A GIS device defect diagnosis model training method and defect diagnosis method

Abstract

The present application relates to the technical field of power equipment monitoring, and specifically provides a GIS equipment defect diagnosis model training method and a defect diagnosis method, comprising: acquiring historical multi-modal defect symptom signals of GIS equipment; performing feature extraction, time sequence coding fusion and modal coding fusion on the historical multi-modal defect symptom signals to obtain historical fusion features; performing random mask occlusion on the unlabeled training samples, filling the occluded feature positions into the placeholder vectors to be learned, then inputting the deep fusion network for feature interaction, and reversely updating the network parameters and the placeholder vectors based on the reconstruction loss; connecting the pre-trained deep fusion network to a diagnosis classification head, using labeled training samples to train a GIS equipment defect diagnosis model. The present application method realizes deep fusion of multi-modal data through cross-modal cross-attention, breaks the dependence on complete data through mask reconstruction, and accurately diagnoses GIS equipment defects.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring technology, and in particular to a training method and a defect diagnosis method for a GIS equipment defect diagnosis model. Background Technology

[0002] Gas-insulated gas-insulated switchgear (GIS) is a core hub device in modern smart grid systems, and its reliability and stability directly affect the safe power supply of the entire grid. With the development of power systems towards higher voltage and larger capacity, GIS equipment is subjected to the combined effects of multiple stresses from electricity, heat, mechanical forces, and the environment during long-term operation, inevitably leading to potential defects such as insulation aging, poor contact, and mechanical loosening. Once these defects develop into sudden failures, they will cause large-scale power outages and huge economic losses. Therefore, real-time and accurate health status monitoring and diagnosis of GIS equipment has become a research hotspot and urgent need in the field of power equipment operation and maintenance.

[0003] With the rapid popularization of the Internet of Things (IoT), smart sensing, and edge computing technologies, GIS equipment in modern substations typically deploys multiple types of sensors simultaneously, forming a multi-parameter monitoring matrix covering various physical quantities such as electricity, sound, heat, and chemistry. This multi-dimensional sensor network can capture equipment operating status information from all angles and perspectives, providing a rich data foundation for more accurate fault diagnosis. However, facing this massive, heterogeneous, and multi-dimensional sensor data, how to effectively perform data fusion and feature extraction to achieve accurate assessment of equipment health status is a major technical challenge currently facing smart grid operation and maintenance. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a training method and a defect diagnosis method for a GIS equipment defect diagnosis model, so as to achieve accurate diagnosis of complex defects in GIS equipment.

[0005] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0006] In a first aspect, the present invention provides a training method for a GIS equipment defect diagnosis model, comprising:

[0007] Acquire historical multimodal defect symptom signals of GIS equipment;

[0008] Feature extraction, temporal coding fusion, and modal coding fusion are performed on the historical multimodal defect symptom signals to obtain historical fused features;

[0009] The historical fusion features are divided into unlabeled training samples and labeled training samples;

[0010] The unlabeled training samples are randomly masked, and the masked feature positions are filled into the placeholder vector to be learned. Then, the samples are input into a deep fusion network for feature interaction. The network parameters and the placeholder vector are updated in reverse based on the reconstruction loss to complete the self-supervised pre-training of the deep fusion network.

[0011] The pre-trained deep fusion network is connected to the diagnostic classification head, and the labeled training samples are used to train a GIS equipment defect diagnosis model.

[0012] In some embodiments of the present invention, the historical multimodal defect symptom signals include high-frequency physical field signals and low-frequency physical field signals;

[0013] The high-frequency physical field signals include: ultra-high frequency partial discharge signals and ultrasonic partial discharge signals;

[0014] The low-frequency physical field signals include: temperature signals and SF6 insulating gas decomposition component concentration signals.

[0015] In some embodiments of the present invention, feature extraction, temporal coding fusion, and modal coding fusion are performed on the historical multimodal defect symptom signals to obtain historical fused features, including:

[0016] Perform a short-time Fourier transform on each of the aforementioned high-frequency physical field signals to obtain the corresponding time-frequency energy spectrum;

[0017] Two-dimensional convolution block processing is performed on each of the time-frequency energy spectra to obtain the corresponding high-frequency physical features;

[0018] One-dimensional causal convolution processing is performed on each of the low-frequency physical field signals to obtain the corresponding low-frequency physical features;

[0019] Each of the high-frequency physical features and each of the low-frequency physical features are added and fused element by element with their respective time-series codes and modal codes to obtain historical fused features.

[0020] In some embodiments of the present invention, the unlabeled training samples are randomly masked, and the masked feature positions are filled into placeholder vectors to be learned; then, they are input into a deep fusion network for feature interaction, and the network parameters and the placeholder vectors are updated in reverse based on the reconstruction loss to complete the self-supervised pre-training of the deep fusion network, including:

[0021] The unlabeled training samples are randomly masked, and the masked feature positions are filled into the placeholder vector to be learned to obtain incomplete feature samples.

[0022] The incomplete feature samples are input into a deep fusion network for feature interaction, and the deep fusion features are output.

[0023] Using the deep fusion features, complete features are reconstructed, and the reconstruction loss of the complete features and the unlabeled training samples is calculated.

[0024] Based on the reconstruction loss, the network parameters of the deep fusion network and the placeholder vector are updated in reverse, and the self-supervised pre-training of the deep fusion network is completed iteratively.

[0025] In some embodiments of the present invention, the unlabeled training samples are randomly masked, and the masked feature positions are filled into placeholder vectors to be learned to obtain incomplete feature samples, including:

[0026] ;

[0027] in, Indicates a sample with incomplete features; This represents unlabeled training samples; This represents a binary mask indicator matrix, where an element 0 indicates that the corresponding feature position is occluded, and an element 1 indicates that the corresponding feature position is preserved. Represents the Hadama product; This represents the placeholder vector to be learned.

[0028] In some embodiments of the present invention, the deep fusion network is a cross-modal Transformer encoder;

[0029] The incomplete feature samples are input into a deep fusion network for feature interaction, and the output deep fusion features include:

[0030] By employing a self-attention mechanism, the temporal evolution patterns of each modal feature sequence in the incomplete feature samples are mined.

[0031] The cross-modal cross-attention mechanism is used to mine the cross-modal correlation between modal feature sequences in the incomplete feature samples; wherein, the unoccluded modal feature sequence is used as the query vector of the cross-attention mechanism.

[0032] After multiple interactions between self-attention and cross-modal cross-attention mechanisms, deep fusion features are output.

[0033] In some embodiments of the present invention, the formula for calculating the reconstruction loss is as follows:

[0034] ;

[0035] in, Indicates the losses incurred during reconstruction; This represents the set of feature location indices that are obscured by a random mask. Indicates the number of feature locations that are occluded; Indicates the first Deeply fused features output by a deep fusion network from each feature location; This represents the function for reconstructing the decoder; Indicates the first Reconstructed features at each feature location; Indicates the first The true features at each feature location; This represents the square of the L2 norm.

[0036] Secondly, the present invention also provides a method for diagnosing defects in GIS equipment, comprising:

[0037] Real-time acquisition of multimodal defect signs from GIS equipment;

[0038] Feature extraction, temporal coding fusion, and modal coding fusion are performed on the multimodal defect symptom signals to obtain fused features; wherein, in the case of missing features in the fused features, placeholder vectors are used to fill the missing feature positions;

[0039] The fused features are input into the GIS equipment defect diagnosis model to obtain the defect diagnosis results of the GIS equipment;

[0040] The placeholder vector and the GIS equipment defect diagnosis model are pre-trained using the training method described above.

[0041] Thirdly, the present invention also provides an electronic device, including: a processor, and a memory storing a program, the program including instructions that, when executed by the processor, cause the processor to perform the training method or defect diagnosis method described above.

[0042] Fourthly, the present invention also provides a non-transitory machine-readable medium storing computer instructions for causing the computer to perform the training method or defect diagnosis method described above.

[0043] Compared with the prior art, the above-described technical solution of the present invention has the following advantages:

[0044] This invention proposes a training method and a defect diagnosis method for a GIS equipment defect diagnosis model. By introducing a cross-modal cross-attention mechanism and abandoning the traditional shallow feature stitching method, it deeply mines the nonlinear causal relationships between multiple physical modes such as electricity, sound, heat, and chemistry, significantly improving the diagnostic effect of complex defects in GIS equipment. Utilizing a multimodal mask reconstruction mechanism, the model is forced to learn the intrinsic correlation rules across modes during the pre-training stage. This enables the model to automatically infer and compensate for missing information when field sensors fail or data is missing, breaking the dependence on complete data input and ensuring the stable operation of the diagnostic system around the clock. Furthermore, this invention constructs a new paradigm for industrial large-scale models: "large-scale unlabeled self-supervised pre-training combined with supervised fine-tuning using minimal samples," effectively solving the problems of scarce fault samples and difficult labeling. Through a high- and low-frequency differentiated feature mapping network, integrating temporal location encoding and modal attribute encoding, heterogeneous multi-parameter data is uniformly mapped into an input form understandable by the large model, breaking down the input barriers between heterogeneous industrial data and the large model architecture. Attached Figure Description

[0045] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other embodiments based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart illustrating a training method for a GIS equipment defect diagnosis model provided in an embodiment of the present invention;

[0047] Figure 2 This is a schematic diagram of the training architecture of the GIS equipment defect diagnosis model in an embodiment of the present invention;

[0048] Figure 3 This is a schematic diagram of the self-supervised pre-training architecture of the deep fusion network in an embodiment of the present invention;

[0049] Figure 4 This is a flowchart illustrating a method for diagnosing defects in GIS equipment provided in an embodiment of the present invention;

[0050] Figure 5 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0051] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0052] like Figure 1 As shown, this embodiment of the invention provides a training method for a GIS equipment defect diagnosis model. Figure 1 This is a flowchart illustrating the training method for a GIS equipment defect diagnosis model. This flowchart only shows the logical sequence of the method in this embodiment. In other possible embodiments of the invention, different methods may be used, provided they do not conflict with each other. Figure 1 Complete the steps shown or described in the order indicated.

[0053] See Figure 1 and Figure 2 The method of this invention specifically includes the following steps:

[0054] Step S101: Obtain historical multimodal defect symptom signals of GIS equipment.

[0055] This invention comprehensively captures symptom information of defects in GIS equipment based on four physical modes: electrical, acoustic, thermal, and chemical. Each physical mode typically includes data from multiple measuring points or multiple sensing channels in actual engineering. For example, the same type of sensor may be deployed in different air chambers, different intervals, and different key parts of the GIS.

[0056] In this embodiment of the invention, the historical multimodal defect symptom signals include: ultra-high frequency (UHF) partial discharge signals, ultrasonic (AE) partial discharge signals, temperature signals, and SF6 insulating gas decomposition component concentration signals.

[0057] The UHF partial discharge signal corresponds to the electrical mode. It is collected by UHF sensors installed in each air chamber of GIS. It can accurately capture transient electromagnetic wave signals generated by insulation defects. It has the characteristics of fast response speed and high sensitivity, and is the core indicator for early warning of partial discharge in GIS.

[0058] Ultrasonic (AE) partial discharge signals correspond to acoustic modes. Ultrasonic sensors are mounted on key parts of GIS housings such as flanges, circuit breakers, and disconnect switches to detect elastic vibration waves caused by partial discharge or mechanical loosening. They have strong resistance to on-site electromagnetic interference and can assist in the spatial location of defects.

[0059] Temperature signals correspond to thermal modes. Contact thermocouple sensors are installed in locations prone to heat generation, such as GIS housings, busbar contacts, and cable joints, to reflect localized overheating caused by poor contact or excessive current in real time.

[0060] The concentration signals of SF6 insulating gas decomposition components correspond to chemical modes. Electrochemical sensors installed at the gas sampling ports of each gas chamber are used to detect the concentrations of characteristic decomposition products such as SO2 and H2S, which are used to characterize the long-term deterioration of insulating materials and identify latent insulation defects.

[0061] Step S102: Perform feature extraction, temporal coding fusion, and modal coding fusion on the historical multimodal defect symptom signals to obtain historical fusion features.

[0062] Ultra-high frequency partial discharge signals and ultrasonic partial discharge signals are high-frequency physical field signals, which have a fast time-varying speed and a high sampling rate; temperature signals and SF6 insulating gas decomposition component concentration signals are low-frequency physical field signals, which change slowly and have a long sampling interval.

[0063] In this embodiment of the invention, step S102 may include:

[0064] Step S1021: After performing a short-term Fourier transform (STFT) on each high-frequency physical field signal, the corresponding time-frequency energy spectrum is obtained.

[0065] For ultra-high frequency waveforms and ultrasonic signals with extremely high sampling rates, traditional time-domain analysis is difficult to capture the energy distribution of transient discharges. Therefore, this invention converts one-dimensional high-frequency time series signals into two-dimensional time-frequency energy maps through short-time Fourier transform.

[0066] Step S1022: Perform two-dimensional convolution block processing on each time-frequency energy spectrum to obtain the corresponding high-frequency physical characteristics.

[0067] The time-frequency energy spectrum is processed by two-dimensional convolution block (2D-Patch segmentation) and flattened into a one-dimensional token sequence to adapt to the input specification of the Transformer encoder.

[0068] Specifically, the expression for the high-frequency physical characteristics is as follows:

[0069] ;

[0070] in, This represents a two-dimensional convolution operation where the kernel size and stride are set to the same block size. ; This represents the two-dimensional convolution kernel weight matrix. The hidden dimension of the Transformer encoder; This represents the number of channels in the time-frequency energy spectrum; This is the bias vector. This is a flattening function. For the generated high-frequency physical characteristics, The number of tokens after segmentation based on high-frequency physical features.

[0071] Step S1023: Perform one-dimensional causal convolution processing on each low-frequency physical field signal to obtain the corresponding low-frequency physical features.

[0072] To prevent future information leakage from disrupting temporal causality, a one-dimensional causal convolution is performed on the low-frequency physical field signal to obtain low-frequency physical features.

[0073] Specifically, the expression for low-frequency physical characteristics is as follows:

[0074] ;

[0075] in, This represents a one-dimensional causal convolution operation; The input is a low-frequency time series containing The chemical gas concentration or temperature at multiple measurement points, with a length of . For one-dimensional causal convolution kernel weights, is the kernel length. This is a one-dimensional causal convolution bias. For the generated low-frequency physical characteristics, The number of tokens after segmentation based on low-frequency physical features.

[0076] By adjusting the convolution stride, This allows low-frequency physical features to align with high-frequency physical features in the time dimension.

[0077] Step S1024: Add and fuse each high-frequency physical feature and each low-frequency physical feature with their respective time-series code and modal code element by element to obtain historical fused features.

[0078] The construction of historical fusion features is intended to address the shortcomings of the Transformer architecture itself, such as its lack of sequence awareness and inability to distinguish between different physical modalities.

[0079] Among them, the timing coding adopts sine or cosine timing position coding, which gives the model the ability to perceive the absolute temporal sequence of signals; the modal coding is a learnable attribute coding set for different physical modes, which ensures that the model can accurately distinguish the type of signal being processed.

[0080] The method of this invention is designed with a multi-physics feature mapping network tailored to the characteristics of industrial time-series data. Through STFT time-frequency conversion, 2D-Patch and 1D causal convolution segmentation, and fusion of temporal location and modal attribute encoding, it successfully transforms raw sensor data of varying lengths and physical representations into a unified token sequence that can be adapted to large models, thus breaking down the input barriers between heterogeneous industrial data and large models.

[0081] Step S103: Divide the historical fusion features into unlabeled training samples and labeled training samples.

[0082] The historical multimodal defect symptom signals used in this invention are derived from the historical operation monitoring data of GIS equipment accumulated by the power grid over many years. Due to the limitations of actual conditions in industrial sites, most daily operation data have not been manually labeled.

[0083] In this embodiment of the invention, 95% of the historical fusion features are classified as unlabeled training samples for subsequent self-supervised pre-training processes; only 5% of the historical fusion features, which have been verified on-site and / or confirmed by expert consultation, are classified as labeled training samples. The labeled training samples are required to cover typical categories such as air gaps inside insulators, poor busbar contact, suspended metal particles, surface contamination of basin insulators, and normal operating conditions.

[0084] This classification method makes full use of the value of massive amounts of unlabeled data, fundamentally breaking through the bottleneck of the scarcity of high-quality labeled data in the industrial field.

[0085] Step S104: Randomly mask the unlabeled training samples and fill the masked feature positions into the placeholder vector to be learned; then input them into the deep fusion network for feature interaction, and update the network parameters and placeholder vectors in reverse based on the reconstruction loss to complete the self-supervised pre-training of the deep fusion network.

[0086] See Figure 3 In this embodiment of the invention, step S104 may include:

[0087] Step S1041: Randomly mask the unlabeled training samples, fill the masked feature positions into the placeholder vector to be learned, and obtain incomplete feature samples.

[0088] Incomplete feature samples can be represented as:

[0089] ;

[0090] in, Indicates a sample with incomplete features; This represents unlabeled training samples; This represents a binary mask indicator matrix. Element 0 in the matrix indicates that the corresponding feature position is occluded and the token at these feature positions is replaced with a placeholder. Element 1 indicates that the corresponding feature position is retained. Represents the Hadama product; This represents a placeholder vector to be learned, used to fill in damaged locations.

[0091] During the training phase, a joint masking strategy is implemented for 30% of the tokens in the random occlusion sequence, including random dot masks and full-modal masks. The full-modal mask is executed with a 10% probability to specifically enhance the model's robustness against defects.

[0092] Step S1042: Input the incomplete feature samples into the deep fusion network for feature interaction and output the deep fusion features.

[0093] In this embodiment of the invention, the deep fusion network employs a cross-modal Transformer encoder.

[0094] The cross-modal Transformer encoder consists of 12 identical encoder layers stacked together. Each encoder layer consists of a self-attention sub-layer, a cross-modal cross-attention sub-layer, and a feedforward neural network (FFN) sub-layer. Each sub-layer is followed by a residual connection and a layer normalization module to ensure stable gradient propagation and feature representation capability.

[0095] In the self-attention computation stage, incomplete feature samples are split into independent feature sequences according to physical modalities, and each sequence is input into a multi-head self-attention sub-layer. Feature vectors of the same modality are multiplied by three independent learnable projection matrices to generate query, key, and value vectors for the corresponding dimensions. Attention weights are calculated within each attention head, and the value vectors are weighted and summed. Finally, the outputs of all heads are concatenated and mapped through a linear layer. After residual connections and layer normalization, the degradation trends of each physical quantity over time are extracted, such as the pulse frequency changes of partial discharge signals and the slow increase in temperature signals.

[0096] The cross-modal attention sublayer is the core of achieving deep multiphysics coupling. Unlike the self-attention sublayer, the query vector in this sublayer comes only from the feature sequence of the unoccluded modality, while the key and value vectors come from the feature sequences of all modalities (including incomplete modes filled with placeholders). By calculating the similarity weights between the query vector and all key vectors, the model can actively address and focus on features in other modalities related to the current defect, based on reliable known features. For example, it can use UHF transient pulse features to associate the hysteresis generation features of SF6 chemical components.

[0097] Nonlinear feature mapping is performed on the feature sequence after interaction between the self-attention sublayer and the cross-modal cross-attention sublayer through the feedforward neural network sublayer.

[0098] During the encoding process, the first encoder layer takes the incomplete feature samples as input, and the second to twelfth encoder layers take the output features of the previous encoder layer as input. Through a progressive serial encoding method, single-modal temporal correlation features and cross-modal coupling features are continuously mined.

[0099] After 12 layers of iterative encoding, the complete feature sequence output by the 12th encoder layer is used as the final deep fusion feature.

[0100] Step S1043: Utilize deep fusion features to reconstruct complete features, and calculate the reconstruction loss between the complete features and the unlabeled training samples.

[0101] Step S1044: Based on the reconstruction loss, update the network parameters and placeholder vectors of the deep fusion network in reverse order, and iteratively complete the self-supervised pre-training of the deep fusion network.

[0102] A lightweight reconstruction decoder is connected at the end of the deep fusion network. In this embodiment, the reconstruction decoder adopts a shallow linear mapping network.

[0103] The reconstruction decoder inverts and restores the deep fusion features output by the deep fusion network, generating reconstructed features corresponding to each mask occlusion position. All reconstructed features are then added to the incomplete feature samples to obtain complete features. The complete features are compared with the unlabeled training samples, and the reconstruction loss is calculated. This loss is iteratively optimized through self-supervised backpropagation, completing the self-supervised pre-training of the deep fusion network.

[0104] In this embodiment of the invention, the reconstruction loss is the mean squared error (MSE) loss, calculated as follows:

[0105] ;

[0106] in, Indicates the losses incurred during reconstruction; This represents the set of feature location indices that are obscured by a random mask. Indicates the number of feature locations that are occluded; Indicates the first Deep fusion features at each feature location; This represents the function for reconstructing the decoder; Indicates the first The reconstructed features at each feature location are then added one by one to the occluded feature locations in the incomplete feature sample to obtain complete features. Indicates the first The true features at each feature location; This represents the square of the L2 norm.

[0107] This invention employs a unified labeling and cross-modal cross-attention mechanism, abandoning the traditional surface feature splicing method. It achieves underlying logical alignment and deep feature interaction of heterogeneous electrical, acoustic, thermal, and chemical sensor signals in a unified high-dimensional space, accurately uncovering nonlinear causal relationships in the defect evolution process, and significantly improving the comprehensive identification and diagnostic accuracy of latent and complex composite defects. Utilizing a missing-tolerance mechanism in multimodal mask reconstruction, the model can automatically complete the inference and compensation of missing sensor data based on intact modal signals, effectively coping with harsh conditions such as sensor failure and communication interruptions. This breaks the strict dependence of traditional diagnostic models on data integrity, ensuring stable and reliable system operation around the clock.

[0108] Step S105: Connect the pre-trained deep fusion network to the diagnostic classification head, and use labeled training samples to train a GIS equipment defect diagnosis model.

[0109] The GIS equipment defect diagnosis model consists of a deep fusion network and a diagnostic classification head.

[0110] The diagnostic classification head is a lightweight fully connected network structure, consisting of two fully connected layers. Non-linear activation functions are used between the layers to achieve feature mapping, and a Softmax normalization layer is connected at the end.

[0111] In the small sample fine-tuning stage, a small number of manually labeled typical defect samples are used as training samples. After the training samples are processed by the feature interaction of the deep fusion network, the output vector corresponding to the CLS global feature label of the Token sequence head is extracted, and the output vector is input into the diagnostic classification head.

[0112] The CLS global feature token is a special type of classification token in the Transformer architecture, set at the beginning of the input sequence. In the supervised fine-tuning and diagnosis stage of this embodiment, the output vector corresponding to the CLS global feature token is extracted to perform classification operations. The output vector is used to represent the global information of the entire sequence.

[0113] We employ the Cross-Entropy Loss function to perform supervised fine-tuning of the diagnostic classification head and some deep fusion network parameters. At the same time, we introduce an L2 regularization term to suppress the overfitting problem caused by small sample training, thereby improving the model's generalization ability and diagnostic accuracy.

[0114] The formula for calculating the fine-tuning loss function is as follows:

[0115] ;

[0116] in, Indicates fine-tuning loss; Indicates the number of samples in the mini-batch during fine-tuning; This indicates the total number of preset GIS equipment defect status categories, including normal operation status and various defect types, such as normal, insulator discharge, poor busbar contact, and compound defects. For the first The true label indicator variable of the nth sample, if the nth sample The true defect type of the sample is the first If the status is a defect, the value is 1; otherwise, the value is 0. This represents the predicted probability output by the GIS equipment defect diagnosis model, i.e., the... The sample was determined to be the first The probability value of a defect-like state; This represents all trainable parameters in the model, including some parameters of the deep fusion network and all parameters of the diagnostic classification head; This represents the squared L2 norm of the trainable parameters, used to constrain parameter magnitudes and improve model generalization ability. is the regularization coefficient.

[0117] The embodiments of the present invention also provide a GIS equipment defect diagnosis method. The GIS equipment defect diagnosis method is based on the GIS equipment defect diagnosis model trained by the above embodiments, and realizes accurate diagnosis of multimodal defect symptom signals of real-time acquired GIS equipment.

[0118] See Figure 4 The method of this invention specifically includes the following steps:

[0119] Step S201: Real-time acquisition of multimodal defect symptom signals from GIS equipment.

[0120] In the actual operation scenario of the substation, the on-site real-time online monitoring system collects multimodal defect symptom signals of GIS equipment. The types of signals collected are consistent with the historical multimodal defect symptom signals in step S101, including UHF partial discharge signals, ultrasonic partial discharge signals, temperature signals, and SF6 insulating gas decomposition component concentration signals.

[0121] Step S202: Perform feature extraction, temporal coding fusion, and modal coding fusion on the multimodal defect symptom signals to obtain fused features.

[0122] Step S203: Input the fused features into the GIS equipment defect diagnosis model to obtain the defect diagnosis results of the GIS equipment.

[0123] When issues such as sensor loss or communication interruption lead to missing features in the fused features, pre-trained placeholder vectors are used to fill the missing feature positions. The deep fusion network utilizes learned cross-modal coupling relationships and relies on the remaining intact signals to implicitly compensate for the missing information. Subsequently, the CLS global feature markers at the head of the token sequence are extracted, and their corresponding output vectors are used as global representation vectors. These global representation vectors highly condense the overall state information of the GIS equipment across multiple dimensions, including electrical, acoustic, thermal, and chemical aspects.

[0124] The global representation vector is input into the diagnostic classification head. After Softmax probability normalization, the classification mapping of GIS equipment defect status is completed, and the probability of each preset defect status category is output, so as to realize real-time and accurate diagnosis of GIS equipment defects.

[0125] The embodiments of the present invention also provide a non-transitory machine-readable medium storing a computer program, wherein the computer program, when executed by a computer processor, is used to cause the computer to perform a training method for a GIS equipment defect diagnosis model or a GIS equipment defect diagnosis method according to the embodiments of the present invention.

[0126] Embodiments of the present invention also provide a computer program product, including a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to execute a training method for a GIS equipment defect diagnosis model or a GIS equipment defect diagnosis method according to embodiments of the present invention.

[0127] Embodiments of this invention also provide an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, which, when executed by the at least one processor, causes the electronic device to perform a training method for a GIS device defect diagnosis model or a GIS device defect diagnosis method according to embodiments of this invention.

[0128] refer to Figure 5The present invention will now describe a structural block diagram of an electronic device that can serve as an embodiment of the present invention, serving as an example of a hardware device applicable to various aspects of the present invention. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present invention described and / or claimed herein.

[0129] like Figure 5 As shown, the electronic device includes a computing unit 101, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 102 or a computer program loaded from a storage unit 108 into a random access memory (RAM) 103. The RAM 103 may also store various programs and data required for the operation of the electronic device. The computing unit 101, ROM 102, and RAM 103 are interconnected via a bus 104. An input / output (I / O) interface 105 is also connected to the bus 104.

[0130] Multiple components in the electronic device are connected to I / O interface 105, including: input unit 106, output unit 107, storage unit 108, and communication unit 109. Input unit 106 can be any type of device capable of inputting information into the electronic device. Input unit 106 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of the electronic device. Output unit 107 can be any type of device capable of presenting information and may include, but is not limited to, a display, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 108 may include, but is not limited to, disks and optical discs. Communication unit 109 allows the electronic device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks, and may include, but is not limited to, modems, network cards, infrared communication devices, and / or wireless communication transceivers, such as Bluetooth devices, WiFi devices, WiMax devices, cellular communication devices, and / or the like.

[0131] The computing unit 101 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 101 include, but are not limited to, CPUs, graphics processing units (GPUs), various special-purpose artificial intelligence (AI) computing units, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. The computing unit 101 performs the various methods and processes described above. For example, in some embodiments, the method embodiments of the present invention can be implemented as computer programs tangibly contained in a machine-readable medium, such as storage unit 108. In some embodiments, part or all of the computer program can be loaded and / or installed on an electronic device via ROM 102 and / or communication unit 109. In some embodiments, the computing unit 101 can be configured to perform the methods described above by any other suitable means (e.g., by means of firmware).

[0132] Computer programs for implementing the methods of embodiments of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0133] In the context of embodiments of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable signal medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, or infrared systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0134] It should be noted that the term "comprising" and its variations used in the embodiments of this invention are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The modifications of "one" and "a plurality" mentioned in the embodiments of this invention are illustrative and not restrictive, and those skilled in the art should understand that unless explicitly indicated otherwise in the context, they should be understood as "one or more".

[0135] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of this invention are subject to strict compliance with relevant laws, regulations, and regulatory requirements in their collection, storage, use, processing, transmission, provision, and disclosure, and adhere to the principles of legality, legitimacy, necessity, and good faith. The acquisition of relevant information and data is premised on the user's explicit consent or other legitimate reasons, and a clear and convenient authorization management approach is provided to the user, allowing the user to independently choose to consent, withdraw consent, or refuse to provide relevant information. For functions that rely on user information, if the user does not authorize or withdraws authorization, the corresponding technical function cannot be implemented, and the technical solution of this invention is not applicable in this scenario.

[0136] The steps described in the method embodiments provided by the present invention can be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of protection of the present invention is not limited in this respect.

[0137] The term "embodiment" in this specification refers to a specific feature, structure, or characteristic described in connection with an embodiment that may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily imply the same embodiment, nor does it imply independence or alternativeity from other embodiments. The various embodiments in this specification are described in a related manner, with reference to each other for similar or identical parts. In particular, for apparatus, device, and system embodiments, since they are substantially similar to method embodiments, the description is relatively simple, and relevant details are referred to in the description of the method embodiments.

[0138] The above embodiments merely illustrate several implementation methods of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of protection. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A training method for a GIS equipment defect diagnosis model, characterized in that, include: The historical multimodal defect symptom signals of GIS equipment are acquired, including high-frequency physical field signals and low-frequency physical field signals. Feature extraction, temporal coding fusion, and modal coding fusion are performed on the historical multimodal defect symptom signals to obtain historical fused features, including: Perform a short-time Fourier transform on each of the aforementioned high-frequency physical field signals to obtain the corresponding time-frequency energy spectrum; Two-dimensional convolution block processing is performed on each of the time-frequency energy spectra to obtain the corresponding high-frequency physical features; One-dimensional causal convolution processing is performed on each of the low-frequency physical field signals to obtain the corresponding low-frequency physical features; Each of the high-frequency physical features and each of the low-frequency physical features are added and fused element by element with their respective time-series codes and modal codes to obtain historical fused features; The historical fusion features are divided into unlabeled training samples and labeled training samples; The unlabeled training samples are randomly masked, and the masked feature positions are filled into the placeholder vector to be learned. Then, the samples are input into a deep fusion network for feature interaction. The network parameters and the placeholder vector are updated in reverse based on the reconstruction loss to complete the self-supervised pre-training of the deep fusion network. The pre-trained deep fusion network is connected to the diagnostic classification head, and the labeled training samples are used to train a GIS equipment defect diagnosis model. The unlabeled training samples are randomly masked, and the masked feature locations are filled into placeholder vectors to be learned; then, they are input into a deep fusion network for feature interaction, including: The unlabeled training samples are randomly masked, and the masked feature positions are filled into the placeholder vector to be learned to obtain incomplete feature samples. The incomplete feature samples are input into a deep fusion network for feature interaction, and the deep fusion features are output. The deep fusion network is a cross-modal Transformer encoder; The incomplete feature samples are input into a deep fusion network for feature interaction, and the output deep fusion features include: By employing a self-attention mechanism, the temporal evolution patterns of each modal feature sequence in the incomplete feature samples are mined. The cross-modal cross-attention mechanism is used to mine the cross-modal correlation between modal feature sequences in the incomplete feature samples; wherein, the unoccluded modal feature sequence is used as the query vector of the cross-attention mechanism. After multiple interactions between self-attention and cross-modal cross-attention mechanisms, deep fusion features are output.

2. The training method for the GIS equipment defect diagnosis model according to claim 1, characterized in that, The high-frequency physical field signals include: ultra-high frequency partial discharge signals and ultrasonic partial discharge signals; The low-frequency physical field signals include: temperature signals and SF6 insulating gas decomposition component concentration signals.

3. The training method for the GIS equipment defect diagnosis model according to claim 1, characterized in that, The network parameters and the placeholder vector are updated in reverse based on the reconstruction loss to complete the self-supervised pre-training of the deep fusion network, including: Using the deep fusion features, complete features are reconstructed, and the reconstruction loss of the complete features and the unlabeled training samples is calculated. Based on the reconstruction loss, the network parameters of the deep fusion network and the placeholder vector are updated in reverse, and the self-supervised pre-training of the deep fusion network is completed iteratively.

4. The training method for the GIS equipment defect diagnosis model according to claim 1, characterized in that, The unlabeled training samples are randomly masked, and the masked feature locations are filled into the placeholder vector to be learned, resulting in incomplete feature samples, including: ； in, Indicates a sample with incomplete features; This represents unlabeled training samples; This represents a binary mask indicator matrix, where an element 0 indicates that the corresponding feature position is occluded, and an element 1 indicates that the corresponding feature position is preserved. Represents the Hadama product; This represents the placeholder vector to be learned.

5. The training method for the GIS equipment defect diagnosis model according to claim 3, characterized in that, The formula for calculating the reconstruction loss is as follows: ； in, Indicates the losses incurred during reconstruction; This represents the set of feature location indices that are obscured by a random mask. Indicates the number of feature locations that are occluded; Indicates the first Deeply fused features output by a deep fusion network from each feature location; This represents the function for reconstructing the decoder; Indicates the first Reconstructed features at each feature location; Indicates the first The true features at each feature location; This represents the square of the L2 norm.

6. A method for diagnosing defects in GIS equipment, characterized in that, include: Real-time acquisition of multimodal defect signs from GIS equipment; Feature extraction, temporal coding fusion, and modal coding fusion are performed on the multimodal defect symptom signals to obtain fused features; wherein, in the case of missing features in the fused features, placeholder vectors are used to fill the missing feature positions; The fused features are input into the GIS equipment defect diagnosis model to obtain the defect diagnosis results of the GIS equipment; The placeholder vector and the GIS equipment defect diagnosis model are pre-trained using the training method described in any one of claims 1 to 5.

7. An electronic device, comprising: A processor and a memory storing a program, characterized in that the program includes instructions that, when executed by the processor, cause the processor to perform the training method of any one of claims 1 to 5 or the GIS equipment defect diagnosis method of claim 6.

8. A non-transitory machine-readable medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute the training method of any one of claims 1 to 5 or the GIS equipment defect diagnosis method of claim 6.

Images