GIS equipment partial discharge detection system based on deep learning

Abstract

The invention provides a GIS equipment partial discharge detection system based on deep learning, and the system comprises an edge layer which is used for collecting sensing data collected by each sensor node, extracting a sensing feature vector, and transmitting a compressed waveform fragment; the fog calculation layer is used for determining defect positions by using a time difference of arrival algorithm according to the sensing data uploaded by the edge layer, and obtaining defect types and confidence coefficients according to the compressed waveform fragments and the sensing feature vectors in combination with a double-flow deformable attention convolutional network model; and the cloud platform is used for updating parameters of the double-flow deformable attention convolutional network model by adopting a meta-learning method according to the defect position, the defect type and the confidence which are uploaded by the fog computing layer in combination with virtual training data generated by the digital twin simulation model, and sending the updated parameters to the fog computing layer. The GIS equipment partial discharge detection system provided by the invention is light in weight and low in cost.

Description

Technical Field

[0001] The present invention relates to the technical field of partial discharge identification, and particularly to a partial discharge detection system for GIS equipment based on deep learning. Background Art

[0002] Gas Insulated Switchgear (GIS) equipment is widely used in the power system due to its many advantages. GIS equipment plays an important role in the power system. GIS equipment is affected by various factors such as electricity, heat, machinery, and environment, and complex chemical and physical changes will occur inside it, resulting in gradual performance degradation. Partial discharge is an important sign of early faults of GIS equipment. Timely and accurate detection of partial discharge is of great significance for ensuring the safe and stable operation of GIS equipment.

[0003] In the renovation project of old substations, there may be a large number of GIS equipment to be detected. On the one hand, the traditional, complex and expensive partial discharge detection system for GIS equipment will increase the cost of the renovation project. If a high-cost detection system is used, the cost of large-scale detection will be very high. On the other hand, the spatial layout of old substations may be relatively complex, and there may be installation and commissioning work of other equipment carried out simultaneously during the renovation process. The traditional, complex and expensive partial discharge detection system for GIS equipment is not suitable for use in this environment.

[0004] Therefore, there is an urgent need for a lightweight and low-cost partial discharge detection system for GIS equipment. Summary of the Invention

[0005] The present invention provides a partial discharge detection system for GIS equipment based on deep learning, and its main purpose is to provide a lightweight and low-cost partial discharge detection system for GIS equipment.

[0006] An embodiment of the present invention provides a partial discharge detection system for GIS equipment based on deep learning. The detection system includes an edge layer, a fog computing layer, and a cloud platform. The edge layer includes N sensor nodes installed inside the GIS equipment. The fog computing layer includes M fog nodes installed in the substation control room. Each sensor node includes an ultra-high frequency sensor and an ultrasonic sensor. The fog nodes communicate with the sensor nodes through LoRA technology, and the fog nodes communicate with the cloud platform through 5G slicing technology, where M and N are both positive integers, and M is less than N;

[0007] The edge layer is used to collect the sensing data collected by each sensor node, extract the sensing feature vectors, and send compressed waveform segments, and the compressed waveform segments are reconstructed from the compressed sensing data;

[0008] The fog computing layer is used to determine the defect location according to the sensing data uploaded by the edge layer by using the time difference of arrival algorithm, and obtain the defect type and confidence level according to the compressed waveform segment and the sensing feature vector, in combination with the dual-stream deformable attention convolutional network model;

[0009] Among them, the model parameters of the dual-stream deformable attention convolutional network are received from the cloud platform;

[0010] The cloud platform is used to train and update the parameters of the dual-stream deformable attention convolutional network model by using the meta-learning method according to the defect location, defect type, and confidence level uploaded by the fog computing layer, in combination with the virtual training data generated by the digital twin simulation model, and send the updated parameters to the fog computing layer.

[0011] Further, the compressed waveform segment includes an electromagnetic wave signal and an ultrasonic wave signal. The dual-stream deformable attention convolutional network model includes a deformable convolutional layer, a polarization attention module, a three-dimensional capsule network layer, an adaptive convolutional layer, a deformable LSTM network, a cross-modal attention layer, a global average pooling layer, and a fully connected classification layer. The steps of obtaining the defect type and confidence level according to the compressed waveform segment and the sensing feature vector, in combination with the dual-stream deformable attention convolutional network model, include:

[0012] Input the electromagnetic wave signal into the deformable convolutional layer, and determine the dynamic offset according to the gradient of the electromagnetic wave signal to obtain the electromagnetic convolution feature;

[0013] Input the electromagnetic convolution feature into the polarization attention module to obtain the horizontal polarization feature;

[0014] Input the horizontal polarization feature into the three-dimensional capsule network layer, first generate primary capsules, and then perform dynamic routing and iteration multiple times to obtain 10 types of discharge modes;

[0015] Input the time-frequency matrix of the ultrasonic wave signal into the adaptive convolutional layer, and dynamically adjust the convolution kernel weights of the adaptive convolutional layer according to the frequency band energy entropy of the ultrasonic wave signal to obtain the ultrasonic convolution feature;

[0016] Input the ultrasonic convolution feature into the deformable LSTM network to obtain the time series feature encoding;

[0017] Input the 10 types of discharge modes and the time series feature encoding into the cross-modal attention layer, calculate the cross-attention matrix and feature interaction to obtain the fusion feature matrix;

[0018] Input the fusion feature matrix into the global average pooling layer and the fully connected classification layer to obtain the defect type and the confidence level.

[0019] Furthermore, the cloud platform is also used to determine the loss function of the two-stream deformable attention convolutional network model during the training process by using the finite-difference time-domain method. The specific steps are as follows:

[0020] Using the finite-difference time-domain method, calculate the theoretical magnetic field signal and the theoretical ultrasonic signal generated by the power source at the defect position;

[0021] According to the electromagnetic wave signal, the theoretical magnetic field signal, the ultrasonic signal, and the theoretical ultrasonic signal, calculate the waveform difference loss;

[0022] According to the gradient of the waveform difference loss, backpropagate through the finite-difference time-domain method to the parameters of the two-stream deformable attention convolutional network model to update the parameters, so that the two-stream deformable attention convolutional network model reaches convergence faster.

[0023] Furthermore, the fog computing layer also performs emergency level assessment and transmission strategy selection according to the defect position, the defect type, the confidence level, and the device historical status database;

[0024] If the confidence level is greater than 90% and the emergency level is greater than or equal to 4, the fog computing layer sends the sensing data and the alarm signal to the cloud platform through 5G;

[0025] If the confidence level is between 70% and 90% and the emergency level is greater than or equal to 3, the fog computing layer sends the sensing feature vector and the request for cloud verification signal to the cloud platform through 5G;

[0026] If the confidence level is less than 70%, the fog computing layer locally stores the sensing data.

[0027] Furthermore, the cloud platform is also used to output a typical defect feature library, and according to the defect type and the typical defect feature library, determine whether the defect type belongs to a typical defect;

[0028] If it is determined that the defect type belongs to a typical defect, output a red warning instruction and generate an escape route map.

[0029] Furthermore, the cloud platform also compares the discharge patterns of multiple GIS devices in the same substation according to the single discharge event report sent by the fog computing layer and the device full-life cycle data to identify systematic risks, where the single discharge event report includes the defect position, the defect type, and the confidence level.

[0030] Furthermore, the cloud platform also generates virtual samples using a digital twin simulation model based on the abnormal data and diagnosis results sent by the fog computing layer, and uses a generative adversarial network to align the simulation and measured data distributions to construct a mixed training set. It adopts a meta-learning method to update the parameters of the dual-stream deformable attention convolutional network model, and retains the historical version model of the dual-stream deformable attention convolutional network model, supporting testing and rollback of both the historical version model and the updated model.

[0031] Furthermore, the cloud platform is also used to output compliance inspection reports and structured case knowledge bases based on complete discharge event records, maintenance process records and result feedback, defect attribution analysis based on knowledge graphs and key operation records stored on blockchain.

[0032] Furthermore, the edge layer includes a sensing unit, a signal conditioning circuit, an edge processor and a wireless communication module.

[0033] This paper proposes a deep learning-based partial discharge detection system for GIS equipment. The system adopts a three-layer collaborative architecture design consisting of an edge layer, a fog computing layer, and a cloud platform. The edge layer adopts ultra-low-power hardware modules and only runs signal preprocessing and simple feature extraction algorithms, eliminating complex computing modules. The sensor also integrates a LoRa communication module, replacing high-power solutions such as 5G / Wi-Fi.

[0034] The fog computing layer disassembles the two-stream deformable attention convolutional network model, and the edge layer performs feature extraction. The fog computing layer only runs a lightweight decision head, reducing computing power requirements by 70%; the nodes of the multi-fog computing layer share computing power through 5G direct connection, avoiding the high deployment cost of centralized servers; and directly make local decisions for high-confidence events, reducing data transmission on the cloud platform by 80%.

[0035] The cloud platform generates 70% of the training data through simulation using the finite-difference time-domain method, reducing reliance on expensive measured data. It also only needs to deliver model difference parameters rather than the complete model, saving bandwidth costs.

[0036] In summary, the present invention has the characteristics of extreme edge simplification, which can sacrifice local accuracy in exchange for global energy efficiency; it also has the characteristics of dynamic collaboration in the fog layer, replacing centralized strong computing power with distributed weak computing power; it also has the characteristics of virtual and real fusion of cloud platforms, reducing the demand for measured data with simulated data. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A schematic diagram of the structure of a deep learning-based partial discharge detection system for GIS equipment provided by an embodiment of the present invention;

[0038] Figure 2It is a schematic structural diagram of a two-stream deformable attention convolutional network model provided in an embodiment of the present invention;

[0039] Figure 3 It is a schematic diagram of the determination process of the training loss function of a DSDA-CNN model provided in an embodiment of the present invention.

[0040] The implementation, functional features and advantages of the present invention will be further described in conjunction with embodiments with reference to the accompanying drawings. Specific Embodiments

[0041] The following details the embodiments of the present application. The examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions from beginning to end. The embodiments described below by referring to the accompanying drawings are exemplary only for explaining the present application and should not be construed as limiting the present application.

[0042] In order to enable those skilled in the art of the present technology to better understand the solution of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present application.

[0043] In the embodiments of the present application, "at least one" means one or more; "a plurality" means two or more. In the description of the present application, terms such as "first", "second", "third", etc. are only used for the purpose of distinguishing descriptions and cannot be understood as indicating or implying relative importance, nor can they be understood as indicating or implying an order. In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present application, the meaning of "a plurality" is two or more unless otherwise specifically defined.

[0044] The reference to "an embodiment" or "some embodiments" etc. in this specification means that specific features, structures or characteristics described in connection with the embodiment are included in one or more embodiments of the present application. Thus, the terms "including", "comprising", "having" and their variants in this specification all mean "including but not limited to" unless otherwise specifically emphasized in other ways.

[0045] The GIS equipment integrates primary equipment in a substation except for transformers, including circuit breakers, disconnectors, earthing switches, instrument transformers, arresters, busbars, connectors, outgoing line terminals, etc. After optimized design, they are organically combined into a whole, completely enclosed in a metal earthed shell, and filled with SF6 insulating gas at a certain pressure.

[0046] Partial discharges in GIS equipment generally include the following four types:

[0047] 1. Corona discharge. Usually, due to the existence of metal burrs, rough conductor surfaces, etc. inside the GIS equipment, the electric field intensity at these burrs will increase significantly. When the electric field intensity exceeds the breakdown field strength of the gas, corona discharge will be triggered.

[0048] 2. Surface discharge. On the one hand, the pollution on the surface of the insulator of the GIS equipment is a common cause of surface discharge. This pollution will reduce the surface insulation performance of the insulator in a humid environment, resulting in uneven surface electric field distribution and thus triggering surface discharge. On the other hand, the existence of air gaps or defects inside the insulator of the GIS equipment will also distort the electric field and trigger surface discharge at the air gaps or defects.

[0049] 3. Floating discharge. If there are metal components with floating potentials in the GIS equipment, floating discharge is likely to occur. Under the action of the electric field, a potential difference will be formed between the floating metal component and the surrounding conductors. When the potential difference is large enough, a discharge phenomenon will occur between the floating metal component and the surrounding conductors, that is, floating discharge.

[0050] 4. Discharge due to internal insulation defects. During long-term operation, the performance of insulation materials will gradually deteriorate due to the action of various stresses such as electricity, heat, and mechanics.

[0051] In view of the above reasons for partial discharges in GIS equipment, in the embodiments of the present invention, sensor nodes are reasonably arranged around the GIS equipment. Each sensor node includes an ultra-high frequency sensor and an ultrasonic sensor, and all sensor nodes belong to the edge layer. Among them, the positions and quantities of the ultra-high frequency sensors and ultrasonic sensors in each sensor node need to be determined according to the size, shape of the GIS equipment and the desired positioning accuracy. The ultra-high frequency sensors and ultrasonic sensors are responsible for collecting ultra-high frequency electromagnetic wave signals and ultrasonic signals generated by partial discharges. During the collection process, parameters such as the sensitivity and sampling frequency of the ultra-high frequency sensors and ultrasonic sensors need to be set appropriately to obtain high-quality signals.

[0052] Figure 1 The structural schematic diagram of a partial discharge detection system for GIS equipment based on deep learning provided by the embodiments of the present invention is as Figure 1As shown in the figure, the detection system includes an edge layer, a fog computing layer, and a cloud platform. The edge layer includes N sensor nodes installed inside the GIS device, and the fog computing layer includes M fog nodes installed in the substation control room. Each sensor node includes an ultra-high frequency sensor and an ultrasonic sensor. The fog nodes communicate with the sensor nodes through LoRA technology, and the fog nodes communicate with the cloud platform through 5G slicing technology. Here, both M and N are positive integers, and M is less than N;

[0053] The edge layer is used to collect the sensing data collected by each sensor node, extract the sensing feature vector, and send the compressed waveform segment, which is reconstructed from the compressed sensing data;

[0054] The fog computing layer is used to determine the defect location according to the sensing data uploaded by the edge layer by using the time difference of arrival algorithm, and obtain the defect type and confidence level according to the compressed waveform segment and the sensing feature vector, in combination with the dual-stream deformable attention convolutional network model;

[0055] Among them, the model parameters of the dual-stream deformable attention convolutional network are received from the cloud platform;

[0056] The cloud platform is used to update the parameters of the dual-stream deformable attention convolutional network model by using the meta-learning method according to the defect location, defect type, and confidence level uploaded by the fog computing layer, combined with the virtual training data generated by the digital twin simulation model, and send the updated parameters to the fog computing layer.

[0057] In the embodiment of the present invention, the edge layer collects the electromagnetic wave signal and ultrasonic wave signal of each sensor node. After the edge layer collects the electromagnetic wave signal and ultrasonic wave signal, the signal can be simply processed by the microprocessor inside the edge layer itself, reducing the hardware requirements of the edge layer, so that the detection system is lightweight.

[0058] In the specific implementation process, the hardware structure of the edge layer includes a sensing unit, a signal conditioning circuit, an edge processor, and a wireless communication module. Specifically, the sensing unit collects ultrasonic wave signals and electromagnetic wave signals. The signal conditioning circuit preprocesses the signals, performs band-pass filtering on the electromagnetic wave signals through envelope detection, and performs wavelet noise reduction on the ultrasonic wave signals through time-frequency analysis; then performs event detection on the noise-reduced electromagnetic wave signals and noise-reduced ultrasonic wave signals, which can specifically be based on threshold comparison or SNN to identify effective discharge pulses. Then the edge processor extracts the features of the detected effective discharge pulses, extracts time-domain features such as pulse peak value, rise time, and repetition frequency, so as to form a sensing feature vector; uses the compressed sensing algorithm to reduce the amount of data. And uploads the sensing data to the fog computing node through the wireless communication module.

[0059] In this embodiment, wireless communication is achieved between the edge layer and the fog computing layer through LoRA technology. Compared with 5G communication technology, it has lower hardware requirements and is more lightweight. Moreover, only simple operations such as data acquisition, preprocessing, feature extraction, and compression are performed in the edge layer, and the simplest and cheapest processor or circuit processing module can achieve this.

[0060] A fog node is a distributed computing unit deployed in a local network, usually located between Internet of Things terminal devices (edge layer) and a cloud computing center, constituting a key intermediate layer in the "edge-fog-cloud" three-tier architecture. In a GIS partial discharge detection system, fog nodes are usually high-performance embedded devices deployed in substations, such as industrial gateways, edge servers, etc.

[0061] Generally, there are more sensor nodes in the edge layer and fewer fog nodes. For example, 50 sensor nodes are installed in this substation, while only 2 fog nodes need to be installed.

[0062] In this embodiment, after the fog computing layer receives the sensing feature vector and compressed waveform segment uploaded by the edge layer, it queries the pre-stored three-dimensional coordinates according to the sensor ID and constructs a sensor spatial relationship matrix, applies dynamic time warping (DTW) to compensate for the propagation delay of the ultrasonic signal, and outputs a sensor spatial coordinate matrix; then, in combination with the electromagnetic wave signal, the defect location is determined.

[0063] In this embodiment, the defect location refers to the position coordinates of partial discharge in the GIS device, and the defect type refers to the specific discharge type, which can specifically refer to the 4 discharge types above; and the compressed waveform segment and the sensing feature vector are input into a two-stream deformable attention convolutional network model, and the specific defect type and confidence level of this discharge can be obtained.

[0064] Specifically, Figure 2 is a schematic structural diagram of a two-stream deformable attention convolutional network model provided in an embodiment of the present invention. As Figure 2 shown, the compressed waveform segment includes an electromagnetic wave signal and an ultrasonic signal. The two-stream deformable attention convolutional network model includes a deformable convolutional layer, a polarization attention module, a three-dimensional capsule network layer, an adaptive convolutional layer, a deformable LSTM network, a cross-modal attention layer, a global average pooling layer, and a fully connected classification layer. According to the compressed waveform segment and the sensing feature vector, in combination with the two-stream deformable attention convolutional network model, obtaining the defect type and the confidence level includes the steps of:

[0065] Input the electromagnetic wave signal into the deformable convolutional layer, and determine the dynamic offset according to the gradient of the electromagnetic wave signal to obtain an electromagnetic convolution feature;

[0066] Input the electromagnetic convolution features into the polarization attention module to obtain horizontal polarization features;

[0067] Input the horizontal polarization features into the three-dimensional capsule network layer. First, generate primary capsules, and then perform dynamic routing and iteration multiple times to obtain 10 types of discharge patterns;

[0068] Input the time-frequency matrix of the ultrasonic signal into the adaptive convolution layer, and dynamically adjust the convolution kernel weights of the adaptive convolution layer according to the frequency band energy entropy of the ultrasonic signal to obtain ultrasonic convolution features;

[0069] Input the ultrasonic convolution features into the deformable LSTM network to obtain temporal feature encoding;

[0070] Input the 10 types of discharge patterns and the temporal feature encoding into the cross-modal attention layer, calculate the cross-attention matrix and feature interaction to obtain a fused feature matrix;

[0071] Input the fused feature matrix into the global average pooling layer and the fully connected classification layer to obtain the defect type and the confidence level.

[0072] In this embodiment, the dual-stream deformable attention convolutional network model is abbreviated as the DSDA-CNN model. The DSDA-CNN model first preprocesses the electromagnetic wave signal and the ultrasonic signal to obtain the time-frequency features of the preprocessed electromagnetic wave signal and ultrasonic signal.

[0073] For the electromagnetic wave signal, input the preprocessed electromagnetic wave signal into the deformable convolution layer. The deformable convolution layer includes 3 groups of deformable convolution kernels, and the dynamic offsets of these 3 groups of deformable convolution kernels are calculated from the input signal gradient, and then electromagnetic convolution features can be obtained.

[0074] In this embodiment, the dynamic offset parameter can reduce the number of parameters of the DSDA-CNN model by 30%.

[0075] Then input the electromagnetic convolution features into the polarization attention module. First, decompose the TE / TM mode components, then generate attention weights, and finally perform feature weighting to output enhanced horizontal polarization features. In this embodiment, the sampling position of the convolution kernel is dynamically adjusted by the deformable convolution layer, and the polarization components related to discharge are adaptively enhanced by the polarization attention module, thereby enhancing the robustness to sensor position deviation and signal deformation. Then input the horizontal polarization features into the three-dimensional capsule network layer. The three-dimensional capsule network layer is abbreviated as 3D-Capsule. The 3D-Capsule first generates primary capsules, and then performs dynamic routing and iteration multiple times to obtain 10 types of discharge patterns.

[0076] For the ultrasonic signal, input the time-frequency matrix of the ultrasonic signal into the adaptive convolutional layer, and dynamically adjust the convolutional kernel weights of the adaptive convolutional layer according to the band energy entropy of the ultrasonic signal, then the ultrasonic convolutional features can be obtained; then input the ultrasonic convolutional features into the deformable LSTM network, first calculate the gating offset and time-varying convolution to obtain the temporal feature encoding.

[0077] Finally, input the 10 types of discharge patterns and the temporal feature encoding into the cross-modal attention layer together, calculate the cross-attention matrix and feature interaction, then the fused feature matrix can be obtained; and input the fused feature matrix into the global average pooling layer and the fully connected classification layer in sequence, then the defect type and confidence can be obtained.

[0078] In this embodiment, through deformable convolution, the number of parameters can be reduced by 30% through dynamic offset; through the capsule network, and the routing protocol is only used at the top layer, the computational amount is reduced by 83%; by using the adaptive convolutional layer, the frequency-domain weight sharing reduces the number of parameters by 80%.

[0079] In this embodiment, by deploying fog nodes inside the distribution substation and deploying artificial intelligence methods inside the fog nodes, the fog nodes can handle relatively complex algorithms or methods, and place the complex processing processes in all sensor nodes in the fog computing layer for processing; the fog nodes are faster than the cloud platform in processing, meeting the real-time requirements of the power system, and can still operate the basic diagnosis function independently when the network is disconnected, only uploading valuable data, reducing the upstream traffic by more than 70%.

[0080] It should be noted that the model parameters of the dual-stream deformable attention convolutional network are received from the cloud platform.

[0081] In the embodiment of the present invention, the dual-stream deformable attention convolutional network is not trained in the fog computing layer, but in the cloud platform. After the cloud platform finishes training, it sends the trained model parameters to the fog computing layer, so that the fog computing layer uses the received model parameters for calculation.

[0082] The cloud platform takes the defect location, defect type, and confidence sent by the fog computing layer as real training data, combines the virtual training data generated by the digital twin simulation model, updates the parameters of the DSDA-CNN model by combining the real training data and the virtual training data, and since it is small-sample data during the update process, the meta-learning method can be used to achieve rapid update of the model parameters, and send the updated model parameters to the fog computing layer. Here, only the changed model parameters are sent to the fog computing layer, which can reduce the amount of data transmitted.

[0083] In this embodiment, the cloud platform can be deployed inside the distribution network center. The hardware of the cloud platform includes computing nodes, GPU accelerators, storage nodes, network switches, and security gateways. Among them, the computing nodes are used for FDTD simulation and the training of the DSDA-CNN model, the GPU accelerators are used for deep learning, the storage nodes are used to store the data uploaded by the fog computing layer, and the network switches are used to achieve backbone switching. Compared with the edge layer and the fog computing layer, the hardware conditions of the cloud platform are better, which can handle the complex training process of the DSDA-CNN model and the complex data processing process, thereby reducing the requirements for the edge layer and the fog computing layer.

[0084] Among them, Figure 3 FIG. is a schematic diagram of the process for determining the training loss function of a DSDA-CNN model provided by an embodiment of the present invention. As Figure 3 shown, the cloud platform is also used to determine the loss function of the dual-stream deformable attention convolutional network model during the training process by using the finite-difference time-domain method. The specific steps are as follows:

[0085] S310, using the finite-difference time-domain method, calculate the theoretical magnetic field signal and the theoretical ultrasonic signal generated by the power source at the defect location;

[0086] S320, calculate the waveform difference loss according to the electromagnetic wave signal and the theoretical magnetic field signal, and the ultrasonic wave signal and the theoretical ultrasonic wave signal;

[0087] S330, according to the gradient of the waveform difference loss, backpropagate through the finite-difference time-domain method to the parameters of the dual-stream deformable attention convolutional network model to update the parameters, so that the dual-stream deformable attention convolutional network model reaches convergence faster.

[0088] In the embodiment of the present invention, the finite-difference time-domain method (abbreviated as FDTD) in electromagnetic field simulation is transformed into a differentiable neural network layer, so that Maxwell's equations become an implicit constraint for model training.

[0089] In the embodiment of the present invention, the finite-difference time-domain method and the CNN model are fused, and the physical simulation is transformed into a differentiable calculation layer and embedded in the neural network to form a closed-loop optimization. The core logic of the fusion is: make the prediction result of the neural network conform to both the data law (CNN learning) and the electromagnetic-acoustic physical law (FDTD constraint); the specific implementation method is: turn the FDTD simulation process into a special "layer" of the network, which can not only forward-compute the physical field but also backpropagate the gradient.

[0090] The specific fusion steps are as follows:

[0091] The sensing signal is input into the DSDA-CNN dual-stream network to extract features, and the discharge source position, discharge intensity, and discharge waveform are output; then, based on the discharge source position, discharge intensity, and discharge waveform, the electromagnetic field distribution generated by the discharge source is calculated using FDTD, and the electromagnetic energy is converted into a sound source to calculate the ultrasonic signal. Finally, the theoretical magnetic field signal and theoretical ultrasonic signal are output; then, based on the theoretical signal and the measured signal, the waveform difference loss is calculated; finally, based on this waveform difference loss, combined with backpropagation, the gradient of the waveform difference loss is backpropagated through the FDTD layer to the prediction parameters of the DSDA-CNN. The differential operator of the FDTD is converted into a sparse matrix to support automatic differentiation. Thus, the DSDA-CNN is forced to adjust its parameters so that the prediction results not only fit the data but also satisfy the physical constraints of the FDTD simulation.

[0092] For example, the CNN predicts that there is a floating discharge at a certain position, but the FDTD simulation finds that the frequency domain difference between the theoretical signal and the measured signal at this position is large, resulting in an increase in the waveform difference loss. Through backpropagation, the CNN is forced to adjust the prediction, which may correct the position coordinates or adjust the discharge intensity, and finally make the prediction match both the data and the physical laws.

[0093] The FDTD layer can be imagined as a physical examiner: the CNN student first submits an answer (predicts the discharge source parameters); the examiner uses FDTD to check whether the answer is physically reasonable; when it is unreasonable, the examiner tells the student how to correct it (gradient backpropagation); the student (CNN) learns both the data and the physics, and finally gives an accurate and reasonable answer. This fusion reduces the false alarm rate by 58% in the detection of high-voltage equipment and enables the model to generalize even with a small amount of data.

[0094] The loss function constraints include physical consistency loss and joint optimization. The physical consistency loss is the theoretical signal waveform minus the original network output of the CNN model; when performing joint optimization, the total loss function = classification loss + 0.3 * physical loss. Then, gradient backpropagation is performed, and gradient reversal is calculated through automatic differentiation (AutoDiff) to update the network parameters.

[0095] A differentiable FDTD solution provided in this embodiment forcibly satisfies Maxwell's equations, can be pre-trained using simulation data (reducing the need for measured data by 70%), and the output can be associated with specific dielectric parameter changes (such as abnormal dielectric constant), so that the DSDA-CNN model reaches convergence faster and improves the training efficiency.

[0096] Furthermore, the fog computing layer also performs emergency level assessment and transmission strategy selection based on the defect position, the defect type, the confidence level, and the device historical status database;

[0097] If the confidence level is greater than 90% and the urgency level is greater than or equal to 4, the fog computing layer sends the sensing data and alarm signals to the cloud platform via 5G;

[0098] If the confidence level is between 70% and 90% and the urgency level is greater than or equal to 3, the fog computing layer sends the sensing feature vectors and requests for cloud review signals to the cloud platform via 5G;

[0099] If the confidence level is less than 70%, the fog computing layer locally stores the sensing data.

[0100] In this embodiment, according to the confidence level and the urgency level, a suitable transmission strategy is selected according to the specific situation. In the case of a higher urgency level, the 5G method with a faster transmission speed is selected to send the data to the cloud platform; in the case of a lower urgency level, only the sensing feature vectors are sent; for data with a low confidence level, the sensing data is directly stored locally.

[0101] Further, the cloud platform is also used to output a typical defect feature library, and based on the defect type and the typical defect feature library, determine whether the defect type belongs to a typical defect;

[0102] If it is determined that the defect type belongs to a typical defect, a red warning instruction is output and an escape route map is generated.

[0103] In this embodiment, the cloud platform also outputs a typical defect feature library, and based on the defect type and the typical defect feature library, determines whether the defect type belongs to a typical defect; if the defect type belongs to a typical defect, the cloud platform issues a red warning instruction and generates an escape route map.

[0104] Further, the cloud platform also compares the discharge patterns of multiple GIS devices in the same substation based on the single discharge event report and the device full - life - cycle data sent by the fog computing layer, and identifies systematic risks, where the single discharge event report includes the defect location, the defect type, and the confidence level.

[0105] Further, the cloud platform also generates virtual samples using the digital twin simulation model based on the abnormal data and diagnostic results sent by the fog computing layer, uses a generative adversarial network to align the simulation and measured data distributions, constructs a mixed training set, adopts a meta - learning method to update the parameters of the dual - stream deformable attention convolutional network model, and retains the historical version model of the dual - stream deformable attention convolutional network model to support the testing and rollback of both the historical version model and the updated model.

[0106] Further, the cloud platform is also used to output a compliance detection report and a structured case knowledge base based on the complete discharge event record, the repair process record and the result feedback, the defect attribution analysis based on the knowledge graph, and the key operation records stored on the blockchain.

[0107] Each module in the above-mentioned GIS device partial discharge detection system based on deep learning can be implemented in whole or in part by software, hardware, or a combination thereof. Each of the above modules can be embedded in the processor of the computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to each of the above modules.

[0108] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to the memory, storage, database, or other media used in the various embodiments provided in the present application can include non-volatile and / or volatile memories. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0109] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the above-mentioned division of functional units and modules is used as an example. In actual applications, the above functions can be allocated to different functional units and modules according to needs, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0110] The above-described embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A GIS equipment partial discharge detection system based on deep learning, characterized in that, The detection system includes an edge layer, a fog computing layer, and a cloud platform. The edge layer includes N sensor nodes installed inside GIS devices. The fog computing layer includes M fog nodes installed in the substation control room. Each sensor node includes an ultra-high frequency sensor and an ultrasonic sensor. The fog nodes communicate with the sensor nodes through LoRA technology. The fog nodes communicate with the cloud platform through 5G slicing technology. Herein, both M and N are positive integers, and M is less than N; The edge layer is used to collect the sensing data collected by each sensor node, extract the sensing feature vectors, and send the compressed waveform segments, which are reconstructed from the compressed sensing data; The fog computing layer is used to determine the defect location according to the sensing data uploaded by the edge layer by using the time difference of arrival algorithm, and obtain the defect type and confidence level according to the compressed waveform segments and the sensing feature vectors, in combination with the dual-stream deformable attention convolutional network model; Among them, the model parameters of the dual-stream deformable attention convolutional network are received from the cloud platform; The cloud platform is used to train and update the parameters of the dual-stream deformable attention convolutional network model by using the meta-learning method according to the defect location, defect type, and confidence level uploaded by the fog computing layer, in combination with the virtual training data generated by the digital twin simulation model, and send the updated parameters to the fog computing layer.

2. The partial discharge detection system for GIS equipment based on deep learning according to claim 1, wherein, The compressed waveform segments include electromagnetic wave signals and ultrasonic wave signals. The dual-stream deformable attention convolutional network model includes a deformable convolutional layer, a polarization attention module, a three-dimensional capsule network layer, an adaptive convolutional layer, a deformable LSTM network, a cross-modal attention layer, a global average pooling layer, and a fully connected classification layer. The steps of obtaining the defect type and confidence level according to the compressed waveform segments and the sensing feature vectors, in combination with the dual-stream deformable attention convolutional network model, include: Input the electromagnetic wave signals into the deformable convolutional layer, and determine the dynamic offset according to the gradient of the electromagnetic wave signals to obtain electromagnetic convolution features; Input the electromagnetic convolution features into the polarization attention module to obtain horizontal polarization features; Input the horizontal polarization features into the three-dimensional capsule network layer, first generate primary capsules, and then perform dynamic routing and iteration multiple times to obtain 10 discharge modes; Input the time-frequency matrix of the ultrasonic wave signals into the adaptive convolutional layer, and dynamically adjust the convolution kernel weights of the adaptive convolutional layer according to the band energy entropy of the ultrasonic wave signals to obtain ultrasonic convolution features; Input the ultrasonic convolution features into the deformable LSTM network to obtain time series feature encodings; Input the 10 discharge modes and the time series feature encodings into the cross-modal attention layer, calculate the cross-attention matrix and feature interaction to obtain a fusion feature matrix; Input the fusion feature matrix into the global average pooling layer and the fully connected classification layer to obtain the defect type and the confidence level.

3. The GIS equipment partial discharge detection system based on deep learning according to claim 2 is characterized in that: The cloud platform is further used to determine the loss function of the two-stream deformable attention convolutional network model during training using a finite difference time domain method, and the specific steps are as follows: Calculating the theoretical magnetic field signal and the theoretical ultrasonic signal generated by the discharge source at the defect location using the finite difference time domain method; Calculating waveform difference loss based on the electromagnetic wave signal and the theoretical magnetic field signal, and the ultrasonic signal and the theoretical ultrasonic signal; According to the gradient of the waveform difference loss, the parameters of the two-stream deformable attention convolutional network model are back-propagated through the time-domain finite difference method to update the parameters, so that the two-stream deformable attention convolutional network model reaches convergence faster.

4. The GIS equipment partial discharge detection system based on deep learning according to claim 1, characterized in that The fog computing layer also performs urgency assessment and transmission strategy selection based on the defect location, the defect type, the confidence level, and the equipment historical status database; If the confidence level is greater than 90% and the urgency level is greater than or equal to 4, the fog computing layer sends the sensing data and alarm signal to the cloud platform via 5G; If the confidence level is between 70% and 90% and the urgency level is greater than or equal to 3, the fog computing layer sends the sensing feature vector and a cloud-review request signal to the cloud platform via 5G. If the confidence level is less than 70%, the fog computing layer stores the sensing data locally.

5. The partial discharge detection system for GIS equipment based on deep learning according to claim 1, characterized in that, The cloud platform is further configured to output a typical defect feature library and determine whether the defect type is a typical defect according to the defect type and the typical defect feature library; If it is determined that the defect type is a typical defect, a red warning instruction is output and an escape route map is generated.

6. The GIS equipment partial discharge detection system based on deep learning according to claim 1, wherein, The cloud platform also compares the discharge patterns of multiple GIS devices in the same substation based on the single discharge event report sent by the fog computing layer and the equipment life cycle data to identify systemic risks. The single discharge event report includes the defect location, the defect type and the confidence level.

7. The partial discharge detection system for GIS equipment based on deep learning according to claim 1, characterized in that, The cloud platform also generates virtual samples using a digital twin simulation model based on the abnormal data and diagnostic results sent by the fog computing layer, and uses a generative adversarial network to align the distribution of simulation and measured data to construct a mixed training set. It adopts a meta-learning method to update the parameters of the dual-stream deformable attention convolutional network model, and retains the historical version model of the dual-stream deformable attention convolutional network model, supporting testing and rollback of both the historical version model and the updated model.

8. The GIS equipment partial discharge detection system based on deep learning according to claim 1 is characterized in that: The cloud platform is also used to output compliance inspection reports and structured case knowledge bases based on complete discharge event records, maintenance process records and result feedback, defect attribution analysis based on knowledge graphs and key operation records stored on blockchain.

9. The GIS equipment partial discharge detection system based on deep learning according to claim 1 is characterized in that: The edge layer includes a sensing unit, a signal conditioning circuit, an edge processor and a wireless communication module.

Images