Intelligent evaluation method for sensitivity of built-in partial discharge sensor in extra-high voltage GIS

Abstract

The embodiment of the application provides a kind of ultra-high voltage GIS built-in partial discharge sensor sensitivity intelligent evaluation method, the method comprises: obtaining the current voltage pulse signal of the output of the ultra-high voltage gas insulated switchgear (GIS) built-in partial discharge sensor;Signal reconstruction network is used to signal decoding and reconstruction are carried out to current voltage pulse signal, obtain the voltage signal to be compared;Wherein, signal reconstruction network is obtained by using multiple calibrated voltage pulse signals, time series convolution network with learnable inflation rate is trained using time-frequency joint multi-scale reconstruction loss;Based on current voltage pulse signal and the voltage signal to be compared, the sensitivity score of built-in partial discharge sensor is determined.The application can improve the accuracy of the sensitivity evaluation of built-in partial discharge sensor.

Description

Technical Field

[0001] This application relates to the field of electrical equipment condition monitoring technology, and includes, but is not limited to, an intelligent assessment method for the sensitivity of a built-in partial discharge sensor in ultra-high voltage GIS. Background Technology

[0002] Gas-insulated switchgear (GIS) is a critical piece of equipment in power systems. Its built-in partial discharge sensor is the core component for monitoring internal insulation defects. However, the sensitivity of the sensor typically degrades due to aging, mechanical vibration, and environmental influences, leading to missed or false alarms for internal insulation defects. Therefore, it is essential to regularly and accurately assess its sensitivity.

[0003] In existing technologies, the sensitivity evaluation of UHV GIS built-in partial discharge sensors is mainly divided into two categories: 1) Laboratory calibration method: In a shielded environment, sensitivity is evaluated by directly comparing the sensor output with parameters such as energy, amplitude, or rise time of the standard signal by injecting a standard pulse. This method is highly accurate, but it relies on an ideal experimental environment and absolute standards, and cannot be applied to the field evaluation of already operational GIS equipment.

[0004] 2) Field evaluation method based on traditional signal processing: A signal is injected in the field, and the evaluation is carried out indirectly by calculating indicators such as signal-to-noise ratio and signal energy. This type of method is suitable for field use, but it is susceptible to interference from complex electromagnetic noise, and the evaluation results are unstable, singular, and unable to accurately characterize the sensor's response to key features of pulse waveforms, such as rise time and oscillation frequency.

[0005] Therefore, the core deficiency of existing technologies lies in the lack of an intelligent evaluation method that can adaptively extract multi-scale features of pulse signals, separate key waveform information from noise, and quantitatively output sensitivity scores in real GIS operating environments with strong noise and multiple interferences. Summary of the Invention

[0006] Based on the above problems, this application provides an intelligent sensitivity evaluation method for UHV GIS built-in partial discharge sensors. The method aims to accurately capture key features of signal fidelity through learnable dilatation rate convolution, and to learn the complete physical characteristics of a pure signal through a time-frequency joint reconstruction loss-driven network. This allows the signal reconstruction network to simultaneously possess adaptive multi-scale feature extraction capabilities and sensitivity to frequency distortion, enabling it to accurately characterize the key pulse waveform features of the current voltage pulse signal output by the built-in partial discharge sensor. This, in turn, improves the accuracy of sensitivity evaluation for UHV GIS built-in partial discharge sensors in adaptive complex environments.

[0007] The technical solution of this application embodiment is implemented as follows: This application provides an intelligent evaluation method for the sensitivity of an ultra-high voltage GIS built-in partial discharge sensor. The method includes: acquiring the current voltage pulse signal output by the built-in partial discharge sensor of the ultra-high voltage gas-insulated switchgear (GIS); using a signal reconstruction network to decode and reconstruct the current voltage pulse signal to obtain a voltage signal to be compared; wherein the signal reconstruction network is obtained by training a time-series convolutional network with a learnable expansion rate using multiple calibrated voltage pulse signals and employing a time-frequency joint multi-scale reconstruction loss; and determining the detection sensitivity score of the built-in partial discharge sensor based on the current voltage pulse signal and the voltage signal to be compared.

[0008] In some embodiments, the signal reconstruction network includes: an encoder and a decoder consisting of four cascaded enhanced temporal convolutional blocks; using the signal reconstruction network to decode and reconstruct the current voltage pulse signal to obtain a voltage signal to be compared, including: using the four cascaded enhanced temporal convolutional blocks to perform layer-by-layer deep feature encoding on the current voltage pulse signal to obtain a low-dimensional feature vector; and using the decoder to recover the signal from the low-dimensional feature vector to obtain the voltage signal to be compared.

[0009] In some embodiments, each level of the enhanced temporal convolutional block includes: a temporal convolutional layer, an upsampling layer, a feature concatenation layer, and a fusion convolutional layer; the temporal convolutional layer includes: multiple parallel causal convolutional layers with different dilation rates; the dilation rates within the enhanced temporal convolutional blocks of different levels are different; using four cascaded enhanced temporal convolutional blocks, deep feature encoding is performed layer by layer on the current voltage pulse signal to obtain a low-dimensional feature vector, including: for each level of enhanced temporal convolutional block, the input signal is input into multiple parallel causal convolutional layers in the level of enhanced temporal convolutional block to extract temporal features, obtaining multiple temporal features corresponding to the level; using the upsampling layer in the level of enhanced temporal convolutional block, the multiple temporal features corresponding to the level are extracted. Linear interpolation is performed on each of the levels to obtain multiple virtual feature sequences corresponding to the level. The feature concatenation layer in the enhanced temporal convolutional block of the level is used to concatenate the multiple virtual feature sequences corresponding to the level, resulting in the concatenated features corresponding to the level. The fusion convolutional layer in the enhanced temporal convolutional block of the level is used to fuse the concatenated features corresponding to the level and compress the channels, resulting in the output features corresponding to the level. The input signal of the first-level enhanced temporal convolutional block is the current voltage pulse signal; the input signal of the Nth-level enhanced temporal convolutional block is the output signal of the (N-1)th-level enhanced temporal convolutional block; N is greater than or equal to 2 and less than or equal to 4; the output signal corresponding to the 4th-level enhanced temporal convolutional block is a low-dimensional feature vector.

[0010] In some embodiments, a decoder is used to recover the signal from the low-dimensional feature vector to obtain the voltage signal to be compared, including: expanding the low-dimensional feature vector into an initial feature map using a fully connected layer in the decoder; progressively upsampling the initial feature map using multiple cascaded deconvolutional layers in the decoder to obtain intermediate features; and fusing and reconstructing the intermediate features using convolutional layers in the decoder to obtain the voltage signal to be compared.

[0011] In some embodiments, the temporal convolutional network includes: an initial encoding module and an initial decoding module consisting of four cascaded initial temporal convolutional blocks; the initial temporal convolutional block includes: a temporal convolutional layer consisting of multiple parallel causal convolutional layers with learnable dilation rates; the construction process of the signal reconstruction network includes: building a training sample set corresponding to multiple calibrated voltage pulse signals; using time-frequency joint multi-scale reconstruction loss, based on the training sample set, adjusting multiple learnable dilation rates in the temporal convolutional layers of at least each level of the initial temporal convolutional block to obtain the signal reconstruction network.

[0012] In some embodiments, constructing a training sample set corresponding to multiple calibrated voltage pulse signals includes: performing data augmentation on each calibrated voltage pulse signal to obtain multiple noisy signals corresponding to the calibrated voltage pulse signal; wherein, the data augmentation includes at least: changing different pulse parameters, injecting signals at different locations, and injecting signals with different signal-to-noise ratios; and constructing a training sample set composed of multiple calibrated voltage pulse signals and multiple noisy signals corresponding to each calibrated voltage pulse signal.

[0013] In some embodiments, the time-frequency joint multi-scale reconstruction loss is: ; in, For time-frequency joint multi-scale reconstruction loss; This represents the total number of calibrated voltage pulse signals. The balancing coefficient for temporal reconstruction loss; For the first A calibrated voltage pulse signal; For the reason and Composed of multiple corresponding noisy signals The first in the signal subset One signal; , These are the initial encoding module and the initial decoding module of the temporal convolutional network, respectively. For the output of a temporal convolutional network The reconstructed signal; for The total number of corresponding noisy signals; The balance coefficients for frequency domain reconstruction loss; To make the first Frequency domain characteristics of a calibrated voltage pulse signal; This is a transform function that converts a time-domain signal into frequency-domain features. for Frequency domain characteristics; The regularization coefficient is used. Model parameters for temporal convolutional networks The regularization term.

[0014] In some embodiments, determining the detection sensitivity score of the built-in partial discharge sensor based on the current voltage pulse signal and the voltage signal to be compared includes: determining the normalized mean square error between the current voltage pulse signal and the voltage signal to be compared using an error calculation formula; wherein the error calculation formula is: ; in, This is the normalized mean square error; and These are the current voltage pulse signal and the voltage signal to be compared, respectively. The value is the Euclidean norm; and the difference between 1 and the normalized mean square error is determined as the detection sensitivity score of the built-in partial discharge sensor.

[0015] The beneficial effects of the technical solutions provided in this application include at least the following: The intelligent sensitivity evaluation method for UHV GIS built-in partial discharge sensors provided in this application involves the following steps: First, the current voltage pulse signal output by the UHV gas-insulated switchgear (GIS) built-in partial discharge sensor is acquired. Then, a signal reconstruction network is used to decode and reconstruct the current voltage pulse signal to obtain a comparison voltage signal. The signal reconstruction network is trained using multiple calibrated voltage pulse signals on a time-series convolutional network with a learnable dilatation rate, employing a time-frequency joint multi-scale reconstruction loss. Finally, based on the current voltage pulse signal and the comparison voltage signal, the detection sensitivity score of the built-in partial discharge sensor is determined. This method achieves accurate capture of key signal fidelity features through learnable dilatation rate convolution and drives the network to learn the complete physical characteristics of a pure signal through the time-frequency joint reconstruction loss. This allows the signal reconstruction network to simultaneously possess adaptive multi-scale feature extraction capabilities and sensitivity to frequency distortion, enabling it to accurately characterize the key pulse waveform features of the current voltage pulse signal output by the built-in partial discharge sensor. This improves the accuracy of sensitivity evaluation for UHV GIS built-in partial discharge sensors in adaptive complex environments.

[0016] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the technical solutions provided in the embodiments of this application. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 A flowchart illustrating the intelligent sensitivity assessment method for an ultra-high voltage GIS built-in partial discharge sensor provided in this application; Figure 2 This is a schematic diagram of the composition of a signal reconstruction network provided in this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0020] It should be noted that the terms "first, second, and third" used in the embodiments of this application are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0021] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0022] Example 1: See Figure 1 The diagram shown is a flowchart illustrating an intelligent sensitivity assessment method for an embedded partial discharge sensor in an ultra-high voltage GIS provided in this application embodiment. This method can be executed by an electronic device, such as a computer or server. Here, in conjunction with... Figure 1 The following explanation is provided: Step 101: Obtain the current voltage pulse signal output by the built-in partial discharge sensor of the UHV gas-insulated switchgear (GIS).

[0023] In some embodiments, when partial discharge occurs inside an ultra-high voltage GIS, a series of physical phenomena are generated. Its built-in partial discharge sensor mainly captures the following two types: ultra-high frequency electromagnetic waves excited at the moment of discharge; and high-frequency pulse current on the grounding wire during discharge, which is converted into a high-frequency voltage pulse signal, i.e., the current voltage pulse signal mentioned in the embodiments of this application.

[0024] Here, the current voltage pulse signal has the following signal characteristics: pulsed, periodic, and high frequency.

[0025] It should be noted that a partial discharge sensor (PDS) is a key device used to detect partial discharge phenomena inside electrical equipment.

[0026] Step 102: Use a signal reconstruction network to decode and reconstruct the current voltage pulse signal to obtain the voltage signal to be compared.

[0027] The signal reconstruction network is obtained by training a temporal convolutional network with a learnable dilation rate using multiple calibrated voltage pulse signals and employing a time-frequency joint multi-scale reconstruction loss.

[0028] In some embodiments, a temporal convolutional network with a learnable dilation rate includes: an initial encoding module and an initial decoding module consisting of four cascaded initial temporal convolutional blocks; the initial temporal convolutional block includes: a temporal convolutional layer consisting of multiple parallel causal convolutional layers with learnable dilation rates; wherein, the construction process of the signal reconstruction network includes: Step A1: Construct a training sample set corresponding to multiple calibrated voltage pulse signals.

[0029] Step A2: Using time-frequency joint multi-scale reconstruction loss, based on the training sample set, adjust multiple learnable dilation rates in the temporal convolutional layers of the initial temporal convolutional block at least for each level to obtain the signal reconstruction network.

[0030] In some embodiments, the signal output by the UHV GIS built-in partial discharge sensor is a short-duration, pulsed high-frequency electromagnetic wave signal. This signal has a very high frequency, mainly concentrated in the range of 300MHz to 1.5GHz, or even up to 3GHz. Given the characteristics of this signal—high-frequency pulse, discontinuous nature, and the need to extract temporal features such as phase, frequency, and amplitude—temporal convolutional networks (TCNs) are suitable for processing it. This is because the core design of TCNs includes causal convolution, dilated convolution, and residual connections. These core designs are highly compatible with the current voltage pulse signal output by the partial discharge sensor, for the following reasons: 1. Causal convolution for handling temporal relationships: Since the pulse occurrence time of the current voltage pulse signal output by the partial discharge sensor is closely related to the phase of the power frequency voltage, this causal convolution can ensure that when the temporal convolutional network predicts or extracts features at the current moment, it only depends on past signals and does not involve future data.

[0031] 2. Capturing long-range dependent dilated convolutions: This dilated convolution can exponentially expand the receptive field. That is, a temporal convolutional network with dilated convolution can cover signals of tens or even hundreds of power frequency cycles at once to extract the phase distribution pattern of the current voltage pulse signal.

[0032] 3. Gradient stability: Since the current voltage pulse signal is non-stationary and the background noise and interference are large, the residual connections inside the temporal convolutional network can avoid the gradient vanishing or exploding problem in deep networks, thus enabling the temporal convolutional network to stably learn the weak discharge characteristics under strong noise background.

[0033] In some embodiments, step A1 can be implemented by steps A11 and A12: Step A11: Perform data enhancement on each calibrated voltage pulse signal to obtain multiple noisy signals corresponding to the calibrated voltage pulse signal.

[0034] Data augmentation includes at least: changing different pulse parameters, injecting signals at different locations, and injecting signals with different signal-to-noise ratios.

[0035] In some embodiments, the plurality of calibrated voltage pulse signals may be generated using a standard pulse source in an experimental platform simulating a GIS operating environment or a high-fidelity simulation environment. Inject a partial discharge analog signal with known characteristics, or the output signal of a reference sensor (or high-fidelity simulation model) with known excellent performance.

[0036] Here, in order to enhance the generalization ability of the temporal convolutional network, a variety of noisy signals are generated by simulating different injection point positions, different signal-to-noise ratios (with controllable noise added), different pulse parameters (i.e., modulation pulse waveform parameters, randomized signal injection positions, and applying multi-level signal-to-noise ratio background noise) in each calibrated voltage pulse signal. Each calibrated voltage pulse signal and its corresponding multiple noisy signals are then used as training samples.

[0037] Step A12: Construct a training sample set consisting of multiple calibrated voltage pulse signals and multiple noisy signals corresponding to each calibrated voltage pulse signal.

[0038] In some embodiments, before constructing a training sample set based on multiple calibrated voltage pulse signals and multiple noisy signals corresponding to each calibrated voltage pulse signal, preprocessing can be performed on each calibrated voltage pulse signal and the multiple noisy signals corresponding to each calibrated voltage pulse signal, such as: truncating a fixed length centered on the pulse peak point. The fragments were processed and their amplitudes were normalized to obtain the training sample set.

[0039] It should be noted that the training sample set can be further divided into: a training subset for each calibrated voltage pulse signal, which consists of each calibrated voltage pulse signal and its corresponding multiple noisy signals.

[0040] In some embodiments, the time-frequency joint multi-scale reconstruction loss is: Formula (1); in, For time-frequency joint multi-scale reconstruction loss; This represents the total number of calibrated voltage pulse signals. The balancing coefficient for temporal reconstruction loss; For the first A calibrated voltage pulse signal; For the reason and Composed of multiple corresponding noisy signals The first in the signal subset One signal; , These are the initial encoding module and the initial decoding module of the temporal convolutional network, respectively. For the output of a temporal convolutional network The reconstructed signal; for The total number of corresponding noisy signals; The balance coefficients for frequency domain reconstruction loss; To make the first Frequency domain characteristics of a calibrated voltage pulse signal; This is a transform function that converts a time-domain signal into frequency-domain features. for Frequency domain characteristics; The regularization coefficient is used. Model parameters for temporal convolutional networks The regularization term.

[0041] here, As a temporal reconstruction loss, it forces the temporal convolutional network to reconstruct a signal whose waveform shape is consistent with the corresponding calibrated voltage pulse signal during training. As a frequency domain reconstruction loss, it forces the temporal convolutional network to reconstruct signals during training that match the frequency characteristics, such as resonant frequency and bandwidth energy, with the corresponding calibrated voltage pulse signal, thus compensating for the insensitivity of temporal domain losses to frequency domain distortions. Furthermore, This is used to constrain the complexity of temporal convolutional networks in order to improve the generalization ability of the subsequent signal reconstruction networks.

[0042] Correspondingly, Used to adjust the weight of the temporal error term in the time-frequency joint multi-scale reconstruction loss; Used to adjust the weight of the frequency domain error term in the time-frequency joint multi-scale reconstruction loss; Used to control the strength of the regularization term; To prevent overfitting, and This represents all trainable parameters of a temporal convolutional network, including kernel weights and learnable dilation rate.

[0043] In some embodiments, It can be the Fast Fourier Transform (FFT) operator, whose amplitude spectrum is used to extract the frequency domain features of the signal.

[0044] In some embodiments, the core component of the temporal convolutional network is: four cascaded initial temporal convolutional blocks; correspondingly, in step A2, the training sample set is... Associated with a calibrated voltage pulse signal Signal subset (including: and (corresponding to multiple noisy signals) the first One signal (It can be a noisy signal, or it can be) After being input into a temporal convolutional network: First, ... The initial encoding module consists of four cascaded initial temporal convolutional blocks. Compressed into low-dimensional feature vectors Then, via the initial decoding module from To recover the original signal, that is, to obtain This compression-recovery process forces the initial encoding module... It is necessary to learn a compact representation that can fully describe the key features of the signal, thus obtaining the initial coding module. It can automatically focus on the multi-scale features of the signal, while discarding irrelevant information such as noise and redundant details in the reconstruction error.

[0045] In some embodiments, during the construction phase of the signal reconstruction network, multiple calibrated voltage pulse signals are used, enabling the encoder within the signal reconstruction network to learn the intrinsic distribution of the ideal sensor response, i.e., the signal manifold. Here, during the construction phase of the signal reconstruction network, when the input signal is multiple calibrated voltage pulse signals (signals from a sensor in good working order), the encoder within the signal reconstruction network focuses on the signal manifold, resulting in minimal time-frequency joint multi-scale reconstruction loss. However, when the input signal is a noisy signal (which can be considered as the source of the signal distortion caused by sensor aging or failure), the input signal deviates from the manifold, and the time-frequency joint multi-scale reconstruction loss increases significantly. This quantization adaptation based on the degree of deviation is the core idea of ​​this scheme, namely, sensitivity degradation monitoring.

[0046] Simultaneously, during the construction phase of the signal reconstruction network, by introducing noisy signals with different noise levels and propagation paths, the initial encoding module in training can learn to remove noise and extract core pulse features from the noisy signal. Thus, after the signal reconstruction network is constructed and applied in the field, even if the input signal contains background interference, the encoder within the signal reconstruction network can still extract the main features, and the decoder within the network can reconstruct a near-ideal waveform after denoising. The noise and distortion components are then represented as reconstruction errors, thereby separating noise from true degradation and improving evaluation robustness.

[0047] In some embodiments, a time-frequency joint multi-scale reconstruction loss is employed, based on the training sample set, to perform at least the following: multiple learnable dilation rates within the temporal convolutional layers of the initial temporal convolutional block at each level (each initial temporal convolutional block has multiple independent learnable dilation rates, such as: the first...). The first convolutional block in the initial time of level Learnable dilation rate of branch causal convolutional layers ), and adjust the corresponding convolution kernel weights (learning).

[0048] It should be noted that the dilation rate in traditional temporal convolutional networks is a manually set discrete positive integer (e.g., 1, 2, 3), with a fixed sampling interval. This means it cannot adaptively adjust based on task data, which is a core limitation for adapting to tasks like sensor sensitivity assessment, which require capturing multi-scale signal features. This approach focuses on the parameterized reconstruction of the dilation rate, parameterizing each dilation rate as a continuous positive real number, which can then be used as a trainable parameter of the network. This parameter, along with conventional trainable parameters such as convolutional kernel weights, is optimized through gradient descent. The optimization goal is to minimize the time-frequency joint multi-scale reconstruction loss in sensor sensitivity assessment, thereby transforming the dilation rate from a manually set hyperparameter into an adaptively learned network parameter.

[0049] In some embodiments, during the construction of the signal reconstruction network: 1. Initialization phase: It can be set to follow a uniform distribution. The core purpose of using small positive random numbers is to avoid an excessively large initial inflation rate and prevent the internal floating-point sampling positions from exceeding the length boundary of the input sequence, thereby ensuring the effectiveness of convolution operations in the early stages of training.

[0050] 2. Training Phase: The core objective of optimization is to minimize the time-frequency joint multi-scale reconstruction loss used to distinguish signal fidelity, and adjustments will be made simultaneously. And the corresponding convolutional kernel weights; whereby the convolutional kernel weights learn how to weight the sampled signal values ​​to extract features. The goal is to learn the optimal sampling interval for the input signal to capture the most discriminative features. The optimization direction is entirely task-driven; for example, if sensor sensitivity evaluation requires focusing on fine, short-timescale features such as the rising edge of a pulse, then… It will be optimized to a small value close to 1, reducing the sampling interval for high-frequency sampling; if it is necessary to capture macro-timescale features such as oscillation decay modes, etc. It will then be optimized to a larger value (e.g., 5.7), and the sampling interval will be expanded to sample at a lower frequency.

[0051] 3. After training convergence, The final continuous value (e.g.: , , The corresponding convolutional kernel weights form a one-to-one optimal signal analysis time scale. The numerical value directly defines the continuous signal sampling interval of the corresponding convolutional branch, that is, "the time scale at which the input signal is sampled non-uniformly." This scale is not preset manually, but is a specific scale learned from a large amount of sensor signal data, which is best suited to distinguish between "high-fidelity / low-fidelity signals." Different convolutional kernel weights correspond to different dimensions of key feature scales in the task, which together constitute a multi-scale feature extraction system adapted to sensor sensitivity evaluation.

[0052] Furthermore, the sensitivity degradation of built-in partial discharge sensors can manifest in various ways, such as decreased high-frequency response (slower rising edge) and resonant point shift (change in oscillation frequency). Treating all differences equally makes it impossible to distinguish which differences originate from key features and which from secondary features. Therefore, by using convolutions with learnable dilation rates, temporal convolutional networks automatically learn during training which time scales (corresponding to which features) are most important for distinguishing "high-fidelity / low-fidelity" signals and how to assign different weights to different features. When reconstructing signals, temporal convolutional networks prioritize the accuracy of key features (such as rising edge and main oscillation frequency), while the accuracy of restoring non-critical details is relatively lower.

[0053] In some embodiments, the encoder of the signal reconstruction network may include four cascaded residual blocks, i.e., enhanced temporal convolutional blocks, which enable the construction of a hierarchical receptive field. Specifically, the first enhanced temporal convolutional block focuses on details within a very short time frame, such as the rising edge of a pulse; the second enhanced temporal convolutional block focuses on a slightly longer time window, such as the oscillation of a complete pulse; the third enhanced temporal convolutional block focuses on an even longer time window, such as the envelope of a pulse train; and the fourth enhanced temporal convolutional block focuses on features over a longer time scale, such as the periodic patterns or overall trends of the pulse sequence. This helps to capture low-frequency trends or periodic patterns in the current voltage pulse signal. Thus, the signal reconstruction network includes an encoder and a decoder composed of four cascaded enhanced temporal convolutional blocks. Step 102 can be implemented through steps 1021 and 1022. Step 1021: Using four cascaded enhanced temporal convolutional blocks, perform deep feature encoding on the current voltage pulse signal layer by layer to obtain a low-dimensional feature vector.

[0054] In some embodiments, the basic time scale of the signal processed by each level of the enhanced temporal convolutional block determines the basic time scale of the signal processed by that level of the enhanced temporal convolutional block. Here, each level of the enhanced temporal convolutional block contains multiple parallel causal convolutional layers with different dilation rates. Specifically, there may be four parallel causal convolutional layers with different dilation rates, wherein a processing branch may be assigned to each of the rising edge, main oscillation, decay envelope, and a redundant / interactive feature of the input signal. Correspondingly, each level of the enhanced temporal convolutional block includes: a temporal convolutional layer, an upsampling layer, a feature concatenation layer, and a fusion convolutional layer; the temporal convolutional layer includes: multiple parallel causal convolutional layers with different dilation rates; the dilation rates are different in different levels of the enhanced temporal convolutional blocks; reference Figure 2 As shown, step 1021 above can be achieved through the following steps B1 to B4: Step B1: For each level of enhanced temporal convolutional block, the input signal is input into multiple parallel causal convolutional layers in the level of enhanced temporal convolutional block to extract temporal features, thereby obtaining multiple temporal features corresponding to the level.

[0055] Step B2: Using the upsampling layer in the level-enhanced temporal convolutional block, perform linear interpolation on the multiple temporal features corresponding to the level to obtain multiple virtual feature sequences corresponding to the level.

[0056] Step B3: Using the feature concatenation layer in the level-enhanced temporal convolutional block, perform feature concatenation on multiple virtual feature sequences corresponding to the level to obtain the concatenated features corresponding to the level.

[0057] Step B4: Using the fusion convolutional layer in the level-enhanced temporal convolutional block, perform feature fusion and channel compression on the spliced ​​features corresponding to the level to obtain the output features corresponding to the level.

[0058] The input signal of the first-level enhanced temporal convolutional block is the current voltage pulse signal; the input signal of the Nth-level enhanced temporal convolutional block is the output signal of the (N-1)th-level enhanced temporal convolutional block; N is greater than or equal to 2 and less than or equal to 4; the output signal of the fourth-level enhanced temporal convolutional block is a low-dimensional feature vector.

[0059] In some embodiments, the encoder, composed of four cascaded enhanced temporal convolutional blocks, receives a single-channel timing signal as its input, namely the current voltage pulse signal. When the current voltage pulse signal is input to the first-stage enhanced temporal convolutional block: First, the input is a temporal convolutional layer consisting of multiple parallel causal convolutional layers with different dilation rates. Each causal convolutional layer has an independent learnable dilation rate and corresponding kernel weights. For the input current voltage pulse signal, each causal convolutional layer extracts temporal features to obtain the corresponding temporal features.

[0060] Secondly, using the upsampling layer, linear interpolation is performed on the temporal features output by each causal convolutional layer to obtain multiple corresponding virtual feature sequences.

[0061] Then, using the feature concatenation layer, multiple virtual feature sequences output by the upsampling layer are concatenated to obtain the concatenated features of the first-level enhanced temporal convolution block.

[0062] Finally, by using a fusion convolutional layer, the spliced ​​features of the first-level enhanced temporal convolutional block are fused and channel compressed to obtain the output features of the first level.

[0063] Similarly, in the four cascaded enhanced temporal convolutional blocks, deep feature extraction is performed layer by layer on the current voltage pulse signal to obtain the low-dimensional feature vector of the current voltage pulse signal.

[0064] Step 1022: Using a decoder, perform signal recovery on the low-dimensional feature vector to obtain the voltage signal to be compared.

[0065] In some embodiments, the decoder within the signal reconstruction network is a symmetrical mirror image of the encoder, and it may consist of: fully connected layers and several deconvolutional layers (transposed convolutions).

[0066] Here, continue to refer to Figure 2 As shown, step 1022 can be achieved through the following steps C1 to C3: Step C1: Expand the low-dimensional feature vector into an initial feature map using the fully connected layer in the decoder.

[0067] Step C2: Using multiple cascaded deconvolutional layers within the decoder, the initial feature map is progressively upsampled to obtain intermediate features.

[0068] Step C3: Use the convolutional layer in the decoder to fuse and reconstruct the intermediate features to obtain the voltage signal to be compared.

[0069] In some embodiments, the fully connected layer in the decoder can expand the low-dimensional feature vector into an initial feature map; then, the initial feature map is progressively upsampled through multiple deconvolutional layers, each of which can be connected to a nonlinear activation function and a normalization layer to progressively recover the temporal structure and details of the current voltage pulse signal; finally, the number of channels is reduced to 1 through a 1x1 convolutional layer to output the reconstructed voltage signal to be compared.

[0070] Step 103: Based on the current voltage pulse signal and the voltage signal to be compared, determine the detection sensitivity score of the built-in partial discharge sensor.

[0071] In some embodiments, the current voltage pulse signal may be preprocessed first, and then the preprocessed signal may be input into a signal reconstruction network for signal reconstruction to obtain the voltage signal to be compared.

[0072] Here, step 103 above can be achieved through the following steps 1031 and 1032: Step 1031: Using the error calculation formula, determine the normalized mean square error between the current voltage pulse signal and the voltage signal to be compared.

[0073] The error calculation formula is as follows: Formula (2); in, This is the normalized mean square error; and These are the current voltage pulse signal and the voltage signal to be compared, respectively. It is the Euclidean norm.

[0074] Step 1032: The difference between 1 and the normalized mean square error is determined as the detection sensitivity score of the built-in partial discharge sensor.

[0075] In some embodiments, the detection sensitivity score of the built-in partial discharge sensor ;Should The range of values ​​is The closer the value is to 1, the more consistent the intrinsic characteristics of the signal to be evaluated, i.e. the current voltage pulse signal, are with the signal generated when partial discharge occurs inside the UHV GIS. In other words, the higher the fidelity of the built-in partial discharge sensor and the better the sensitivity performance is maintained.

[0076] In this scheme, the normalized mean square error between the current voltage pulse signal and the voltage signal to be compared is calculated after network reconstruction, instead of directly calculating the normalized mean square error between the current voltage pulse signal and the signal generated when partial discharge occurs inside the UHV GIS. The signals generated when partial discharge occurs inside an UHV GIS are inherently unknown and variable. In field applications, it is unlikely that a clear, reliable reference signal perfectly matches the physical conditions of the current voltage pulse signal will exist. Furthermore, the signal reconstruction network learns the inherent manifold characteristics of the clean signal and multiple calibrated voltage pulse signals during the training phase (simulating different operating conditions by changing different pulse parameters, injecting signals at different locations, and injecting signals with different signal-to-noise ratios). When the current voltage pulse signal is input, the encoder within the signal reconstruction network extracts its core features, and the decoder reconstructs the comparison voltage signal based on these features. This process is equivalent to the signal reconstruction network, based on its understanding of the ideal signal and guided by the characteristics of the current voltage pulse signal, outputting the signal that the ideal built-in partial discharge sensor should output—the reconstructed comparison voltage signal. This comparison voltage signal is dynamically generated and adaptively adapted to the physical conditions of the current voltage pulse signal, rather than a fixed, static template that may not match the actual operating conditions.

[0077] Based on the above description, directly comparing the current voltage pulse signal with the signal generated when partial discharge occurs inside the Inter-HV GIS will mix all differences (including noise, interference, and sensor degradation). When the signal-to-noise ratio is low, noise will dominate, and the inherent problems of the built-in partial discharge sensor will be overshadowed. The signal reconstruction network mentioned in this solution, having seen clean signals during training, has learned to map the input current voltage pulse signal onto a clean signal manifold. Here, when the current voltage pulse signal is input to the signal reconstruction network, the encoder can extract the parts that conform to the characteristics of a clean signal, such as the main oscillation and rising edge shape of the pulse, and encode and decode them to reconstruct the clean signal corresponding to the current voltage pulse signal. This process is essentially implicit denoising and feature enhancement; that is, the signal reconstruction network only reconstructs the pulse body that the network believes should exist, while leaving noise, random interference, and distortion caused by sensor degradation in the time-frequency joint multi-scale reconstruction error.

[0078] Therefore, the normalized mean square error between the current voltage pulse signal and the voltage signal to be compared is actually a comprehensive fidelity index after feature weighting. Differences in key features are amplified in the error, while differences in non-key features are compressed. This intelligent weighting capability makes the sensitivity score of the built-in partial discharge sensor more reflective of the true performance degradation of the built-in partial discharge sensor.

[0079] Correspondingly, as shown in Table 1, a comparison is given between the normalized mean square error 1 calculated using this application for the current voltage pulse signal and the voltage signal to be compared, and the normalized mean square error 2 calculated for the current voltage pulse signal and the signal generated when partial discharge occurs inside the UHV GIS. Table 1 Comparative Description In some embodiments, the calculated quantitative sensitivity score can be further transformed into intuitive equipment status conclusions and maintenance guidance. That is, based on the sensitivity score... [-1, 1] and a performance threshold preset based on historical data or standards, such as: Compare the results and generate a clear evaluation conclusion accordingly. If... If the sensitivity of the built-in partial discharge sensor is satisfactory under the current test conditions, it is determined that the sensor meets the operational requirements; otherwise, if... If the status is deemed suspicious or unqualified, an automatic warning signal will be issued, prompting maintenance or further inspection. Simultaneously, the sensitivity score from this evaluation can be recorded and stored as a data point in the long-term performance monitoring database of the built-in partial discharge sensor. By tracking the trend of sensitivity score changes in each evaluation, continuous and predictive monitoring of the sensitivity performance degradation of the built-in partial discharge sensor can be achieved.

[0080] The intelligent sensitivity evaluation method for UHV GIS built-in partial discharge sensors provided in this application involves the following steps: First, the current voltage pulse signal output by the UHV gas-insulated switchgear (GIS) built-in partial discharge sensor is acquired. Then, a signal reconstruction network is used to decode and reconstruct the current voltage pulse signal to obtain a comparison voltage signal. The signal reconstruction network is trained using multiple calibrated voltage pulse signals on a time-series convolutional network with a learnable dilatation rate, employing a time-frequency joint multi-scale reconstruction loss. Finally, based on the current voltage pulse signal and the comparison voltage signal, the detection sensitivity score of the built-in partial discharge sensor is determined. This method achieves accurate capture of key signal fidelity features through learnable dilatation rate convolution and drives the network to learn the complete physical characteristics of a pure signal through the time-frequency joint reconstruction loss. This allows the signal reconstruction network to simultaneously possess adaptive multi-scale feature extraction capabilities and sensitivity to frequency distortion, enabling it to accurately characterize the key pulse waveform features of the current voltage pulse signal output by the built-in partial discharge sensor. This improves the accuracy of sensitivity evaluation for UHV GIS built-in partial discharge sensors in adaptive complex environments.

[0081] The following describes the intelligent sensitivity evaluation method for the built-in partial discharge sensor of the UHV GIS with a specific embodiment. However, it is worth noting that this specific embodiment is only for better illustration of this application and does not constitute an improper limitation of this application.

[0082] In practical applications, strictly adhering to standards in the field of power equipment condition monitoring, a high-fidelity simulation and physical testing environment is constructed to verify the intelligent sensitivity evaluation method for the UHV GIS built-in partial discharge sensor provided in this application, and its evaluation performance and engineering practicality under complex electromagnetic interference. The high-fidelity simulation experimental platform employs a dual-modal verification system combining a scaled-down model of the real GIS cavity and high-frequency electromagnetic simulation software, such as CST Studio Suite and Ansys HFSS, to ensure that the results are both reproducible and physically realistic. Simultaneously, the following operations are performed: 1. Signal and Interference Simulation: 1.1. Standard Pulse: A programmable high-voltage pulse source is used to generate an adjustable standard partial discharge analog pulse P with a rise time of 2ns, an amplitude of 100mV-1V, and a frequency of 100kHz-10MHz. std .

[0083] 1.2. Background Noise and Interference: Three typical types of on-site interference are simulated through signal injection: A. Gaussian white noise, with an adjustable signal-to-noise ratio (SNR) from 0 dB to 20 dB.

[0084] B. Periodic carrier interference, such as 200kHz or 400kHz fixed frequency sine waves.

[0085] C. Random pulse-type interference to mimic switch operation.

[0086] 1.3. Sensor simulation: A set of built-in sensors of the same model that have been precisely calibrated in the laboratory are used as the evaluation objects, and different levels of sensitivity degradation are artificially introduced, including: high frequency response attenuation (simulated by low-pass filtering), sensitivity decrease (simulated by attenuator) and baseline drift.

[0087] 2. Comparison method settings: 2.1. Traditional time-domain comparison method: The signal to be tested is compared with the laboratory standard reference signal, the root mean square error and cross-correlation coefficient are calculated, and the weighted sum of the two is taken as the sensitivity score, which serves as the performance baseline. This represents the most intuitive comparison approach.

[0088] 2.2. Frequency Domain Energy Ratio Method: The energy ratio between the signal under test and the laboratory standard reference signal in the key frequency band, such as 1-10MHz, is calculated and used as a sensitivity score. This represents a simple evaluation method based on frequency domain analysis commonly used in the prior art.

[0089] 2.3. Deep Learning Regression Method: This method uses a conventional autoencoder composed of ordinary convolutional layers (with a fixed dilation rate for the convolutional kernels, no learnable dilation rate, and no frequency domain loss). It only uses the temporal reconstruction error as the evaluation metric, representing a deep learning method without adaptive capabilities.

[0090] 2.4. This application: A method for intelligently evaluating the sensitivity of a built-in partial discharge sensor in UHV GIS, namely, an autoencoder based on a learnable dilatation rate temporal convolutional network, trained using a time-frequency joint multi-scale reconstruction loss. In application, no reference signal is required; sensitivity is evaluated solely through the reconstruction error of the signal under test.

[0091] 3. Set evaluation indicators: 3.1. Assessment Accuracy: Pearson Correlation Coefficient (PCC) and Mean Squared Error (MSE) between the scores (or indicators) output by each method and the actual attenuation coefficient of the sensor at different degradation levels; where, the higher the PCC and the lower the MSE, the more accurate the assessment.

[0092] 3.2. Anti-interference robustness: Under a fixed sensor degradation level, the background noise intensity is gradually increased, i.e., the SNR is reduced, and the standard deviation (Std) of the evaluation results of each method is observed; the smaller the standard deviation, the stronger the anti-interference ability of the method and the more stable the results.

[0093] 3.3. Computational efficiency: Average inference time (ms) required for a single assessment to reflect the feasibility of field deployment.

[0094] The experimental results are shown in Table 2.

[0095] Table 2 Summary of Experimental Results Based on the above description, traditional methods (2.1. traditional time-domain comparison method and 2.2. frequency-domain energy ratio method) have poor evaluation accuracy (PCC in the range of 0.65-0.72) and anti-interference ability (Std>0.13), proving that simple time-frequency domain indicators are unreliable under complex noise conditions and rely on laboratory standard reference signals, limiting their field applications. However, they are extremely fast (<1ms), making them suitable for rapid screening scenarios.

[0096] The accuracy of the deep learning regression method in 2.3 (PCC=0.89) is significantly improved, but due to the fixed convolution kernel expansion rate, it cannot adaptively match the multi-scale features of the pulse (e.g., fast rising edges and slow oscillations), and it is still relatively sensitive to noise (Std=0.072). The inference time is 18ms, which meets the real-time requirements.

[0097] This application achieves optimal results in both evaluation accuracy (PCC=0.95) and robustness against interference (Std=0.048), significantly outperforming other methods. Its core advantage lies in: 1. The learnable dilation rate enables the network to adaptively extract key time-scale features of pulses, such as rising edge, oscillation frequency, and decay rate.

[0098] 2. The time-frequency joint multi-scale reconstruction loss simultaneously constrains the time-domain waveform and frequency-domain energy distribution, making the reconstructed signal closer to the complete physical characteristics of the pure signal.

[0099] 3. The architecture of temporal convolutional networks naturally achieves the separation of noise and signal, and noise and distortion components are effectively included in the reconstruction error.

[0100] Furthermore, although the inference time (24ms) of this application is slightly higher than that of the deep learning regression method, it fully meets the real-time requirements of field testing (second-level response) and still has an overwhelming accuracy advantage compared with traditional methods (traditional time-domain comparison method and frequency-domain energy ratio method).

[0101] Correspondingly, this application further conducts ablation experiments, namely, rigorous ablation experiments were designed to clarify the independent contributions of each aspect of the signal reconstruction network in this application. Based on the full model, key components were removed or replaced sequentially, as shown in Table 3: Table 3 Ablation Experiment Data Table 3 shows that the learnable dilation rate mechanism is the core of improving the discriminative power of feature extraction, contributing the most significant accuracy gain (PCC improvement of 0.05). Frequency domain loss effectively complements the sensitivity of time domain loss to frequency characteristic changes; the combination of the two forms a joint time-frequency constraint, further improving the evaluation accuracy (PCC improvement of 0.03). The encoder reconstruction architecture is more stable and has a clearer physical meaning than simple feature distance, providing a fundamental guarantee for the reliability of the evaluation metrics.

[0102] Through comparative experiments and ablation analysis of the above-mentioned systems, the beneficial effects of this application have been fully demonstrated: 1. High evaluation accuracy: Compared with traditional time-frequency domain methods and standard autoencoders, this application has significant advantages in evaluation accuracy (PCC improvement of 7%-31%) and resistance to complex electromagnetic interference (result volatility reduction of more than 33%). Its core lies in achieving accurate capture of key features of signal fidelity through learnable dilation rate convolution, and driving the network to learn the complete physical characteristics of pure signals through time-frequency joint reconstruction loss.

[0103] 2. Clear module contributions and effective innovations: Ablation experiments confirm that the learnable dilation rate mechanism contributes the most significant accuracy gain (PCC improvement of 0.05), while the frequency domain loss provides an effective supplement (PCC improvement of 0.03). The synergy between the two enables the network to simultaneously possess adaptive multi-scale feature extraction capabilities and sensitivity to frequency distortion. The autoencoder reconstruction architecture provides a stable and physically meaningful evaluation benchmark.

[0104] 3. Excellent Engineering Practicality: This solution takes approximately 24ms per evaluation, meeting the needs of rapid on-site testing. Its reference signal-free characteristic completely eliminates the dependence on an ideal calibration environment, allowing direct application to on-site verification of already operational GIS equipment. It maintains an evaluation capability of PCC>0.78 even in 0dB high-noise environments and maintains linear response within a severely degraded range of -20dB, providing a reliable and efficient next-generation solution for intelligent operation and maintenance and accurate condition assessment of UHV GIS with built-in partial discharge sensors.

[0105] 4. Significant technical and economic advantages: Compared with traditional methods that require periodic equipment shutdown and sensor disassembly for laboratory calibration, this solution can complete rapid evaluation while the equipment is running normally, significantly reducing maintenance costs and time; compared with standard self-encoders, although the inference time of this solution is slightly increased (24ms vs 18ms), the improvement in evaluation accuracy and anti-interference capability is enough to offset this slight cost, resulting in a higher overall cost-effectiveness.

[0106] In other words, this application provides an intelligent sensitivity evaluation method for UHV GIS built-in partial discharge sensors that can accurately characterize the sensor's response capability to key features of pulse waveforms and adapt to complex environments.

[0107] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0108] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0109] In the several embodiments provided in this application, it should be understood that the disclosed methods can be implemented in other ways.

[0110] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0111] The features disclosed in the methods provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0112] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligently evaluating the sensitivity of an embedded partial discharge sensor in ultra-high voltage GIS, characterized in that, The method includes: Acquire the current voltage pulse signal output by the built-in partial discharge sensor of the UHV gas-insulated switchgear (GIS); A signal reconstruction network is used to decode and reconstruct the current voltage pulse signal to obtain the voltage signal to be compared. The signal reconstruction network is trained on a temporal convolutional network with a learnable dilation rate using multiple calibrated voltage pulse signals and employing a time-frequency joint multi-scale reconstruction loss. Based on the current voltage pulse signal and the voltage signal to be compared, the detection sensitivity score of the built-in partial discharge sensor is determined.

2. The method according to claim 1, characterized in that, The signal reconstruction network includes an encoder and a decoder consisting of four cascaded enhanced temporal convolutional blocks; Using a signal reconstruction network, the current voltage pulse signal is decoded and reconstructed to obtain the voltage signal to be compared, including: By using four cascaded enhanced temporal convolutional blocks, deep feature encoding is performed layer by layer on the current voltage pulse signal to obtain a low-dimensional feature vector; The low-dimensional feature vector is recovered using a decoder to obtain the voltage signal to be compared.

3. The method according to claim 2, characterized in that, Each level of the enhanced temporal convolutional block includes: a temporal convolutional layer, an upsampling layer, a feature concatenation layer, and a fusion convolutional layer; the temporal convolutional layer includes: multiple parallel causal convolutional layers with different dilation rates; the dilation rates are different in different levels of the enhanced temporal convolutional block. By utilizing four cascaded enhanced temporal convolutional blocks, deep feature encoding is performed layer by layer on the current voltage pulse signal to obtain a low-dimensional feature vector, including: For each level of enhanced temporal convolutional block, the input signal is input into multiple parallel causal convolutional layers in the level of enhanced temporal convolutional block to extract temporal features, thereby obtaining multiple temporal features corresponding to the level. By utilizing the upsampling layer in the level-enhanced temporal convolutional block, linear interpolation is performed on multiple temporal features corresponding to the level to obtain multiple virtual feature sequences corresponding to the level. By utilizing the feature concatenation layer in the enhanced temporal convolutional block of the level, multiple virtual feature sequences corresponding to the level are concatenated to obtain the concatenated features corresponding to the level. By utilizing the fusion convolutional layer in the enhanced temporal convolutional block of the level, feature fusion and channel compression are performed on the spliced ​​features corresponding to the level to obtain the output features corresponding to the level. The input signal of the first-level enhanced temporal convolutional block is the current voltage pulse signal; the input signal of the Nth-level enhanced temporal convolutional block is the output signal of the (N-1)th-level enhanced temporal convolutional block; N is greater than or equal to 2 and less than or equal to 4; the output signal of the fourth-level enhanced temporal convolutional block is a low-dimensional feature vector.

4. The method according to claim 2 or 3, characterized in that, Using a decoder, signal recovery is performed on the low-dimensional feature vector to obtain the voltage signal to be compared, including: The low-dimensional feature vectors are expanded into an initial feature map using a fully connected layer within the decoder. By using multiple cascaded deconvolutional layers within the decoder, the initial feature map is progressively upsampled to obtain intermediate features; The intermediate features are fused and reconstructed using the convolutional layer within the decoder to obtain the voltage signal to be compared.

5. The method according to claim 1, characterized in that, The temporal convolutional network includes: an initial encoding module and an initial decoding module consisting of four cascaded initial temporal convolutional blocks; the initial temporal convolutional block consists of: a temporal convolutional layer consisting of multiple parallel causal convolutional layers with learnable dilation rates; The construction process of a signal reconstruction network includes: Construct training sample sets corresponding to multiple calibrated voltage pulse signals; By employing a time-frequency joint multi-scale reconstruction loss, and based on the training sample set, the signal reconstruction network is obtained by adjusting multiple learnable dilation rates within the temporal convolutional layers of the initial temporal convolutional block at least for each level.

6. The method according to claim 5, characterized in that, Construct multiple calibrated training sample sets corresponding to voltage pulse signals, including: Data enhancement is performed on each calibrated voltage pulse signal to obtain multiple noisy signals corresponding to the calibrated voltage pulse signal; wherein, data enhancement includes at least: changing different pulse parameters, injecting signals at different locations, and injecting signals with different signal-to-noise ratios; Construct a training sample set consisting of multiple calibrated voltage pulse signals and multiple noisy signals corresponding to each calibrated voltage pulse signal.

7. The method according to claim 1 or 5, 6, characterized in that, The time-frequency joint multi-scale reconstruction loss is: ； in, For time-frequency joint multi-scale reconstruction loss; This represents the total number of calibrated voltage pulse signals. The balancing coefficient for temporal reconstruction loss; For the first A calibrated voltage pulse signal; For the reason and Composed of multiple corresponding noisy signals The first in the signal subset One signal; , These are the initial encoding module and the initial decoding module of the temporal convolutional network, respectively. For the output of a temporal convolutional network The reconstructed signal; for The total number of corresponding noisy signals; The balance coefficients for frequency domain reconstruction loss; To make the first Frequency domain characteristics of a calibrated voltage pulse signal; This is a transform function that converts a time-domain signal into frequency-domain features. for Frequency domain characteristics; The regularization coefficient is used. Model parameters for temporal convolutional networks The regularization term.

8. The method according to claim 1, characterized in that, Based on the current voltage pulse signal and the voltage signal to be compared, the detection sensitivity score of the built-in partial discharge sensor is determined, including: The normalized mean square error between the current voltage pulse signal and the voltage signal to be compared is determined using the error calculation formula; the error calculation formula is as follows: ； in, This is the normalized mean square error; and These are the current voltage pulse signal and the voltage signal to be compared, respectively. It is the Euclidean norm; The difference between 1 and the normalized mean square error is determined as the detection sensitivity score of the built-in partial discharge sensor.

Images