Cable partial discharge signal noise separation mode identification method and system

Abstract

The invention discloses a cable partial discharge signal noise separation mode identification method and system, and particularly relates to the technical field of cable partial discharge detection, in a multi-cable stacking environment, partial discharge signals are synchronously collected through multiple types of sensors such as a high-frequency current sensor, an ultrasonic sensor and an electromagnetic antenna; performing band-pass filtering, wavelet denoising and multi-channel synchronous alignment on the acquired signals to suppress background noise and retain key features of partial discharge pulses; separating the mixed signals by adopting independent component analysis; calculating a decision coefficient based on an inter-channel correlation coefficient, a signal-noise power ratio and an amplitude dynamic range, and judging whether to introduce a denoising method based on deep learning; for the waveform after noise separation, executing an amplitude re-calibration process so as to correct amplitude scaling uncertainty, and extracting multi-dimensional time domain, frequency domain and phase features; and a support vector machine and other machine learning algorithms are combined to realize automatic identification of partial discharge types.

Description

Technical Field

[0001] This invention relates to the field of cable partial discharge detection technology, and more specifically, to a method and system for identifying the noise separation pattern of cable partial discharge signals. Background Technology

[0002] As a crucial transmission equipment in power systems, the insulation condition of cables directly affects the safe and stable operation of the power grid. Partial discharge is a typical early manifestation of cable insulation defects, and detecting partial discharge signals can enable early diagnosis of potential cable faults. However, cable operating environments are often complex, especially in chemical, power, and rail transportation scenarios, where multiple cables are often laid side-by-side in close proximity, leading to significant electromagnetic coupling effects during the propagation of partial discharge signals. Furthermore, the area surrounding cables is subject to substantial power frequency and harmonic interference, radio frequency noise, and system background noise. These factors cause the actual acquired partial discharge signals to be mixed with noise and crosstalk signals, resulting in a low signal-to-noise ratio and complex waveforms.

[0003] In existing technologies, high-frequency current sensors (HFCTs), ultrasonic sensors, or electromagnetic antennas are typically used to acquire partial discharge signals, and noise suppression is achieved through bandpass filtering, wavelet transform, and other techniques. However, in environments with multiple cables laid in parallel, partial discharge signals on different cables couple with each other through spatial electromagnetic fields or metal components, resulting in high correlation between the signals acquired from each channel. This causes traditional noise separation techniques, such as independent component analysis (ICA), to suffer from the failure of the independence assumption during signal separation, limiting the separation accuracy. Furthermore, blind source separation algorithms also have limitations when processing non-stationary, low signal-to-noise ratio signals, and the separation results may exhibit problems such as uncertain amplitude scaling, polarity reversal, and inconsistent signal arrangement order, affecting the accuracy of subsequent feature extraction and discharge type identification.

[0004] Therefore, there is an urgent need for a partial discharge signal noise separation method that can combine strong noise suppression capability and high separation accuracy in multi-cable stacking environments, and achieve accurate identification of partial discharge modes, so as to improve the reliability and intelligence level of cable insulation condition monitoring. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a method and system for identifying the noise separation mode of cable partial discharge signals to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for identifying noise separation patterns in cable partial discharge signals includes the following steps: The partial discharge signal of the cable is synchronously acquired by a high-frequency current sensor, an ultrasonic sensor, and an electromagnetic antenna; the acquired signals from each channel are subjected to bandpass filtering, wavelet denoising, and multi-channel synchronous alignment. Independent component analysis is used to separate noise from the preprocessed multi-channel signal; the decision coefficient is calculated by weighted summation based on the correlation coefficient between channels, the partial discharge event signal-to-noise power ratio and the amplitude dynamic range, to determine whether to introduce a deep learning-based denoising method. If the decision coefficient is greater than the system threshold Th, a deep learning-based denoising method is introduced to further separate noise from the preprocessed multi-channel signal using independent component analysis. For the noise-separated signal, if a deep learning-based denoising method is introduced, then K=1; otherwise, K=0. Based on the K value and the pulse shape stability index, it is determined whether an amplitude recalibration process needs to be introduced. If so, the amplitude scaling factor is calculated using a standard pulse injection experiment, and the signal amplitude is corrected. Multidimensional features are extracted from the separated and calibrated partial discharge pulse waveforms; based on the extracted multidimensional features, machine learning algorithms are used to automatically identify the type of partial discharge.

[0007] In a preferred embodiment, the inter-channel correlation coefficient is used to quantitatively assess the degree of residual coupling between multiple cables. First, the correlation coefficient between each channel is calculated. The calculation formula is as follows: ;in , These are the independent component waveforms obtained from ICA separation, where Cov is the covariance and σ is the standard deviation. If there are M channels, then the pairwise combinations contain... Combination number; calculate the correlation coefficient between all channels. Then take the average: i and j represent labels for different channels; This represents the correlation coefficient between channels.

[0008] In a preferred embodiment, the partial discharge event signal-to-noise power ratio is used to evaluate the clarity of the partial discharge pulse signal relative to residual noise; it is defined as: ; The average power within a short window near the pulse peak; SNR represents the average power of the waveform region outside the pulse, and SNR represents the signal-to-noise power ratio of the partial discharge event.

[0009] In a preferred embodiment, the amplitude dynamic range is used to describe the contrast between the partial discharge pulse and the background noise, and is defined as: ;in This represents the peak value of the partial discharge pulse. DR represents the dynamic range of the amplitude, where RMS is the background noise value.

[0010] In a preferred embodiment, the recalibration determination coefficient is calculated by weighted summation based on the K value and the pulse shape stability index. When the recalibration determination coefficient is greater than the system preset threshold Bh, an amplitude recalibration process needs to be introduced. When the recalibration determination coefficient is not greater than the system preset threshold Bh, an amplitude recalibration process does not need to be introduced.

[0011] In a preferred embodiment, the amplitude recalibration process is as follows: First, standard pulses with known amplitudes are injected simultaneously or sequentially into the multi-channel system; then, the standard pulses are acquired and separated; next, the separated amplitudes are compared with the injected amplitudes, and the amplitude scaling factor is calculated; finally, the amplitude of the partial discharge pulse waveform is corrected based on the scaling factor.

[0012] In a preferred embodiment, the feature extraction includes: time-domain feature extraction, frequency-domain feature extraction, and phase feature extraction.

[0013] In a preferred embodiment, the machine learning algorithm is a support vector machine.

[0014] In a preferred embodiment, the following modules are included: The signal acquisition module is used to synchronously acquire partial discharge signals from the cable and input the signals from each channel into the data acquisition system; The signal preprocessing module is used to preprocess the acquired multi-channel raw signals; The noise separation module is used to perform noise separation on the preprocessed multi-channel signal using independent component analysis; it calculates the decision coefficient based on the correlation coefficient between channels, the partial discharge event signal-to-noise power ratio, and the amplitude dynamic range to determine whether to introduce a deep learning-based denoising method. The amplitude recalibration module is used to correct for potential amplitude scaling uncertainties after noise separation; The feature extraction module is used to extract multidimensional time-domain, frequency-domain, and phase features of partial discharge signals; The pattern recognition module is used to automatically identify and classify different partial discharge types based on the extracted multidimensional features using a support vector machine.

[0015] In a preferred embodiment, the amplitude recalibration module includes a decision module, which is used to determine whether an amplitude recalibration process needs to be introduced based on the K value and the pulse shape stability index.

[0016] The technical effects and advantages of this invention are as follows: In the noise separation stage, this invention innovatively proposes a strategy combining adaptive separation and amplitude recalibration. First, it calculates the inter-channel correlation coefficient, the partial discharge event signal-to-noise power ratio, and the amplitude dynamic range. After normalizing these parameters, a decision coefficient is used to determine whether a deep learning-based denoising method needs to be introduced. When the decision coefficient exceeds a preset system threshold, a deep learning model such as a denoising autoencoder or convolutional neural network is automatically introduced. This leverages the advantages of deep learning models in processing highly nonlinear mixed signals to effectively extract partial discharge pulse characteristics. Even in environments with low signal-to-noise ratios or severe multi-cable coupling, good separation results can be achieved. This overcomes the problem of decreased separation accuracy in traditional blind source separation (ICA) when the assumption of statistical independence of source signals fails, significantly improving both noise suppression and separation accuracy.

[0017] Furthermore, this invention proposes an amplitude recalibration process to address the common problems in blind source separation, such as amplitude scaling uncertainty, signal polarity reversal, and inconsistent order of independent components. Specifically, this invention generates standard pulses with known amplitudes using a signal generator in a laboratory or field setting, and injects them simultaneously or sequentially into a multi-channel acquisition system. After noise separation processing, the peak amplitude of the acquired standard pulse waveform is measured, and this amplitude is compared with the actual amplitude at the time of injection to calculate the amplitude scaling factor. Using this amplitude scaling factor, the partial discharge pulse waveform is corrected, restoring the amplitude of the separated waveform to its true physical value. This process not only improves the physical accuracy of the separated waveform but also ensures the comparability and reliability of subsequent amplitude-sensitive feature extraction results.

[0018] Furthermore, the amplitude recalibration process also makes decisions based on whether to introduce deep learning denoising (i.e., the K-value) and pulse shape stability indicators. When ICA is used and the pulse waveform has large shape fluctuations, recalibration is more likely to be triggered to ensure the accuracy of subsequent analysis data. Through the above-mentioned integrated adaptive separation and amplitude recalibration techniques, this invention not only achieves high-precision separation of partial discharge signals in complex multi-cable stacking environments, but also effectively improves the physical authenticity of the detection results and the accuracy of quantitative analysis. Attached Figure Description

[0019] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is a flowchart illustrating the cable partial discharge signal noise separation pattern recognition method of the present invention. Figure 2 This is a schematic diagram of the amplitude recalibration process; Figure 3 This is a schematic diagram of the cable partial discharge signal noise separation pattern recognition system of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1: The cable partial discharge signal noise separation mode recognition method of the present invention, such as... Figure 1 As shown, it includes the following steps: First, signal acquisition is performed to ensure that all characteristic information of partial discharge can be fully captured in a multi-cable stacking environment, laying the foundation for subsequent signal processing. During implementation, considering the actual conditions of the cable laying site, high-frequency current sensors (HFCTs) are installed at both ends of each cable to be monitored or at the cable joint. These sensors are fixed to the grounding lead of the cable sheath or the lead-out end of the metal shielding layer using a clamping method to sense the transient high-frequency pulse current generated along the cable sheath or grounding circuit when partial discharge occurs. This pulse typically has a frequency range between hundreds of kilohertz and tens of megahertz. HFCTs can detect such fast transient signals with high sensitivity and retain key information such as amplitude, waveform, and occurrence time.

[0022] In addition, to enhance the spatial dimension of partial discharge signal capture and the ability to acquire multi-physical information, ultrasonic sensors or electromagnetic antennas are arranged along the axial direction on the outer surface of the cable or on the outer shell of cable accessories. The ultrasonic sensors, using a surface-mount installation, can detect weak mechanical vibrations and acoustic signals generated by ionization and air gap breakdown within the discharge channel during partial discharge. These ultrasonic signals have the characteristic of penetrating cable insulation and surrounding air, providing another dimension of perception for partial discharge events. The electromagnetic antennas are used to receive the ultra-high frequency electromagnetic waves radiated by partial discharge, and are particularly suitable for areas with relatively weak local metal shielding, such as cable terminals and joints. By receiving electromagnetic radiation pulses, they further enhance the detection sensitivity and location determination capabilities of partial discharge events.

[0023] To ensure precise time correlation between the data collected by each sensor, facilitating subsequent steps, a data acquisition system with high-precision synchronization is employed. This typically utilizes fiber optic distributed clock synchronization technology, connecting all sensors to a multi-channel high-speed acquisition device operating on the same clock reference. Each data stream is acquired at a high sampling rate (generally no less than 50 MS / s, preferably above 100 MS / s) to ensure complete capture of waveform details and amplitude information of partial discharge spikes. Simultaneously, to minimize external electromagnetic interference, the entire data acquisition system is enclosed in a shielded cabinet, and differential input is used to reduce the impact of power frequency interference and ground potential drift on sampling accuracy.

[0024] During the acquisition process, the system stores the signals from each channel in real time in a high-speed cache and marks the data blocks with a unified time tag to facilitate subsequent synchronization and amplitude correlation analysis among the multi-channel signals. Through this spatial collaborative arrangement of multiple types of sensors and strictly synchronized high-speed acquisition, the system can effectively acquire full waveform data, temporal characteristics, and multi-physics information of partial discharge signals in multi-cable stacking environments, providing a solid data foundation for subsequent steps.

[0025] After acquiring the raw signal, the obtained signal often contains both real partial discharge pulses and various noise components, including power frequency and its harmonics, external radio frequency interference, inherent system noise, and cross-interference caused by electromagnetic coupling between cables. Without effective preprocessing of these raw signals, the accuracy of subsequent noise separation, feature extraction, and discharge type identification will be severely limited. Therefore, a complete signal preprocessing procedure is established before the data enters the noise separation stage to minimize various interferences while preserving the effective information of the partial discharge pulses, ensuring the integrity of the signal's physical meaning and time alignment. The specific steps are as follows: First, in the digital signal processing stage, bandpass filtering is used to perform frequency domain screening on the acquired raw signals from each channel. Based on extensive field measurements and literature analysis, cable partial discharge pulses typically have energy concentrated between hundreds of kilohertz and tens of megahertz, while power frequency interference mostly occurs in the 50 Hz and its integer multiples of harmonics, and high-frequency radio frequency noise is distributed in the tens of megahertz and above frequency bands. Based on this, a digital bandpass filter with adjustable center frequency and bandwidth is designed to fully preserve the frequency band of interest (e.g., 0.5 MHz to 20 MHz) while effectively suppressing low-frequency and excessively high-frequency non-partial discharge components. During the filter design process, its amplitude-frequency response is experimentally verified to ensure that it does not attenuate or distort the partial discharge pulse waveform, especially the steep rising edge.

[0026] Secondly, considering the non-stationary and transient characteristics of partial discharge (PD) signals, wavelet transform technology is introduced to perform multi-scale analysis of the bandpass signal. Specifically, the preprocessed waveform is decomposed into wavelet coefficients at different scales. Utilizing the localization properties of wavelets in the time-frequency domain, high-amplitude coefficients generated by transient PD pulses are distinguished from low-amplitude random coefficients generated by background noise. Through threshold elimination, low-amplitude, non-abrupt wavelet coefficients are weakened or zeroed out, retaining only the significant energy concentration portion of the PD pulse at specific scales. The purified time waveform is then reconstructed through inverse wavelet transform. This method not only effectively preserves the sharp waveform and high-frequency components of the PD pulse but also avoids the spectral leakage and waveform tailing problems that traditional linear filtering suffers from when dealing with non-stationary signals.

[0027] Furthermore, considering that multiple synchronous signals will be collected in a multi-cable stacking environment, the sampling circuits, transmission lines, and sensor response characteristics of different channels may all cause time delay deviations on the order of microseconds or smaller. These minute time differences can cause errors in subsequent amplitude comparison, phase feature extraction, or discharge source location. Therefore, a synchronization alignment step is set up to accurately match the time reference of multi-channel data through calibration pulse injection experiments or algorithm correction. In implementation, standard test pulses can be injected simultaneously into each channel in the laboratory or on-site to measure the inherent delay between channels, and the time axis of each channel signal can be corrected during digital processing; alternatively, a cross-correlation algorithm can be used to automatically detect and align the delay of multiple signals when similar pulses occur. This step ensures that the waveforms recorded by the same partial discharge event in different channels are under a strictly consistent time reference throughout the entire analysis process, thereby guaranteeing the effectiveness and accuracy of subsequent blind source separation, deep denoising, and location algorithms.

[0028] Through the above-mentioned triple preprocessing of bandpass filtering, wavelet denoising and synchronization alignment, background noise and coupling interference can be suppressed to the maximum extent in complex multi-cable stacking environments, while retaining key physical characteristics such as amplitude, shape, timing and phase of partial discharge signals, laying a solid data foundation for subsequent noise separation, feature extraction and discharge pattern recognition.

[0029] Because multiple cables are laid side-by-side with small spacing, when a partial discharge occurs within any cable, the electromagnetic field not only propagates along the sheath-to-ground loop but also couples to adjacent cables through spatial or metallic components. This results in the signal received by the multi-channel acquisition system being a superposition of partial discharge pulses, coupled crosstalk signals, and various environmental noises. This superposition not only reduces the signal-to-noise ratio of a single cable partial discharge event but can also cause discharge signals from different cables to confuse each other, severely affecting subsequent defect localization and type identification. Therefore, to accurately separate partial discharge signals in a multi-cable stacked environment, two combined separation strategies were designed in the noise separation step: Independent Component Analysis (ICA) and a deep learning-based denoising method.

[0030] Firstly, blind source separation technology is represented by Independent Component Analysis (ICA). Its core idea is that although the original signals emitted by different cables or noise sources are mixed at the observation end, if these original signals are statistically independent, the mixed waveform can be deconstructed into several independent components by performing linear or matrix transformations on the observed signals, thereby extracting each source signal. In practice, all synchronously acquired multi-channel waveforms are combined into an observation matrix, which is then input into the ICA algorithm for demixing. Mathematically, ICA uses iterative optimization to ensure that the demixing matrix satisfies high-order statistical independence between the components, thus achieving source signal separation. In multi-cable environments, if the partial discharge signals of each cable are sufficiently different in time or amplitude characteristics, ICA can, to a certain extent, separate the partial discharge signal of each cable from the mixed waveform and effectively suppress spatial background noise. The advantage of this method is that it does not require pre-trained models, is an unsupervised algorithm, and can be quickly applied in the field. However, the limitation of ICA lies in its theoretical basis, which assumes that all source signals are independent of each other. In a multi-cable stacking environment, due to strong coupling between cables, partial discharge pulses may appear simultaneously on multiple cables and between sensors. This coupling will undermine the independence assumption, thereby reducing the separation accuracy.

[0031] In addition, the ICA separation results have problems with amplitude scaling, signal arrangement order and polarity reversal uncertainty. If amplitude calibration is not performed, it will directly affect the subsequent feature extraction and recognition that are sensitive to amplitude.

[0032] To overcome the limitations of ICA under highly coupled conditions, an alternative separation technique is proposed: a deep learning-based denoising method. Its basic principle is to train a deep neural network using a large number of samples, enabling the network to learn the nonlinear distinguishing features between partial discharge pulses and noise / crosstalk, thereby achieving automatic separation of the partial discharge signal. Typical implementations include using a denoising autoencoder (DAE) or a convolutional neural network (CNN).

[0033] In a denoising autoencoder, a noisy mixed signal is input into the encoder and compressed into a low-dimensional representation. The decoder then reconstructs the output to approximate a clean partial discharge waveform. During training, by minimizing the reconstruction error between the input and the desired net signal, the network learns to preserve the morphology and spectral characteristics of the partial discharge pulse while suppressing noise components. Convolutional neural networks, on the other hand, automatically extract local features in the time and frequency domains through multiple layers of convolutional filters, and are particularly adept at identifying the spike characteristics and specific waveform shapes of partial discharge pulses.

[0034] The biggest advantage of deep learning-based denoising methods is their ability to handle highly nonlinear signal aliasing. Even with coupling effects, they can gradually extract the unique patterns of partial discharge signals through network training, and the output signal amplitude no longer suffers from the scaling problem seen in ICA (Inductively Coupled Algorithm). However, their disadvantages include the need for a large amount of labeled data for training, resulting in a long training cycle, and the model's generalization ability depending on whether the training data can cover a sufficient range of field conditions. Furthermore, if the goal of minimizing the mean square error is excessively pursued during training, deep networks may exhibit waveform smoothing during denoising, weakening the pulse rising edge and affecting the subsequent extraction of physical parameters related to peak features.

[0035] The decision to introduce a deep learning-based denoising method for further denoising is determined based on the inter-channel correlation coefficient, the partial discharge event signal-to-noise power ratio, and the amplitude dynamic range. First, the inter-channel correlation coefficient, the partial discharge event signal-to-noise power ratio, and the amplitude dynamic range are normalized. Then, the decision coefficient is calculated using the following formula.

[0036] The specific formula is as follows: Where Comp represents the introduced decision coefficient; These are the weighting coefficients for inter-channel correlation coefficient, partial discharge event signal-to-noise power ratio, and amplitude dynamic range, respectively.

[0037] When Comp is greater than the system threshold Th (e.g., Th=0.6), a deep learning-based denoising method needs to be introduced. When Comp is not greater than the system threshold Th, a deep learning-based denoising method is not needed.

[0038] If introduced, let K=1; otherwise, let K=0.

[0039] The inter-channel correlation coefficient is used to quantitatively assess the degree of residual coupling between multiple cables. First, the correlation coefficient between each channel is calculated. The calculation formula is as follows: ;in , These are the independent component waveforms obtained from ICA separation, where Cov is the covariance and σ is the standard deviation. A channel refers to each independent measurement path used to independently acquire partial discharge signals from the cable; that is, the sampling line corresponding to each sensor and the input interface allocated to that sensor in the data acquisition system.

[0040] If there are M channels, then the total number of channels in each pair is... Combination number; calculate the correlation coefficient between all channels. Then take the average: i and j represent labels for different channels; This represents the correlation coefficient between channels.

[0041] If the separation results of different channels are highly correlated, it indicates the presence of strong coupling components in the signals of each channel. For example, if the same partial discharge event occurs simultaneously on multiple cables due to spatial coupling, the independence condition of the ICA assumption is violated, resulting in poor separation performance.

[0042] When the correlation coefficient between channels is high, it indicates that the signal is severely mixed. In this case, ICA may not be able to distinguish between the real independent signal and the coupled noise. In this case, it is more necessary to rely on deep learning for denoising and use the nonlinear ability of deep networks to identify partial discharge characteristics.

[0043] The partial discharge event signal-to-noise power ratio (SNR) is used to evaluate the clarity of the partial discharge pulse signal relative to residual noise. It is defined as: ; The average power within a short window near the pulse peak; This represents the average power in the waveform region outside the pulse.

[0044] If the energy of a partial discharge pulse is very low compared to the surrounding noise (low signal-to-noise ratio), blind source separation algorithms may easily filter out the partial discharge pulse as noise or fail to accurately separate the pulse details. Deep learning-based denoising methods can be trained to recognize partial discharge pulse waveforms even when submerged in noise, thereby improving the ability to extract partial discharge signals in low SNR environments.

[0045] The amplitude dynamic range (DR) is used to describe the contrast between the partial discharge pulse and the background noise, and is defined as follows: ;in This represents the peak value of the partial discharge pulse. This represents the RMS value of the background noise.

[0046] The smaller the amplitude dynamic range, the lower the contrast between the partial discharge pulse and the noise, making it difficult for the waveform to stand out from the noise. ICA is not sensitive to amplitude differences and easily fails to distinguish weak partial discharge pulses, while deep learning can identify partial discharge pulses even if the amplitude is not significant by learning the pulse shape, spectrum and other features.

[0047] After noise separation, a relatively clean waveform signal containing partial discharge information is obtained. However, due to different technical approaches used in the noise separation stage, the resulting waveform signals differ in amplitude accuracy, waveform detail fidelity, and frequency domain feature preservation, thus posing different technical requirements for subsequent feature extraction steps. Therefore, to ensure the accuracy and applicability of feature extraction, the feature extraction process needs to be adapted to different noise separation schemes.

[0048] While noise separation can effectively preserve the sharp waveform characteristics and high-frequency details of the partial discharge pulse, making it particularly suitable for extracting shape-sensitive features such as pulse rise edge, spike coefficient, and high-frequency energy, the separated output may suffer from amplitude scaling uncertainty, polarity reversal, and inconsistent order of independent components. Therefore, the K-value and pulse shape stability index are used to determine whether an amplitude recalibration process is necessary. The specific formula is as follows: FZ represents the coefficient of determination for recalibration; zb represents the pulse shape stability index. , These are the weighting coefficients for the K value and the pulse shape stability index, respectively.

[0049] When the recalibration determination coefficient is greater than the system preset threshold Bh (e.g., Bh=0.5), an amplitude recalibration process needs to be introduced. When the recalibration determination coefficient is not greater than the system preset threshold Bh, an amplitude recalibration process does not need to be introduced.

[0050] Simultaneously, polarity unification processing is required before feature extraction, adjusting all pulses to the same direction to avoid affecting subsequent phase feature analysis. Furthermore, by comparing the energy distribution, spectral characteristics, or phase distribution of each component, channel attribution is determined for the separated signals, ensuring that each feature vector correctly corresponds to a specific cable or sensor location. Only after the above processing can the extraction of time-domain, frequency-domain, and phase features continue in the aforementioned process.

[0051] The amplitude recalibration process is as follows: Figure 2 As shown, the specific steps are as follows: First, in the laboratory or on-site, a standard pulse signal with a known amplitude is generated using a signal generator. The amplitude, waveform characteristics, and spectral characteristics of the standard pulse are all calibrated to ensure the accuracy of its physical parameters.

[0052] Secondly, the standard pulse signal is injected simultaneously into each acquisition channel, or sequentially into each acquisition channel, and acquired through the deployed sensor system and data acquisition system. During the acquisition process, the position of the injected standard pulse signal in the time domain is clearly defined to correspond with the acquired data.

[0053] Next, the collected waveform data containing standard pulses is input into a noise separation algorithm for processing to obtain the separated standard pulse waveform, and its peak amplitude is measured.

[0054] Then, the measured amplitude after separation is compared with the known true amplitude at the time of injection, and the amplitude scaling factor S is calculated using the following formula: ; The actual amplitude of the standard pulse signal injected into the system, i.e., the amplitude value generated by the calibrated signal source; The peak amplitude of a standard pulse signal is measured in the separation result after it has been processed by a sensor system, a data acquisition system, and a noise separation algorithm. Finally, using the amplitude scaling factor S, the amplitude of all partial discharge pulse waveforms to be corrected is adjusted using the following formula: ; The amplitude of the partial discharge pulse waveform after amplitude recalibration correction at time t; t represents the time variable, used to describe the change of the pulse waveform over time.

[0055] This allows the amplitude of the separated partial discharge pulse waveform to be restored to the true physical quantity, ensuring the accuracy of subsequent feature extraction and quantitative analysis.

[0056] If deviations still exist after correction, the above injection and correction steps can be repeated until the amplitude error is controlled within the preset accuracy range.

[0057] The K value indicates whether a deep learning-based denoising method was introduced (K=1) or only ICA was used (K=0). If K=0, recalibration is more necessary due to amplitude scaling uncertainty. If a deep learning-based denoising method was introduced (K=1), although calibration may still be needed, the amplitude accuracy is usually higher than that of ICA, and recalibration is not always required.

[0058] The pulse shape stability index refers to the fluctuation of different pulse waveforms in shape characteristics (such as rising edge, peak shape, and waveform energy distribution). If the fluctuation is large, it indicates that the independent components of ICA are unstable or aliased. In this case, recalibration is even more necessary to ensure that the extracted amplitude characteristics are physically comparable.

[0059] After amplitude recalibration and polarity unification, the corrected partial discharge pulse waveform is further analyzed to extract multidimensional parameters characterizing the physical properties of partial discharge, which are used for discharge type identification, defect localization, and partial discharge intensity assessment. The feature extraction includes, but is not limited to, the following: Time-domain feature extraction: Extracting characteristic parameters of the pulse waveform in the time domain, such as the peak amplitude, rise time, duration, spike coefficient, pulse area, and pulse occurrence time interval, to describe the shape characteristics and discharge activity of the partial discharge pulse.

[0060] Frequency domain feature extraction: By using Fourier transform or power spectrum analysis, the spectral energy distribution, main frequency components, and frequency band energy ratio of the pulse signal are extracted to distinguish different types of partial discharge modes and identify external interference signals.

[0061] Phase feature extraction: Based on the power frequency cycle, the phase angle of the pulse occurrence time is calculated to obtain the phase distribution characteristics of partial discharge within the voltage cycle, which is used to identify different types of discharge mechanisms and assess the location of defects.

[0062] Through the above feature extraction, a set of feature vectors describing partial discharge events can be formed, providing input for subsequent pattern recognition, machine learning, or deep learning algorithms to achieve automatic classification and diagnosis of different types of partial discharge.

[0063] After feature extraction is completed, a set of feature vectors describing the multidimensional physical characteristics of partial discharge events is obtained. To achieve automatic identification of different types of partial discharge, various classification strategies based on machine learning and deep learning are designed in the pattern recognition and classification steps, one of which is Support Vector Machine (SVM).

[0064] Support Vector Machine (SVM) is a supervised learning algorithm that is particularly suitable for processing small, nonlinear, and overlapping datasets in high-dimensional space. Because of its good generalization ability and high classification accuracy in engineering, it is preferred for partial discharge type identification.

[0065] Example 2: The design of the cable partial discharge signal noise separation pattern recognition system of the present invention is based on the method in Example 1, specifically as follows: Figure 3 The following modules are shown: The signal acquisition module is used to synchronously acquire partial discharge signals of cables in a multi-cable stacking environment using multiple types of sensors such as high-frequency current sensors, ultrasonic sensors, and electromagnetic antennas, and input the signals from each channel into the data acquisition system; the data acquisition system has a high-precision clock synchronization function and supports multi-channel high-speed sampling and signal buffering.

[0066] The signal preprocessing module is used to preprocess the acquired multi-channel raw signals, including: The bandpass filter unit is used to filter out power frequency and high frequency noise, while retaining the frequency band where the partial discharge signal energy is concentrated. Wavelet denoising unit is used to preserve the sharp features of partial discharge pulses and suppress background noise; The synchronization alignment unit achieves precise matching of multi-channel signals under a time base through pulse injection experiments or algorithm correction.

[0067] The noise separation module is used to separate noise and partial discharge signals from the preprocessed signal. First, it uses independent component analysis (ICA) to quickly separate the mixed signal. Then, based on the inter-channel correlation coefficient, the partial discharge event signal-to-noise power ratio, and the amplitude dynamic range, it determines whether to introduce a deep learning-based denoising method to further separate the processed mixed signal. The amplitude recalibration module is used to correct for potential amplitude scaling uncertainties after noise separation.

[0068] The amplitude recalibration module includes a decision module, which is used to determine whether an amplitude recalibration process needs to be introduced based on the K value and the pulse shape stability index.

[0069] The feature extraction module is used to extract multidimensional features from the separated and calibrated partial discharge pulse waveforms.

[0070] The pattern recognition module is used to automatically identify and classify different partial discharge types based on the extracted feature vectors using a support vector machine.

[0071] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0072] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0073] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0074] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for identifying the noise separation mode of partial discharge signals in cables, characterized in that: Includes the following steps: The partial discharge signal of the cable is synchronously acquired by a high-frequency current sensor, an ultrasonic sensor, and an electromagnetic antenna; the acquired signals from each channel are subjected to bandpass filtering, wavelet denoising, and multi-channel synchronous alignment. Independent component analysis is used to separate noise from the preprocessed multi-channel signal; the decision coefficient is calculated by weighted summation based on the correlation coefficient between channels, the partial discharge event signal-to-noise power ratio and the amplitude dynamic range, to determine whether to introduce a deep learning-based denoising method. If the decision coefficient is greater than the system threshold Th, a deep learning-based denoising method is introduced to further separate noise from the preprocessed multi-channel signal using independent component analysis. For the noise-separated signal, if a deep learning-based denoising method is introduced, then K=1; otherwise, K=0. Based on the K value and the pulse shape stability index, it is determined whether an amplitude recalibration process needs to be introduced. If so, the amplitude scaling factor is calculated using a standard pulse injection experiment, and the signal amplitude is corrected. Multidimensional features are extracted from the separated and calibrated partial discharge pulse waveforms; Based on the extracted multidimensional features, machine learning algorithms are used to automatically identify the type of partial discharge.

2. The cable partial discharge signal noise separation pattern recognition method according to claim 1, characterized in that: The inter-channel correlation coefficient is used to quantitatively assess the degree of residual coupling between multiple cables. First, the correlation coefficient between each channel is calculated. The calculation formula is as follows: ;in , These are the independent component waveforms obtained from ICA separation, where Cov is the covariance and σ is the standard deviation. If there are M channels, then the pairwise combinations contain... Combination number; calculate the correlation coefficient between all channels. Then take the average: ; i and j represent labels for different channels; This represents the correlation coefficient between channels.

3. The cable partial discharge signal noise separation pattern recognition method according to claim 1, characterized in that: The partial discharge event signal-to-noise power ratio is used to evaluate the clarity of the partial discharge pulse signal relative to residual noise; it is defined as: ; The average power within a short window near the pulse peak; SNR represents the average power of the waveform region outside the pulse, and SNR represents the signal-to-noise power ratio of the partial discharge event.

4. The cable partial discharge signal noise separation mode recognition method according to claim 1, characterized in that: The amplitude dynamic range is used to describe the contrast between the partial discharge pulse and the background noise, and is defined as follows: ;in This represents the peak value of the partial discharge pulse. DR represents the dynamic range of the amplitude, where RMS is the background noise value.

5. The cable partial discharge signal noise separation mode recognition method according to claim 1, characterized in that: The recalibration determination coefficient is calculated by weighted summation based on the K value and the pulse shape stability index. When the recalibration determination coefficient is greater than the system preset threshold Bh, an amplitude recalibration process needs to be introduced. When the recalibration determination coefficient is not greater than the system preset threshold Bh, an amplitude recalibration process does not need to be introduced.

6. The cable partial discharge signal noise separation pattern recognition method according to claim 5, characterized in that: The amplitude recalibration process is as follows: First, standard pulses with known amplitudes are injected into the multi-channel system simultaneously or sequentially; then, the standard pulses are acquired and separated; next, the separated amplitude is compared with the injected amplitude, and the amplitude scaling factor is calculated. Finally, based on the scaling factor, the amplitude of the partial discharge pulse waveform is corrected.

7. The cable partial discharge signal noise separation pattern recognition method according to claim 1, characterized in that: The feature extraction includes: time-domain feature extraction, frequency-domain feature extraction, and phase feature extraction.

8. The cable partial discharge signal noise separation pattern recognition method according to claim 1, characterized in that: The machine learning algorithm is a support vector machine.

9. A cable partial discharge signal noise separation pattern recognition system, characterized in that, The identification system, based on the method described in any one of claims 1-8, includes the following modules: The signal acquisition module is used to synchronously acquire partial discharge signals from the cable and input the signals from each channel into the data acquisition system; The signal preprocessing module is used to preprocess the acquired multi-channel raw signals; The noise separation module is used to perform noise separation on the preprocessed multi-channel signal using independent component analysis; it calculates the decision coefficient based on the correlation coefficient between channels, the partial discharge event signal-to-noise power ratio, and the amplitude dynamic range to determine whether to introduce a deep learning-based denoising method. The amplitude recalibration module is used to correct for potential amplitude scaling uncertainties after noise separation; The feature extraction module is used to extract multidimensional time-domain, frequency-domain, and phase features of partial discharge signals; The pattern recognition module is used to automatically identify and classify different partial discharge types based on the extracted multidimensional features using a support vector machine.

10. The cable partial discharge signal noise separation pattern recognition system according to claim 9, characterized in that: The amplitude recalibration module includes a decision module, which is used to determine whether an amplitude recalibration process needs to be introduced based on the K value and the pulse shape stability index.

Images