GIS equipment fault diagnosis method based on sub-model training and joint training

Abstract

The application discloses a GIS equipment fault diagnosis method based on sub-model training and joint training, comprising the following steps: obtaining historical operation data of GIS equipment, and extracting features of multi-dimensional operation data through a traditional signal processing method combined with a pre-training process of a deep learning network. A correlation analysis method assisted by virtual simulation technology is used to analyze the representation relationship between each typical fault and operation characteristics, and a GIS equipment fault set is constructed. A GIS fault diagnosis model is constructed. A fault diagnosis sub-model is constructed for each typical fault, and a sub-model training data set is constructed, and each fault feature is learned through sub-model individual training. The training data sets of all sub-models are combined, and the entire GIS fault diagnosis model is optimized through joint training. The method can make full use of multi-source operation data and simulation data, improve the accuracy and robustness of GIS equipment fault diagnosis, and realize efficient and reliable operation state monitoring and fault early warning.

Description

Technical Field

[0001] This invention belongs to the field of GIS equipment fault diagnosis technology, specifically relating to a GIS equipment fault diagnosis method based on sub-model training and joint training. Background Technology

[0002] Gas-insulated switchgear (GIS) is widely used in high-voltage transmission and substation systems due to its compact structure, reliable operation, and low maintenance requirements. However, during long-term operation, GIS equipment may still fail due to insulation aging, mechanical wear, loose electrical connections, etc., leading to performance degradation or even major accidents. Therefore, fault diagnosis and condition monitoring of GIS equipment are of significant engineering importance.

[0003] Existing fault diagnosis methods for GIS equipment mainly include traditional signal processing methods and experience-based diagnostic methods. Traditional signal processing methods typically analyze UHF partial discharge signals, ultrasonic signals, and mechanical vibration signals collected by the equipment, extracting feature indicators and applying threshold judgments. However, these methods have poor adaptability to fault modes under complex operating conditions and struggle to accurately diagnose multiple typical faults. Experience-based diagnostic methods rely on human experience and rule-making, are highly subjective, and cannot adapt to changes in the equipment's operating environment, making automated and real-time fault diagnosis difficult.

[0004] In recent years, deep learning methods have been gradually applied in the field of equipment fault diagnosis. By performing feature learning on equipment operating data, the accuracy of diagnosis can be improved to a certain extent.

[0005] For example, in the prior art, CN117829264A discloses a method, device, equipment and storage medium for GIS insulation and mechanical fault diagnosis. This method achieves real-time diagnosis by collecting operating data, cleaning data, extracting features, labeling samples, constructing a lightweight convolutional neural network model, and training the model using a transfer learning strategy.

[0006] However, existing GIS equipment fault diagnosis methods based on deep neural networks have the following problems: on the one hand, directly modeling multiple types of faults in a unified manner can easily lead to insufficient ability of the model to distinguish different fault modes; on the other hand, the model lacks constraints on negative samples during training, which may result in misjudgment when inputting non-target fault samples, affecting the reliability of diagnosis.

[0007] Therefore, there is an urgent need for a GIS equipment fault diagnosis method that can make full use of historical operation data and simulation data, while taking into account the training of typical fault sub-models and global joint optimization, so as to improve the accuracy and reliability of fault diagnosis and realize intelligent monitoring and early warning of equipment status. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a GIS equipment fault diagnosis method based on sub-model training and joint training;

[0009] The objective of this invention can be achieved through the following technical solutions:

[0010] This invention provides a GIS equipment fault diagnosis method based on sub-model training and joint training, comprising the following steps:

[0011] Acquire historical operational data of GIS equipment, including multi-dimensional operational data corresponding to various typical faults of GIS equipment;

[0012] Based on the pre-training process of traditional signal processing methods combined with deep learning networks, feature extraction is performed on the multi-dimensional operational data to obtain operational feature sets corresponding to each typical fault; the operational feature set includes multiple operational features.

[0013] A correlation analysis method based on virtual simulation technology is used to analyze the representational relationship between each typical fault and the operational features through the operational feature sets corresponding to each typical fault.

[0014] Construct a GIS equipment fault set based on the representational relationships and operational feature set;

[0015] Construct a GIS fault diagnosis model based on CvT neural network;

[0016] Based on the GIS equipment fault set, the GIS fault diagnosis model is trained by sub-model training and joint training respectively to obtain the trained GIS fault diagnosis model.

[0017] Fault diagnosis of GIS equipment is achieved through a trained GIS fault diagnosis model.

[0018] Furthermore, the typical faults include insulation faults, mechanical structure faults, and electrical connection faults;

[0019] The multidimensional operating data includes various types of operating data, including UHF partial discharge signals, ultrasonic partial discharge signals, ultraviolet spectral signals, mechanical vibration signals, mechanical stroke signals, operating coil current signals, and GIS equipment temperature rise data.

[0020] The ultra-high frequency partial discharge signal is obtained by real-time monitoring and acquisition of the partial discharge phenomenon inside the GIS equipment through an ultra-high frequency UHF sensor installed in the GIS equipment.

[0021] The ultrasonic partial discharge signal is detected by an ultrasonic sensor installed in the GIS equipment and acquired in real time.

[0022] The ultraviolet spectral signal is acquired in real time by an ultraviolet spectral sensor installed inside the GIS equipment;

[0023] The mechanical vibration signal is acquired in real time by installing vibration sensors on key mechanical components of the GIS equipment.

[0024] The mechanical travel signal is acquired in real time by displacement sensors installed on moving parts such as circuit breakers in GIS equipment;

[0025] The operating coil current signal is acquired in real time by a current sensor installed in the control circuit of the GIS equipment;

[0026] The temperature rise data of the GIS equipment is collected in real time by temperature sensors installed in various key parts of the GIS equipment.

[0027] Furthermore, the pre-training process based on traditional signal processing methods combined with deep learning networks extracts features from the multi-dimensional operational data to obtain operational feature sets corresponding to each typical fault, specifically including:

[0028] Traditional signal processing methods are used to extract features from each operational data point in the multidimensional operational data corresponding to each typical fault, thereby obtaining multiple primary features of each typical fault.

[0029] The pre-training process of deep learning networks is used to process multiple primary features to obtain multiple operational features, which are then used as the operational feature set corresponding to each typical fault.

[0030] Furthermore, the traditional signal processing method is used to extract features from each piece of operational data in the multidimensional operational data corresponding to each typical fault, obtaining multiple primary features for each typical fault, specifically including:

[0031] For UHF partial discharge signals in multidimensional operational data, two-dimensional or three-dimensional spectral analysis methods are used to extract primary features of the UHF partial discharge signals, including average discharge power, initial discharge voltage, extinction discharge voltage, and average discharge current.

[0032] For ultrasonic partial discharge signals in multidimensional operational data, time-domain features are extracted by combining time-domain analysis and frequency-domain analysis. These time-domain features are used as primary features of the ultrasonic partial discharge signals. The time-domain features include root mean square, variance, absolute integral mean, kurtosis, and skewness.

[0033] For the ultraviolet spectral signal in the multidimensional running data, the Burg recursive algorithm is applied to estimate the power spectrum and extract the power spectral density, which is then used as the primary feature of the ultraviolet spectral signal.

[0034] For mechanical vibration signals in multidimensional operational data, time-domain analysis and frequency-domain analysis based on Fourier transform are used, combined with time-frequency envelope analysis based on Hilbert-Huang transform, to extract primary features of the mechanical vibration signals, including amplitude, root mean square, peak value, and variance.

[0035] For the mechanical travel signal in the multidimensional operation data, based on the physical motion process of the circuit breaker, the motion speed, acceleration, opening time, closing time, and contact overtravel parameters are extracted as primary features of the mechanical travel signal.

[0036] For the operating coil current signal in the multidimensional operating data, the peak current, effective current value, and curve secant slope are extracted by analyzing the current waveform as primary features of the operating coil current signal.

[0037] For the temperature rise data of GIS equipment in multidimensional operational data, time series analysis is applied to extract the temperature change trend and obtain the temperature rise rate, temperature fluctuation amplitude and temperature peak value as the primary features of the temperature rise data of GIS equipment.

[0038] Furthermore, the pre-training process using a deep learning network processes the obtained primary features to obtain multiple operational features, which are then used as the operational feature set corresponding to each typical fault. Specifically, this includes:

[0039] Multiple primary features from multidimensional operational data are input into a deep neural network model. Pre-training is used to initialize network parameters. Lower-level primary features are combined to form more abstract higher-level weight representations. The hierarchical structure of the deep neural network is utilized to automatically learn and output the corresponding higher-level feature representations, specifically including:

[0040] The primary features of UHF partial discharge signals, ultrasonic partial discharge signals, ultraviolet spectral signals, and mechanical vibration signals are input into a convolutional neural network (CNN). The time-frequency local features are extracted through convolution operations to generate corresponding high-level feature representations.

[0041] The primary features of the mechanical stroke signal and the operating coil current signal are input into a network structure that combines a convolutional neural network (CNN) and a recurrent neural network (RNN) to extract temporal and periodic features and generate a high-level feature representation of the mechanical stroke and the operating coil current.

[0042] The primary features of GIS equipment temperature rise data are input into a structure combining a convolutional neural network (CNN) and a long short-term memory network (LSTM) to extract the temporal dynamic features of the temperature rise signal and generate a high-level feature representation of the GIS equipment temperature rise data.

[0043] The obtained high-level features are used as operational features to obtain the operational feature set corresponding to each typical fault.

[0044] Furthermore, the correlation analysis method based on virtual simulation technology analyzes the representational relationship between each typical fault and the operational features through the operational feature set corresponding to each typical fault, specifically including:

[0045] For each typical fault, a simulation model of the GIS equipment under the typical fault is constructed, and the simulation model of the GIS equipment under each fault is simulated and run on the simulation platform to obtain multi-dimensional operation data during the simulation process.

[0046] Feature extraction is performed on the multi-dimensional operational data obtained during the simulation to obtain operational feature sets under each typical fault.

[0047] The set of operational features extracted from historical operational data will be merged with the set of operational features obtained from simulation to construct a correlation analysis dataset;

[0048] Based on the Pearson correlation coefficient analysis method, the correlation coefficient between each typical fault and each operating feature is calculated using the correlation analysis dataset, thereby obtaining the characterization relationship between each typical fault and the operating feature.

[0049] Furthermore, the construction of the GIS equipment fault set based on the representation relationship and the operational feature set specifically includes:

[0050] The correlation coefficients between each typical fault and each operational feature in the characterization relationship are sorted according to the magnitude of the correlation coefficients to obtain the correlation ranking results corresponding to each typical fault.

[0051] For each typical fault, the top K operational features with the highest correlation coefficients are selected from the correlation ranking results, and the corresponding operational features are screened from the correlation analysis dataset obtained from the simulation to construct the fault set corresponding to the typical fault; K is a constant.

[0052] Perform the filtering operation on each typical fault in turn to obtain the fault set corresponding to each typical fault;

[0053] The fault sets corresponding to all typical faults are summarized to construct the GIS equipment fault set. The GIS equipment fault set includes the fault set corresponding to each typical fault. The fault set of each typical fault includes the top K operating features with the highest correlation coefficients selected from the correlation analysis dataset.

[0054] Furthermore, the construction of the GIS fault diagnosis model based on the CvT neural network specifically includes:

[0055] For each typical fault, K operational features selected from the fault set corresponding to the typical fault are used as input to construct a fault diagnosis sub-model based on a CvT neural network structure. The fault diagnosis sub-model includes a three-level cascaded CvT feature extraction structure, which includes a convolutional token embedding layer and a convolutional Transformer module. The convolutional token embedding layer is used to perform convolutional projection on the K operational features, converting the input single-channel feature matrix into a token sequence. The convolutional Transformer module includes a self-attention layer and a feedforward network layer, which extract the global correlation of high-dimensional operational features through an attention mechanism. The input of the fault diagnosis sub-model is a feature matrix composed of K operational features, and the output is the probability distribution corresponding to each typical fault and normal mode.

[0056] The output probability distributions of the fault diagnosis sub-models corresponding to all typical faults are integrated, and the probability outputs of each fault diagnosis sub-model are input into the integrated decision module.

[0057] The integrated decision module calculates the probability distribution of each fault diagnosis sub-model based on weighted fusion, confidence calibration or Softmax normalization mechanism, and outputs the final fault classification result and the corresponding confidence level.

[0058] The fault diagnosis sub-models of each typical fault are combined with the integrated decision module to construct a GIS fault diagnosis model based on CvT neural network.

[0059] Furthermore, the step of performing sub-model training and joint training on the GIS fault diagnosis model based on the GIS equipment fault set to obtain the trained GIS fault diagnosis model specifically includes:

[0060] Step A1: For each fault diagnosis sub-model in the GIS fault diagnosis model The fault set corresponding to typical faults in GIS equipment fault set. Compared with several normal samples The faults were merged, and samples from other typical fault sets of GIS equipment fault sets were also included. Randomly select a number of samples as the negative sample set Merge the datasets to construct the first training dataset. :

[0061]

[0062] in, Indicates the first i The first training dataset for a typical fault; This indicates the first fault in the GIS equipment fault set. i A set of typical faults This represents the normal sample set. Indicates the first i A negative sample set of typical faults; the normal samples are the operating characteristics corresponding to the normal mode of GIS equipment;

[0063] Step A2: Transfer the first training dataset Input the corresponding fault diagnosis sub-model The first loss function is used for the fault diagnosis sub-model. Conduct individual training:

[0064] Step A3: Perform steps A1-A2 sequentially for each fault diagnosis sub-model until all fault diagnosis sub-models have completed individual training;

[0065] Step A4: Merge the first training datasets corresponding to all fault diagnosis sub-models to construct the second training dataset;

[0066] Step A5: Input the second training dataset into the entire GIS fault diagnosis model based on CvT neural network, and use the second loss function to jointly train the GIS fault diagnosis model to obtain the trained GIS fault diagnosis model.

[0067] Furthermore, the first loss function is formulated as follows:

[0068]

[0069] in, Indicates sample The label, with a value of 1, indicates that it belongs to the fault diagnosis sub-model. Typical faults; Represents the fault diagnosis sub-model For the sample The probability of the output; This represents the maximum allowed output probability threshold for negative samples, used to constrain misclassification of negative samples; These are the weighting coefficients; Represents the first loss function;

[0070] The second loss function is expressed as follows:

[0071]

[0072] in, For the second loss function, For the second training dataset, This is a set of labels, including various typical faults and normal modes; For the sample The true label, if it belongs to the category c but ,otherwise ; For the GIS fault diagnosis model in the joint training phase, input Output category The predicted probability.

[0073] Compared with the prior art, the present invention has the following advantages:

[0074] (1) In the prior art, deep learning methods that directly model multiple types of faults in a unified manner are prone to insufficient ability to distinguish different fault modes, resulting in the risk of misjudgment. This invention combines sub-model training with joint training, ensuring that the training set of each fault diagnosis sub-model contains only positive samples, normal samples, and negative samples of other faults corresponding to that fault. This enhances the sensitivity of each sub-model to its target fault, while constraining misjudgment of non-target faults, thereby achieving the specificity of each fault sub-model and improving the accuracy of the overall model.

[0075] (2) In the prior art, deep learning models lack constraints on negative samples during training, which may lead to misjudgment when inputting non-target fault samples, affecting the reliability of diagnosis. In the training process of the fault diagnosis sub-model, the present invention introduces a first loss function, which imposes constraints on the output probability of negative samples to ensure that the output probability of negative samples is lower than a set threshold, thereby reducing the probability of misjudgment and improving the reliability of the model in practical applications.

[0076] (3) In the existing technology, traditional feature extraction methods have limited ability to process high-dimensional nonlinear and multi-source sensor data, making it difficult to fully characterize the complex operating state of GIS equipment under different typical faults, resulting in insufficient accuracy of fault diagnosis. This invention extracts features from multi-dimensional operating data by combining traditional signal processing methods with the pre-training process of deep learning networks; it realizes high-level abstract characterization of state information such as UHF partial discharge, ultrasonic partial discharge, ultraviolet spectroscopy, mechanical vibration, mechanical stroke, coil current, and infrared temperature rise, improving the feature expression ability of high-dimensional nonlinear data, and providing accurate and stable input features for subsequent fault diagnosis of ultra-high voltage GIS equipment based on multi-dimensional operating data, thereby significantly improving the accuracy and robustness of the fault diagnosis model.

[0077] (4) In the prior art, due to the limited number of GIS equipment fault samples, relying solely on measured operating data for the characterization and analysis of typical faults and operating characteristics may result in biases or omissions, leading to the fault diagnosis model failing to fully reflect the fault characteristics. This invention utilizes a correlation analysis method based on virtual simulation technology to construct a corresponding GIS equipment simulation model for each typical fault occurrence, and acquires multidimensional operating data during the simulation process to analyze the relationship between GIS temperature characteristics and other operating characteristics and each typical fault mode. This technical feature effectively compensates for the limited number of measured data samples, improves the comprehensiveness and accuracy of the characterization relationship between typical faults and operating characteristics, thereby providing a reliable basis for constructing a high-quality GIS equipment fault set, enhancing the accuracy and stability of the fault diagnosis model, and reducing the risk of misjudgment due to insufficient samples.

[0078] (5) In the prior art, feature selection and representation are arbitrary, making it difficult to guarantee the correlation between the extracted features and typical faults. This invention uses a correlation analysis method based on virtual simulation technology to analyze the operating characteristics of historical operating data and simulation data, and selects the top K most representative operating features of each typical fault based on the Pearson correlation coefficient, ensuring a high correlation between fault features and typical faults, thereby improving the effectiveness and accuracy of the input features of the diagnostic model.

[0079] (6) In the existing technology, there is a lack of means to uniformly model multi-type sensor data, which makes it difficult to take into account both local and global features. This invention adopts a CvT neural network structure, which uses a convolutional token embedding layer to extract local features in each sub-model and a convolutional Transformer module to extract global features, thereby realizing efficient feature fusion and information representation of single-channel multi-dimensional operating features and improving the fault detection capability of the model. Attached Figure Description

[0080] Figure 1 This is a flowchart of a GIS equipment fault diagnosis method according to an embodiment of the present invention;

[0081] Figure 2 This is a schematic diagram of the fault diagnosis sub-model structure according to an embodiment of the present invention. Detailed Implementation

[0082] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0083] Example 1:

[0084] This embodiment specifically provides a GIS equipment fault diagnosis method based on sub-model training and joint training, such as Figure 1 As shown, it includes the following steps:

[0085] Step S1: Obtain historical operating data of the GIS equipment. The historical operating data includes multi-dimensional operating data corresponding to various typical faults of the GIS equipment.

[0086] Typical faults include insulation faults, mechanical structure faults, and electrical connection faults. Multidimensional operational data includes various types of data, such as UHF partial discharge signals, ultrasonic partial discharge signals, ultraviolet spectral signals, mechanical vibration signals, mechanical travel signals, operating coil current signals, and GIS equipment temperature rise data. UHF partial discharge signals are obtained through real-time monitoring and acquisition of internal partial discharge phenomena using UHF sensors installed in the GIS equipment. Ultrasonic partial discharge signals are obtained in real-time through acoustic wave detection using ultrasonic sensors installed in the GIS equipment. Ultraviolet spectral signals are obtained in real-time through ultraviolet spectral sensors installed inside the GIS equipment. Mechanical vibration signals are obtained in real-time through vibration sensors installed on key mechanical components of the GIS equipment. Mechanical travel signals are obtained in real-time through displacement sensors installed on moving parts such as circuit breakers in the GIS equipment. Operating coil current signals are obtained in real-time through current sensors installed in the control circuit of the GIS equipment. GIS equipment temperature rise data is obtained in real-time through temperature sensors installed in various key parts of the GIS equipment.

[0087] Step S2: Based on the pre-training process of traditional signal processing methods combined with deep learning networks, feature extraction is performed on the multi-dimensional operational data to obtain the operational feature sets corresponding to each typical fault, specifically including:

[0088] Traditional signal processing methods are used to extract features from each operational data point in the multidimensional operational data corresponding to each typical fault, thereby obtaining multiple primary features of each typical fault.

[0089] The pre-training process of deep learning networks is used to process multiple primary features to obtain multiple operational features, which are then used as the operational feature set corresponding to each typical fault.

[0090] Step S3: Based on the correlation analysis method assisted by virtual simulation technology, the representational relationship between each typical fault and the operational characteristics is analyzed through the operational feature set corresponding to each typical fault. Specifically, this includes:

[0091] For each typical fault, a simulation model of the GIS equipment under the typical fault is constructed, and the simulation model of the GIS equipment under each fault is simulated and run on the simulation platform to obtain multi-dimensional operational data during the simulation process; specifically, the construction of the GIS equipment simulation model for each typical fault includes:

[0092] For insulation faults, a partial discharge power source model is introduced into the insulation system inside the GIS equipment. The partial discharge location, discharge gap, discharge voltage and dielectric parameters are adjusted so that the simulation model can present partial discharge phenomena of different degrees, thereby simulating the changes of UHF signals, ultrasonic signals and temperature rise characteristics under insulation fault conditions.

[0093] For mechanical structure faults, mechanical parameter changes are introduced into the circuit breaker and operating mechanism model, such as contact wear, spring force decay, guide rail deformation or lubrication coefficient changes. By modifying the mechanical kinematic parameters, the mechanical stroke, operating coil current and vibration signal exhibit the characteristics of abnormal mechanical structure.

[0094] For electrical connection faults, the contact resistance, conductor connection tightness, or wire aging coefficient are adjusted in the conductive circuit and contact point model of the GIS equipment. This allows the current, voltage, and local temperature rise signals during the simulation to reflect the state of poor contact or discontinuous conductivity, thereby generating electrical connection fault characteristics. Feature extraction is performed on the multi-dimensional operational data obtained during the simulation to obtain operational feature sets for each typical fault.

[0095] The set of operational features extracted from historical operational data will be merged with the set of operational features obtained from simulation to construct a correlation analysis dataset;

[0096] Based on the Pearson correlation coefficient analysis method, the correlation coefficient between each typical fault and each operating characteristic is calculated through the correlation analysis dataset, thereby obtaining the characterization relationship between each typical fault and the operating characteristic.

[0097] Step S4: Construct a GIS equipment fault set based on the representation relationship and operational feature set, specifically including:

[0098] The correlation coefficients between each typical fault and each operational feature in the characterization relationship are sorted according to the magnitude of the correlation coefficients to obtain the correlation ranking results corresponding to each typical fault.

[0099] For each typical fault, the top K operational features with the highest correlation coefficients are selected from the correlation ranking results, and the corresponding operational features are screened from the correlation analysis dataset obtained from the simulation to construct the fault set corresponding to the typical fault; K is a constant.

[0100] Perform the filtering operation on each typical fault in turn to obtain the fault set corresponding to each typical fault;

[0101] The fault sets corresponding to all typical faults are summarized to construct the GIS equipment fault set. The GIS equipment fault set includes the fault set corresponding to each typical fault. The fault set of each typical fault includes the top K operating features with the highest correlation coefficients selected from the correlation analysis dataset.

[0102] Step S5: Construct a GIS fault diagnosis model based on CvT neural networks, specifically including:

[0103] For each typical fault, K operational features selected from the fault set corresponding to the typical fault are used as input to construct a fault diagnosis sub-model based on the CvT neural network structure; such as Figure 2 As shown, Figure 2 (a) in the figure is the fault diagnosis sub-model, which includes a three-level cascaded CvT feature extraction structure. The CvT feature extraction structure includes a convolutional token embedding layer and a convolutional Transformer module. The convolutional token embedding layer is used to perform convolutional projection on K running features and convert the input single-channel feature matrix into a token sequence. Figure 2 (b) is the convolutional Transformer module, which includes a self-attention layer and a feedforward network layer. It extracts the global correlation of high-dimensional operational features through the attention mechanism. The input of the fault diagnosis sub-model is a feature matrix composed of K operational features, and the output is the probability distribution corresponding to each typical fault and normal mode.

[0104] The output probability distributions of the fault diagnosis sub-models corresponding to all typical faults are integrated, and the probability outputs of each fault diagnosis sub-model are input into the integrated decision module.

[0105] The integrated decision module calculates the probability distribution of each fault diagnosis sub-model based on weighted fusion, confidence calibration or Softmax normalization mechanism, and outputs the final fault classification result and the corresponding confidence level.

[0106] By combining the fault diagnosis sub-models of each typical fault with the integrated decision module, a GIS fault diagnosis model based on CvT neural network is constructed.

[0107] Step S6: Based on the GIS equipment fault set, perform sub-model training and joint training on the GIS fault diagnosis model to obtain the trained GIS fault diagnosis model, specifically including:

[0108] Step A1: For each fault diagnosis sub-model in the GIS fault diagnosis model The fault set corresponding to typical faults in GIS equipment fault set. Compared with several normal samples The faults were merged, and samples from other typical fault sets of GIS equipment fault sets were also included. Randomly select a number of samples as the negative sample set Merge the datasets to construct the first training dataset. :

[0109]

[0110] in, Indicates the first i The first training dataset for a typical fault; This indicates the first fault in the GIS equipment fault set. i A set of typical faults This represents the normal sample set. Indicates the first i The negative sample set represents a typical fault; the normal sample set represents the operating characteristics of GIS equipment under normal operating conditions.

[0111] Step A2: Transfer the first training dataset Input the corresponding fault diagnosis sub-model The first loss function is used for the fault diagnosis sub-model. For separate training, the first loss function is:

[0112]

[0113] in, Indicates sample The label, with a value of 1, indicates that it belongs to the fault diagnosis sub-model. Typical faults; Represents the fault diagnosis sub-model For the sample The probability of the output; This represents the maximum allowed output probability threshold for negative samples, used to constrain misclassification of negative samples; These are the weighting coefficients; Represents the first loss function;

[0114] Step A3: Perform steps A1-A2 sequentially for each fault diagnosis sub-model until all fault diagnosis sub-models have completed individual training;

[0115] Step A4: Merge the first training datasets corresponding to all fault diagnosis sub-models to construct the second training dataset;

[0116] Step A5: Input the second training dataset into the entire CvT neural network-based GIS fault diagnosis model, and jointly train the GIS fault diagnosis model using the second loss function to obtain the trained GIS fault diagnosis model. The second loss function is formulated as follows:

[0117]

[0118] in, For the second loss function, For the second training dataset, This is a set of labels, including various typical faults and normal modes; For the sample The true label, if it belongs to the category c but ,otherwise ; For the GIS fault diagnosis model in the joint training phase, input Output category The predicted probability.

[0119] Step S7: Implement fault diagnosis of GIS equipment using the trained GIS fault diagnosis model, specifically including:

[0120] Acquire real-time multidimensional operational data of the GIS equipment to be tested;

[0121] Feature extraction is performed on multidimensional operational data to obtain the corresponding operational feature matrix, which is then input into each fault diagnosis sub-model to obtain the probability output of each sub-model for each typical fault and normal mode.

[0122] The output probability distribution of each fault diagnosis sub-model is input into the integrated decision module, and the final probability of each typical fault and normal mode is calculated comprehensively based on weighted fusion, confidence calibration or Softmax normalization mechanism.

[0123] Based on the final probability distribution, the fault type or normal state of the GIS equipment is determined, and the corresponding confidence value is output, providing a basis for equipment operation status assessment and fault early warning.

[0124] This invention proposes a GIS equipment fault diagnosis method based on sub-model training and joint training. First, it acquires historical multi-dimensional operational data of the GIS equipment during various typical faults to ensure a sufficient and diverse data foundation for fault diagnosis. Then, it uses traditional signal processing methods combined with the pre-training process of deep learning networks to extract features from this multi-dimensional operational data, obtaining operational feature sets corresponding to each typical fault. This step can combine low-level features to form more abstract high-level feature representations, enabling automatic learning of complex nonlinear state information and improving feature representation capabilities and diagnostic accuracy. To overcome the potential problems of insufficient samples and biases in analysis based solely on measured data, this invention further utilizes a correlation analysis method assisted by virtual simulation technology to construct a GIS equipment simulation model under typical faults. Multi-dimensional operational data is acquired through simulation operation, and correlation analysis is performed by combining the simulation data with historical data to accurately assess the representational relationship between each typical fault and operational features, ensuring that the constructed GIS equipment fault set comprehensively covers the characteristics of various typical faults. Based on this fault set, a three-level cascaded CvT neural network fault diagnosis model is constructed. Each typical fault corresponds to a sub-model. Local and global features are extracted through convolutional token embedding layers and convolutional Transformer modules to accurately identify high-dimensional operational features. Simultaneously, the outputs of each sub-model are comprehensively calculated through an integrated decision module to obtain the final fault classification result and confidence level, thereby improving the reliability of the diagnosis. To further optimize diagnostic performance, this invention adopts a strategy combining sub-model training and joint training. First, each fault diagnosis sub-model is trained separately using an amplified fault set, normal samples, and negative samples, ensuring that each sub-model is most sensitive to its corresponding fault while avoiding misclassification of other faults and normal samples. Then, the training data of each sub-model are merged, and the entire GIS fault diagnosis model is optimized through joint training to ensure the stability and accuracy of the overall model performance. Finally, the trained GIS fault diagnosis model is used to analyze real-time multi-dimensional operational data of GIS equipment, outputting the fault type and corresponding confidence level of the equipment. This achieves efficient, accurate, and reliable diagnosis of the operational status of GIS equipment, providing a scientific basis for equipment maintenance, fault early warning, and operation management.

[0125] Example 2:

[0126] This embodiment provides a GIS equipment fault diagnosis system based on sub-model training and joint training, including: a data acquisition module, a feature extraction module, a correlation analysis module, a fault set construction module, a GIS fault diagnosis model module, a sub-model training module, a joint training module, and a fault diagnosis output module. The system comprises the following modules: a data acquisition module to obtain multi-dimensional operational data of GIS equipment during normal operation and under various typical faults; a feature extraction module to process the acquired multi-dimensional operational data based on traditional signal processing methods combined with a deep learning network pre-training process to obtain operational feature sets corresponding to each typical fault; a correlation analysis module to construct a GIS equipment simulation model under typical faults and acquire simulation operational data using a correlation analysis method assisted by virtual simulation technology, combining the simulation data with historical data to perform feature correlation analysis, thereby evaluating the representational relationship between each typical fault and operational features; a fault set construction module to construct a fault set corresponding to each typical fault based on the analysis results, forming a complete GIS equipment fault set; a sub-model training module to train each fault diagnosis sub-model separately using corresponding fault samples, normal samples, and negative samples, making the sub-model sensitive to its corresponding fault and reducing misjudgments; a joint training module to merge the training data of each sub-model and jointly train the entire GIS fault diagnosis model to optimize overall performance; and a fault diagnosis output module to apply the trained model to the real-time operational data of the GIS equipment, achieving efficient and accurate diagnosis of equipment fault types and confidence levels, providing a scientific basis for equipment operation and maintenance.

[0127] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0128] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for fault diagnosis of GIS equipment based on sub-model training and joint training, characterized in that, Includes the following steps: Acquire historical operational data of GIS equipment, including multi-dimensional operational data corresponding to various typical faults of GIS equipment; Based on the pre-training process of traditional signal processing methods combined with deep learning networks, feature extraction is performed on the multi-dimensional operational data to obtain operational feature sets corresponding to each typical fault; the operational feature set includes multiple operational features. A correlation analysis method based on virtual simulation technology is used to analyze the representational relationship between each typical fault and the operational features through the operational feature sets corresponding to each typical fault. Construct a GIS equipment fault set based on the representational relationships and operational feature set; Construct a GIS fault diagnosis model based on CvT neural network; Based on the GIS equipment fault set, the GIS fault diagnosis model is trained by sub-model training and joint training respectively to obtain the trained GIS fault diagnosis model. Fault diagnosis of GIS equipment is achieved through a trained GIS fault diagnosis model. The pre-training process based on traditional signal processing methods combined with deep learning networks extracts features from the multi-dimensional operational data to obtain operational feature sets corresponding to each typical fault, specifically including: Traditional signal processing methods are used to extract features from each operational data point in the multidimensional operational data corresponding to each typical fault, thereby obtaining multiple primary features of each typical fault. The pre-training process of deep learning networks is used to process multiple primary features to obtain multiple operational features, which are then used as the operational feature set corresponding to each typical fault. The pre-training process using a deep learning network processes multiple primary features to obtain multiple operational features, which are then used as the operational feature set corresponding to each typical fault. Specifically, this includes: Multiple primary features from multidimensional operational data are input into a deep neural network model. Pre-training is used to initialize network parameters. Lower-level primary features are combined to form more abstract higher-level weight representations. The hierarchical structure of the deep neural network is utilized to automatically learn and output the corresponding higher-level feature representations, specifically including: The primary features of UHF partial discharge signals, ultrasonic partial discharge signals, ultraviolet spectral signals, and mechanical vibration signals are input into a convolutional neural network (CNN). The time-frequency local features are extracted through convolution operations to generate corresponding high-level feature representations. The primary features of the mechanical travel signal and the operating coil current signal are input into a network structure that combines a convolutional neural network (CNN) and a recurrent neural network (RNN) to extract temporal and periodic features and generate a high-level feature representation of the mechanical travel and the operating coil current. The primary features of GIS equipment temperature rise data are input into a structure combining a convolutional neural network (CNN) and a long short-term memory network (LSTM) to extract the temporal dynamic features of the temperature rise signal and generate a high-level feature representation of the GIS equipment temperature rise data. The obtained high-level feature representations are used as operational features to obtain the operational feature sets corresponding to each typical fault. The correlation analysis method based on virtual simulation technology analyzes the representational relationship between each typical fault and the operational features through the operational feature set corresponding to each typical fault, specifically including: For each typical fault, a simulation model of the GIS equipment under the typical fault is constructed, and the simulation model of the GIS equipment under each fault is simulated and run on the simulation platform to obtain multi-dimensional operation data during the simulation process. Feature extraction is performed on the multi-dimensional operational data obtained during the simulation to obtain operational feature sets under each typical fault. The set of operational features extracted from historical operational data will be merged with the set of operational features obtained from simulation to construct a correlation analysis dataset; Based on the Pearson correlation coefficient analysis method, the correlation coefficient between each typical fault and each operating feature is calculated through the correlation analysis dataset, thereby obtaining the characterization relationship between each typical fault and the operating feature. The step of training the GIS fault diagnosis model by performing sub-model training and joint training on the GIS equipment fault set to obtain the trained GIS fault diagnosis model specifically includes: Step A1: For each fault diagnosis sub-model in the GIS fault diagnosis model The fault set corresponding to typical faults in GIS equipment fault set. Compared with several normal samples The faults were merged, and samples from other typical fault sets of GIS equipment fault sets were also included. Randomly select a number of samples as the negative sample set Merge the datasets to construct the first training dataset. : in, Indicates the first i The first training dataset for a typical fault; This indicates the first fault in the GIS equipment fault set. i A set of typical faults This represents the normal sample set. Indicates the first i A negative sample set of typical faults; the normal samples are the operating characteristics corresponding to the normal mode of GIS equipment; Step A2: Transfer the first training dataset Input the corresponding fault diagnosis sub-model The first loss function is used for the fault diagnosis sub-model. Conduct individual training: Step A3: Perform steps A1-A2 sequentially for each fault diagnosis sub-model until all fault diagnosis sub-models have completed individual training; Step A4: Merge the first training datasets corresponding to all fault diagnosis sub-models to construct the second training dataset; Step A5: Input the second training dataset into the entire GIS fault diagnosis model based on CvT neural network, and use the second loss function to jointly train the GIS fault diagnosis model to obtain the trained GIS fault diagnosis model.

2. The GIS equipment fault diagnosis method based on sub-model training and joint training according to claim 1, characterized in that, The typical faults mentioned include insulation faults, mechanical structure faults, and electrical connection faults. The multidimensional operating data includes various types of operating data, including UHF partial discharge signals, ultrasonic partial discharge signals, ultraviolet spectral signals, mechanical vibration signals, mechanical stroke signals, operating coil current signals, and GIS equipment temperature rise data. The ultra-high frequency partial discharge signal is obtained by real-time monitoring and acquisition of the partial discharge phenomenon inside the GIS equipment through an ultra-high frequency UHF sensor installed in the GIS equipment. The ultrasonic partial discharge signal is detected by an ultrasonic sensor installed in the GIS equipment and acquired in real time. The ultraviolet spectral signal is acquired in real time by an ultraviolet spectral sensor installed inside the GIS equipment; The mechanical vibration signal is acquired in real time by installing vibration sensors on key mechanical components of the GIS equipment. The mechanical travel signal is acquired in real time by a displacement sensor installed on the moving parts of the GIS equipment circuit breaker; The operating coil current signal is acquired in real time by a current sensor installed in the control circuit of the GIS equipment; The temperature rise data of the GIS equipment is collected in real time by temperature sensors installed in various key parts of the GIS equipment.

3. The GIS equipment fault diagnosis method based on sub-model training and joint training according to claim 1, characterized in that, The method employs traditional signal processing techniques to extract features from the multidimensional operational data corresponding to each typical fault, obtaining multiple primary features for each typical fault, specifically including: For UHF partial discharge signals in multidimensional operational data, two-dimensional or three-dimensional spectral analysis methods are used to extract primary features of the UHF partial discharge signals, including average discharge power, initial discharge voltage, extinction discharge voltage, and average discharge current. For ultrasonic partial discharge signals in multidimensional operational data, time-domain features are extracted by combining time-domain analysis and frequency-domain analysis. These time-domain features are used as primary features of the ultrasonic partial discharge signals. The time-domain features include root mean square, variance, absolute integral mean, kurtosis, and skewness. For the ultraviolet spectral signal in the multidimensional running data, the Burg recursive algorithm is applied to estimate the power spectrum and extract the power spectral density, which is then used as the primary feature of the ultraviolet spectral signal. For mechanical vibration signals in multidimensional operational data, time-domain analysis and frequency-domain analysis based on Fourier transform are used, combined with time-frequency envelope analysis based on Hilbert-Huang transform, to extract primary features of the mechanical vibration signals, including amplitude, root mean square, peak value, and variance. For the mechanical travel signal in the multidimensional operation data, based on the physical motion process of the circuit breaker, the motion speed, acceleration, opening time, closing time, and contact overtravel parameters are extracted as primary features of the mechanical travel signal. For the operating coil current signal in the multidimensional operating data, the peak current, effective current value, and curve secant slope are extracted by analyzing the current waveform as primary features of the operating coil current signal. For the temperature rise data of GIS equipment in multidimensional operational data, time series analysis is applied to extract the temperature change trend and obtain the temperature rise rate, temperature fluctuation amplitude and temperature peak value as the primary features of the temperature rise data of GIS equipment.

4. The GIS equipment fault diagnosis method based on sub-model training and joint training according to claim 1, characterized in that, The construction of the GIS equipment fault set based on the representation relationship and the operational feature set specifically includes: The correlation coefficients between each typical fault and each operational feature in the characterization relationship are sorted according to the magnitude of the correlation coefficients to obtain the correlation ranking results corresponding to each typical fault. For each typical fault, the top K operational features with the highest correlation coefficients are selected from the correlation ranking results, and the corresponding operational features are screened from the correlation analysis dataset obtained from the simulation to construct the fault set corresponding to the typical fault; K is a constant. Perform the filtering operation on each typical fault in turn to obtain the fault set corresponding to each typical fault; The fault sets corresponding to all typical faults are summarized to construct the GIS equipment fault set. The GIS equipment fault set includes the fault set corresponding to each typical fault. The fault set of each typical fault includes the top K operating features with the highest correlation coefficients selected from the correlation analysis dataset.

5. The GIS equipment fault diagnosis method based on sub-model training and joint training according to claim 1, characterized in that, The construction of the GIS fault diagnosis model based on CvT neural network specifically includes: For each typical fault, K operational features selected from the fault set corresponding to the typical fault are used as input to construct a fault diagnosis sub-model based on a CvT neural network structure. The fault diagnosis sub-model includes a three-level cascaded CvT feature extraction structure, which includes a convolutional token embedding layer and a convolutional Transformer module. The convolutional token embedding layer is used to perform convolutional projection on the K operational features, converting the input single-channel feature matrix into a token sequence. The convolutional Transformer module includes a self-attention layer and a feedforward network layer, which extract the global correlation of high-dimensional operational features through an attention mechanism. The input of the fault diagnosis sub-model is a feature matrix composed of K operational features, and the output is the probability distribution corresponding to each typical fault and normal mode. The output probability distributions of the fault diagnosis sub-models corresponding to all typical faults are integrated, and the probability outputs of each fault diagnosis sub-model are input into the integrated decision module. The integrated decision module, based on weighted fusion, confidence calibration, or Softmax normalization mechanisms, comprehensively calculates the probability distributions output by each fault diagnosis sub-model and outputs the final fault classification result and the corresponding confidence level. The fault diagnosis sub-models of each typical fault are combined with the integrated decision module to construct a GIS fault diagnosis model based on CvT neural network.

6. The GIS equipment fault diagnosis method based on sub-model training and joint training according to claim 1, characterized in that, The first loss function is defined as follows: in, Indicates sample The label, with a value of 1, indicates that it belongs to the fault diagnosis sub-model. Typical faults; Represents the fault diagnosis sub-model For the sample The probability of the output; This represents the maximum allowed output probability threshold for negative samples, used to constrain misclassification of negative samples. These are the weighting coefficients; Represents the first loss function; The second loss function is expressed as follows: in, For the second loss function, For the second training dataset, This is a set of labels, including various typical faults and normal modes; For the sample The true label, if it belongs to the category c but ,otherwise ; For the GIS fault diagnosis model in the joint training phase, input Output category The predicted probability.

Images