceprint signal is obtained as follows:

[0035]

[0036] In the formula, For the first j In layer wavelet decomposition, the first k Wavelet coefficients, For wavelet domain scale adaptive thresholding, , For the first j Layer noise standard deviation l The signal length; This is an adjustment factor used to control the slope of the transition region, which is dynamically adjusted based on the signal-to-noise ratio of the voiceprint signal. This is the denoised voiceprint signal. For the first j In the layer decomposition k Wavelet coefficients.

[0037] It should be noted that the energy of vibration acoustic signals is concentrated in a few wavelet coefficients, while noise energy is dispersed and has smaller coefficients. Therefore, this embodiment uses wavelet threshold denoising technology to denoise the signal by utilizing the different characteristics of signal and noise in the wavelet domain. This is achieved by selecting a suitable wavelet basis for the acoustic signal. conduct j Layer decomposition is performed, and the processed wavelet coefficients are used to reconstruct the signal, resulting in a denoised speaker signal. Traditional wavelet hard / soft thresholding functions are prone to feature information loss during thresholding; therefore, this embodiment designs an adaptive multi-scale thresholding function to avoid the discontinuity of hard thresholding and the constant deviation of soft thresholding. By setting a threshold, the signal correlation coefficient is preserved, noise coefficients are eliminated, and coefficients greater than the adaptive threshold are retained.

[0038] As a further preferred technical solution, step S12': using wavelet-frequency domain joint denoising technology to denoise the ultrasonic signal to obtain a denoised ultrasonic signal, specifically includes the following steps: S121': Wavelet packet decomposition of the ultrasonic signal is performed using the Symlet wavelet basis to obtain several sub-bands; It should be noted that this embodiment selects the Symlet wavelet basis, which, based on its compact support and approximate symmetry, is very suitable for ultrasonic signal analysis. The signal is decomposed into J-level wavelet packets to obtain... Each frequency band corresponds to a different time-frequency resolution.

[0039] S122' Calculate the normalized Shannon entropy for each node, and select the wavelet packet basis determined by the decomposition path with the minimum entropy value as the optimal basis. The formula is as follows:

[0040]

[0041] In the formula, Here is the normalized Shannon entropy of the node, with a one-to-one correspondence between the node and the sub-band. To represent the first node within that node k The energy probability distribution of each wavelet coefficient relative to the total energy of that node. k For wavelet coefficient indexing, For the first j In the layer decomposition i The node of the first kWavelet coefficients; It should be noted that the energy of low-entropy nodes is concentrated, corresponding to the characteristic frequency band of partial discharge, while the energy of high-entropy nodes is dispersed, usually the frequency band dominated by noise. The optimal basis selection of wavelet packet decomposition can accurately locate the frequency band of partial discharge and avoid invalid decomposition.

[0042] S123' Perform a short-time Fourier transform on the optimal basis to calculate the power spectral density, expressed by the formula:

[0043] In the formula, The frequency domain signal obtained by short-time Fourier transform of the time-domain signal with optimal basis. This is the denoised frequency domain signal. The power spectral density of the denoised frequency domain signal. Soft thresholding in the frequency domain;

[0044] In the formula, No. j The noise standard deviation of each frequency sub-band l This is the signal length.

[0045] It should be noted that this method closely reflects real-world physics, as the energy distribution of noise may be uneven across different frequency bands, and the signal strength also varies across these bands. This approach enables more precise noise reduction, preserving more detail in strong signal bands and achieving more thorough suppression in weak signal or pure noise bands.

[0046] S124': Perform inverse wavelet packet operation on the denoised frequency domain signal to obtain the denoised ultrasonic signal.

[0047] It should be noted that this embodiment calculates the power spectral density by performing a short-time Fourier transform on the optimal base node, thereby determining the main frequency band range of the partial discharge signal. By employing wavelet-frequency domain joint denoising technology, only the noise frequency band can be suppressed, preserving pulse integrity.

[0048] As a further preferred technical solution, step S13': using adaptive Kalman filtering denoising technology to denoise the UHF signal to obtain the denoised UHF signal, specifically includes the following steps: S231', Constructing the state equation and observation equation:

[0049]

[0050] In the formula, It is an ultra-high frequency partial discharge signal; This is the state transition matrix, which describes the dynamic changes of the signal; The noise is the process noise, which has a mean of 0 and a covariance matrix of... The multidimensional Gaussian distribution; It is a noisy ultra-high frequency partial discharge signal; The observation matrix represents the mapping of signal measurements; The observed noise follows a mean of 0 and a covariance matrix of... The multidimensional Gaussian distribution; , Sampling interval, attenuation coefficient , This represents the partial discharge pulse decay time constant. , This is the noise attenuation time constant.

[0051] S232': The ultra-high frequency signal is modeled as an exponentially decaying oscillating signal, and the state equation and observation equation are constructed as follows:

[0052] In the formula, These are the pulse amplitude, the rate of change of amplitude, and the decay time constant, respectively. T This is the matrix transpose symbol.

[0053] S233', Calculation :

[0054] In the formula, It reflects the deviation between observation and prediction and is used for noise statistics estimation.

[0055] S234' Update the noise covariance matrix based on the deviation between observed and predicted values. and :

[0056]

[0057]

[0058] In the formula, This is the forgetting factor, which controls the weight of historical data. Here is the Kalman gain matrix. For the posterior estimation of the error covariance matrix, Let $\mathbf{a}$ be the posterior estimation error covariance matrix of the previous time step. b Forgetting factor, The forgetting factor b raised to the power of (D+1) For prior state estimation, , This is an estimate of the process noise covariance matrix. , This is an estimate of the observation noise covariance matrix; S235' At each time step, perform prediction and update steps to progressively denoise the signal and obtain the denoised UHF signal.

[0059] It should be noted that this embodiment utilizes Kalman filtering to recursively estimate the signal through a state-space model, adaptively and dynamically adjusts the model parameters to adapt to non-stationary signals, achieves denoising processing of ultra-high frequency signals, and improves robustness by adjusting the noise estimate.

[0060] As a further preferred technical solution, such as Figure 2 As shown, step S31: Processing the voiceprint signal using the Transformer model to obtain fault analysis result one, specifically includes: The probability distribution of fault analysis results for voiceprint signals is predicted using a pre-trained Transformer model. The Transformer model is a deep learning-based sequence modeling method that captures long-range dependencies in voiceprint signals through a self-attention mechanism and maps them to the probability space of fault categories. The Transformer model includes an embedding layer, a Transformer encoder layer, and a softmax classifier. The Transformer encoder layer includes multi-head self-attention, residual connections and normalization, and a feedforward network. The specific implementation process is as follows: (1) Input representation and preprocessing The acoustic signature signal of GIS equipment is usually a time-frequency domain feature, which is denoted as... Where T is the time step and d is the feature dimension. Then, the voiceprint signal is standardized as follows:

[0061] in, Standard deviation The mean squared error is denoted as .

[0062] (2) Input embedding layer Transformers lack time-aware capabilities and require the injection of positional information for positional encoding, given the position of the input sequence. and feature dimension index Each element of the position encoding vector is defined as:

[0063]

[0064] in, For time location index; Index for feature dimensions; The dimension of the input embedding; Used to control frequencies in different dimensions, creating a gradient from high to low frequencies, with the input embedding being... .

[0065] (3) Transformer encoder layer Projecting the input E onto the multi-head self-attention layer:

[0066]

[0067]

[0068] Among them, the query matrix, key matrix, and value matrix .

[0069] Calculate the attention weights and sum them into a weighted vector:

[0070] In this context, Q, K, and V represent Query, Key, and Value, respectively. This represents the scaling factor.

[0071] Combine the outputs of h attention heads:

[0072] In the formula, For the output of multi-head attention, For the output of each attention head, To output the projection weight matrix, This is for splicing operations.

[0073] Next, the attention head output is calculated using residual connections and normalization layers, as expressed by the formula:

[0074] In the formula, The output is after multi-head attention, residual connections, and layer normalization. This is a layer normalization layer.

[0075] Next, in the feedforward network, two fully connected layers and an activation function are used:

[0076] In the formula, As input to the feedforward network, This is the weight matrix of the first fully connected layer. This is the weight matrix of the second fully connected layer. For the bias term of the first fully connected layer, For the bias term of the first fully connected layer, This is the output of the feedforward network.

[0077] Performing residual join again yields:

[0078] Repeating the above steps will yield the deep encoding.

[0079] (4) Sequence aggregation and classification output Output the probability of the fault category using a softmax classifier:

[0080]

[0081] Where C represents the fault category. Given an input sequence X At that time, the predicted fault category is... y The probability distribution, This represents the feature representation after sequence aggregation. This is the weight matrix of the classification layer. For the bias term of the classification layer, For exponential operations, For exponential operations, y Indicates the fault category label, i This indicates a category index.

[0082] As a further preferred technical solution, step S32, processing the UHF signal using a graph neural network to obtain fault analysis result two, specifically includes the following steps: (1) Input data modeling as a graph structure: In this embodiment, the feature extraction vector of each partial discharge signal is used as the node feature. ,in For the first Each sample has 17-dimensional features. An adjacency matrix A is constructed based on the similarity between samples, where... =1 indicates a node With nodes Connected, otherwise 0.

[0083] (2) Forward propagation of graph neural networks Next, a graph convolutional network (GCN) or a graph attention network (GAT) is used to aggregate neighbor information. The key steps are as follows: When using graph convolutional layers, the node representations of each layer are updated as follows:

[0084] in, For degree matrix, , For the first Layer node characteristics, , For learnable weight matrix, For activation function, For a self-loop enhanced adjacency matrix, To enhance the elements of the adjacency matrix.

[0085] When using a graph attention layer, nodes Features are aggregated into neighbors through an attention mechanism:

[0086] Calculate the attention coefficient:

[0087] In the formula, For nodes i In the l The new feature representation of layer +1, Attention coefficient For neighboring nodes j In the l Layer feature representation, The weight matrix is ​​a learnable linear transformation. , , For the feature vector of the node, For learnable attention vectors, This is a vector concatenation operation. For the leakage linear rectification activation function, It is an exponential function.

[0088] (3) Graph-level output and probability distribution prediction Generate graph-level representations using graph pooling layers:

[0089] Will The inputs are fed into a fully connected layer and a softmax layer, and the output is a probability distribution of fault categories:

[0090] Where K is the number of fault categories, W o b o For output layer parameters, Represents given graph dataX and adjacency matrix A At that time, the sample belongs to the fault category. k The probability, Let be the global representation vector of the graph. , For the bias term of the output layer, , The weight vector of the output layer. For the first L The last layer (the third layer) of a graph neural network n The feature vector of each node For the first L The feature matrix of all nodes in the layer, For graph pooling functions, N This represents the total number of nodes in the graph.

[0091] As a further preferred technical solution, step S33, processing the ultrasonic signal using a generative adversarial network to obtain fault analysis result three, specifically includes: The generator design for generative adversarial networks (GANs) requires that the generator output satisfies probability distribution constraints.

[0092] in, For neural networks, z This is a random noise vector input to the generator. x For conditional information, this is the ultrasonic signal. The output of the generator, The Softmax function ensures that the generator output follows a probability distribution.

[0093] The discriminator is optimized to distinguish between the real label distribution and the generated distribution:

[0094] in, For the Sigmoid function, This indicates the need to learn the differences between real data and generated data. y For fault category labels, x Given an ultrasonic signal. This is the output of the discriminator.

[0095] The fault probability is estimated using a trained GAN network via generator Monte Carlo sampling:

[0096] In the formula, Given an ultrasonic signal x At that time, the predicted failure probability distribution, For the first i The generator's output at the next sampling time For the first b Random noise vector from the sampled subsamples x S represents the ultrasonic signal, and S represents the number of Monte Carlo samplings.

[0097] It should be noted that, in this embodiment, when training the model, the corresponding model is trained based on the enhanced voiceprint signal, ultrasonic signal and ultra-high frequency signal to obtain the fault prediction results corresponding to the voiceprint signal, ultrasonic signal and ultra-high frequency signal.

[0098] As a preferred technical solution, before performing the fusion diagnostic analysis based on acoustic signature signals, ultrasonic signals, ultra-high frequency signals, and the components and content of characteristic decomposition products, the method further includes the following steps: S01. Collect acoustic signal samples, ultrasonic signal samples, and ultra-high frequency signal samples when GIS equipment malfunctions; S02. Denoise the voiceprint signal samples, ultrasonic signal samples, and ultra-high frequency signal samples respectively. S03. Perform data augmentation processing on the denoised voiceprint signal sample, ultrasonic signal sample and ultra-high frequency signal sample to obtain the enhanced voiceprint signal sample, ultrasonic signal sample and ultra-high frequency signal sample. S04. Based on the enhanced voiceprint signal samples, ultrasonic signal samples, and UHF signal samples, the Transformer model, graph neural network, and generative adversarial network are trained respectively to obtain the trained Transformer model, graph neural network, and generative adversarial network.

[0099] As a further preferred technical solution, step S03: performing data enhancement processing on the denoised voiceprint signal sample, ultrasonic signal sample, and ultra-high frequency signal sample to obtain enhanced voiceprint signal sample, ultrasonic signal sample, and ultra-high frequency signal sample, specifically including the following steps: Feature extraction is performed on the denoised voiceprint signal, the denoised ultrasonic signal, and the denoised UHF signal respectively to obtain the feature vectors of the voiceprint signal, the ultrasonic signal, and the UHF signal, as expressed by the formula:

[0100]

[0101]

[0102] In the formula, This is the weight matrix for extracting voiceprint signal features. For bias terms, It is the ReLU activation function. These are the denoised speakerprint signal samples. This is the feature vector of the low-frequency vibration acoustic signature signal; This is the weight matrix for feature extraction from ultrasonic signals. For pulse position adaptive weights, For bias terms, This is a sample of the denoised ultrasonic signal. This is the feature vector of the ultrasonic signal; These are the weight matrices for extracting features from the real and imaginary parts, respectively. For bias terms, This is the feature vector of the ultra-high frequency signal. and These represent the real part and imaginary part of the denoised UHF signal sample, respectively; e represents the input feature vector index, and g represents the output feature vector index. Enhanced samples for the voiceprint signal, ultrasonic signal, and UHF signal are generated based on the feature vectors of the voiceprint signal, ultrasonic signal, and UHF signal, respectively. The formulas are as follows:

[0103]

[0104]

[0105]

[0106]

[0107] In the formula, This is the decoding weight matrix for the voiceprint signal. The in-mode perturbation coefficient, For baseline adjustment bias term, Samples for enhancing voiceprint signals; This is the decoding weight matrix for the ultrasonic signal. This is the intramodal pulse indication function. For baseline bias term, This is a sample for enhancing ultrasonic signals. This is the decoding weight matrix for ultra-high frequency signals. For the baseline bias term of the ultra-high frequency complex signal, For enhanced samples of ultra-high frequency signals, Let be the real part and the imaginary part of the denoised UHF signal, respectively. It is the imaginary unit.

[0108] It should be noted that this embodiment can use a generative adversarial network for data augmentation, including a multi-branch encoder and a multi-branch decoder. The multi-branch encoder is used to generate feature vectors corresponding to each modality of data, and the multi-branch decoder is used to generate corresponding augmented samples based on the feature vectors corresponding to each modality of data.

[0109] In this embodiment, the generation of enhanced samples for the three signals each independently depends on their respective modal feature vectors. The mapping from features to signals is achieved through a weight matrix, while preserving the unique physical properties of each mode. This ensures that the generated samples are consistent with the original modal features and also possess new fault evolution morphologies.

[0110] As a further preferred technical solution, the characteristic decomposition product components of the mixed gas are sampled and detected in step S20, specifically using a mixed gas purity detection device. The mixed gas purity detection device includes a gas chromatography detection system with 4 valves and 3 columns using dual (2) helium ionization detectors (PDHID) and a gas trace component detection device based on the high-resolution mass spectrometry detection principle.

[0111] The gas chromatography detection device can detect more than 10 decomposition characteristic components in gases such as C4F7N / CO2 and CF3SO2F / N2, with a detection limit not exceeding 1 μL / L and repeatability ≤ ±3%. Switching valve I uses an automatic injection method. Fluorocarbons in the C4F7N / CO2 mixed gas after passing through quantitative tube 1 are separated using column 3 and then analyzed by detector 1. C4F7N and CO2 are vented. Switching valve II uses a ten-way backflush injection method. The C4F7N / CO2 mixed gas after passing through quantitative tube 2 first undergoes pre-separation of the sample through column 1, venting CO2 gas, allowing air, CO, CF4, CO2, and C2F6 to enter column 2 for analysis by detector 2.

[0112] Taking C4F7N / CO2 as an example, by using a dedicated gas chromatograph for C4F7N / CO2 detection, its appearance is as follows: Figure 5 As shown, the use of a high-precision PDHID detector significantly improves the detection accuracy of trace components. This device detects gaseous components, except for C4F7N / CO2 or impurities in C4F7N gas; the detection limit for other gases is no higher than 1 μL / L, and the repeatability is ≤ ±3%.

[0113] As a further preferred technical solution, this example analyzes the characteristic decomposition products generated by C4F7N / CO2 under faults such as partial discharge and overheating, and calculates the concentration ratios between these products to construct fault analysis logic. The specific process is as follows: (1) Based on the decomposition pathway of C4F7N molecules and the formation pathway of C4F7N gas decomposition products, and considering the decomposition characteristics of C4F7N / CO2 mixed gas with discharge defects and overheating defects, four ratio characteristics are constructed: content ratios c[CO] / c[C3F6], c[C2F6] / c[CF4], and chromatographic peak area ratios v[C2F4] / v[CF4+C2F6], v[C2F4] / v[C3F6], to identify discharge and overheating defects in the experiment. On this basis, a three-ratio "characteristic quantity group" is constructed using the content ratios c[C2F6] / c[CF4], v[C2F4] / v[C3F6], and v[C2F4] / v[CF4+C2F6].

[0114] Figure 6 The experimental data for discharge and overheating defects are distributed in a triangular coordinate system composed of three ratio "characteristic quantity sets". High-temperature overheating defects and discharge defects have distinct distribution areas. Overheating defects are mainly distributed in the region where v[C2F4] / v[C3F6]=0 and c[C2F6] / c[CF4]=1. Discharge defects are distributed in the region where v[C2F4] / v[C3F6]>0.25 and c[C2F6] / c[CF4]<0.75. This is mainly because under high-temperature overheating defects (T≥650℃), although C2F4, CF4, and C2F6 gases are produced, the content of these three gases is much lower than that of C3F6, making the ratio v[C2F4] / v[C3F6] almost zero. The overheating decomposition experiment shows that the content of the product CF4 is comparable to that of C2F6, and the content ratio c[C2F6] / c[CF4] is almost 1. In addition, the experimental data points for corona discharge, spark discharge, and surface discharge defects are also relatively concentrated.

[0115] Drawing on the David triangle method, the defect distribution area displayed by the three ratio "characteristic quantity group" under triangular coordinates was initially divided, such as... Figure 7As shown in the diagram. Since none of the current experimental data points fall within the gray area DT between discharge and overheating defects, area DT represents the uncertain region where neither discharge nor overheating defects have occurred yet. The distribution region of high-temperature overheating defects is denoted as T, where area T1 represents overheating defects at 650℃, area T2 represents overheating defects at 700℃, and area T3 represents overheating defects at temperatures greater than 700℃. The distribution region of discharge defects is denoted as D, where area D1 represents discharge defects with low discharge energy, and area D2 represents discharge defects in the form of breakdown. Compared to spark discharge, corona discharge and partial discharge along the surface have relatively low discharge energy and relatively stable characteristic ratios. Therefore, corona discharge and partial discharge along the surface are classified as area D1. Compared to corona discharge, spark discharge has higher energy, its characteristic ratio is unstable, and the data points are more dispersed. Furthermore, the data points of surface discharge in the form of breakdown are significantly far from the concentrated area of ​​surface discharge data points in the form of partial discharge. Therefore, spark discharge and surface discharge in the form of breakdown are classified as area D2. The characteristic set composed of three ratios can identify discharge and overheating defects in the experiment, and also has a certain degree of recognition for electrode materials and overheating temperature.

[0116] Specifically, the analysis of the variation patterns of various product concentrations under different fault conditions is similar to that of C4F7N / CO2. The key is not the absolute concentration of a single product, but the relative ratio between different products, because this can better reflect the energy density and nature of the fault.

[0117] Therefore, taking C4F7N / CO2 as an example, when the mixed gas uses C4F7N / CO2, step S34: analyzing the components and contents of characteristic decomposition products using the three-ratio method to obtain fault analysis result four, specifically includes the following steps: S341. Calculate the content ratio c[C2F6] / c[CF4], the peak area ratio v[C2F4] / v[C3F6], and the peak area ratio v[C2F4] / v[CF4+C2F6]. It should be noted that calculating the peak area ratio for GIS equipment insulation defect diagnosis reflects the dynamic process of decomposition reaction and enhances the ability to distinguish complex faults.

[0118] S342. Construct a three-ratio characteristic quantity group using the content ratio c[C2F6] / c[CF4], the peak area ratio v[C2F4] / v[C3F6], and the peak area ratio v[C2F4] / v[CF4+C2F6]. Use the three-ratio method to determine the fourth fault diagnosis result: defects distributed in the region where v[C2F4] / v[C3F6]=0 and c[C2F6] / c[CF4]=1 are overheating defects, and defects distributed in the region where v[C2F4] / v[C3F6]>0.25 and c[C2F6] / c[CF4]<0.75 are discharge defects.

[0119] It should be noted that this embodiment uses a three-ratio "characteristic group" method consisting of the content ratio c[C2F6] / c[CF4], the peak area ratio v[C2F4] / v[C3F6], and the peak area ratio v[C2F4] / v[CF4+C2F6], and converts the three ratios of different faults into probability distributions to achieve fault diagnosis based on the decomposition products of mixed gas.

[0120] The calculation process of the three ratios is further explained here: through previous research on the content of several types of decomposition products under different fault types and severity conditions, a ratio method is constructed to form a corresponding relationship with the fault type and severity, which can be directly called in actual scenarios; the several types of decomposition products are defined as characteristic decomposition products. In this embodiment, the several types of decomposition products are defined as characteristic decomposition products.

[0121] Specifically: Taking C4F7N / CO2 as an example, the three ratios of its characteristic decomposition products are extracted as follows:

[0122]

[0123]

[0124] This represents the ratio of gaseous component content. , This represents the peak area ratio.

[0125] As a further preferred technical solution, step S35: performing multimodal uncertainty fusion diagnostic analysis based on fault analysis result one, fault analysis result two, fault analysis result three, and fault analysis result four to obtain the fault type and severity, specifically including the following steps: S351. Calculation of multimodal time series (1) The low-frequency vibration acoustic signal, ultrasonic signal, and ultra-high frequency signal of partial discharge are processed in the above-mentioned process, and the denoised acoustic signal, ultrasonic signal, and ultra-high frequency signal are finally obtained as follows:

[0126] These are the denoising time sequences for voiceprint signals, ultrasonic signals, and ultra-high frequency signals, respectively. This is a time series of partial discharge acoustic-optical-electrical integrated signals.

[0127] All modal data are aligned to a uniform time scale to form a synchronous multimodal time series:

[0128] in, The ratio of the content of characteristic decomposition product gas components. , The peak area ratio of chromatographic data for different characteristic decomposition products; S352. LSTM Model Design: Traditional neural networks have fully connected layers, but the nodes within each layer are independent, meaning the input and output are related as vectors. However, in prediction problems, the relationships between consecutive terms are not independent, thus traditional neural networks have limitations in handling regression prediction. Long Short-Term Memory (LSTM) networks address the problem of long-term information dependencies by introducing special connection modules. They are suitable for processing time-series signals involving partial discharge data and gas parameter data fusion. LSTM uses gate structures to control the flow of memory cell states, selectively adding or removing information from the memory cells. The LSTM gate structure includes input gates, forget gates, and output gates. The input gate controls the amount of information flowing into the memory cells, the forget gate controls the amount of information flowing from the previous time step to the current time step, and the output gate controls the amount of information flowing from the current time step to the current hidden state. The gate structure acts like a filter during information flow, limiting the amount of information. The calculation formula for LSTM, under the influence of the gate structure and memory cells, is as follows:

[0129]

[0130]

[0131]

[0132]

[0133] In the formula, , , These represent the states of the input gate, forget gate, and output gate, respectively. For the current input, , These represent the states of the memory units at the previous and current time points, respectively. , These are the hidden states from the previous and current time steps, respectively. , , These are the weight matrices between the input gate, forget gate, and output gate and the input, respectively. , , These are the weight matrices between the input gate, forget gate, and output gate and the hidden state of the previous time step, respectively. , , These are the bias vectors for the input gate, forget gate, and output gate, respectively. , , These are the weight matrix between the memory unit and the input, the bias vector of the memory unit, and so on. This represents the activation function. express Activation function The operator indicates element-wise multiplication.

[0134] As can be seen from the above expression, the information flow and gate structure control of LSTM, as well as the data updates, originate from the current input. and the hidden state of the previous moment Historical information is accumulated using memory units. The input gate restricts information at new moments, the forget gate restricts information from the memory units at the previous moment, and the output gate filters to obtain the hidden state at the current moment.

[0135] Each data sample It is a time series consisting of partial discharge data and gas parameter data. After extracting features through an LSTM network, the evaluation results are output through a fully connected layer of Softmax.

[0136] S353. Output Layer Design: From LSTM to Discharge Severity Scoring Physical meaning of output value: Define the output Its size represents the severity of the discharge.

[0137]

[0138] The output calculation process is as follows: 1) Temporal Feature Extraction The final temporal feature vector of the last LSTM layer is:

[0139] Where L represents the input signal The sequence length, where n represents the number of layers in the LSTM network. This represents an abstract feature vector that contains acoustic, optical, and electrical partial discharge data and gas parameter data.

[0140] 2) Fully connected layer transformation

[0141]

[0142]

[0143]

[0144] , Here, represents the weight matrix and bias vector of the first fully connected layer, respectively, and ReLU is the activation function. The first layer of linear activation outputs the feature vector, and Dropout represents random deactivation regularization. The discard probability means that during training, 30% of the neurons' outputs will be randomly set to zero. The regularized feature vector; , These are the weight matrix and bias vector of the second fully connected layer, respectively. The output feature vector of the second-layer linear activation is... The eigenvectors are regularized.

[0145] 3) Severity calculation

[0146] This represents the output layer weight vector. This represents the transpose vector. For the output layer bias, where Sigmoid activation function .

[0147] LSTM model training strategy: using weighted mean squared error loss function

[0148] This indicates the number of samples in the current training batch. This represents the true severity score of the discharge for the i-th sample. This represents the predicted discharge severity score for the i-th sample. This represents the value of the loss function.

[0149] Among them, weight Set according to severity:

[0150] This technical approach has advantages over traditional approaches: the weight vector of the output layer is automatically learned by the neural network, eliminating the need for manual definition of feature weights; this model is a data-driven model for assessing the severity of discharge, and does not require empirical formulas for determination.

[0151] This embodiment addresses the problem of fusing multiple physical signals with different physical mechanisms, spatiotemporal scales, and interference sources. It processes each signal separately to generate corresponding fault diagnosis results. Then, it uses the expected probability distribution of fusion and the total confidence level to determine the fault category and analyze the cause of the fault. By comprehensively utilizing electrical, acoustic, and chemical signals, the chemical signals compensate for the randomness of electrical signal capture. This avoids the potential for misjudgment caused by relying solely on instantaneous electrical and acoustic signals due to their weakness and intermittency in the early stages.

[0152] In addition, such as Figure 8 As shown, the second embodiment of the present invention also proposes a partial discharge diagnostic system for GIS equipment based on multimodal data, the system comprising: The online monitoring module 10 is used for online monitoring of multimodal data of GIS equipment, including acoustic signature signals, ultrasonic signals, and ultra-high frequency signals. The offline detection module 20 is used to sample and detect the components and contents of characteristic decomposition products of the mixed gas when anomalies occur in the multimodal data. The fault diagnosis module 30 is used to perform fusion diagnostic analysis based on acoustic signal, ultrasonic signal, ultra-high frequency signal and the composition and content of characteristic decomposition products, to obtain fault diagnosis results and issue warnings.

[0153] As a further preferred technical solution, the system also includes: The first denoising module is used to denoise the voiceprint signal using wavelet threshold denoising technology to obtain the denoised voiceprint signal. The second denoising module is used to denoise the ultrasonic signal using wavelet-frequency domain joint denoising technology to obtain the denoised ultrasonic signal. The third denoising module is used to denoise the UHF signal using adaptive Kalman filtering denoising technology to obtain the denoised UHF signal.

[0154] As a further preferred technical solution, the fault diagnosis module 30 specifically includes: The first analysis unit is used to process the voiceprint signal using the Transformer model to obtain fault analysis result one; The second analysis unit is used to process the UHF signal using a graph neural network to obtain fault analysis result two. The third analysis unit is used to process ultrasonic signals using a generative adversarial network to obtain fault analysis result three. The fourth analysis unit is used to analyze the components and contents of characteristic decomposition products using the three-ratio method to obtain the fault analysis result four. The fault diagnosis unit is used to perform multimodal uncertainty fusion diagnostic analysis based on fault analysis result 1, fault analysis result 2, fault analysis result 3 and fault analysis result 4, to obtain fault diagnosis results and issue warnings.

[0155] Based on the above description, the composition and content of the acoustic-optical-electric partial discharge signal and its characteristic decomposition products are complementary. Establishing a joint diagnostic model significantly improves the detection rate and accuracy of early faults. The fault case analysis and diagnostic process is enhanced as follows: (1) Only the optical sensor emits a monitoring signal ----- early weak discharge; (2) Only the ultrasonic sensor emits a monitoring signal ----- surface discharge; (3) Only ultra-high frequency signals are emitted for monitoring --- tip discharge; (4) The ultra-high frequency and ultrasonic sensors simultaneously emit monitoring signals ----- levitation discharge; (5) The three ratios reflect the type and content of characteristic decomposition products, and the decomposition products are strongly correlated with faults such as partial discharge and overheating. Therefore, the type and severity of the fault can be judged by the three ratios.

[0156] As a further preferred technical solution, the fault diagnosis unit specifically includes: The evidence conversion subunit is used to convert the results of each fault analysis into Dirichlet parameters; The joint subunit is used to combine the Dirichlet parameters corresponding to each fault analysis result to obtain the joint Dirichlet parameters. Uncertainty fusion subunit is used to calculate the fusion expected probability distribution and total confidence based on joint Dirichlet parameters.

[0157] It should be noted that other embodiments or specific implementation methods of the GIS equipment partial discharge diagnosis system based on multimodal data described in this invention can refer to the above-mentioned method embodiments, and will not be repeated here.

[0158] Furthermore, the present invention also proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the fault diagnosis system and method for mixed gas insulated power equipment as described in the first embodiment above.

[0159] It should be noted that the computer-readable medium disclosed in this embodiment may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, and portable compact disk read-only memory (CD-ROM). ROM, optical storage devices, magnetic storage devices, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0160] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform a zero-sample image anomaly detection method according to the above embodiments.

[0161] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server.

[0162] In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0163] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0164] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0165] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" or "several" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0166] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A fault diagnosis system and method for mixed gas insulated power equipment, characterized in that, include: Multimodal data acquisition module: used to acquire multimodal data from GIS equipment, including acoustic signature signals, ultrasonic signals, and ultra-high frequency signals; Anomaly Analysis Module: When anomalies occur in multimodal data, sample and detect characteristic decomposition products of the mixed gas, and calculate the three ratios. Fault diagnosis module: Based on the fusion diagnosis analysis of acoustic signal, ultrasonic signal, ultra-high frequency signal and three ratios, the fault type and fault severity are obtained.

2. The fault diagnosis system and method for mixed gas insulated power equipment as described in claim 1, characterized in that, The fusion diagnostic analysis includes: The Transformer model was used to process the voiceprint signal, and the fault analysis result was obtained. By using a graph neural network to process the UHF signal, fault analysis result two was obtained. The ultrasonic signal was processed using a generative adversarial network to obtain fault analysis result three; The components and contents of the characteristic decomposition products were analyzed using the three-ratio method, and the failure analysis results were obtained. Based on fault analysis results 1, 2, 3, and 4, a multimodal uncertainty fusion diagnostic analysis is performed to obtain fault diagnosis results and issue early warnings.

3. The fault diagnosis system and method for mixed gas insulated power equipment as described in claim 1 or 2, characterized in that, Before the fusion diagnostic analysis, the acoustic signature signal, ultrasonic signal, and ultra-high frequency signal are cleaned and aligned to obtain... , in, These are the denoising time sequences for voiceprint signals, ultrasonic signals, and ultra-high frequency signals, respectively. This is a time series of partial discharge acoustic-optical-electrical integrated signals.

4. The fault diagnosis system and method for mixed gas insulated power equipment as described in claim 3, characterized in that, The three-ratio method is used to analyze the components and contents of characteristic decomposition products. The three-ratio calculation method is as follows: by studying the contents of several types of decomposition products under different fault types and severity conditions in the early stage, a ratio method is constructed to form a corresponding relationship with the fault type and severity, which can be directly called in actual scenarios; the several types of decomposition products are defined as characteristic decomposition products. The time sequence of the characteristic decomposition products is as follows: ,in, The ratio of the content of characteristic decomposition product gas components. , This represents the peak area ratio of chromatographic data for different characteristic decomposition products.

5. The fault diagnosis system and method for mixed gas insulated power equipment as described in claim 4, characterized in that, The calculation process for the fault type and fault severity is as follows: All modal data are aligned to a uniform time scale to form a synchronous multimodal time series: After extracting features using an LSTM network, the evaluation results are output through a fully connected layer of Softmax.

6. A fault diagnosis system and method for mixed gas insulated power equipment, characterized in that, include: Steps for acquiring multimodal data: This involves acquiring multimodal data from GIS equipment, including acoustic signature signals, ultrasonic signals, and ultra-high frequency signals. The steps of anomaly analysis are as follows: When anomalies occur in multimodal data, sample and detect characteristic decomposition products of the mixed gas, and calculate the three ratios. The steps of fault diagnosis are as follows: Based on the acoustic signal, ultrasonic signal, ultra-high frequency signal and the three ratios, a fusion diagnostic analysis is performed to obtain the fault type and fault severity.

7. The fault diagnosis system and method for mixed gas insulated power equipment as described in claim 6, characterized in that, The fusion diagnostic analysis yields fault diagnosis results and provides early warnings, including: The Transformer model was used to process the voiceprint signal, and the fault analysis result was obtained. By using a graph neural network to process the UHF signal, fault analysis result two was obtained. The ultrasonic signal was processed using a generative adversarial network to obtain fault analysis result three; The components and contents of the characteristic decomposition products were analyzed using the three-ratio method, and the failure analysis results were obtained. Based on fault analysis results 1, 2, 3, and 4, a multimodal uncertainty fusion diagnostic analysis is performed to obtain fault diagnosis results and issue early warnings.

8. The fault diagnosis system and method for mixed gas insulated power equipment as described in claim 6 or 7, characterized in that, Before the fusion diagnostic analysis, the acoustic signature signal, ultrasonic signal, and ultra-high frequency signal are cleaned and aligned to obtain... , in, These are the denoising time sequences for voiceprint signals, ultrasonic signals, and ultra-high frequency signals, respectively. This is a time series of partial discharge acoustic-optical-electrical integrated signals.

9. The fault diagnosis system and method for mixed gas insulated power equipment as described in claim 8, characterized in that, The three-ratio method is used to analyze the components and contents of characteristic decomposition products. The three-ratio calculation method is as follows: by studying the contents of several types of decomposition products under different fault types and severity conditions in the early stage, a ratio method is constructed to form a corresponding relationship with the fault type and severity, which can be directly called in actual scenarios; the several types of decomposition products are defined as characteristic decomposition products. The time sequence of the characteristic decomposition products is as follows: ,in, The ratio of the content of characteristic decomposition product gas components. , This represents the peak area ratio of chromatographic data for different characteristic decomposition products.

10. The fault diagnosis system and method for mixed gas insulated power equipment as described in claim 9, characterized in that, The calculation process for the fault type and fault severity is as follows: All modal data are aligned to a uniform time scale to form a synchronous multimodal time series: After extracting features using an LSTM network, the evaluation results are output through a fully connected layer of Softmax.

Images