A method and system for diagnosing partial discharge faults in GIS based on ultra-high frequency signals.

Abstract

This invention proposes a method and system for GIS partial discharge fault diagnosis based on UHF signals. First, simulated GIS insulation defect fault samples, field fault samples, and environmental noise interference samples are acquired in the laboratory. Then, their 3D volumes are convolved with an anisotropic detection kernel to project them into 2D, and a deep volumetric convolutional neural network structure is added to diagnose the 2D projection. Next, an autoencoder model is trained unsupervised based on the samples, and the initial parameters of the convolutional layers of the deep volumetric convolutional neural network model are initialized using the model parameters obtained from the autoencoder model. The parameters of the deep convolutional neural network model are iteratively updated using the backpropagation algorithm and stochastic gradient descent method to obtain the optimal parameters. Using a volumetric convolutional neural network for fault diagnosis effectively avoids manual feature engineering and grayscale processing of 3D images, improving the diagnosis speed and fault recognition rate compared to traditional machine learning algorithms.

Description

Technical Field

[0001] This invention belongs to the field of power equipment monitoring, specifically relating to a GIS partial discharge fault diagnosis method and system based on ultra-high frequency signals. Background Technology

[0002] Gas-insulated switchgear (GIS) is widely used in power systems due to its reliable operation, small footprint, and electromagnetic environment friendliness. Although its failure rate is low, its compact structure and other characteristics mean that once a failure occurs, the controllability of the fault is extremely low, seriously threatening the safety of the power system and causing significant economic losses. Therefore, online real-time monitoring of it is of great importance.

[0003] According to statistics from the State Grid Corporation of China, insulation faults account for more than half of all GIS (Gas Insulation System) faults, and these faults often manifest as partial discharges. Partial discharges further exacerbate insulation defects. Therefore, online monitoring of partial discharges in GIS is essential. Partial discharge monitoring helps detect latent faults in GIS and allows for targeted maintenance based on insulation classification, thereby ensuring the safety and reliability of the power grid.

[0004] For partial discharge monitoring in GIS (Gas Insulation System), current detection methods include pulsed current method, ultra-high frequency method, ultrasonic method, optical method, and chemical product decomposition method. Among these, the pulsed current method is the only method specified by IEC that can be used for discharge quantity calibration, but due to its stringent experimental conditions, it is difficult to apply to field detection. Optical methods are still in the laboratory research stage, and chemical methods are rarely used in the field due to limitations in detection sensitivity and sensors. Therefore, current partial discharge detection largely relies on the ultra-high frequency detection method for partial discharge according to the IEC 62478 standard.

[0005] Most methods for UHF partial discharge diagnosis in GIS involve analyzing UHF signals using time-domain analysis (TRPD) and phase-domain analysis (PRPD), followed by manual feature extraction to extract UHF statistical features. Key feature parameters are then extracted using algorithms, and machine learning methods are applied for diagnosis. Currently, algorithms using deep convolutional neural networks (DNNs) have emerged for UHF partial discharge diagnosis in GIS. While DNNs, as a 2D image processing method, effectively avoid complex manual feature engineering, they still require processing the data obtained from PRPD methods. Three-dimensional maps are used for grayscale image processing, which involves a large amount of data and makes it difficult to fully utilize the information contained in ultra-high frequency signals, resulting in low processing efficiency.

[0006] Although there is considerable research on ultra-high frequency methods for partial discharge, most of it relies on laboratory data and rarely incorporates field fault data, failing to reflect the multifaceted randomness of field accidents. Taking free metal particle defects as an example, the size, tip curvature radius, and particle position of the free metal particles all affect the partial discharge signal. Simulating faults at the same or a few locations with one or two sizes in the laboratory is insufficient to simulate all the characteristics of free metal particle defects. Summary of the Invention

[0007] To address the problems existing in the background technology, a method and system for GIS partial discharge fault diagnosis based on ultra-high frequency signals is provided. This method and system can effectively perform pattern recognition and fault diagnosis on partial discharge signals from different sources, environments, times, and detection channels. This allows for timely and effective acquisition of the insulation status of equipment, timely elimination of hidden dangers, guidance for on-site maintenance, and prevention of major accidents.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A method for diagnosing partial discharge in GIS based on ultra-high frequency signals includes the following steps:

[0010] Step (1) Obtain UHF partial discharge signal samples from GIS;

[0011] Step (2) Process the UHF partial discharge signal of GIS;

[0012] Step (3) Construct a deep volumetric convolutional neural network model;

[0013] Step (4) Train a deep volumetric convolutional neural network using GIS partial discharge UHF signal samples;

[0014] Step (5) Use the trained deep volumetric convolutional neural network to perform partial discharge fault diagnosis on the signal to be diagnosed.

[0015] Furthermore, the composition of the GIS partial discharge UHF signal samples includes fault samples obtained from laboratory simulations of GIS insulation defects, accumulated GIS fault samples collected during field operation, and environmental noise interference samples. The effective frequency band of the UHF signal is 0.3-3 GHz.

[0016] Further specifying, the aforementioned UHF signal processing refers to performing phase domain analysis (PRPD) on the detected UHF signal to construct... Three-dimensional map.

[0017] Further specifying, the deep volumetric convolutional neural network model includes at least one input layer, at least two volumetric convolutional layers and corresponding two pooling layers, at least two fully connected layers, and one output softmax classification layer.

[0018] Further specifying, the volumetric convolutional neural network projects its 3D volume onto 2D by convolving it with an anisotropic probe kernel. The kernel enables long-term interactions between encoding points. Then, an image convolutional neural network is added to classify the 2D projection. Note that the training of the projection module and the image classification module is end-to-end.

[0019] Further specifying, the volumetric convolutional network includes data augmentation and multi-directional pooling layers. Compared to two-dimensional image datasets, existing three-dimensional shape datasets have limitations in scale and variability. The data augmentation process adds training data with rotations from different orientations and elevation angles, enabling the first network to cover local regions in different directions, and the second network to connect distant points at different relative angles. The multi-directional pooling layer refers to adding a directional pooling stage to aggregate information from different directions.

[0020] Further specifying, the specific implementation of step four includes: constructing an autoencoder model, performing unsupervised training on the autoencoder model based on the partial discharge signal samples, and initializing the initial parameters of the convolutional layers of the deep volumetric convolutional neural network model using the model parameters obtained from the autoencoder model; iteratively updating the parameters of the deep convolutional neural network model using the BP algorithm and stochastic gradient descent method to obtain the optimal parameters.

[0021] A GIS partial discharge fault diagnosis system based on ultra-high frequency (UHF) signals includes: a GIS, an UHF sensor, a signal conditioning circuit, a data acquisition instrument, a fault diagnosis center, and a PC. The UHF sensor is installed on the GIS housing to collect UHF signals. The UHF signals are conditioned by the signal conditioning circuit and then input to the data acquisition instrument. The data acquisition instrument is used to collect UHF signals transmitted from inside the GIS. The fault diagnosis center analyzes and processes the collected UHF signals to diagnose partial discharge faults in the GIS using UHF signals and determine the degree of the partial discharge fault. The PC is used to display the judgment results of the fault diagnosis center.

[0022] Further specifying, the partial discharge signal detection channel also includes a coupling capacitor, a measuring impedance, and a partial discharge amplifier. The coupling capacitor, the measuring impedance, and the partial discharge amplifier are connected together and connected to the IEC60270 partial discharge instrument and the PD portable detection device. The bus in the GIS cavity is connected to the step-up transformer T2 through a current-limiting resistor. The step-up transformer T2 is connected to the autotransformer T1, and the input of the autotransformer T1 is 380V AC.

[0023] Furthermore, the ultra-high frequency sensor has a detection frequency band of 0.3-2GHz and consists of an amplifier, a high-pass filter, a detector, and a shielding housing. The amplifier has an operating bandwidth of 0.3-1.5GHz and a gain of dB.

[0024] In summary, the present invention has the following superior effects:

[0025] 1. The pattern recognition algorithm for fault diagnosis using volumetric convolutional neural networks effectively avoids the massive task of manual feature engineering or the grayscale processing of 3D images. At the same time, it makes fuller use of the information contained in the ultra-high frequency signals of partial discharge. Compared with traditional machine learning algorithms or deep convolutional neural networks, it improves the speed of diagnosis and the fault recognition rate.

[0026] 2. Applying field data to the training of deep volumetric convolutional neural networks can incorporate more field information, compensating for the limitations and incompleteness of laboratory data, thereby improving the generalization performance of the trained volumetric neural network and enabling it to be applied more accurately to field fault diagnosis.

[0027] 3. Data augmentation and multi-directional pooling are employed in deep volumetric convolutional neural networks. The data augmentation process adds training data with different orientations and elevation angles, enabling the first network to cover local regions in different directions, and the second network to connect distant points at different relative angles. Multi-directional pooling refers to adding a directional pooling stage, which aggregates information from different directions, making the diagnostic results of deep volumetric convolutional neural networks more accurate. Attached Figure Description

[0028] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0029] Figure 1 This is a schematic diagram illustrating the principle of fault diagnosis based on a deep volumetric convolutional neural network.

[0030] Figure 2 This is a schematic diagram of the components of a fault diagnosis system based on a deep volumetric convolutional neural network.

[0031] Figure 3 Circuit diagram of a GIS partial discharge UHF detection system;

[0032] Figure 4 This is a diagram of the deep volumetric convolutional neural network structure.

[0033] Figure 5 This is a diagram of an image-based deep convolutional neural network structure. Detailed Implementation

[0034] To better demonstrate the technical solution and advantages of the present invention, the present invention will now be described in further detail with reference to the accompanying drawings.

[0035] A method and system for diagnosing partial discharge faults in GIS based on ultra-high frequency signals, according to Figure 1 Schematic diagram of fault diagnosis based on deep volumetric convolutional neural network and Figure 2 The structure diagram of the fault diagnosis system based on deep volumetric convolutional neural networks can be summarized as follows: GIS partial discharge UHF diagnosis includes the following steps:

[0036] (1) Obtain UHF partial discharge signal samples from GIS;

[0037] (2) Processing the UHF partial discharge signal of GIS;

[0038] (3) Construct a deep volumetric convolutional neural network model;

[0039] (4) Train a deep volumetric convolutional neural network using GIS partial discharge UHF signal samples;

[0040] (5) A trained deep volumetric convolutional neural network is used to diagnose partial discharge faults in the signal to be diagnosed.

[0041] The specific process is as follows:

[0042] Obtain UHF partial discharge signal samples from GIS:

[0043] The circuit diagram of the GIS partial discharge UHF detection system is as follows: Figure 3 As shown. The partial discharge diagnostic system includes: a GIS (Gas Insulator), a UHF sensor, a signal conditioning circuit, a data acquisition instrument, a fault diagnosis center, and a PC. The UHF sensor is installed on the GIS housing to collect UHF signals. The UHF signals are conditioned by the signal conditioning circuit and then input to the data acquisition instrument. The data acquisition instrument collects UHF signals transmitted from inside the GIS. The fault diagnosis center analyzes and processes the collected UHF signals to diagnose partial discharge faults in the GIS using UHF signals and determine the severity of the partial discharge fault. The PC displays the judgment results from the fault diagnosis center.

[0044] In this step, according to Figure 3 A partial discharge UHF signal acquisition system was installed at the substation site. Five typical partial discharge models were set up, including a floating potential defect model, a surface discharge defect model, a tip corona defect model, an air gap discharge defect model, and a free metal particle discharge defect model. Partial discharge samples were acquired using various instruments such as a digital partial discharge instrument, oscilloscope, and portable partial discharge detector through partial discharge simulation experiments, substation field detection, and typical interference experiments. Simultaneously, nearly ten years of accumulated GIS partial discharge UHF data were used as training samples, along with noise signals collected when no partial discharge occurred. Applying field data to the training of a deep volumetric convolutional neural network incorporates more field information, compensating for the limitations and incompleteness of laboratory data, thereby improving the generalization performance of the trained volumetric neural network and enabling more accurate application in field fault diagnosis. A total of 1500 sets of GIS partial discharge UHF samples were obtained.

[0045] Processing of UHF partial discharge signals from GIS:

[0046] For the obtained 1000 training samples, the physical parameters required to construct the PRPD model include the power frequency phase (0-360°) corresponding to the partial discharge pulse, the single PD discharge quantity q, and the number of discharges n, etc. The PRPD model describes the partial discharge type in a certain discharge mode. The relationship between the three. By collecting partial discharge pulse data from multiple power frequency cycles, Divide the plane into several small intervals (q and q), form several grids on the q plane, and count the number of discharges in each grid to plot the corresponding partial discharge. In the three-dimensional statistical spectrum of the model, since the discharge quantity q cannot be accurately calibrated in the UHF method, the partial discharge pulse amplitude u is used instead.

[0047] Constructing a deep volumetric convolutional neural network model:

[0048] The constructed deep volumetric convolutional neural network structure is as follows: Figure 4As shown. The key to this network is the use of a thin, anisotropic kernel, which helps capture the overall structure of a 3D volumetric convolutional neural network. The deep volumetric convolutional neural network has two modules: an anisotropic detection module and the network module within the network module. The anisotropic detection module contains three coiled, thin kernel layers, each followed by a nonlinear ReLU layer. Note that the input and output of each layer are 3D tensors. Compared to traditional anisotropic detection modules, the anisotropic detection module has the advantage of aggregating long-range interactions early in feature learning with fewer parameters. Compared to traditional neural networks composed of isotropic kernels, introducing long-range interactions early on can only be achieved with large kernels, which inevitably introduces more parameters. After anisotropic detection, a suitable NIN network is used to solve the classification problem. The detection network is able to capture the internal structure of objects through its x-ray-like projection mechanism. This is a capability not found in standard rendering. Combining this detection mechanism with multi-directional pooling can capture any 3D structure. Furthermore, this architecture can scale to higher resolutions because all its layers can be viewed as 2D. Although 3D convolution involves computation at cubic resolution locations, it maintains quadratic computation.

[0049] For image-based deep convolutional network structures in deep volumetric convolutional networks, such as... Figure 5As shown. A deep convolutional neural network for GIS partial discharge fault diagnosis based on images is trained using a network similar to the classic LeNet-5 network, with similar parameter settings during training. The input is a 32×32×3 matrix. The first layer uses a 5×5 filter size, a stride of 1, padding of 0, and 6 filters, resulting in an output of 28×28×6. This layer is labeled CONV1, which uses 6 filters, increases the bias, applies the ReLU nonlinear function, and finally outputs the result of CONV1. Then, a pooling layer is constructed, choosing max pooling with parameters of , . Now, the pooling layer is constructed, using a 2×2 filter size and a stride of 2, which means the layer height and width will be halved. Therefore, 28×28 becomes 14×14, while the number of channels remains unchanged, so the final output is 14×14×6, labeled POOL1. CONV1 and POOL1 are combined into a single convolution and labeled Layer1. When counting the number of network layers, only layers with weights are counted, meaning CONV1 and POOL1 are considered as Layer 1. We'll label them CONV1 and POOL1 here; both are part of Layer 1 in the neural network. POOL1 is also included in Layer 1 because it has no weights, and its output is 14×14×6. We then build another convolutional layer for it, with a filter size of 5×5 and a stride of 1. This time, we use 10 filters, resulting in a 10×10×10 matrix, labeled CONV2. Next, we perform max pooling, halving the hyperparameters, height, and width, resulting in a 5×5×10 output, labeled POOL2. This is the second convolutional layer of the neural network, Layer 2.

[0050] If we apply another convolutional layer to Layer 1, with a 5×5 filter (stride 1, padding 0, so omitted here), there are 16 filters, resulting in a 10×10×16 output for CONV2. We then perform max pooling. Max pooling is performed on the 10×10×16 input, halving the parameters, height, and width. The result is 5×5×16, with the same number of channels as before, labeled POOL2. This is a convolution, Layer 2, because it has only one set of weights and one convolutional layer, CONV2. The 5×5×16 matrix contains 400 elements. Now, we flatten POOL2 into a one-dimensional vector of size 400. Imagine the flattened result as a set of neurons, and then use these 400 units to build the next layer. The next layer contains 120 units; this is the first fully connected layer, labeled FC3. These 400 units are tightly connected to the 120 units, forming a fully connected layer. It's very similar to the single neural network layers we discussed in the first and second courses; this is a standard neural network. Its weight matrix has dimensions of 120×400. This is what's called "fully connected" because each of the 400 units is connected to each of the 120 units, with a bias parameter. The final output has 120 dimensions because there are 120 outputs. Then we add another fully connected layer to these 120 units; this layer is smaller, let's say it contains 84 units, labeled FC4. Finally, we fill a softmax unit with these 84 units.

[0051] Training a deep volumetric convolutional neural network using GIS partial discharge UHF signal samples:

[0052] During training, we generate different rotations of the 3D model by randomly sampling and varying the azimuth and elevation angles. A capacity CNN is first trained for a single rotation. Then, the network is decomposed into CNN1 (lower layer) and CNN2 (upper layer) to build a multi-directional version. The weights of MO-VCNN are initialized by the previously trained capacity CNN, and the weights of CNN1 are fixed during fine-tuning. The common practice is to extract the highest-level features from multiple directions (features before the last linear classification layer), average / maximize / concatenate them, and then train a linear SVM on the combined features; this is just a special case of MO-VCNN.

[0053] An autoencoder model B is constructed. Unsupervised training is performed on autoencoder model B using a large dataset of partial discharge signals. The output of the sample data is calculated, and the error Σσ' between the output and the sample label is calculated. The initial parameters of the convolutional layers of the deep volumetric convolutional neural network model A are initialized using the model parameters obtained from autoencoder model B. The deep convolutional neural network A is then trained using the training sample set data. The output of the sample data is calculated, and the error Σσ between the output and the sample label is calculated. The parameters of the deep convolutional neural network model A are iteratively updated using the backpropagation algorithm and stochastic gradient descent to obtain the optimal parameters.

[0054] A trained deep volumetric convolutional neural network is used to diagnose partial discharge faults in the signal to be diagnosed.

[0055] The remaining 500 sets of GIS partial discharge UHF samples were used to test the trained deep volumetric convolutional neural network, and partial discharge fault diagnosis was performed on these 500 sets of signals to be diagnosed. The results were then compared with those obtained using traditional machine learning algorithms.

[0056] It should be emphasized that the embodiments described in this invention are illustrative rather than limiting. Therefore, this invention is not limited to the embodiments described in the specific implementation. Any other implementation methods derived by those skilled in the art based on the technical solutions of this invention also fall within the scope of protection of this invention.

Claims

1. A method for diagnosing partial discharge in GIS based on ultra-high frequency signals, characterized in that, It includes the following steps: Step (1) Obtain UHF partial discharge signal samples from GIS; Step (2) Process the UHF partial discharge signal of GIS; Ultra-high frequency signal processing refers to the phase domain analysis (PRPD) of detected ultra-high frequency signals. The PRPD mode describes the relationship between φ-qn, discharge phase φ, discharge amplitude q, and discharge number n. φ and q are divided into several small intervals, forming several grids on the φ-q plane. The number of discharges in each grid is counted to obtain a three-dimensional map of the φ-qn mode. Step (3) Construct a deep volumetric convolutional neural network model; A deep volumetric convolutional neural network model includes at least one input layer, at least two volumetric convolutional layers and corresponding two pooling layers, at least two fully connected layers, and one output softmax classification layer; Volumetric convolutional neural networks project their 3D volume onto 2D by convolving it with an anisotropic probe kernel. The kernel enables long-term interactions between encoding points. Then, an image convolutional neural network is added to classify the 2D projection. Note that the training of the projection module and the image classification module is end-to-end. Step (4) Train a deep volumetric convolutional neural network using GIS partial discharge UHF signal samples; Step (5) Use the trained deep volumetric convolutional neural network to perform partial discharge fault diagnosis on the signal to be diagnosed.

2. The method for diagnosing partial discharge faults in GIS based on UHF signals according to claim 1, characterized in that, UHF signal samples are obtained through a GIS partial discharge diagnostic system. The collected UHF signal samples are then input into a software model based on a deep volumetric convolutional neural network for diagnosis. The UHF signal samples of GIS partial discharge consist of fault samples obtained from laboratory simulation of GIS insulation defect tests, accumulated GIS fault samples collected in the field, and environmental noise interference samples. The effective frequency band of the UHF signal is 0.3-3GHz.

3. The method for diagnosing partial discharge faults in GIS based on UHF signals according to claim 1, characterized in that, The volumetric convolutional network includes data augmentation and multi-directional pooling layers. Compared with two-dimensional image datasets, existing three-dimensional shape datasets have limitations in scale and variation. The data augmentation process adds training data with different orientations and elevation angles, enabling the first network to cover local regions in different directions, and the second network to connect distant points at different relative angles. The multi-directional pooling layer refers to adding a directional pooling stage to aggregate information from different directions.

4. The method for diagnosing partial discharge faults in GIS based on UHF signals according to claim 1, characterized in that, The specific implementation of training the deep volumetric convolutional neural network includes: constructing an autoencoder model, performing unsupervised training on the autoencoder model based on the partial discharge signal samples, and initializing the initial parameters of the convolutional layers of the deep volumetric convolutional neural network model using the model parameters obtained from the autoencoder model; iteratively updating the parameters of the deep volumetric convolutional neural network model using the BP algorithm and stochastic gradient descent method to obtain the optimal parameters.

5. A GIS partial discharge fault diagnosis system based on UHF signals, employing the method described in claim 1, characterized in that, The partial discharge diagnostic system includes: GIS (Gas Insulation System), UHF (Ultra-High Frequency) sensor, signal conditioning circuit, data acquisition unit, fault diagnosis center, and PC. The UHF sensor, mounted on the GIS housing, collects UHF signals. Electromagnetic waves generated by partial discharge in the GIS are received by the UHF sensor, converted into voltage signals, and then transmitted via coaxial cable to the signal conditioning unit. After conditioning, the partial discharge signal is sent to the data acquisition card for acquisition, storage, and processing. The computer controls the signal conditioning unit through a parallel interface, enabling the selection and control of the center frequency of the system's passband and the bandwidth of the filters. The fault diagnosis center analyzes and processes the collected UHF signals to diagnose partial discharge faults in the GIS and determine the severity of the fault. The PC displays the diagnostic results from the fault diagnosis center.

6. The GIS partial discharge fault diagnosis system based on UHF signals according to claim 5, characterized in that, The partial discharge signal detection channel also includes a coupling capacitor, a measuring impedance, and a partial discharge amplifier. The input of the autotransformer T1 is 380V AC, and the output is connected to the step-up transformer T2. The step-up transformer T2 is connected to the bus in the GIS cavity through a current-limiting resistor. The partial discharge signal detection channel is connected to the IEC60270 partial discharge instrument and the PD portable detection device through the coupling capacitor and the measuring impedance.

7. A GIS partial discharge fault diagnosis system based on ultra-high frequency signals according to claim 5, characterized in that, The ultra-high frequency sensor is installed in the GIS cavity and collects the ultra-high frequency signals into the data acquisition card. The ultra-high frequency sensor has a detection frequency band of 0.3-2GHz and consists of an amplifier, a high-pass filter, a detector and a shielding shell. The amplifier has a working bandwidth of 0.3-1.5GHz and a gain of 40 dB.

Images