Method and device for denoising partial discharge signals

Abstract

The application provides a partial discharge signal denoising method and device, the method comprising: performing amplitude normalization and time domain scale transformation on a noisy first partial discharge signal to obtain a second partial discharge signal; inputting the second partial discharge signal into a stochastic resonance system, using the noise energy contained in the second partial discharge signal to induce particles to jump between three stable potential wells, realizing resonance amplification of the discharge characteristic signal, and obtaining a third partial discharge signal; performing inverse scale transformation on the third partial discharge signal to restore the original time scale, and obtaining a fourth partial discharge signal that retains discharge characteristics and has a preliminary improved signal-to-noise ratio; inputting the fourth partial discharge signal into a pre-trained deep learning model, using a multi-head self-attention mechanism to capture long-distance time dependence and global context information in the enhanced signal sequence, distinguishing noise from real discharge pulse characteristics, and reconstructing through a model decoder to output a partial discharge pulse signal with a high signal-to-noise ratio.

Description

Technical Field

[0001] This application relates to the field of signal denoising technology, and in particular to a method and apparatus for denoising partial discharge signals. Background Technology

[0002] GIS equipment is widely used in high-voltage power transmission systems due to its compact structure and high operational reliability. Insulators, as key insulating components, are often subject to partial discharge (PD) phenomena as their insulation performance deteriorates. The evolution trend of partial discharge signals can reflect the development process of insulation defects and is an important basis for judging the insulation condition.

[0003] Current partial discharge detection methods primarily utilize ultra-high frequency (UHF) sensors. Some initial partial discharge pulses with small amplitudes and low frequencies become extremely weak by the time they reach the sensor after propagation, attenuation, and scattering within the equipment. Compared to the complex electromagnetic environment of the field, the amplitude and frequency of these extremely weak partial discharge signals are often "submerged" in background noise. While these signals contain rich information, current limitations in their capture necessitate a noise reduction method to separate them from ambient interference signals.

[0004] Existing methods for denoising such signals include Empirical Mode Decomposition (EMD), blind source separation, and convolutional neural network (CNN) denoising. Empirical Mode Decomposition and its variants, such as EMD, EEMD, and VMD, are adaptive signal decomposition methods that decompose the signal into intrinsic mode functions (IMFs) and then reconstruct and denoise based on the characteristics of the PD signal and noise in the IMF components. Blind source separation, especially independent component analysis, assumes that the observed signal is a linear mixture of multiple source signals (including PD and various types of noise). Algorithms separate these source signals to extract the pure PD signal, which is particularly effective when processing multi-channel data. Convolutional neural network (CNN) denoising can directly extract and classify features from the time-spectrum or original waveform of the PD signal, enabling automatic identification, denoising, and type diagnosis of the PD signal.

[0005] However, the existing methods described above still have the following drawbacks: Separation algorithms for empirical mode decomposition and its variants require the observation of such partial discharge signals under noise-free conditions, identification of features, and then separation. This places extremely high demands on the initial field data. In practical applications, such algorithms are difficult to implement due to the lack of corresponding detection methods.

[0006] The blind source separation algorithm assumes that the PD and various noises are stable signal sources. In actual applications, the sources of noise in the field are very complex, and it is a typical asynchronous multi-source system. The principle of the PD determines that the occurrence of extremely weak signals is mostly an occasional phenomenon, which does not meet the application conditions of the blind source separation algorithm in terms of timing.

[0007] Convolutional Neural Networks (CNNs), as a new-era artificial intelligence algorithm, possess the ability to extract complex signals based on similar typical features in the time and frequency domains. However, CNNs require a large amount of basic data for training, and obtaining such a large dataset is costly and extremely difficult in practical applications. If the training data does not accurately represent the distribution of the real world (i.e., it contains bias), CNNs will learn and amplify these biases, resulting in poor generalization ability when the model is deployed. For example, if the same person is carrying the same mobile phone, the phone's signal will be trained to be considered a common "abnormal signal" rather than noise. The core of CNNs is local convolution operations, and their receptive field is local, lacking a global perspective. Although the receptive field of deep neurons can be increased by stacking multiple layers of convolutions, the efficiency of capturing long-range dependencies is very low, requiring very deep networks. Summary of the Invention

[0008] In view of this, this application provides a method and apparatus for denoising partial discharge signals to solve at least one of the aforementioned problems.

[0009] To achieve the above objectives, this application adopts the following approach: According to a first aspect of this application, a method for denoising partial discharge signals is provided, the method comprising: The amplitude is normalized and the time-domain scale is transformed on the noisy first partial discharge signal to obtain the second partial discharge signal; The second partial discharge signal is input into a stochastic resonance system. The noise energy contained in the second partial discharge signal is used to induce the particles to transition between the three-stable potential traps, thereby achieving resonant amplification of the discharge characteristic signal and obtaining the third partial discharge signal. The third partial discharge signal is inversely scaled to restore the original time scale, resulting in a fourth partial discharge signal that retains the discharge characteristics and has a preliminary improved signal-to-noise ratio. The fourth partial discharge signal is input into a pre-trained Transformer deep learning model. The multi-head self-attention mechanism in the model encoder is used to capture long-distance temporal dependencies and global context information in the enhanced signal sequence, distinguish noise from real discharge pulse features, and reconstruct the signal through the model decoder to output a partial discharge pulse signal with a high signal-to-noise ratio.

[0010] In one embodiment of this application, the above method for amplitude normalization of the noisy first partial discharge signal includes: The following formula is used to analyze the noisy first partial discharge signal. Perform amplitude linear normalization: ; In the formula: for The normalization result, For the noisiest signal, max() represents the maximum value.

[0011] In one embodiment of this application, the specific steps of the time-domain scaling transformation in the above method include: Define a scale compression factor k, and redefine the time axis of the amplitude-normalized signal as... At the same time, the actual sampling frequency of the signal fs Compress to fs ′= fs / k The waveform of the second partial discharge signal after time-domain scaling remains unchanged in the time domain, while the frequency band shifts to the low-frequency band that the stochastic resonance system can directly act upon in the frequency domain.

[0012] In one embodiment of this application, the model of the stochastic resonance system in the above method adopts a tristable stochastic resonance model, and the dynamic behavior of the tristable stochastic resonance model is described by the Langevin equation: ; In the formula: U(x) is the potential well function of the potential field where the Brownian particle is located. , and Here, γ is the system parameter, τ is the system damping ratio, δ(t) is the time delay, and δ(t) is the pulse signal.

[0013] In one embodiment of this application, the method further includes a step of optimizing the system parameters of the stochastic resonance system using a chaotic sparrow search algorithm before inputting the second partial discharge signal into the stochastic resonance system: The output signal-to-noise ratio of the stochastic resonance system is used as the fitness function to represent the system's output performance, as shown in the following formula: ; In the formula: The characteristic frequency of the system input signal. This is the sequence number corresponding to the characteristic frequency. This represents the energy at the characteristic frequency. For noise energy; The Chaotic Sparrow Search algorithm is used to find the optimal combination of system parameters that maximizes the output signal-to-noise ratio under the current input signal. , , as well as And this optimal parameter combination is applied to the stochastic resonance system.

[0014] In one embodiment of this application, the above method utilizes a multi-head self-attention mechanism in the model encoder to capture long-range temporal dependencies and global contextual information in the enhanced signal sequence, specifically including: Step S1: Construct matrix vectors , , Three weight matrices are created through three fully connected layers. , , For the input signal Linear mapping, forming new matrix vectors , , : ; ; ; Where Linear represents the operation of the fully connected layer. , and These are the bias vectors for different networks; Step S2: , , Divide equally Group, , and Different heads are formed, and each head calculates an attention mechanism in a different space to obtain different features. The calculation process of the attention mechanism for a single head is as follows: ; in yes transpose, represent Dimensions and The dot product calculates the correlation between the two, and the softmax function will... and The matrix after dot product is mapped to the range of 0 to 1. The softmax calculation process is as follows:

[0015] Step S3: Multi-head attention extracts different feature information from different heads, and concatenates these features to obtain richer semantic features. The multi-head feature fusion process is as follows: ; in Matrix is ​​used to Features obtained from multiple attention mechanisms are fused together.

[0016] In one embodiment of this application, the Transformer deep learning model in the above method further includes residual connections, layer normalization, and feedforward neural network modules, and the method further includes the following processing steps: Directly input signals of the multi-head self-attention mechanism The output signal after layer normalization and multi-head self-attention mechanism operation By performing matrix addition, the intermediate feature signal is obtained. : ; The intermediate feature signal Inputting the feedforward neural network module yields the hidden layer output. : ; Where Linear represents fully connected layer operations; The approximate calculation process for the GELU activation function is as follows: ; Where tanh is the hyperbolic tangent function.

[0017] According to a second aspect of this application, a noise reduction device for partial discharge signals is provided, the device comprising: The signal processing unit is used to perform amplitude normalization and time-domain scaling transformation on the noisy first partial discharge signal to obtain the second partial discharge signal. The signal amplification unit is used to input the second partial discharge signal into the stochastic resonance system, and use the noise energy contained in the second partial discharge signal to induce the particles to transition between the tristable state traps, thereby realizing the resonant amplification of the discharge characteristic signal and obtaining the third partial discharge signal. The inverse transformation unit is used to perform an inverse-scale transformation on the third partial discharge signal to restore it to the original time scale, thereby obtaining a fourth partial discharge signal that retains the discharge characteristics and has a preliminary improved signal-to-noise ratio. The denoising unit is used to input the fourth partial discharge signal into the pre-trained Transformer deep learning model, use the multi-head self-attention mechanism in the model encoder to capture long-distance time dependence and global context information in the enhanced signal sequence, distinguish noise from real discharge pulse features, and reconstruct the signal through the model decoder to output a partial discharge pulse signal with a high signal-to-noise ratio.

[0018] According to a third aspect of this application, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0019] According to a fourth aspect of this application, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the above-described method.

[0020] According to a fifth aspect of this application, a computer program product is provided, comprising a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.

[0021] The denoising method and apparatus for partial discharge signals proposed in this application introduce a tristable stochastic resonance system and utilizes noise energy to induce particle transitions to achieve resonant amplification of weak signals, providing a high signal-to-noise ratio input for subsequent processing. This effectively solves the problem of deep learning models struggling to converge in extremely low signal-to-noise ratio environments. Furthermore, by combining the Transformer deep learning model and utilizing its multi-head self-attention mechanism in the encoder to capture long-distance temporal dependencies and global contextual information in the signal sequence, it overcomes the limitations of traditional convolutional neural networks in terms of limited receptive field and poor generalization ability. This architecture, which combines physical mechanisms with data-driven approaches, not only achieves accurate extraction and high-fidelity waveform reconstruction of extremely weak partial discharge signals in strong noise environments but also significantly enhances the system's adaptability and robustness to complex operating conditions, thereby greatly improving the accuracy of online monitoring of the insulation status of power equipment and its early fault warning capability. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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. In the drawings: Figure 1 This is a schematic flowchart of a method for denoising partial discharge signals provided in an embodiment of this application; Figure 2 This is a flowchart of the random resonance signal enhancement provided in the embodiments of this application; Figure 3 This is a schematic diagram of the transformer denoising network provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of a noise reduction device for partial discharge signals provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the embodiments of this application will be further described in detail below with reference to the accompanying drawings. Here, the illustrative embodiments and descriptions of this application are used to explain this application, but are not intended to limit this application.

[0024] like Figure 1 The diagram shown is a flowchart illustrating a method for denoising partial discharge signals according to an embodiment of this application. The method includes the following steps: Step S101: Perform amplitude normalization and time-domain scaling transformation on the noisy first partial discharge signal to obtain the second partial discharge signal.

[0025] Step S101 performs amplitude normalization and time-domain scaling transformation to meet the input conditions of the subsequent stochastic resonance system.

[0026] A stochastic resonance system is a nonlinear dynamic system that is highly sensitive to the amplitude of the input signal. If the amplitude of the input signal is too large or too small, Brownian particles may fail to transition correctly between potential wells, or the transitions may become completely uncontrolled by noise. The partial discharge signals from GIS collected in the field vary greatly in intensity. By linearly normalizing the amplitude to [-1, 1] or a specific range, the differences in signal intensity at different measurement points can be eliminated, ensuring that all signals are within the optimal operating range of the stochastic resonance system parameters.

[0027] Furthermore, GIS partial discharge signals are ultra-high frequency (UHF) signals, typically in the range of 300MHz-3GHz. Stochastic resonance theory is usually based on the adiabatic approximation and is only applicable to low-frequency, small-parameter signals. If the signal frequency is too high, the system cannot respond in time, and particles cannot follow the signal to transition between potential wells, thus the resonance effect fails. Time-domain scaling can disguise ultra-high frequency PD signals as low-frequency signals, bringing them into the frequency range where the stochastic resonance system can directly operate, thereby utilizing stochastic resonance to enhance weak signals.

[0028] Step S102: Input the second partial discharge signal into the stochastic resonance system, and use the noise energy contained in the second partial discharge signal to induce the particles to transition between the tristable state traps, thereby realizing the resonant amplification of the discharge characteristic signal and obtaining the third partial discharge signal.

[0029] The core objective of this step is to leverage the phenomenon of stochastic resonance to overcome the limitation of traditional noise reduction methods that suppress noise while simultaneously diminishing weak signals. Instead, it utilizes the noise energy already present in the signal as a driving force to assist the weak characteristic signal in driving the system to generate a response, thereby enhancing the amplitude of the weak signal.

[0030] Step S103: Perform inverse scaling transformation on the third partial discharge signal to restore it to the original time scale, and obtain a fourth partial discharge signal that retains the discharge characteristics and has a preliminary improved signal-to-noise ratio.

[0031] In step S101, to adapt the ultra-high frequency partial discharge signal to the stochastic resonance system, a time-domain scale transformation was performed on the signal, lengthening the time axis and lowering the frequency. The resonance enhancement in step S102 is performed at this "disguised" low-frequency time scale. Although the signal is enhanced, its time axis no longer represents true physical time, and the frequency is no longer the true ultra-high frequency. To ensure that subsequent processing (Transformer model) and the final analysis reflect the true physical phenomena, the time scale must be restored to its original state.

[0032] This step involves performing the mathematical operation opposite to step S101 to obtain the characteristics of the fourth partial discharge signal.

[0033] After this processing step, the output signal is converted back into an ultra-high frequency (UHF) signal, with the time unit consistent with the original acquired signal.

[0034] Step S104: Input the fourth partial discharge signal into the pre-trained Transformer deep learning model, use the multi-head self-attention mechanism in the model encoder to capture the long-distance temporal dependence and global context information in the enhanced signal sequence, distinguish noise from real discharge pulse features, and reconstruct the partial discharge pulse signal with a high signal-to-noise ratio through the model decoder.

[0035] This step inputs the signal, enhanced by stochastic resonance and restored to its normal timescale, into the Transformer model. The core of the model lies in the multi-head self-attention mechanism in the encoder, which breaks the limitation of traditional convolutional neural networks that can only focus on local features. This mechanism can scan the entire signal sequence in parallel across multiple different feature subspaces, capturing long-range dependencies and global contextual information between any two time points. This means the model not only looks at the current waveform but also combines it with the signal change trends before and after, thereby accurately identifying discharge pulses with specific physical laws and distinguishing them from randomly distributed background noise.

[0036] After feature extraction and noise removal by the encoder, the decoder reconstructs the signal based on the extracted clean features. The final output partial discharge pulse signal not only greatly removes background noise, but also maintains high fidelity in key features such as waveform shape and amplitude, providing a high-quality data foundation for subsequent insulation condition assessment.

[0037] As described above, the denoising method for partial discharge signals proposed in this application introduces a tristable stochastic resonance system and utilizes noise energy to induce particle transitions to achieve resonant amplification of weak signals, providing a high signal-to-noise ratio input for subsequent processing. This effectively solves the problem of deep learning models struggling to converge in extremely low signal-to-noise ratio environments. Furthermore, by combining the Transformer deep learning model and utilizing its multi-head self-attention mechanism in the encoder to capture long-distance temporal dependencies and global contextual information in the signal sequence, it overcomes the limitations of traditional convolutional neural networks in terms of limited receptive field and poor generalization ability. This architecture, combining physical mechanisms with data-driven approaches, not only achieves accurate extraction and high-fidelity waveform reconstruction of extremely weak partial discharge signals under strong noise backgrounds but also significantly enhances the system's adaptability and robustness to complex operating conditions, thereby greatly improving the accuracy of online monitoring of the insulation status of power equipment and its early fault warning capability.

[0038] In some embodiments of this application, the amplitude normalization of the noisy first partial discharge signal in step S101 includes: The following formula is used to analyze the noisy first partial discharge signal. Perform amplitude linear normalization: ; In the formula: for The normalization result, For the noisiest signal, max() represents the maximum value.

[0039] Because the intensity of the ultra-high frequency partial discharge signals collected on-site varies greatly, direct processing may lead to saturation or no response in the subsequent stochastic resonance system. This step uses amplitude linear normalization to map the original signals of different intensities to a standard numerical range (usually the dimensionless range of [-1, 1] or [0, 1]) to eliminate the differences in signal magnitude and ensure that the signal can be adapted to the potential well parameters of the subsequent stochastic resonance system.

[0040] First, the algorithm traverses the entire signal sequence S. n The maximum amplitude (usually the maximum absolute value to ensure that peaks of both positive and negative half-axis are included) is found, and then each data point in the original signal sequence is divided by this maximum amplitude, which is the normalized result of the first partial discharge signal. After this processing, the signal retains the original waveform shape and relative trend, but its amplitude is strictly limited to ±1.

[0041] In some embodiments of this application, the specific steps of time-domain scaling transformation in step S101 above include: defining a scaling compression coefficient k, and redefining the time axis of the amplitude-normalized signal as... At the same time, the actual sampling frequency of the signal fs Compress to fs ′= fs / k The waveform of the second partial discharge signal after time-domain scaling remains unchanged in the time domain, while the frequency band shifts to the low-frequency band that the stochastic resonance system can directly act upon in the frequency domain.

[0042] Since partial discharge signals are typically ultra-high frequency signals, their extremely high frequencies do not satisfy the adiabatic approximation condition and cannot directly drive stochastic resonance systems, which are generally suitable for low-frequency, small-parameter signals. Therefore, it is necessary to stretch the time scale of the signal, artificially compressing its characteristic frequencies to the low-frequency range that the stochastic resonance system can respond to. After normalizing the amplitude of the first partial discharge signal, the following mathematical transformation is performed: a scale compression coefficient k much greater than 1 is set, stretching the original signal's time axis tt to a new time axis t′, and correspondingly, the actual sampling frequency of the signal is... fs Reduce to a new sampling frequency fs ′.

[0043] The second partial discharge signal obtained after the above processing has the following characteristics: the waveform shape of the signal in the time domain remains completely unchanged, only the time scale is changed (equivalent to slowing it down by k times). The frequency band of the signal has shifted as a whole in the frequency domain, from the original ultra-high frequency band to the low frequency band that the stochastic resonance system can directly act on, thus creating the necessary conditions for step S102 to enhance the weak signal with physical noise.

[0044] In some embodiments of this application, the model of the above-mentioned stochastic resonance system adopts a tristable stochastic resonance model, the dynamic behavior of which is described by the Langevin equation: ; In the formula: U(x) is the potential well function of the potential field where the Brownian particle is located. , a, b, c Here, γ is the system parameter, τ is the system damping ratio, δ(t) is the time delay, and δ(t) is the pulse signal.

[0045] To enhance extremely weak partial discharge signals, this embodiment does not employ the traditional bistable model, but instead constructs a tristable stochastic resonance system. This system has three potential wells (stable states), and compared to the bistable system, its structure is more complex, enabling it to generate richer dynamic behaviors. Consequently, it exhibits better enhancement and robustness when processing weak signals under complex background noise.

[0046] In the above equation, the weak input signal s(t) alone is insufficient to propel the particle over the potential barrier. However, with the energy assistance of the background noise n(t), the particle gains enough energy to undergo periodic transitions between the three potential wells. By adjusting the system parameters... a, b, c Or noise intensity, when the transition frequency of a particle is synchronized with the frequency of a weak signal s(t), the energy of the noise is transferred into the signal energy, thereby greatly enhancing the characteristic components in the output signal, achieving signal strength with noise.

[0047] In some embodiments of this application, the method further includes a step of optimizing the system parameters of the stochastic resonance system using a chaotic sparrow search algorithm before inputting the second partial discharge signal into the stochastic resonance system: The output signal-to-noise ratio of the stochastic resonance system is used as the fitness function to represent the system's output performance, as shown in the following formula: ; In the formula: The characteristic frequency of the system input signal. This is the sequence number corresponding to the characteristic frequency. This represents the energy at the characteristic frequency. For noise energy; The Chaotic Sparrow Search algorithm is used to find the optimal combination of system parameters that maximizes the output signal-to-noise ratio under the current input signal. , , as well as And this optimal parameter combination is applied to the stochastic resonance system.

[0048] Because stochastic resonance systems are extremely sensitive to parameters, fixed parameters cannot adapt to complex and ever-changing ambient noise environments. Therefore, before formally processing the signal, this embodiment utilizes an intelligent optimization algorithm to find the optimal operating point of the system. Specifically, this embodiment employs a chaotic sparrow search algorithm as the optimization tool. This algorithm is responsible for automatically searching in a multi-dimensional parameter space to find the parameter combination that best enables the output performance of the stochastic resonance system, i.e., the aforementioned... , , as well as To evaluate the quality of a set of parameters, the algorithm defines a fitness function. In this embodiment, the fitness function is the output signal-to-noise ratio (SNR) of the stochastic resonance system. The algorithm's goal is to maximize the SNR. out The value of SNR was compared with that of other parameters. A higher SNR indicates a higher efficiency in transferring noise energy and a better signal enhancement effect. The final parameter was selected to maximize SNR. out The set of parameters that reaches the maximum value is taken as the optimal system parameters and used for subsequent formal stochastic resonance signal enhancement processing.

[0049] The above-mentioned stochastic resonance signal enhancement process can also be found in [reference needed]. Figure 2 ,Depend on Figure 2As can be seen, the signal is first subjected to a time-domain scaling transformation. By defining a scaling compression coefficient k, the waveform of the high-frequency signal is stretched on the time axis, shifting its characteristic frequencies to the low-frequency band that the stochastic resonance system can respond to. Then, a chaotic sparrow search algorithm is used to automatically find the optimal structural parameters of the tristable stochastic resonance system. The algorithm starts running, randomly initializing one or more sets of system parameters (including potential well parameters a, b, c and damping ratio γ) in a multidimensional parameter space. The current parameter combination is substituted into the tristable stochastic resonance model (Langevin equation) to process the signal, and the output signal-to-noise ratio (SNR) of the processed signal is calculated. It is determined whether the output SNR is optimal. If the current SNR has not reached the convergence condition or is not the optimal value in the current iteration, the algorithm updates the parameter positions according to the sparrow search strategy, and then returns to the previous step to recalculate, forming an iterative optimization closed loop. When the algorithm meets the termination condition (such as reaching the maximum number of iterations or the SNR no longer improving), it exits the loop and locks the current set of parameters as the globally optimal parameters. Then, the parameters of the tristable stochastic resonance system are set as the optimal parameters. Finally, a tristable system with optimal parameters is used to process the scaled signal. At this point, the system exhibits strong random resonance, maximally transferring the energy of the background noise to the weak feature signal, resulting in a significantly enhanced signal sequence. This initial enhancement is complete, and the output signal is ready for the next stage (Transformer denoising).

[0050] In some embodiments of this application, step S104 above utilizes the multi-head self-attention mechanism in the model encoder to capture long-range temporal dependencies and global contextual information in the enhanced signal sequence, specifically including: Step S1: Construct matrix vectors , , Three weight matrices are created through three fully connected layers. , , For the input signal Linear mapping, forming new matrix vectors , , : ; ; ; Where Linear represents the operation of the fully connected layer. , and These are the bias vectors for different networks.

[0051] This step maps the input signal sequence to three different latent feature spaces: query, key, and value spaces, in preparation for subsequent calculations of the correlation between elements within the signal.

[0052] Step S2: , , Divide equally Group, , and Different heads are formed, and each head calculates an attention mechanism in a different space to obtain different features. The calculation process of the attention mechanism for a single head is as follows: ; in yes transpose, represent Dimensions and The dot product calculates the correlation between the two, and the softmax function will... and The matrix after dot product is mapped to the range of 0 to 1. The softmax calculation process is as follows: .

[0053] To allow the model to observe signals from different feature subspaces (e.g., one head focuses on the rising edge of the pulse, while another head focuses on the decaying oscillation of the pulse), the generated Q, K, and V matrices are divided into h groups along the feature dimension: , and .

[0054] In each individual head i, attention weights are calculated and summed using weighted averages. This is a crucial step in distinguishing noise from signal. .

[0055] in, Used to calculate the dot product of the query vector and the key vector, measuring the similarity or correlation between any two positions in a signal sequence; That is, divide by the dimension d k The root of the dot product prevents the Softmax function from entering the saturation region of minimal gradient due to excessively large dot product values. Through Softmax, the model assigns higher weights (close to 1) to the true discharge pulse features and lower weights (close to 0) to random noise. Finally, the calculated weights are used to adjust the value vector V. i Weighting is performed to extract key signal features.

[0056] Step S3: Multi-head attention extracts different feature information from different heads, and concatenates these features to obtain richer semantic features. The multi-head feature fusion process is as follows: ; in Matrix is ​​used to Features obtained from multiple attention mechanisms are fused together.

[0057] This step integrates the feature information extracted from the h heads in different subspaces to form a global feature representation containing rich semantics. Specifically, it integrates the calculation results of all heads [Attention1,…,Attention…]. h The data is concatenated along the feature dimension. This is achieved through a new linear projection layer (weight matrix W). O The concatenated vectors are transformed to output the final attention feature matrix X. Attention .

[0058] Through these three sub-steps, the Transformer model can deeply explore the hidden temporal patterns in partial discharge signals, effectively suppress background noise, and provide high-precision feature support for subsequent signal reconstruction.

[0059] In some embodiments of this application, the Transformer deep learning model described above further includes residual connections, layer normalization, and feedforward neural network modules, and the method further includes the following processing steps: Directly input signals of the multi-head self-attention mechanism The output signal after layer normalization and multi-head self-attention mechanism operation By performing matrix addition, the intermediate feature signal is obtained. : ; The intermediate feature signal Inputting the feedforward neural network module yields the hidden layer output. : ; Where Linear represents fully connected layer operations; The approximate calculation process for the GELU activation function is as follows: ; Where tanh is the hyperbolic tangent function.

[0060] This embodiment describes two key processing steps in step S104 that immediately follow the multi-head self-attention mechanism: Residual Connection and Feed-Forward Network (FFN). These two modules work together to address the training degradation problem of deep networks and further improve the nonlinear feature extraction capability.

[0061] In deep neural networks, gradients often vanish with increasing layers, making the model difficult to train. Residual connections introduce a shortcut, directly transmitting the original input signal to the output, allowing the network to learn only the residual (difference) between the input and output. This significantly alleviates the vanishing gradient problem and accelerates model convergence. This step transforms the original input signal X from the multi-head self-attention mechanism... Embedding The output signal X after layer normalization and multi-head self-attention mechanism operation Attention The matrix addition results in an intermediate feature signal that combines the original information with the attention features. It preserves the integrity of the original signal while superimposing the extracted key features.

[0062] The feedforward neural network module is mainly used for nonlinear transformation and dimension mapping of features. To further improve network performance and prevent overfitting, this embodiment introduces a Gaussian Error Linear Unit (GELU) activation function, fully connected layers, and dropout layers into the FFN. The intermediate feature signal X′ obtained above is then processed... Attention The input is fed into the feedforward neural network to obtain the final hidden layer output X. hidden The entire process involves two linear transformations and one nonlinear activation. By introducing GELU, the model can exhibit stronger fitting ability when dealing with the complex nonlinear characteristics of partial discharge signals.

[0063] In some embodiments of this application, the training process of the convolutional neural network is essentially a process of minimizing the model loss function. This application selects the commonly used mean squared error (MSE) as the loss function to optimize the model parameters.

[0064] In the formula: This represents the total number of training sample pairs. Represents the trainable parameters of the model. This indicates the residual mapping obtained by training the model through residual learning. This represents the L2 norm.

[0065] The above diagram of the transformer denoising network can also be found in [reference needed]. Figure 3 As shown, the network consists of a bottom input layer, a middle deep feature extraction module (the two dashed boxes represent the stacking or cascading of multiple Transformer Blocks), and a top output reconstruction layer. The bottom input box receives the preliminary enhanced signal sequence after the stochastic resonance processing in step S103. ⊕ represents the superposition of positional codes. Since the Transformer processes data in parallel, it does not have a natural sense of time sequence. Therefore, it is necessary to embed and add the positional information with the input signal so that the model can perceive the temporal order of the signal. Figure 3 It contains two cascaded modules with identical structures, representing a stack of multiple Transformer encoders. Each module contains the following sub-layers: Multi-head attention mechanism: This is the core of the network. It maps the input signal to different subspaces and captures long-distance temporal dependencies and global contextual information by calculating the correlation within the signal sequence (self-attention), thereby distinguishing noise from real impulses.

[0066] Residual connections and normalization: The bypass arrows in the diagram represent residual connections, used to solve the vanishing gradient problem in deep networks and preserve the details of the original signal. Layer normalization is used to stabilize the training process and accelerate model convergence. There is one of these modules after the multi-head attention mechanism and another after FFN.

[0067] FFN stands for Feedforward Neural Network. According to the documentation, this module consists of a fully connected layer + GELU activation function + Dropout + fully connected layer, used to perform non-linear transformations and dimensional mapping on features, enhancing the model's expressive power.

[0068] The top output layer consists of a fully connected layer. High-dimensional deep features extracted by multiple Transformer encoders are input into the fully connected layer. This layer acts as a decoder or mapper, restoring the abstract features to a one-dimensional time-series signal. The final output represents the reconstructed high signal-to-noise ratio partial discharge pulse signal with background noise removed.

[0069] like Figure 4 The diagram shown is a structural schematic of a partial discharge signal denoising device provided in an embodiment of this application. The device includes: Signal processing unit 410 is used to perform amplitude normalization and time-domain scaling transformation on the noisy first partial discharge signal to obtain the second partial discharge signal; The signal amplification unit 420 is used to input the second partial discharge signal into the random resonance system, and use the noise energy contained in the second partial discharge signal to induce the particles to transition between the tristable state traps, thereby realizing the resonant amplification of the discharge characteristic signal and obtaining the third partial discharge signal. The inverse transformation unit 430 is used to perform an inverse scale transformation on the third partial discharge signal to restore it to the original time scale, thereby obtaining a fourth partial discharge signal that retains the discharge characteristics and has a preliminary improved signal-to-noise ratio. The denoising unit 440 is used to input the fourth partial discharge signal into a pre-trained Transformer deep learning model, use the multi-head self-attention mechanism in the model encoder to capture long-distance temporal dependencies and global context information in the enhanced signal sequence, distinguish noise from real discharge pulse features, and reconstruct the signal through the model decoder to output a partial discharge pulse signal with a high signal-to-noise ratio.

[0070] In one embodiment of this application, the signal processing unit 410 performs amplitude normalization on the noisy first partial discharge signal, including: The following formula is used to analyze the noisy first partial discharge signal. Perform amplitude linear normalization: ; In the formula: for The normalization result, For the noisiest signal, max() represents the maximum value.

[0071] In one embodiment of this application, the specific steps of the signal processing unit 410 performing time-domain scaling transformation include: Define a scale compression factor k, and redefine the time axis of the amplitude-normalized signal as... At the same time, the actual sampling frequency of the signal fs Compress to fs ′= fs / k The waveform of the second partial discharge signal after time-domain scaling remains unchanged in the time domain, while the frequency band shifts to the low-frequency band that the stochastic resonance system can directly act upon in the frequency domain.

[0072] In one embodiment of this application, the model of the above-mentioned stochastic resonance system adopts a tristable stochastic resonance model, the dynamic behavior of which is described by the Langevin equation: ; In the formula: U(x) is the potential well function of the potential field where the Brownian particle is located. , a、b and c Here, γ is the system parameter, τ is the system damping ratio, δ(t) is the time delay, and δ(t) is the pulse signal.

[0073] In one embodiment of this application, the above-mentioned apparatus further includes: a parameter optimization unit, used to optimize the system parameters of the stochastic resonance system using a chaotic sparrow search algorithm before inputting the second partial discharge signal to the stochastic resonance system. The output signal-to-noise ratio of the stochastic resonance system is used as the fitness function to represent the system's output performance, as shown in the following formula: ; In the formula: The characteristic frequency of the system input signal. This is the sequence number corresponding to the characteristic frequency. This represents the energy at the characteristic frequency. For noise energy; The Chaotic Sparrow Search algorithm is used to find the optimal combination of system parameters that maximizes the output signal-to-noise ratio under the current input signal. , , as well as And this optimal parameter combination is applied to the stochastic resonance system.

[0074] In one embodiment of this application, the denoising unit 440 utilizes the multi-head self-attention mechanism in the model encoder to capture long-range temporal dependencies and global contextual information in the enhanced signal sequence, specifically including: Step S1: Construct matrix vectors , , Three weight matrices are created through three fully connected layers. , , For the input signal Linear mapping, forming new matrix vectors , , : ; ; ; Where Linear represents the operation of the fully connected layer. , and These are the bias vectors for different networks; Step S2: , , Divide equally Group, , and Different heads are formed, and each head calculates an attention mechanism in a different space to obtain different features. The calculation process of the attention mechanism for a single head is as follows: ; in yes transpose, represent Dimensions and The dot product calculates the correlation between the two, and the softmax function will... and The matrix after dot product is mapped to the range of 0 to 1. The softmax calculation process is as follows:

[0075] Step S3: Multi-head attention extracts different feature information from different heads, and concatenates these features to obtain richer semantic features. The multi-head feature fusion process is as follows: ; in Matrix is ​​used to Features obtained from multiple attention mechanisms are fused together.

[0076] In one embodiment of this application, the Transformer deep learning model further includes residual connections, layer normalization, and feedforward neural network modules, and the method further includes the following processing steps: Directly input signals of the multi-head self-attention mechanism The output signal after layer normalization and multi-head self-attention mechanism operation By performing matrix addition, the intermediate feature signal is obtained. : ; The intermediate feature signal Inputting the feedforward neural network module yields the hidden layer output. : ; Where Linear represents fully connected layer operations; The approximate calculation process for the GELU activation function is as follows: ; Where tanh is the hyperbolic tangent function.

[0077] The partial discharge signal denoising device proposed in this application introduces a tristable stochastic resonance system, utilizing noise energy to induce particle transitions to achieve resonant amplification of weak signals, providing a high signal-to-noise ratio input for subsequent processing. This effectively solves the problem of deep learning models struggling to converge in extremely low signal-to-noise ratio environments. Furthermore, it combines a Transformer deep learning model, utilizing the multi-head self-attention mechanism in its encoder to capture long-distance temporal dependencies and global contextual information in the signal sequence, overcoming the limitations of traditional convolutional neural networks in terms of limited receptive field and poor generalization ability. This architecture, combining physical mechanisms with data-driven approaches, not only achieves accurate extraction and high-fidelity waveform reconstruction of extremely weak partial discharge signals in strong noise backgrounds but also significantly enhances the system's adaptability and robustness to complex operating conditions, thereby greatly improving the accuracy of online monitoring of the insulation status of power equipment and its early fault warning capability.

[0078] Figure 5 This is a schematic diagram of the electronic device provided in the embodiments of this application. Figure 5 The illustrated electronic device is a general-purpose data processing apparatus, comprising a general-purpose computer hardware structure, including at least a processor 801 and a memory 802. The processor 801 and memory 802 are connected via a bus 803. The memory 802 is adapted to store one or more instructions or programs executable by the processor 801. These instructions or programs are executed by the processor 801 to implement the steps in the aforementioned method for denoising partial discharge signals.

[0079] The processor 801 described above can be a standalone microprocessor or a collection of one or more microprocessors. Thus, the processor 801 executes commands stored in the memory 802, thereby performing the method flow described in the embodiments of this application to process data and control other devices. The bus 803 connects the aforementioned components together, and also connects these components to the display controller 804, the display device, and the input / output (I / O) device 805. The input / output (I / O) device 805 can be a mouse, keyboard, modem, network interface, touch input device, motion-sensing input device, printer, and other devices known in the art. Typically, the input / output (I / O) device 805 is connected to the system via an input / output (I / O) controller 806.

[0080] The memory 802 can store software components, such as an operating system, a communication module, an interaction module, and application programs. Each of the modules and application programs described above corresponds to a set of executable program instructions that perform one or more functions and the methods described in the embodiments of the invention.

[0081] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method for denoising partial discharge signals.

[0082] The denoising method and apparatus for partial discharge signals proposed in this application introduce a tristable stochastic resonance system and utilizes noise energy to induce particle transitions to achieve resonant amplification of weak signals, providing a high signal-to-noise ratio input for subsequent processing. This effectively solves the problem of deep learning models struggling to converge in extremely low signal-to-noise ratio environments. Furthermore, by combining the Transformer deep learning model and utilizing its multi-head self-attention mechanism in the encoder to capture long-distance temporal dependencies and global contextual information in the signal sequence, it overcomes the limitations of traditional convolutional neural networks in terms of limited receptive field and poor generalization ability. This architecture, which combines physical mechanisms with data-driven approaches, not only achieves accurate extraction and high-fidelity waveform reconstruction of extremely weak partial discharge signals in strong noise environments but also significantly enhances the system's adaptability and robustness to complex operating conditions, thereby greatly improving the accuracy of online monitoring of the insulation status of power equipment and its early fault warning capability.

[0083] Preferred embodiments of this application have been described above with reference to the accompanying drawings. Many features and advantages of these embodiments are apparent from this detailed description, and therefore the claims are intended to cover all such features and advantages of these embodiments that fall within their true spirit and scope. Furthermore, since many modifications and alterations will readily occur to those skilled in the art, the embodiments of this application are not intended to be limited to the precise structures and operations illustrated and described, but rather to encompass all suitable modifications and equivalents falling within their scope.

[0084] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0085] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0086] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0087] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0088] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of this application. It should be understood that the above descriptions are merely specific embodiments of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for denoising partial discharge signals, characterized in that, The method includes: The amplitude is normalized and the time-domain scale is transformed on the noisy first partial discharge signal to obtain the second partial discharge signal; The second partial discharge signal is input into a stochastic resonance system. The noise energy contained in the second partial discharge signal is used to induce the particles to transition between the three-stable potential traps, thereby achieving resonant amplification of the discharge characteristic signal and obtaining the third partial discharge signal. The third partial discharge signal is inversely scaled to restore the original time scale, resulting in a fourth partial discharge signal that retains the discharge characteristics and has a preliminary improved signal-to-noise ratio. The fourth partial discharge signal is input into a pre-trained Transformer deep learning model. The multi-head self-attention mechanism in the model encoder is used to capture long-distance temporal dependencies and global context information in the enhanced signal sequence, distinguish noise from real discharge pulse features, and reconstruct the signal through the model decoder to output a partial discharge pulse signal with a high signal-to-noise ratio.

2. The method for denoising partial discharge signals as described in claim 1, characterized in that, The amplitude normalization of the noisy first partial discharge signal includes: The following formula is used to analyze the noisy first partial discharge signal. Perform amplitude linear normalization: ； In the formula: for The normalization result, For the noisiest signal, max() represents the maximum value.

3. The method for denoising partial discharge signals as described in claim 1, characterized in that, The specific steps of the time-domain scaling transformation include: Define a scale compression factor k, and redefine the time axis of the amplitude-normalized signal as... At the same time, the actual sampling frequency of the signal fs Compress to fs ′= fs / k The waveform of the second partial discharge signal after time-domain scaling remains unchanged in the time domain, while the frequency band shifts to the low-frequency band that the stochastic resonance system can directly act upon in the frequency domain.

4. The method for denoising partial discharge signals as described in claim 1, characterized in that, The stochastic resonance system is modeled using a tristable stochastic resonance model, and its dynamic behavior is described by the Langevin equations: ； In the formula: U(x) is the potential well function of the potential field where the Brownian particle is located. , a、b and c Here, γ is the system parameter, τ is the system damping ratio, δ(t) is the time delay, and δ(t) is the pulse signal.

5. The method for denoising partial discharge signals as described in claim 4, characterized in that, Before inputting the second partial discharge signal into the stochastic resonance system, the method further includes a step of optimizing the system parameters of the stochastic resonance system using a chaotic sparrow search algorithm. The output signal-to-noise ratio of the stochastic resonance system is used as the fitness function to represent the system's output performance, as shown in the following formula: ； In the formula: The characteristic frequency of the system input signal. The sequence number corresponding to the characteristic frequency. This represents the energy at the characteristic frequency. For noise energy; The Chaotic Sparrow Search algorithm is used to find the optimal combination of system parameters that maximizes the output signal-to-noise ratio under the current input signal. , , as well as And this optimal parameter combination is applied to the stochastic resonance system.

6. The method for denoising partial discharge signals as described in claim 1, characterized in that, The method of utilizing the multi-head self-attention mechanism in the model encoder to capture long-range temporal dependencies and global contextual information in the enhanced signal sequence specifically includes: Step S1: Construct matrix vectors , , Three weight matrices are created through three fully connected layers. , , For the input signal Linear mapping, forming new matrix vectors , , : ； ； ； Where Linear represents the operation of the fully connected layer. , and These are the bias vectors for different networks; Step S2: , , Divide equally Group, , and Different heads are formed, and each head calculates an attention mechanism in a different space to obtain different features. The calculation process of the attention mechanism for a single head is as follows: ； in yes transpose, represent Dimensions and The dot product calculates the correlation between the two, and the softmax function will... and The matrix after dot product is mapped to the range of 0 to 1. The softmax calculation process is as follows: Step S3: Multi-head attention extracts different feature information from different heads, and concatenates these features to obtain richer semantic features. The multi-head feature fusion process is as follows: ； in Matrix is ​​used to Features obtained from multiple attention mechanisms are fused together.

7. The method for denoising partial discharge signals as described in claim 6, characterized in that, The Transformer deep learning model further includes residual connections, layer normalization, and feedforward neural network modules, and the method further includes the following processing steps: Directly input signals of the multi-head self-attention mechanism The output signal after layer normalization and multi-head self-attention mechanism operation By performing matrix addition, the intermediate feature signal is obtained. : ； The intermediate feature signal Inputting the feedforward neural network module yields the hidden layer output. : ; Where Linear represents fully connected layer operations; The approximate calculation process for the GELU activation function is as follows: ； Where tanh is the hyperbolic tangent function.

8. A noise reduction device for partial discharge signals, characterized in that, The device includes: The signal processing unit is used to perform amplitude normalization and time-domain scaling transformation on the noisy first partial discharge signal to obtain the second partial discharge signal. The signal amplification unit is used to input the second partial discharge signal into the stochastic resonance system, and use the noise energy contained in the second partial discharge signal to induce the particles to transition between the tristable state traps, thereby realizing the resonant amplification of the discharge characteristic signal and obtaining the third partial discharge signal. The inverse transformation unit is used to perform an inverse-scale transformation on the third partial discharge signal to restore it to the original time scale, thereby obtaining a fourth partial discharge signal that retains the discharge characteristics and has a preliminary improved signal-to-noise ratio. The denoising unit is used to input the fourth partial discharge signal into the pre-trained Transformer deep learning model, use the multi-head self-attention mechanism in the model encoder to capture long-distance time dependence and global context information in the enhanced signal sequence, distinguish noise from real discharge pulse features, and reconstruct the signal through the model decoder to output a partial discharge pulse signal with a high signal-to-noise ratio.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-7.

Images