A method and system for partial discharge signal denoising based on time-frequency domain coordination

Abstract

This application relates to the technical field of partial discharge (PD) detection, and in particular to a PD signal denoising method and system based on time-frequency domain coordination. The method includes: acquiring a prediction model built on an encoder-decoder framework and training it using a composite loss function; acquiring the PD signal and performing a time-frequency transformation to obtain the time spectrum of the noisy signal, the time spectrum of the noisy signal including the clean PD signal and noise components in the time spectrum; inputting the time spectrum of the noisy signal into the trained prediction model and outputting a corresponding prediction signal mask matrix, and multiplying the prediction signal mask matrix element-wise with the time spectrum of the noisy signal to generate the time spectrum of the denoised signal; and performing an inverse time-frequency transformation on the time spectrum of the denoised signal to obtain the denoised time-domain signal. This application aims to improve the noise suppression effect of PD signals in noisy environments.

Description

Technical Field

[0001] This application relates to the technical field of partial discharge detection, and in particular to a method and system for partial discharge signal noise reduction based on time-frequency domain coordination. Background Technology

[0002] Partial discharge (PD) detection is an important means of assessing the insulation status of power equipment, and ultra-high frequency (UHF) detection methods are widely used in the monitoring of critical equipment such as GIS and transformers due to their advantages of strong anti-interference ability and high sensitivity. However, in actual operating environments, UHF signals are highly susceptible to various noise interferences, including background electromagnetic noise in substations, crosstalk of communication signals, and broadband noise caused by mechanical vibration of equipment, which leads to a significant reduction in the signal-to-noise ratio (SNR) and seriously affects the positioning accuracy of the PD source. Traditional noise reduction methods, such as wavelet thresholding and empirical mode decomposition (EMD), can suppress noise to a certain extent, but they often rely on manual parameter settings, making it difficult to adapt to complex and changing field environments. Moreover, in strong noise backgrounds, they are prone to distortion of effective signals or loss of edge information, affecting subsequent PD pulse extraction and time difference of origin (TDOA) calculation.

[0003] In recent years, noise reduction methods combining time-frequency analysis techniques and deep learning have shown considerable potential, with mask learning achieving significant results in speech enhancement. However, partial discharge (PD) UHF signals and speech signals differ fundamentally in their time-frequency characteristics: PD signals typically manifest as transient pulse groups with nanosecond-level durations in the time domain and exhibit non-uniform distribution characteristics in the frequency domain, while noise often covers the entire frequency band and has no significant time correlation.

[0004] If existing mask-based denoising algorithms are directly applied to partial discharge detection, they will fail to adequately consider the transient nature and bandwidth sparsity of the partial discharge signal, resulting in time delays or loss of frequency components during signal reconstruction.

[0005] In addition, the traditional STFT fixed window length design makes it difficult to balance time and frequency resolution, and is prone to time ambiguity in partial discharge pulse start point detection, which further restricts the positioning accuracy.

[0006] Currently, the field of partial discharge (PD) UHF signal processing still lacks an efficient noise reduction method that can adaptively distinguish the time-frequency distribution characteristics of signals and noise. Existing technologies either rely excessively on prior knowledge (such as noise template libraries) or have high computational complexity that makes them difficult to meet the needs of online monitoring. In particular, in scenarios with mixed multi-source PD and non-stationary noise, how to achieve robust noise suppression without destroying the phase information of the original signal remains a pressing technical challenge. Summary of the Invention

[0007] To improve the noise suppression effect of partial discharge signals in noisy environments, this application provides a partial discharge signal noise reduction method and system based on time-frequency domain coordination.

[0008] Firstly, this application provides a partial discharge signal noise reduction method based on time-frequency domain coordination, employing the following technical solution:

[0009] A partial discharge signal denoising method based on time-frequency domain coordination includes the following steps:

[0010] Obtain the prediction model built based on the encoder-decoder framework and train it using a composite loss function;

[0011] The partial discharge signal is acquired and time-frequency transformed to obtain the time spectrum of the noisy signal, wherein the time spectrum of the noisy signal includes the clean partial discharge signal and the noise component in the time spectrum;

[0012] The noisy signal time spectrum is input into the trained prediction model and the corresponding prediction signal mask matrix is ​​output. The prediction signal mask matrix is ​​then multiplied element-wise with the noisy signal time spectrum to generate the denoised signal time spectrum.

[0013] The time-frequency inverse transform of the denoised signal's time spectrum is performed to obtain the denoised time-domain signal.

[0014] In some embodiments, the construction of the prediction model includes:

[0015] An encoder is obtained by cascading several spatiotemporally separated convolutional modules. Each spatiotemporally separated convolutional module includes a sensing path. The output of the sensing path is connected to a 1×1 convolution and a multi-scale receptive field module. The spatiotemporally separated convolutional modules are downsampled by 2×2 average pooling.

[0016] The encoder end also includes a time-frequency feature reconstruction module, which is used to calculate vector correlation.

[0017] In some embodiments, the sensing path includes a frequency domain sensing path composed of 1×7 vertical convolutional kernels and a time domain sensing path composed of 7×1 horizontal convolutional kernels, wherein the frequency domain sensing path and the time domain sensing path run in parallel and their outputs are spliced ​​together in the channel dimension.

[0018] In some embodiments, the construction of the prediction model includes:

[0019] The decoder is composed of the same number of transposed convolutions. The transposed convolutions upsample the encoder end sequentially, and each transposed convolution is connected to the spatiotemporal separation convolution module of the corresponding level for splicing and reconstruction.

[0020] The transposed convolution also includes a feature alignment module, which is used to perform channel and spatial calibration on the features passed to the transposed convolution by each spatiotemporal separation convolution module through skip connections.

[0021] In some embodiments, the prediction model is trained using a composite loss function, including the following steps:

[0022] Obtain the training dataset;

[0023] The training noisy signal time spectrum in the training dataset is input into the prediction model and the training signal mask matrix is ​​obtained. The training noisy signal time spectrum and the training signal mask matrix are multiplied element by element to obtain the predicted denoised signal time spectrum.

[0024] Obtain the time spectrum of the target clean signal corresponding to the time spectrum of the training noisy signal in the training dataset;

[0025] The composite loss function is obtained based on the mean square error of the time spectrum of the predicted denoised signal and the time spectrum of the target clean signal, as well as the similarity loss term.

[0026] In some of these embodiments, the training dataset includes simulated noisy partial discharge signals, field noisy partial discharge signals, and mixed partial discharge noise superimposed from simulated standard partial discharge signals and real background noise collected in the field.

[0027] In some embodiments, the calculation of the similarity loss term includes the following steps:

[0028] The time spectrum of the predicted denoised signal and the time spectrum of the target clean signal are decomposed into several local windows, and the pixel values ​​of the time spectrum of the predicted denoised signal and the time spectrum of the target clean signal in the corresponding local windows are obtained.

[0029] Brightness similarity is calculated based on the mean of the pixel values ​​in the two windows, contrast similarity is calculated based on the standard deviation of the pixel values ​​in the two windows, and structural similarity is calculated based on the covariance of the pixel values ​​in the two windows.

[0030] The average similarity index is obtained by calculating the product of the brightness similarity, the contrast similarity, and the structural similarity, and combining the product with the number of local windows.

[0031] The similarity loss term is calculated based on the average similarity index.

[0032] In some embodiments, training the prediction model further includes the following steps:

[0033] The training noise mask matrix is ​​calculated based on the training signal mask matrix, and the time spectrum of the training noisy signal is multiplied element by element by the training noise mask matrix to obtain the amount of noise suppressed.

[0034] Calculate the total noise energy after denoising based on the amount of noise suppression;

[0035] The total energy before denoising is calculated based on the actual noise corresponding to the frequency spectrum of the training noisy signal.

[0036] The noise suppression ratio is calculated based on the total noise energy after denoising and the total noise energy before denoising, and the parameter optimization of the prediction model is determined based on the noise suppression ratio.

[0037] In some embodiments, both the predicted signal mask matrix and the training signal mask matrix correspond to soft masks, wherein the soft mask generation method is based on the following formula:

[0038] , ;

[0039] in, Represented as a signal soft mask matrix, Characterized as a noise soft mask matrix, Characterized as noise components, It is characterized as a clean partial discharge signal.

[0040] Secondly, this application provides a partial discharge signal noise reduction system based on time-frequency domain coordination, which adopts the following technical solution:

[0041] A partial discharge signal noise reduction system based on time-frequency domain coordination is provided to implement the above method.

[0042] The technical solutions provided by the embodiments of this application have the following technical effects:

[0043] By combining traditional time-frequency analysis techniques with modern deep learning methods, an efficient and robust method for partial discharge (PD) UHF signal denoising is proposed, solving the problem of extracting weak PD signals against strong noise backgrounds. Specifically:

[0044] First, time-frequency domain analysis can intuitively display the local characteristics of the signal, providing a clearer processing space for the separation of noise and partial discharge signals. Second, the adaptive learning capability of deep learning models enables them to handle various complex noise types, avoiding the tedious parameter adjustment process in traditional methods. Third, the mask prediction method has good interpretability, allowing users to intuitively understand the degree of noise suppression and signal preservation by observing the mask matrix. Finally, the entire processing flow has high computational efficiency, and after appropriate optimization, it can meet the real-time requirements of online monitoring. Attached Figure Description

[0045] Figure 1 This is a schematic diagram illustrating the steps of a partial discharge signal noise reduction method based on time-frequency domain coordination provided in this embodiment.

[0046] Figure 2 This is a schematic diagram of the overall framework of the prediction model in the embodiments of this application.

[0047] Figure 3 This is a schematic diagram of the spatiotemporal separation convolution module in an embodiment of this application.

[0048] Figure 4 This is a schematic diagram of the framework of the multi-scale receptive field module in the embodiments of this application.

[0049] Figure 5 This is a schematic diagram of the time-frequency feature reconstruction module in the embodiments of this application.

[0050] Figure 6 This is a schematic diagram of the feature alignment module provided in this embodiment. Detailed Implementation

[0051] To better understand the purpose, technical solutions, and advantages of this application, it has been described and illustrated below with reference to the accompanying drawings and embodiments. However, those skilled in the art should understand that this application can be implemented without these details. In some cases, to avoid obscuring various aspects of this application due to unnecessary description, well-known methods, processes, systems, components, and / or circuits already described at a higher level will not be elaborated upon. It will be apparent to those skilled in the art that various modifications can be made to the embodiments disclosed in this application, and the general principles defined in this application can be applied to other embodiments and application scenarios without departing from the principles and scope of this application. Therefore, this application is not limited to the illustrated embodiments, but conforms to the broadest scope consistent with the scope of protection claimed in this application.

[0052] It should be noted that the descriptions of these embodiments are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0053] In the description of this application, "several" means one or more, "more than" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0054] In the description of this application, the terms "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples.

[0055] like Figure 1 As shown in the figure, this application discloses a partial discharge signal noise reduction method based on time-frequency domain coordination, including the following steps:

[0056] S100: Obtain the prediction model built based on the encoder-decoder framework and train it using a composite loss function.

[0057] The network used in this application is a deep convolutional neural network based on an encoder-decoder framework, the core of which is a U-Net structure with numerous targeted improvements. The overall framework can be understood as a "shrinking (encoder) - expanding (decoder)" path, where the encoder gradually extracts and compresses features, while the decoder gradually restores spatial resolution and reconstructs details.

[0058] Meanwhile, compared with the original U-Net, the model network of this application has been innovatively improved and enhanced in every key component, making it very suitable for processing the special two-dimensional signal in the time-frequency domain, which will be disclosed later. The framework structure of the model network will be described in detail later.

[0059] Furthermore, the training of the neural network is a crucial step for the success of this method. In the design of the loss function, in addition to the commonly used Mean Square Error (MSE) loss, a composite loss function that integrates Time-Frequency Structural Similarity (TF-SSIM) is introduced to maintain the authenticity of the reconstructed signal in terms of time-frequency structure.

[0060] The specific details of model training will be explained later.

[0061] After the prediction model is built and trained, it can be used to denoise the partial discharge signal.

[0062] S200: Acquire the partial discharge signal and perform time-frequency transformation to obtain the time spectrum of the noisy signal. The time spectrum of the noisy signal includes the clean partial discharge signal and the noise component in the time spectrum.

[0063] Time-frequency domain representation is a method of analyzing signals simultaneously from both time and frequency dimensions, effectively revealing the dynamic characteristics of signals in these two directions. Converting time-domain UHF partial discharge signals (transient, difficult to distinguish directly from noise) into a two-dimensional "time-frequency" graph helps the model fully utilize the discriminative information in the frequency dimension when distinguishing partial discharge characteristics from noise interference, showing the time-frequency distribution differences between exposed signals and noise.

[0064] In this embodiment, short-time Fourier transform (STFT) is used to achieve time-frequency conversion. This method is particularly suitable for processing non-stationary, transient partial discharge signals. STFT divides the signal into segments on the time axis and applies a Fourier transform to each segment to obtain the signal's representation in the joint time-frequency domain. This process can clearly capture the spectral distribution and energy changes of the partial discharge signal at different times, providing an important basis for subsequent feature extraction and pattern recognition.

[0065] The mathematical expression for STFT is as follows:

[0066] .

[0067] in, This is the result of STFT, where t is the center time of the window function. It is the input signal. It is a window function. It is the kernel function of the Fourier transform. It refers to frequency.

[0068] Substituting the mathematical expression into the technical solution of this application, the final noisy partial discharge time spectrum can be calculated as follows: Meanwhile, in the time-frequency domain, This can be represented as a clean partial discharge signal. and noise components The linear superposition of can be expressed by the following relationship:

[0069] .

[0070] This model helps to effectively separate and distinguish the true discharge characteristics from interference components in partial discharge signals.

[0071] The purpose of inputting the time spectrum of the noisy signal into the prediction model for denoising is to analyze the noise component in the noisy partial discharge time spectrum as accurately as possible. The magnitude of the signal is adjusted and removed to make the final predicted partial discharge signal as close as possible to the original real signal.

[0072] S300 inputs the time spectrum of the noisy signal into the trained prediction model and outputs the corresponding prediction signal mask matrix. Then, it multiplies the prediction signal mask matrix element by element with the time spectrum of the noisy signal to generate the time spectrum of the denoised signal.

[0073] Since noise is typically distributed across all frequencies, while signals are concentrated in specific frequency regions, masks can be used to selectively preserve or suppress certain parts of the time-frequency representation.

[0074] Therefore, by inputting the noisy time spectrum into the trained prediction model, we can obtain the signal mask output by the prediction model, as well as the noise mask. The amplitude-time-frequency masking method is used, and the resulting mask can adaptively distinguish between noise components and the real partial discharge signal.

[0075] Furthermore, the prediction signal mask matrix in this application and the training signal mask matrix to be disclosed later both correspond to soft masks with values ​​continuously taking values ​​in the interval [0, 1], and are not hard masks with only two values, 0 and 1.

[0076] The method for generating the predicted signal mask matrix is ​​based on the following formula:

[0077] .

[0078] The noise mask is achieved using the following formula:

[0079] .

[0080] Mask prediction is essentially a pixel-level classification or regression problem. Neural networks learn from a large amount of labeled time-frequency data to establish a mapping from noisy time-frequency maps to ideal masks. The size of each mask is related to the input time-frequency domain signal. Consistent, among which Indicates a clean signal In noisy signals The proportion it accounts for.

[0081] Traditional masks use a two-endpoint decision method. If the mask of a time-frequency unit is 1, it is determined to be a signal and is retained. If its mask is 0, it is determined to be noise and needs to be completely removed.

[0082] This method can lead to the loss of signal edge information or result in residual noise.

[0083] The soft mask is actually a two-dimensional matrix, where the value of each element represents the probability or confidence that the time-frequency cell belongs to a clean partial discharge signal.

[0084] A value of 1 indicates that the time-frequency unit is a clean partial discharge signal and should be retained based on high confidence; if If the value is 0, it indicates that the time-frequency unit is very likely noise and should be suppressed with high confidence; if Within the 0-1 range, the time-frequency unit assumes the presence of a mixture of signal and noise, and performs partial preservation and partial suppression based on the size of the mask value.

[0085] The method for removing noise components from the time spectrum is to multiply the prediction signal mask matrix element-wise with the noisy signal time spectrum, as shown in the formula:

[0086] ⊙ .

[0087] in, When represented as the denoised signal's spectrum, The matrix represented by ⊙ is the actual predicted signal mask matrix output by the prediction model after the noisy signal's frequency spectrum is input to the prediction model. ⊙ represents element-wise multiplication.

[0088] The multiplication process is essentially a weighted retention process. For high-confidence signal units (mask value of 1), multiplication indicates that the signal is completely retained; for high-confidence noise units (mask value of 0), multiplication indicates that the signal is completely weakened and eliminated; for ambiguous mixed units (mask value of 0.5), multiplication indicates that half of the signal is retained and the other half of the noise is suppressed.

[0089] The above method can handle ambiguous regions where signal and noise are mixed, achieving "progressive" preservation and suppression, avoiding signal edge distortion caused by hard cutting. In particular, for transient pulses with steep rising edges, such as partial discharge signals, it can better preserve their physical characteristics. Since the mask itself is smooth, the spectrum of the denoised signal after multiplication is also smooth. The signal converted back to the time domain by inverse STFT will not produce ringing or artifacts.

[0090] S400 performs an inverse time-frequency transformation on the time spectrum of the denoised signal to obtain the denoised time-domain signal.

[0091] After obtaining the denoised clean time-domain spectrum through the prediction model, the denoised clean time-domain signal is obtained based on the short-time Fourier inverse transform.

[0092] By combining traditional time-frequency analysis techniques with modern deep learning methods, the above scheme proposes an efficient and robust method for partial discharge (PD) UHF signal denoising, solving the problem of extracting weak PD signals in strong noise backgrounds. Specifically:

[0093] First, time-frequency domain analysis can intuitively display the local characteristics of the signal, providing a clearer processing space for the separation of noise and partial discharge signals. Second, the adaptive learning capability of deep learning models enables them to handle various complex noise types, avoiding the tedious parameter adjustment process in traditional methods. Third, the mask prediction method has good interpretability, allowing users to intuitively understand the degree of noise suppression and signal preservation by observing the mask matrix. Finally, the entire processing flow has high computational efficiency, and after appropriate optimization, it can meet the real-time requirements of online monitoring.

[0094] like Figure 2 As shown, in some other embodiments, the construction of the prediction model includes:

[0095] S110 is an encoder obtained by cascading several spatiotemporally separated convolutional modules. The spatiotemporally separated convolutional module includes a perception path. The output of the perception path is connected to a 1×1 convolution and a multi-scale receptive field module. The spatiotemporally separated convolutional modules are downsampled through 2×2 average pooling.

[0096] The encoder consists of five cascaded spatio-temporal separable convolution blocks (ST-Conv blocks), as shown in the structure below. Figure 3 As shown.

[0097] Each spatiotemporal separation convolutional module performs downsampling once, progressively converting the input high-resolution temporal spectrum into a low-resolution, high-dimensional feature representation.

[0098] Each spatiotemporal separation convolutional module does not employ standard two-dimensional convolution, but instead uses two paths in parallel, including a frequency-domain-aware path and a temporal-domain-aware path:

[0099] Path 1 (Frequency Domain Sensing Path): Uses a 1×7 vertical convolution kernel. This kernel has a large receptive field on the frequency axis (vertical direction) and is specifically designed to capture long-range dependencies and global spectral patterns in the frequency dimension (e.g., the distribution of a narrowband interference across the entire frequency band).

[0100] Path 2 (Temporal Awareness Path): Uses a 7×1 horizontal convolutional kernel. This kernel has a large receptive field on the time axis (horizontal direction), specifically designed to capture dynamic changes and the temporal duration characteristics of impulses in the time dimension.

[0101] The receptive field is represented as the size of the region of the original input data (time-frequency map) corresponding to "one pixel on the feature map" in a convolutional neural network. Essentially, it is the "field of vision" that the model "sees" the original input. The larger the field of vision (the larger the receptive field), the more pixels / data points in the input can be focused on at the same time. The smaller the field of vision (the smaller the receptive field), the more local points can be focused on.

[0102] In path one, the convolution kernel size is "1 time-domain point × 7 frequency-domain points". Therefore, the receptive field covers 7 frequency units in the frequency domain (vertical direction) (such as 7 consecutive frequency points in the original time-frequency map, such as 300MHz, 400MHz...900MHz), but only 1 time unit in the time domain (horizontal direction). This means that a single frequency-domain feature point on the feature map is calculated from "7 adjacent frequency points at the same time" in the original time-frequency map.

[0103] In path two, the convolution kernel size is "7 time-domain points × 1 frequency-domain point". The receptive field covers 7 time units in the time domain (horizontally) (such as 7 consecutive time points in the original time-frequency plot, for example, 10ns, 11ns...16ns), and only covers 1 frequency unit in the frequency domain. A single time-domain feature point on the feature map is calculated from "7 adjacent time points of the same frequency" in the original time-frequency plot.

[0104] The long-range dependency in Path 1 is characterized as the correlation between "frequently distant frequency points" in the frequency domain. The distribution of partial discharge signals or noise in the frequency domain is often not an "isolated single frequency point", but a "correlated distribution across multiple long-distance frequency points". For example, the partial discharge signal of a certain GIS equipment may have energy peaks at 1.2GHz (high frequency band) and 0.8GHz (mid frequency band) at the same time due to the resonance characteristics of insulation defects. These two frequency points are far apart in the frequency domain (spanning 0.4GHz), but they are both characteristic frequency bands of partial discharge signals, and there is a "long-range dependency" between them (both are part of the effective signal).

[0105] By using a large receptive field in the frequency domain (covering 7 frequency points), this "correlation of distant frequency points" is captured, avoiding the situation where only adjacent frequency points can be seen due to a small receptive field, which may lead to the omission of "effective signal / noise correlation at a distance" and result in the accidental deletion of effective signals or the omission of noise during noise reduction.

[0106] The global spectrum pattern in Path 1 represents the overall distribution pattern of signal and noise across the entire frequency range, that is, the global pattern of "where the characteristic frequency band of the effective signal is and where the concentrated frequency band of noise is". Partial discharge UHF signals have the characteristics of "local concentration and non-uniformity" in the frequency domain (with clearly defined characteristic frequency bands), while noise may have "wideband coverage" (e.g., mechanical vibration noise covering 500MHz-2GHz) or "narrowband concentration" (e.g., communication crosstalk concentrated around 2GHz). This "distribution pattern of signal and noise across the entire frequency band" constitutes the global spectrum pattern.

[0107] By using a larger frequency domain receptive field, the model can "perceive a wider frequency range" and gradually learn this global spectrum pattern, rather than focusing on only a few local frequency points. This allows it to "globally identify" which frequency bands are signals and which are noise during noise reduction, avoiding the mistake of treating noise as a signal or vice versa due to a "local perspective".

[0108] In path two, the dynamic changes in the time dimension are characterized as the process of signal change over time. Partial discharge signals are "transient pulse groups" that are not stationary in time but have obvious dynamic changes; noise may be "stationary noise" (such as background electromagnetic noise, with small changes in intensity over time) or "sudden noise" (such as lightning interference, which appears suddenly and disappears quickly).

[0109] By using a large receptive field in the time domain (covering 7 time points), this "dynamic change over time" is captured, avoiding the situation where a small receptive field can only see a few time points and cannot fully identify the "dynamic process of the pulse from its appearance to its disappearance", which would lead to the "rising edge / falling edge of the pulse" being mistakenly deleted as noise during noise reduction.

[0110] The time duration characteristics of the pulse in path two are characterized by the "duration length" and "characteristics during the duration" of the partial discharge pulse in time.

[0111] The pulse of a partial discharge signal is not "a point in time", but "lasts for multiple nanoseconds" and has a fixed continuity characteristic within the duration. These "duration length + continuity characteristics" are the time continuity characteristics of the pulse.

[0112] The temporal receptive field is matched with the "nanosecond-level duration" of the partial discharge pulse, thereby capturing the pulse's temporal continuity characteristics. This avoids the situation where the receptive field is too small to cover only a part of the pulse (such as only the peak moment), thus failing to identify the "complete pulse continuity process." This can lead to the "rising edge / falling edge" of the pulse being misjudged as noise during noise reduction, or the "continuous pulse" being split into multiple isolated points, affecting subsequent partial discharge pulse extraction and TDOA localization.

[0113] The outputs of the two paths are concatenated along the channel dimension and then fused and dimensionality reduced by a 1×1 convolution. This design greatly reduces the number of parameters, lowers the risk of overfitting, and explicitly separates and enhances the network's ability to learn features in both the time and frequency dimensions.

[0114] At the end of each spatiotemporal separation convolutional module, a multi-scale receptive field module (MSRF) is also introduced, with the structure as follows: Figure 4 As shown.

[0115] It captures multi-scale contextual information by paralleling multiple atrous convolutions with different dilation rates (e.g., 1×1 conv, 3×3 dilation=2, 5×5 dilation=3), ensuring that the network can both focus on local subtle impulses and perceive large-scale noise backgrounds, effectively avoiding the loss of important features during downsampling.

[0116] In a regular convolution, the kernel elements are directly convolved with the adjacent pixels of the input feature map (without gaps), while dilated convolution adds "holes" (zero-padding gaps) between the kernel elements—expanding the "receptive field" (i.e. the range covering the input feature map) of the convolution without increasing the kernel size or the number of parameters.

[0117] The dilatation rate is characterized by the number of intervals between kernel elements in dilated convolution, and is represented by an integer, where...

[0118] When d=1: there is no gap, which is exactly the same as ordinary convolution, such as 1×1 conv;

[0119] When d=2, the convolution kernels are spaced one pixel apart, such as 3×3 with dilation=2.

[0120] The core steps of the multi-scale receptive field module in this application are as follows:

[0121] For the input time-spectral features, three independent dilated convolutions are launched simultaneously, with each branch processing information at different scales.

[0122] 1×1 ordinary convolution (hole rate = 1) has no holes, the effective receptive field is only 1×1, the number of parameters is very small, and the calculation is fast. It focuses on the "most subtle local features" - such as the nanosecond-level peak point of the partial discharge pulse, the signal energy change of a single pixel in the time-frequency graph, and avoids "blurring" the subtle pulse due to the excessively large receptive field.

[0123] 3×3 dilated convolution (dilation rate = 2), with kernel elements spaced 1 pixel apart, provides an effective receptive field of 5×5, covering a "medium-scale range"—capturing both the complete outline of the partial discharge pulse (such as the rising edge, peak, and falling edge of the pulse) and perceiving local noise around the pulse (such as small-scale background noise near the pulse), thus achieving "preliminary distinction between signal and local noise".

[0124] A 5×5 dilated convolution (dilation rate = 3) with kernel elements spaced 2 pixels apart provides an effective receptive field of 13×13, covering "large-scale global information"—specifically capturing broadband noise backgrounds where partial discharge signals are located (such as electromagnetic noise covering multiple frequency bands in substations, broadband interference from mechanical vibrations), avoiding "ignoring the global noise distribution pattern" by focusing only on local areas.

[0125] After parallel processing based on three convolutions, three feature maps of the same size but different numbers of channels are output, and then dimensionality is reduced by concatenating the channel dimensions and performing 1×1 convolution.

[0126] Each spatiotemporal separation convolutional module is further downsampled using 2×2 average pooling with a stride of 2 between its ends. Compared to max pooling, average pooling better preserves energy information in the time-frequency domain, which is crucial for quantitative analysis.

[0127] S111, the encoder end also includes a time-frequency feature reconstruction module, which is used to calculate vector correlation.

[0128] Furthermore, at the end of the encoder, the feature map is compressed to its lowest resolution. Here, a Time-Frequency Recalibration Module (TFRM) based on Multi-Head Self-Attention (MHSA) is introduced, with the following structure: Figure 5 As shown.

[0129] The multi-head self-attention mechanism learns which regions in the time-frequency domain are important (e.g., pulse start points, specific resonant frequency bands) and which are secondary (e.g., stationary noise regions) by calculating the correlation between these vectors, and then reweights the features according to their importance. This module enables the network to have global modeling capabilities, establishing correlations between distant time-frequency points (e.g., linking the start and end of a pulse), completely overcoming the locality limitation of convolution operations. It is key to the network's accurate locality locking of weak partial discharge pulses in strong noise.

[0130] The components and implementation process of the time-frequency feature reconstruction module are as follows:

[0131] The input terminal receives low-resolution time-frequency features from the encoder output; LayerNorm normalizes the input features to stabilize the training process and avoid gradient anomalies; Flatten(H,W,C)

[0132] The representation is as follows: the two-dimensional feature map (H=height, W=width, C=number of channels) is transformed into a one-dimensional vector to prepare for attention calculation; Multi-Head Attention is a multi-head image attention that calculates the correlation between time-frequency feature vectors and learns the weights of "important feature regions" and "minor feature regions"; Reshape(H,W,C) is represented as restoring the one-dimensional vector output by multi-head attention to a two-dimensional time-frequency feature map with the same size as the input feature map; Conv(1×1)+GeLU represents adjusting the channel dimension and performing nonlinear transformation on the two-dimensional time-frequency feature map through 1×1 convolution and GeLU activation function; finally, the output is passed through Output.

[0133] In other embodiments, the construction of the prediction model includes:

[0134] S120 uses the same number of transposed convolutions to form a decoder. The transposed convolutions sequentially upsample the encoder end, and each transposed convolution is connected to the corresponding spatiotemporal separation convolution module for splicing and reconstruction.

[0135] The decoder first upsamples the feature maps using the same number of transposed convolutions as the spatiotemporal separation convolution modules, doubling the resolution. The upsampled features are then skip-connected with features from the corresponding encoder layer. When the decoder upsamples using transposed convolutions, these skip connections directly pass the original high-resolution feature map from the encoder layer to the decoder. These high-resolution features from the encoder provide spatial detail lost during compression, ensuring accurate local reconstruction by the decoder.

[0136] S121, the transposed convolution also includes a feature alignment module, which is used to perform channel and spatial calibration on the features passed to the transposed convolution by each spatiotemporal separation convolution module through skip connections.

[0137] Because the features of the encoder and decoder may have semantic and scale differences, direct concatenation may not yield good results. Therefore, a Feature Alignment Module (FAM) is added to each skip connection, with the following structure: Figure 6 As shown.

[0138] It uses a lightweight set of convolutional layers to recalibrate the channels and spaces of the features passed from the encoder, ensuring seamless integration with the features of the decoder.

[0139] Specifically, the input feature map is first subjected to (AdaptiveAvgPool2D) average pooling, which compresses the spatial information of each channel into a single global feature value, resulting in a 1×1 vector. This vector is then passed through a 1×1 convolution and activation function in sequence.

[0140] Spatial features are refined using 1×1 channel-wise convolutions (Depthwise), connecting channels and spatial attention. The weights are mapped to the [0, 1] interval using an activation function. The final output is spatially calibrated and then concatenated with the feature map upsampled by the decoder to provide adaptive features for subsequent mask prediction.

[0141] The decoder's final output maps the number of channels to 1 using a 1×1 convolution, and then passes it through a sigmoid activation function to generate a soft mask matrix with a value range of [0,1]. This matrix has the same spectral size as the input, and each pixel value represents the probability that the corresponding time-frequency unit is a "real partial discharge signal". The advantage of the soft mask is that it provides more granular confidence information, rather than a hard judgment of either 0 or 1, making the signal reconstruction process smoother.

[0142] In other embodiments, the prediction model is trained using a composite loss function, including the following steps:

[0143] S130, Obtain the training dataset.

[0144] The training dataset contains simulated noisy partial discharge signals, field noisy partial discharge signals, and mixed partial discharge noise superimposed from simulated standard partial discharge signals and real background noise collected in the field.

[0145] To ensure the robustness of the model in complex noisy environments, the training data needs to cover various possible noise types and signal-to-noise ratio levels. This application employs a hybrid data generation strategy, which goes beyond simply using field-collected or laboratory-simulated data. Instead, it superimposes a standard partial discharge signal simulated in the laboratory with real background noise collected in the field, and further expands the dataset using data augmentation techniques (such as additive noise, time shift, frequency shift, amplitude scaling, etc.).

[0146] S131, input the time spectrum of the training noisy signal in the training dataset into the prediction model and obtain the training signal mask matrix, multiply the time spectrum of the training noisy signal and the training signal mask matrix element by element to obtain the time spectrum of the predicted denoised signal.

[0147] This step is the same as the subsequent process of obtaining the mask matrix through the prediction model and multiplying it with the time spectrum of the noisy signal to obtain the time spectrum of the denoised signal.

[0148] The only difference is that the model is trained continuously based on the training dataset during training, and subsequently trained using the actual detected partial discharge signals.

[0149] S132, obtain the target clean signal time spectrum corresponding to the noisy signal time spectrum in the training dataset.

[0150] S133, a composite loss function is obtained based on the mean square error of the time spectrum of the predicted denoised signal and the time spectrum of the target clean signal, as well as the similarity loss term.

[0151] The composite loss function compares the spectrum of the predicted denoised signal. Spectrum of the target clean signal To calculate, where the spectrum of the target clean signal is... It can be obtained directly from the training dataset, and its expression is:

[0152] .

[0153] in To determine the pointwise mean square error between the predicted time spectrum and the target time spectrum, To measure the structural similarity between two temporal spectra, The hyperparameters are used to balance the weights of the two components.

[0154] In other embodiments, the calculation of the similarity loss term includes the following steps:

[0155] S1331, decompose the time spectrum of the predicted denoised signal and the time spectrum of the target clean signal into several local windows, and obtain the pixel values ​​of the time spectrum of the predicted denoised signal and the time spectrum of the target clean signal in the corresponding local windows.

[0156] S1332 calculates brightness similarity based on the mean of pixel values ​​in two windows, calculates contrast similarity based on the standard deviation of pixel values ​​in two windows, and calculates structural similarity based on the covariance of pixel values ​​in two windows.

[0157] S1333 calculates the product of brightness similarity, contrast similarity, and structural similarity, and combines it with the number of local windows to calculate the average value to obtain the average similarity index.

[0158] S1334, calculates the similarity loss term based on the average similarity index.

[0159] The structural similarity loss term, by comparing the brightness, contrast, and structural information of local regions in time-frequency images, can better preserve the detailed features of partial discharge signals. Specifically, it is defined as follows:

[0160] .

[0161] in The time spectrum of the predicted denoised signal obtained by sliding window calculation. Spectrum of the target clean signal The average similarity index is used to comprehensively compare the brightness, contrast, and structural information of the spectra of two times within a local window.

[0162] The Mean Similarity Index (MSSIM) is the core of the time-frequency structural similarity (TF-SSIM) loss term calculation. Its calculation process is as follows:

[0163] First, predict the time spectrum of the denoised signal. Spectrum of the target clean signal The process is decomposed into multiple corresponding local windows; then, for each corresponding window, its brightness similarity, contrast similarity, and structural similarity are calculated; finally, the MSSIM value is the average of the similarity indices of all local windows. The similarity index of a single window is defined by the following formula:

[0164] ;

[0165] Where x and y represent the sources from and The corresponding local window.

[0166] Brightness comparison function The average of the pixel values ​​in the two windows The calculation is performed using the following formula:

[0167] ;

[0168] Contrast comparison function Standard values ​​of pixel values ​​from two windows The calculation is performed using the following formula:

[0169] ;

[0170] Structural comparison function Covariance of pixel values ​​in two windows The calculation is performed using the following formula:

[0171] ;

[0172] constants in the formula , , A tiny positive constant is included to prevent the denominator from being zero. The product of the three components—luminance, contrast, and structure—ultimately constitutes a similarity index that can comprehensively evaluate the two time-frequency matrices.

[0173] The core purpose of introducing the TF-SSIM loss term is to guide the neural network to not only minimize pixel-by-pixel numerical errors during training, but also to maintain the overall structure and visual fidelity of the time-frequency representation of the denoised signal. This loss function effectively constrains the structural features (such as pulse edges, oscillation modes, and energy distribution) of the reconstructed time-frequency spectrum in both the time and frequency domains to remain consistent with the real clean signal, thus overcoming the shortcomings of traditional mean square error loss functions, which tend to lead to overly smooth output results and loss of detail and texture. Its technical effect is a significant improvement in the integrity of the denoised partial discharge pulse waveform, the accuracy of time positioning, and the accuracy of subsequent quantitative analysis. Especially under extremely low signal-to-noise ratio conditions, its ability to recover weak discharge signals is significantly improved compared to a single mean square error loss function.

[0174] In other embodiments, training the prediction model further includes the following steps:

[0175] S140: Calculate the training noise mask matrix based on the training signal mask matrix, and multiply the time spectrum of the training noisy signal with the training noise mask matrix element by element to obtain the amount of suppressed noise.

[0176] S141, calculate the total noise energy after denoising based on the amount of noise suppressed.

[0177] S142, calculate the total noise energy before denoising based on the real noise corresponding to the spectrum of the training noisy signal.

[0178] S143, calculate the noise suppression ratio based on the total noise energy after denoising and the total noise energy before denoising, and determine whether to optimize the parameters of the prediction model based on the noise suppression ratio.

[0179] In this embodiment of the application, when training the prediction model, in addition to the composite loss function, the model parameters are further optimized by the noise suppression ratio.

[0180] The noise suppression ratio (NSR) is represented as a logarithmic expression of "noise energy before noise reduction / residual noise energy after noise reduction" and is used to determine whether noise has been effectively suppressed.

[0181] First, the corresponding training noise mask matrix is ​​obtained during training. and compared with the spectrum of the training noisy signal. Perform element-wise multiplication.

[0182] The final result of the multiplication is represented by the amount of noise that needs to be suppressed or removed. and to The absolute value is moduloed to obtain the noise amplitude in each "time-frequency unit". Since the amplitude is positively correlated with the energy, it can be approximately characterized as the noise energy intensity of the unit. The total noise energy after denoising is obtained by summing all time-frequency units.

[0183] Similarly, the actual noise is obtained from the spectrum of the training noisy signal in the training dataset and used as reference data before denoising. The amplitude is still calculated by taking the modulus and then accumulated to finally calculate the total noise energy before denoising.

[0184] The noise suppression ratio is calculated based on the following formula:

[0185] .

[0186] The larger the NSR, the more thoroughly the noise is suppressed. If, after noise reduction of a partial discharge signal, the energy in the "non-specific frequency domain" covered by the noise mask increases from... Down to If the NSR is 20dB, it means that the noise has been effectively suppressed.

[0187] Furthermore, different model optimization methods can be selected based on the different noise suppression ratios, for example:

[0188] Supplement target noise samples to generate dedicated training sets for different scenarios, ensuring that the model fully learns the time-frequency distribution characteristics of this type of noise;

[0189] Optimize the kernel size of spatiotemporal separable convolution (ST-Conv);

[0190] Adjust the void ratio combination of the multiscale receptive field (MSRF);

[0191] Strengthen the attention weights of Time-Frequency Feature Reconstruction (TFRM);

[0192] Adjust the weights of mean squared error or video structure similarity loss in the composite loss function to ensure that the model dynamically focuses on different parameter features for different situations.

[0193] The above methods are some of the schemes for model optimization based on the noise suppression ratio. Other schemes and the specific schemes to be selected can be chosen and adjusted based on the actual situation.

[0194] The implementation principle is as follows:

[0195] By combining traditional time-frequency analysis techniques with modern deep learning methods, an efficient and robust method for partial discharge (PD) UHF signal denoising is proposed, solving the problem of extracting weak PD signals against strong noise backgrounds. Specifically:

[0196] First, time-frequency domain analysis can intuitively display the local characteristics of the signal, providing a clearer processing space for the separation of noise and partial discharge signals. Second, the adaptive learning capability of deep learning models enables them to handle various complex noise types, avoiding the tedious parameter adjustment process in traditional methods. Third, the mask prediction method has good interpretability, allowing users to intuitively understand the degree of noise suppression and signal preservation by observing the mask matrix. Finally, the entire processing flow has high computational efficiency, and after appropriate optimization, it can meet the real-time requirements of online monitoring.

[0197] It should be understood that although the steps in the flowcharts in the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily performed in the order indicated by the arrows. Unless otherwise expressly stated herein, there is no strict order in which these steps are performed, and they may be performed in other orders.

[0198] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A partial discharge signal denoising method based on time-frequency domain coordination, characterized in that, Includes the following steps: Obtain the prediction model built based on the encoder-decoder framework and train it using a composite loss function; The partial discharge signal is acquired and time-frequency transformed to obtain the time spectrum of the noisy signal, wherein the time spectrum of the noisy signal includes the clean partial discharge signal and the noise component in the time spectrum; The noisy signal time spectrum is input into the trained prediction model and the corresponding prediction signal mask matrix is ​​output. The prediction signal mask matrix is ​​then multiplied element-wise with the noisy signal time spectrum to generate the denoised signal time spectrum. The construction of the prediction model includes: An encoder is obtained by cascading several spatiotemporally separated convolutional modules. Each spatiotemporally separated convolutional module includes a sensing path. The output of the sensing path is connected to a 1×1 convolution and a multi-scale receptive field module. The spatiotemporally separated convolutional modules are downsampled by 2×2 average pooling. The encoder end also includes a time-frequency feature reconstruction module, which is used to calculate vector correlation. The components and implementation process of the time-frequency feature reconstruction module are as follows: The input end receives low-resolution time-frequency features output from the encoder; the input features are normalized based on layer normalization; the two-dimensional feature map is transformed into a one-dimensional vector; the correlation between time-frequency feature vectors is calculated based on multi-head graph attention; the one-dimensional vector output by multi-head attention is restored to a two-dimensional time-frequency feature map with the same size as the input feature map; the two-dimensional time-frequency feature map is adjusted for channel dimension and nonlinearly transformed by 1×1 convolution and GeLU activation function; finally, the output is output. The time-frequency inverse transform of the denoised signal's time spectrum is performed to obtain the denoised time-domain signal.

2. The partial discharge signal denoising method based on time-frequency domain coordination according to claim 1, characterized in that, The sensing path includes a frequency domain sensing path composed of 1×7 vertical convolutional kernels and a time domain sensing path composed of 7×1 horizontal convolutional kernels. The frequency domain sensing path and the time domain sensing path run in parallel and their outputs are spliced ​​together in the channel dimension.

3. The partial discharge signal denoising method based on time-frequency domain coordination according to claim 1, characterized in that, The construction of the prediction model includes: The decoder is composed of the same number of transposed convolutions. The transposed convolutions upsample the encoder end sequentially, and each transposed convolution is connected to the spatiotemporal separation convolution module of the corresponding level for splicing and reconstruction. The transposed convolution also includes a feature alignment module, which is used to perform channel and spatial calibration on the features passed to the transposed convolution by each spatiotemporal separation convolution module through skip connections.

4. The partial discharge signal denoising method based on time-frequency domain coordination according to claim 1, characterized in that, The prediction model is trained using a composite loss function, including the following steps: Obtain the training dataset; The training noisy signal time spectrum in the training dataset is input into the prediction model and the training signal mask matrix is ​​obtained. The training noisy signal time spectrum and the training signal mask matrix are multiplied element by element to obtain the predicted denoised signal time spectrum. Obtain the time spectrum of the target clean signal corresponding to the time spectrum of the training noisy signal in the training dataset; The composite loss function is obtained based on the mean square error of the time spectrum of the predicted denoised signal and the time spectrum of the target clean signal, as well as the similarity loss term.

5. The partial discharge signal denoising method based on time-frequency domain coordination according to claim 4, characterized in that, The training dataset includes simulated noisy partial discharge signals, field noisy partial discharge signals, and mixed partial discharge noise superimposed from simulated standard partial discharge signals and real background noise collected on-site.

6. The partial discharge signal denoising method based on time-frequency domain coordination according to claim 4, characterized in that, The calculation of the similarity loss term includes the following steps: The time spectrum of the predicted denoised signal and the time spectrum of the target clean signal are decomposed into several local windows, and the pixel values ​​of the time spectrum of the predicted denoised signal and the time spectrum of the target clean signal in the corresponding local windows are obtained. Brightness similarity is calculated based on the mean of the pixel values ​​in the two windows, contrast similarity is calculated based on the standard deviation of the pixel values ​​in the two windows, and structural similarity is calculated based on the covariance of the pixel values ​​in the two windows. The average similarity index is obtained by calculating the product of the brightness similarity, the contrast similarity, and the structural similarity, and combining the product with the number of local windows. The similarity loss term is calculated based on the average similarity index.

7. The partial discharge signal denoising method based on time-frequency domain coordination according to claim 4, characterized in that, Training the prediction model further includes the following steps: The training noise mask matrix is ​​calculated based on the training signal mask matrix, and the time spectrum of the training noisy signal is multiplied element by element by the training noise mask matrix to obtain the amount of noise suppressed. Calculate the total noise energy after denoising based on the amount of noise suppression; Calculate the total noise energy before denoising based on the actual noise corresponding to the time spectrum of the training noisy signal; The noise suppression ratio is calculated based on the total noise energy after denoising and the total noise energy before denoising. Based on the noise suppression ratio, it is determined whether to optimize the parameters of the prediction model.

8. The partial discharge signal denoising method based on time-frequency domain coordination according to claim 4, characterized in that, The mask matrices all correspond to soft masks, wherein the soft mask generation method is based on the following formula: ， ； in, Represented as a signal soft mask matrix, Characterized as a noise soft mask matrix, Characterized as noise components, It is characterized as a clean partial discharge signal.

9. A partial discharge signal noise reduction system based on time-frequency domain coordination, characterized in that, Used to implement the method as described in any one of claims 1-8.

Images