A high-voltage cable multi-source heterogeneous sensing signal preprocessing method and system

Abstract

The present application relates to the technical field of power equipment monitoring data, in particular to a high-voltage cable multi-source heterogeneous sensing signal preprocessing method and system; the method comprises the following steps: S1: multiple sensors are arranged in the high-voltage cable monitoring scene, and based on the sensors, multi-source heterogeneous signals of the high-voltage cable are collected, time-space synchronization compensation is performed on the multi-source heterogeneous signals of the high-voltage cable, and time-aligned signals are obtained; S2: a wavelet-depth learning hybrid model is established by comprehensively using wavelet transform and a deep learning network, the time-aligned signals are denoised, and denoised time-domain signals are obtained; S3: the denoised time-domain signals are fed back to step S1 to obtain time-aligned signals as optimized signals, and based on the optimized signals, step S2 is executed for collaborative optimization iteration, and the preprocessing of the multi-source heterogeneous signals of the high-voltage cable is completed. The existing data resources are fully utilized, time alignment and signal denoising are performed, and collaborative optimization iteration is performed to avoid the data island effect.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring data analysis technology, specifically to a method and system for preprocessing multi-source heterogeneous sensor signals of high-voltage cables. Background Technology

[0002] High-voltage cables are the energy arteries of urban power grids, and their operational status directly affects power supply reliability. Currently, by deploying various sensors such as distributed fiber optic vibration sensors, wireless temperature sensors, and high-frequency current transformers, multi-dimensional information on cable vibration, temperature, partial discharge, and electromagnetic environment can be collected, forming a multi-modal monitoring system. However, in practical applications, the sampling start time, sampling frequency, data transmission path, and physical installation location of different sensors are all different, resulting in asynchronous temporal signals and spatial misalignment of signals. This creates a data silo effect, preventing collaborative analysis of multi-source data and wasting data resources.

[0003] Existing technologies typically employ simple global interpolation or hardware GPS (Global Positioning System) clock synchronization. However, global interpolation cannot handle transmission delay jitter, while hardware GPS clock synchronization is costly and suffers from signal limitations in environments such as tunnels and underground utility tunnels, resulting in poor adaptability. Furthermore, the cable field environment is complex, with electromagnetic interference, mechanical vibration noise, and other noises overlapping with the frequency bands of characteristic signals from actual cable partial discharge, abnormal temperature rise, and mechanical deformation. Using traditional denoising methods such as low-pass filtering will result in the loss of high-frequency fault characteristics, and wavelet denoising with fixed thresholds will be drastically less effective under non-stationary noise conditions. Summary of the Invention

[0004] To address the technical problems of spatiotemporal asynchrony of multi-source signals and the difficulty in simultaneously achieving noise suppression and feature preservation in existing technologies, the present invention aims to provide a preprocessing method for multi-source heterogeneous sensing signals from high-voltage cables. The specific technical solution adopted is as follows: Step S1: In the high-voltage cable monitoring scenario, various sensors are deployed, and multi-source heterogeneous signals of the high-voltage cable are collected based on the sensors. Spatiotemporal synchronization compensation is performed on the multi-source heterogeneous signals of the high-voltage cable to obtain time-aligned signals. Step S2: Establish a wavelet-deep learning hybrid model by combining wavelet transform and deep learning network, and perform denoising processing on the time-aligned signal to obtain the denoised time domain signal; Step S3: Feed the denoised time-domain signal back to step S1 to obtain a time-aligned signal as the optimized signal, and perform step S2 on the optimized signal to perform collaborative optimization iteration, thereby completing the preprocessing of the multi-source heterogeneous signal of the high-voltage cable.

[0005] Preferably, any one sensor is defined as the target sensor, wherein the sensing signal collected by the target sensor is the time reference signal, and the sensing signals collected by the remaining sensors are the signals to be aligned. The time reference signal and the signals to be aligned are integrated to form a multi-source heterogeneous signal for the high-voltage cable.

[0006] Preferably, step S1 includes: Step S11: Model the spatiotemporal offset between the signal to be aligned and the time reference signal as a state vector containing the absolute time delay and its rate of change, and establish the state equation; Step S12: Determine the cross-correlation function based on the signal to be aligned and the time reference signal, define the peak position in the cross-correlation function as the observed value, determine the observation noise covariance matrix of the real-time signal-to-noise ratio adjusted Kalman filter, and iteratively estimate the optimal spatiotemporal offset. Step S13: Resample and align the signal to be aligned based on the estimated optimal spatiotemporal offset to obtain a time-aligned signal.

[0007] Preferably, step S12 includes: Step S121: Determine the cross-correlation function based on the signal to be aligned and the time reference signal, and define the peak position in the cross-correlation function as the observed value; Step S122: Introduce observation noise, establish the observation equation in conjunction with the state equation, and determine the observation matrix; Step S123: Obtain the signal-to-noise ratio estimate based on the observation window corresponding to the cross-correlation function, and obtain the basic observation noise variance, then adjust the observation noise covariance matrix of the Kalman filter; Step S124: Iterate the state equation using standard Kalman filtering to obtain the predicted state estimation error covariance matrix. Then, use the cross-correlation function, observations, and observation matrix to correct the predicted state estimation error covariance matrix and obtain the estimated optimal spatiotemporal offset.

[0008] Preferably, in step S13, sinc interpolation is used to resample and align the signal to be aligned.

[0009] Preferably, step S2 includes: Step S21: Wavelet transform is used to decompose the time-aligned signal into wavelet packets to initially separate noise and useful signals, obtain multiple signals, and extract low-frequency approximation coefficients and high-frequency detail coefficients. Step S22: Introduce a deep learning network to optimize the high-frequency detail coefficients in the multi-channel signals; Step S23: Perform wavelet reconstruction on the optimized high-frequency detail coefficients and low-frequency approximation coefficients to obtain the denoised time-domain signal.

[0010] Preferably, step S22 includes: Step S221: Introduce a deep learning network and train it; Step S222: Standardize the high-frequency detail coefficients, stack them along the channel dimension to form a multidimensional tensor, input it into the trained deep learning network, and output the adjustment coefficient matrix corresponding to the high-frequency detail coefficients; Step S223: Integrate the high-frequency detail coefficients and adjustment coefficient matrix to complete the optimization of the high-frequency detail coefficients.

[0011] Preferably, step S221 specifically includes: The deep learning network is a one-dimensional convolutional neural network with an encoder-bottleneck-decoder architecture, and it is trained by minimizing a composite loss function.

[0012] Preferably, in step S3, the iteration termination condition for collaborative optimization iteration is reaching the preset maximum number of iterations, and / or the difference between the delay estimates of two adjacent iterations is less than a preset threshold.

[0013] To address the aforementioned problems, this invention also provides a high-voltage cable multi-source heterogeneous sensor signal preprocessing system, the system comprising: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus, and the processor calls logical instructions in the memory to execute the high-voltage cable multi-source heterogeneous sensor signal preprocessing method described in any of the preceding claims.

[0014] The present invention has the following beneficial effects: 1. Spatiotemporal synchronization compensation based on multi-source heterogeneous signals from high-voltage cables: Adaptive Kalman filtering dynamically tracks and compensates for random delays, achieving microsecond-level or even higher precision synchronization of multi-source heterogeneous signals from high-voltage cables, superior to fixed interpolation methods. Combining wavelet transform with deep learning networks leverages the multi-resolution analysis advantages of wavelets and learns complex noise patterns through data-driven approaches, suppressing strong noise while preserving the characteristics of weak fault impacts or abnormal waveforms to the greatest extent. Furthermore, the deep learning network can be trained to adapt to different lines and noise environments, avoiding the drawbacks of manually adjusting parameters required by traditional threshold methods. Finally, backtracking step S1 yields an optimized signal, and step S2 is performed on the optimized signal for collaborative optimization iteration, completing the preprocessing of multi-source heterogeneous signals from high-voltage cables. The output high-quality, highly synchronized multimodal data fully utilizes existing data resources, avoids data silos, and can be directly used for subsequent feature fusion and fault diagnosis by deep learning networks, significantly improving recognition accuracy and reliability.

[0015] 2. The high-voltage cable multi-source heterogeneous sensor signal preprocessing system provided by the present invention has the same beneficial effects as the high-voltage cable multi-source heterogeneous sensor signal preprocessing method provided by the present invention, and will not be described in detail here. Attached Figure Description

[0016] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 The flowchart illustrates the steps of a high-voltage cable multi-source heterogeneous sensor signal preprocessing method according to an embodiment of the present invention. Detailed Implementation

[0018] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a high-voltage cable multi-source heterogeneous sensing signal preprocessing method and system proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0020] The following description, in conjunction with the accompanying drawings, details the specific scheme of the high-voltage cable multi-source heterogeneous sensor signal preprocessing method and system provided by the present invention.

[0021] To better illustrate this, during the operation of high-voltage cables, due to the complex electromagnetic environment and diverse sensor equipment arrangements, multi-source heterogeneous sensor signals are generated. That is, throughout the monitoring process, because sensors based on different principles are used, such as detection devices for various physical quantities such as temperature, current, vibration, and partial discharge, their output data formats, sampling frequencies, and signal characteristics are significantly different. Furthermore, the actual operating conditions of the cable, such as load changes, external interference, and insulation aging, are different, resulting in diverse and complex sensor signals.

[0022] Directly analyzing heterogeneous signals after acquisition is prone to errors. Unprocessed heterogeneous signals often contain a large amount of redundant and erroneous information. Direct analysis not only increases the computational burden but may also affect the accuracy of judgment. Therefore, preprocessing multi-source heterogeneous sensor signals can improve signal data quality, eliminate noise interference, and unify data format to extract valuable feature information, laying a reliable foundation for subsequent signal analysis, condition assessment, and fault diagnosis.

[0023] Please see Figure 1 The diagram illustrates a flowchart of the steps in a high-voltage cable multi-source heterogeneous sensing signal preprocessing method provided in the first embodiment of the present invention, the method comprising: Step S1: In the high-voltage cable monitoring scenario, various sensors are deployed, and multi-source heterogeneous signals of the high-voltage cable are collected based on the sensors. Spatiotemporal synchronization compensation is performed on the multi-source heterogeneous signals of the high-voltage cable to obtain time-aligned signals. Step S2: Establish a wavelet-deep learning hybrid model by combining wavelet transform and deep learning network, and perform denoising processing on the time-aligned signal to obtain the denoised time domain signal; Step S3: Feed the denoised time-domain signal back to step S1 to obtain a time-aligned signal as the optimized signal, and perform step S2 on the optimized signal to perform collaborative optimization iteration, thereby completing the preprocessing of the multi-source heterogeneous signal of the high-voltage cable.

[0024] As an optional implementation method, in this embodiment, the high-voltage cable monitoring scenario is taken as a typical scenario, such as the high-voltage cable operation status monitoring scenario in urban underground high-voltage cable tunnels or cable trenches.

[0025] Furthermore, any sensor is defined as the target sensor, where the sensing signal collected by the target sensor is the time reference signal, and the sensing signals collected by the remaining sensors are the signals to be aligned. The time reference signal and the signals to be aligned are integrated to form a multi-source heterogeneous signal for the high-voltage cable.

[0026] To clarify, in a high-voltage cable monitoring scenario, it is assumed that the system includes three types of sensors: a high-frequency current transformer (HFCT), a distributed fiber optic vibration sensor (DVS), and a wireless temperature sensor. The HFCT is used to collect broadband electromagnetic signals, effectively capturing high-frequency components that traditional current transformers cannot cover. The DVS collects vibration signals, utilizing Rayleigh scattering, Brillouin scattering, or Raman scattering effects of light propagating in optical fibers. By analyzing the backscattered light signals, real-time monitoring of vibration information at any location along the fiber optic cable is achieved. The WTS collects temperature signals to promptly detect abnormal temperature signals and prevent power outages. In actual operation, the sampling frequency, initial phase, and transmission delay of the three signals are all different.

[0027] Specifically, in this embodiment, a high-frequency current transformer is defined as the target sensor, and the current signal collected by the target sensor is used as a time reference signal, denoted as . Its power frequency current signal is strictly synchronized with the power grid frequency, possessing globally consistent and stable time reference. In subsequent data analysis, the clarity of the time reference signal as a time reference is enhanced through preprocessing methods such as extracting the power frequency component and bandpass filtering. Therefore, the remaining vibration and temperature signals are signals to be aligned, and are respectively denoted as... and The three signals are integrated to form a multi-source heterogeneous signal for the high-voltage cable.

[0028] Further, step S1 includes: Step S11: Model the spatiotemporal offset between the signal to be aligned and the time reference signal as a state vector containing the absolute time delay and its rate of change, and establish the state equation.

[0029] Specifically, firstly, based on any signal in the signals to be aligned, denoted as the th... A signal to be aligned, at discrete time points Define its state vector, and the corresponding calculation formula is:

[0030] in, Represents at discrete time points Place, No. A state vector of a signal to be aligned; Represents at discrete time points Place, No. Estimated absolute time delay (in seconds) of each signal to be aligned; This represents the rate of change of time delay.

[0031] What can be explained is the rate of change of time delay. The delay rate, or latency, reflects the dynamic characteristics of delay over time during signal transmission. It is used to model delay jitter or trend changes, such as significant changes in the rate of change of delay under scenarios like sudden load increases, signal switching, or link quality fluctuations.

[0032] Next, the state equations are established, and the corresponding calculation formulas are as follows:

[0033] in, , They represent the discrete time points respectively. and Place, No. The state of a signal to be aligned; Represents the state transition matrix; Represents at discrete time points Process noise at the point.

[0034] It can be explained that the state transition matrix , This represents the estimated period of the Kalman filter, and is not equal to the signal sampling period; process noise. This refers to a value with a mean of zero and a covariance matrix of... , which represents the delay mutation not considered in the Kalman filtering process. In the mathematical model of Kalman filtering, the evolution of the state is usually modeled as a linear dynamic system. That is, the current state and the state at the next adjacent discrete time point are linearly transformed through the state transition matrix, and process noise is superimposed. The auxiliary covariance matrix describes the energy magnitude of the process noise and the correlation between the components.

[0035] Step S12: Determine the cross-correlation function based on the signal to be aligned and the time reference signal, define the peak position in the cross-correlation function as the observed value, determine the observation noise covariance matrix of the real-time signal-to-noise ratio adjusted Kalman filter, and iteratively estimate the optimal spatiotemporal offset.

[0036] To clarify, in this step, adaptive Kalman filtering estimation, i.e., dynamic and optimal estimation of the state vector, is performed. Traditional Kalman filtering requires that the covariance matrices of process noise and observation noise be known and constant. However, in practice, noise is difficult to be accurate and is usually transformed. Therefore, an adaptive mechanism is used to adjust the state vector.

[0037] Further, step S12 includes: Step S121: Determine the cross-correlation function based on the signal to be aligned and the time reference signal, and define the peak position in the cross-correlation function as the observed value.

[0038] Specifically, the estimation period is defined by defining discrete time points, denoted as . Extract lengths from the time reference signal and the signal to be aligned, respectively. For the signal segments within the time window, determine the normalized cross-correlation function between the two signals. The corresponding calculation formula is:

[0039] in, Represents the time reference signal and the first Normalized cross-correlation function of the signals to be aligned; Indicates based on estimated period The time index corresponding to the extracted signal segment; This indicates the current time point within the signal segment of the time reference signal. The corresponding signal data; Indicates the first The current time point in the signal segment of the signal to be aligned The corresponding signal data; Indicates the length of the signal segment; Indicates the time point index within the signal segment.

[0040] By finding the cross-correlation function The peak position in the middle is denoted as Multiply it by the sampling interval to determine the original time delay observation value for the current observation, i.e. .

[0041] Step S122: Introduce observation noise, establish the observation equation in conjunction with the state equation, and determine the observation matrix.

[0042] Specifically, it is still based on discrete time points within the current estimation period. The analysis was performed, and the observed noise was denoted as... Its variance is denoted as That is, the observation noise covariance matrix. The observation equation is established by combining the state equation, and the corresponding calculation formula is:

[0043] in, Represents at discrete time points Place, No. Observations of the signal to be aligned; Represents the observation matrix; Represents at discrete time points Place, No. The state of a signal to be aligned; Represents at discrete time points Observation noise at the location.

[0044] It can be explained that the observation matrix It directly reflects the degree of influence of the parameters on the observed values, representing the directly observed delay, i.e. .

[0045] Step S123: Obtain the signal-to-noise ratio estimate based on the observation window corresponding to the cross-correlation function, and obtain the basic observation noise variance, then adjust the observation noise covariance matrix of the Kalman filter.

[0046] Specifically, based on the current observation window corresponding to the observation noise, i.e. based on the estimated period. The captured signal segment is used to determine the first... The estimated local signal-to-noise ratio of the signals to be aligned is denoted as . It is a cross-correlation function Mid-peak Compared with the average of the second peak The ratio, expressed in decibels; and the basic observation noise variance, i.e., by repeatedly observing the same physical quantity, the noise variance is determined by calculating the average of the squared deviations between the observed values ​​and the average value; the observation noise covariance matrix is ​​adjusted to the Kalman filter, and the corresponding calculation formula is:

[0047] in, Represents at discrete time points At this point, the adjusted observation noise covariance matrix; This represents the variance of the basic observation noise; Indicates An exponential function with base 0; Indicates the regulating factor; Represents at discrete time points Place, No. The local signal-to-noise ratio estimate of the signal to be aligned.

[0048] It can be noted that, in this embodiment, the adjustment factor It can be adjusted according to the actual situation; when the local signal-to-noise ratio estimate is... A high value indicates good observation quality; in this case, the adjusted observation noise covariance matrix... Decreasing the Kalman filter increases its gain, indicating greater confidence in the current observations. This enables rapid tracking; conversely, it reduces the estimated signal-to-noise ratio for local signals. A low value indicates potential strong noise interference and poor observation quality. In this case, the adjusted observation noise covariance matrix... As the filter gain increases, it decreases and becomes more dependent on the prediction results of the state equation, exhibiting stronger smoothing and anti-interference capabilities.

[0049] Step S124: Iterate the state equation using standard Kalman filtering to obtain the predicted state estimation error covariance matrix. Then, use the cross-correlation function, observations, and observation matrix to correct the predicted state estimation error covariance matrix and obtain the estimated optimal spatiotemporal offset.

[0050] Specifically, the standard Kalman filter formula is used for iteration to obtain the predicted state estimation error covariance matrix, and the corresponding calculation formula is as follows:

[0051]

[0052] in, , They represent the discrete time points respectively. and Place, No. The predicted state of a signal to be aligned; Represents the state transition matrix; Represents at discrete time points The covariance matrix of the predicted state estimation error at the location; This represents the covariance matrix corresponding to the process noise.

[0053] Next, observations are made based on the cross-correlation function to determine the predicted time-delay observations, i.e. Then, by combining the observed values ​​and the observation matrix, the predicted state estimation error covariance matrix is ​​corrected, i.e., the predicted state estimation error covariance matrix is ​​updated. The corresponding calculation formula is:

[0054]

[0055]

[0056] in, Represents at discrete time points The state estimation error covariance matrix at the location; Represents at discrete time points Kalman gain at the location; Represents the observation matrix; Represents at discrete time points At this point, the adjusted observation noise covariance matrix; Represents at discrete time points Place, No. The predicted state vector of the signal to be aligned; Represents the identity matrix.

[0057] Finally, after the current iteration is completed, the estimated optimal spatiotemporal offset is output, i.e. After several iterations, the Kalman filter converges to obtain the optimal delay estimation sequence; based on the estimated period... Similarly, determine the optimal delay estimation sequence for all signal segments corresponding to the estimated periods in the time base signal and the signal to be aligned.

[0058] Step S13: Resample and align the signal to be aligned based on the estimated optimal spatiotemporal offset to obtain a time-aligned signal.

[0059] As explained, based on all the optimal delay estimation sequences obtained in step S12, the average delay estimation value of the last stable segment is selected as the optimal spatiotemporal offset, denoted as . This is done to perform resampling alignment; where the stable segment refers to a continuous data segment in the delay estimation sequence that, after initial fluctuations, has a stable delay value, a small change amplitude, and a long duration.

[0060] Furthermore, in step S13, sinc interpolation is used to resample and align the signals to be aligned; that is, chord interpolation is used for analysis, which has ideal low-pass filtering characteristics and can theoretically achieve distortion-free signal reconstruction. In other words, when faced with two or more signals to be aligned that have time offsets or frequency differences, the sinc interpolation method can accurately resample one or more of the signals to eliminate time deviations and achieve signal alignment on the time axis.

[0061] Specifically, for each original sampling point in the signal to be aligned, i.e. Its timescale is: , Indicates the signal to be aligned The original sampling rate is used to determine the time marker for the acquisition of the original sampling points; then the estimated optimal spatiotemporal offset is used. The corrected and aligned timescale is calculated using the following formula:

[0062] in, This indicates the time stamp of the signal after alignment.

[0063] It can be explained that, due to the time scale Typically not aligned with the time axis of the time base signal, resampling is performed using sinc interpolation to generate a time base signal. Data on the same equally spaced time grid, i.e. The corresponding calculation formula is:

[0064] in, Indicates time reference signal Aligned A time-aligned signal; Indicates the first One signal to be aligned; Indicates the first The index of the original sampling points in the signal to be aligned; Indicates the signal to be aligned The original sampling rate; A time scale representing a time reference signal; This indicates an interpolation operation.

[0065] It can be explained that the time scale of the time base signal , Indicates the sampling rate of the time reference signal; through Interpolation is used to minimize spectral aliasing introduced by resampling and to ensure the integrity of the high-frequency characteristics of the signal.

[0066] By using adaptive Kalman filtering for synchronization compensation, the signal to be aligned is aligned with the time reference signal on the time axis. This reduces the asynchronous error from the original millisecond level to the microsecond or even nanosecond level and effectively suppresses estimation abrupt changes caused by network jitter or transient interference, providing a strictly synchronized data foundation for subsequent joint analysis.

[0067] Further, step S2 includes: Step S21: Wavelet transform is used to decompose the time-aligned signal into wavelet packets to initially separate noise and useful signals, obtain multiple signals, and extract low-frequency approximation coefficients and high-frequency detail coefficients.

[0068] As explained, wavelet transform is used to analyze the time-frequency characteristics of signals or data. By using wavelet basis functions such as Haar wavelets, Daubechies wavelets, and Morlet wavelets with finite durations, it can perform localized analysis of signals at different scales (corresponding to frequencies) and locations (corresponding to times), achieving precise temporal localization of transient features, abrupt changes, or different frequency components in the signal. Specifically, the signal after spatiotemporal alignment in step S1 undergoes deep denoising to map it to a more suitable time-frequency domain for feature separation using wavelet transform. The time-aligned signal obtained after step S1 is denoted as: Furthermore, all signals have been standardized to the same sampling rate. .

[0069] Specifically, in this embodiment, the Daubechies 4 (db4) wavelet is selected as the mother wavelet, which has moderate tight support and smoothness, and can achieve a good balance between time resolution and frequency resolution when analyzing non-stationary cable vibration and electromagnetic pulse signals; that is, for each signal... A five-level wavelet packet decomposition is performed, which differs from standard wavelet decomposition. Wavelet packet decomposition further decomposes the high-frequency details, providing a finer frequency band division. It can decompose each time-domain signal into different time-frequency sub-bands, so that noise and useful signal features can be initially separated at different scales, resulting in multiple signals.

[0070] After decomposition, for the signals within... We obtain a set of wavelet packet coefficient trees, and extract all node coefficients from the 5th layer, resulting in a total of 2^5 = 32 subbands, which includes one set of low-frequency approximation coefficients. and 31 sets of high-frequency detail coefficients To balance computational efficiency and noise reduction, in practical processing, we can focus on selecting mid-to-high frequency subbands related to known fault characteristics such as partial discharge pulses and mechanical vibration resonance bands, for example, specific detail coefficients corresponding to layers 3 to 5 for subsequent in-depth processing.

[0071] Step S22: Introduce a deep learning network to optimize the high-frequency detail coefficients in the multi-channel signals.

[0072] Understandably, the high-frequency detail coefficients obtained based on the aforementioned steps are not directly processed using a fixed threshold. Instead, a neural network is trained to intelligently assign appropriate adjustment weights to each coefficient. This fully considers the correlation between different coefficients and their actual contribution to the overall signal, avoiding information loss or noise residue caused by a uniform threshold. This facilitates more accurate preservation of useful information and suppression of irrelevant interference during denoising or feature enhancement, thereby improving the adaptability and effectiveness of the overall processing.

[0073] Further, step S22 includes: Step S221: Introduce a deep learning network and train it.

[0074] To explain, deep learning networks are machine learning models based on multi-layer nonlinear transformations. They mimic the neural network structure of the human brain and process complex data patterns through multi-layer neuron connections. Compared with traditional neural networks, they have more hidden layers and can automatically extract high-level feature representations from raw data.

[0075] Furthermore, in step S221, specifically: Deep learning networks are one-dimensional convolutional neural networks with an encoder-bottleneck-decoder architecture, and are trained by minimizing a composite loss function.

[0076] It can be explained that the encoder is used to extract features from the input. It contains 2 to 3 one-dimensional convolutional layers, each followed by a ReLU activation function and a max-pooling layer, to abstract layer by layer and extract deep joint representations related to noise and fault features from the input. The bottleneck layer is used for information compression. It compresses and fuses the encoded features through one or more convolutional layers with a kernel of 1 to form high-dimensional semantic features. The decoder is used to predict the weights of high-frequency detail coefficients. It maps the bottleneck layer features back to the same spatial resolution as the original input coefficient tensor through deconvolutional layers or upsampling layers in conjunction with convolutional layers. In this embodiment, the input is a cross-modal coefficient tensor with a spatial resolution of 3×256. Then, the last convolutional layer in the deep learning network uses the Sigmoid activation function to make its output value between 0 and 1. The output spatial resolution is a 3×256 matrix, which is the adjustment coefficient or preservation probability predicted by the deep learning network for the high-frequency detail coefficients of each input.

[0077] Specifically, training based on this deep learning network involves first determining the training data by injecting typical faults such as simulated partial discharge and mechanical impact into a cable model in a laboratory environment, while simultaneously collecting controlled light and strong noise data. For the relatively clean signals collected on-site and marked by experts at the fault time, synthetic noise such as white noise, periodic interference, and impulse noise with similar characteristics to actual noise is artificially added to generate noisy signals, resulting in a large number of pairs of noisy signal-relatively clean signal data, which are then used to generate training data.

[0078] Next, a composite loss function combining reconstruction error and sparsity constraints is used to train the deep learning network. The corresponding calculation formula is as follows:

[0079] in, Represents the composite loss function; , All represent equilibrium hyperparameters; Wavelet detail coefficients representing relatively clean signals; Wavelet detail coefficients representing noisy signals; Indicates the adjustment factor; Represents norm operations.

[0080] It can be explained that the reconstruction error That is, the MSE (Mean Squared Error) loss, which forces the adjustment coefficient of the deep learning network's prediction. It can restore the purity coefficient to the maximum extent; sparsity constraint L1 regularization encourages adjustment of coefficients. Tendency to sparsity means that deep learning networks learn to push most irrelevant noise coefficients toward 0, retaining only a few coefficients that are relevant to the features.

[0081] Step S222: Standardize the high-frequency detail coefficients, stack them along the channel dimension to form a multidimensional tensor, input it into the trained deep learning network, and output the adjustment coefficient matrix corresponding to the high-frequency detail coefficients.

[0082] Specifically, based on the aforementioned step S21, high-frequency detail coefficients are extracted, that is, high-frequency detail coefficients corresponding to the same frequency band range, such as three sub-bands of the fourth layer, are extracted from the same time segment of the three signals (time reference signal, signal to be aligned, etc.). The length of each coefficient segment is set to... Each coefficient segment is standardized; then stacked along the channel dimension to form a 3 (channel) × 256 (coefficient length) two-dimensional tensor, which is the original input coefficient tensor for constructing the deep learning network, denoted as . This enables deep learning networks to simultaneously learn the time-frequency characteristics of different physical quantities at the same time, providing the possibility of using cross-modal correlation to suppress common-mode noise and enhance real fault characteristics.

[0083] In the application of deep learning networks, specifically during the network forward propagation, the detail coefficient tensor of the signal to be denoised, i.e. The input is fed into the trained deep learning network, and the output is an adjustment matrix, which is the adjustment coefficient matrix corresponding to the high-frequency detail coefficients based on the composite loss function. .

[0084] Step S223: Integrate the high-frequency detail coefficients and adjustment coefficient matrix to complete the optimization of the high-frequency detail coefficients.

[0085] Specifically, the corresponding calculation formula is:

[0086] in, This represents the optimized high-frequency detail coefficients; Represents the adjustment coefficient matrix; This indicates element-wise multiplication. This represents a multidimensional tensor formed after standardization based on high-frequency detail coefficients.

[0087] It can be explained that optimizing high-frequency detail coefficients is equivalent to a data-driven, context-aware, cross-modal correlation soft threshold function. Through its full perception of contextual information and its deep cross-modal correlation capability, the deep learning network can adaptively perform filtering based on the local features of high-frequency detail coefficients.

[0088] Step S23: Perform wavelet reconstruction on the optimized high-frequency detail coefficients and low-frequency approximation coefficients to obtain the denoised time-domain signal.

[0089] Specifically, the high-frequency detail coefficients optimized in step S21, i.e. Low-frequency approximation coefficients without deep learning network processing Perform inverse wavelet packet transform to reconstruct the denoised time-domain signal, denoted as . Among them, the low-frequency approximation coefficient The main energy and slowly varying information are concentrated in the low frequency range, where noise usually has little impact.

[0090] It can be explained that the denoised time-domain signal The signal-to-noise ratio is significantly improved, meaning that weak partial discharge pulses are completely submerged in the original electromagnetic signal containing strong periodic industrial interference and random white noise. After processing in step S2, background noise is greatly suppressed, while the amplitude and phase information of the partial discharge pulses are recovered with high fidelity. Compared with traditional wavelet threshold denoising such as general threshold and Sure threshold, the method in this step avoids feature passivation caused by over-thresholding or residual noise left by insufficient threshold. It is particularly good at handling non-stationary signals where noise and feature spectra overlap, providing extremely clean and reliable data input for subsequent accurate identification of cable defects and fault type classification based on multi-mode signal fusion.

[0091] The explanation is that the cascaded processing of steps S1 and S2 significantly improves signal quality, and the two are inherently dependent on each other. Specifically, the estimated spatiotemporal offset obtained in step S1 is affected by the noise level of the original signal; the denoising effect in step S2 depends on the high-precision alignment between signals. Therefore, in step S3, the denoised time-domain signal obtained in step S2 is fed back to step S1 as a new and more reliable observation signal source. The time-aligned signal obtained by re-analysis is the optimized signal, i.e., the optimized estimated spatiotemporal offset is obtained. Then, based on the optimized signal, step S2 is executed for collaborative optimization iteration to break the unidirectional dependency and form a closed-loop feedback system. This allows synchronization and denoising to promote each other in multiple iterations and converge to a better global solution, thus completing the preprocessing of multi-source heterogeneous signals of high-voltage cables.

[0092] Furthermore, in step S3, the iteration termination condition for collaborative optimization iteration is reaching the preset maximum number of iterations, and / or the difference between the delay estimates of two adjacent iterations is less than a preset threshold.

[0093] Specifically, first, complete the iteration initialization operation, assuming the iteration number index is... Perform basic processing That is, the original signal In this embodiment, the time reference signal Signal to be aligned and The initial synchronization based on adaptive Kalman filtering is performed to obtain the initial alignment signal; then basic processing is performed. That is, performing initial hybrid denoising using a wavelet-deep learning hybrid model on the initial aligned signal.

[0094] Then, for the first The denoised time-domain signal obtained in the next iteration With time reference signal Step S3, time reference signal The noise level is low during operation, or noise has been synchronously denoised; that is, the cross-correlation function of the two signals is recalculated. Due to the denoising of the time domain signal The random noise in the signal has been significantly suppressed, and its true correlation with the time base signal has been highlighted, hence the obtained cross-correlation function The peak value will be higher than the first cross-correlation function calculated using the original noisy signal. The observations are sharper and more precise, and less affected by noise-induced sidelobe interference, providing higher quality observations for delay estimation.

[0095] The peak position in the updated, higher-quality cross-correlation function, i.e. As a new observation An adaptive Kalman filter is used to estimate the state based on the final state estimate from the previous iteration. With the state estimation error covariance matrix Continue running to generate new observation sequences. This allows for a new round of filtering updates, resulting in a more refined delay estimate. Due to the improved quality of the observation input, the convergence speed and final accuracy of this round of estimation are expected to be better than the previous one.

[0096] Next, refined delay estimates are used. According to step S13, the original signal Realignment is performed, which involves resampling and aligning the original signal. In actual operation, resampling and alignment are always based on the original signal to avoid introducing distortion signals through multiple interpolations; this then generates a new time-aligned signal. The signal is input into the wavelet-deep learning hybrid model to perform step S2, resulting in a new round of denoised time-domain signal. That is, the time-aligned signal after iteration. It has higher synchronization accuracy, better alignment of different modal signal features in the time and frequency domains, and deep learning networks can more effectively utilize cross-modal information for joint denoising.

[0097] Finally, an iteration termination condition is set to balance the processing effect and computational cost. Preferably, in this embodiment, a maximum number of iterations is set. It can be adjusted according to the actual situation; at the same time, the difference between the delay estimates of two adjacent iterations or the change in normalized mean square error between the denoised time-domain signals is calculated, and the corresponding calculation formula is:

[0098] in, Indicates the first The second iteration and the first The difference between the delay estimates of each iteration; , They represent the first The second iteration and the first The delay estimate for the next iteration; This represents absolute value operations.

[0099] The preset threshold is denoted as In this embodiment, it is 1% of the sampling interval, which can be adjusted according to the actual situation. If the collaborative optimization iteration has converged, the iteration is terminated; otherwise, when... Continue iterating until the preset maximum number of iterations is reached.

[0100] Understandably, spatiotemporal synchronization compensation based on multi-source heterogeneous signals from high-voltage cables—that is, through adaptive Kalman filtering, dynamically tracking and compensating for random delays—achieves microsecond-level or even higher precision synchronization of multi-source heterogeneous signals from high-voltage cables, which is superior to fixed interpolation methods. Furthermore, combining wavelet transform with deep learning networks leverages the multi-resolution analysis advantages of wavelets and learns complex noise patterns through a data-driven approach, suppressing strong noise while preserving the characteristics of weak fault impacts or abnormal waveforms to the greatest extent possible. Moreover, the deep learning network can be trained to adapt to different lines and noise environments, avoiding the drawbacks of manually adjusting parameters required by traditional threshold methods. Finally, by backtracking step S1 to obtain the optimized signal, step S2 is executed to perform collaborative optimization iteration, completing the preprocessing of multi-source heterogeneous signals from high-voltage cables. The output high-quality, highly synchronized multimodal data fully utilizes existing data resources, avoids data silos, and can be directly used for subsequent feature fusion and fault diagnosis by deep learning networks, significantly improving recognition accuracy and reliability.

[0101] The second embodiment of the present invention provides a preprocessing system for multi-source heterogeneous sensor signals of high-voltage cables. The system includes a processor, a communication interface, a memory, and a communication bus. The processor, the communication interface, and the memory communicate with each other through the communication bus. The processor calls logical instructions in the memory to execute the preprocessing method for multi-source heterogeneous sensor signals of high-voltage cables described in any embodiment of the present invention.

[0102] During its operation, it requires a high-voltage cable multi-source heterogeneous sensor signal preprocessing method. Therefore, whether the system and program data are integrated or different hardware is configured to produce a function with similar effect to that achieved by the present invention, it is within the protection scope of the present invention. The system has the same beneficial effect as the aforementioned high-voltage cable multi-source heterogeneous sensor signal preprocessing method, and will not be elaborated here.

[0103] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0104] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A method for preprocessing multi-source heterogeneous sensor signals of high-voltage cables, characterized in that, The method includes: Step S1: In the high-voltage cable monitoring scenario, various sensors are deployed, and multi-source heterogeneous signals of the high-voltage cable are collected based on the sensors. Spatiotemporal synchronization compensation is performed on the multi-source heterogeneous signals of the high-voltage cable to obtain time-aligned signals. Any sensor is defined as the target sensor, where the sensing signal collected by the target sensor is the time reference signal, and the sensing signals collected by the remaining sensors are the signals to be aligned. The time reference signal and the signals to be aligned are integrated to form a multi-source heterogeneous signal for high-voltage cables. Step S1 includes: Step S11: Model the spatiotemporal offset between the signal to be aligned and the time reference signal as a state vector containing the absolute time delay and its rate of change, and establish the state equation; Step S12: Determine the cross-correlation function based on the signal to be aligned and the time reference signal, define the peak position in the cross-correlation function as the observed value, determine the observation noise covariance matrix of the real-time signal-to-noise ratio adjusted Kalman filter, and iteratively estimate the optimal spatiotemporal offset. Step S12 includes: Step S121: Determine the cross-correlation function based on the signal to be aligned and the time reference signal, and define the peak position in the cross-correlation function as the observed value; Step S122: Introduce observation noise, establish the observation equation in conjunction with the state equation, and determine the observation matrix; Step S123: Obtain the signal-to-noise ratio estimate based on the observation window corresponding to the cross-correlation function, and obtain the basic observation noise variance, then adjust the observation noise covariance matrix of the Kalman filter; Step S124: Iterate the state equation using standard Kalman filtering to obtain the predicted state estimation error covariance matrix. Then, use the cross-correlation function, observations, and observation matrix to correct the predicted state estimation error covariance matrix and obtain the estimated optimal spatiotemporal offset. Step S13: Resample and align the signal to be aligned based on the estimated optimal spatiotemporal offset to obtain a time-aligned signal; Step S2: Establish a wavelet-deep learning hybrid model by combining wavelet transform and deep learning network, and perform denoising processing on the time-aligned signal to obtain the denoised time domain signal; Step S3: Feed the denoised time-domain signal back to step S1 to obtain a time-aligned signal as the optimized signal, and perform step S2 on the optimized signal to perform collaborative optimization iteration, thereby completing the preprocessing of the multi-source heterogeneous signal of the high-voltage cable.

2. The method for preprocessing multi-source heterogeneous sensor signals of high-voltage cables according to claim 1, characterized in that, In step S13, sinc interpolation is used to resample and align the signal to be aligned.

3. The method for preprocessing multi-source heterogeneous sensor signals of high-voltage cables according to claim 1, characterized in that, Step S2 includes: Step S21: Wavelet transform is used to decompose the time-aligned signal into wavelet packets to initially separate noise and useful signals, obtain multiple signals, and extract low-frequency approximation coefficients and high-frequency detail coefficients. Step S22: Introduce a deep learning network to optimize the high-frequency detail coefficients in the multi-channel signals; Step S23: Perform wavelet reconstruction on the optimized high-frequency detail coefficients and low-frequency approximation coefficients to obtain the denoised time-domain signal.

4. The method for preprocessing multi-source heterogeneous sensor signals of high-voltage cables according to claim 3, characterized in that, Step S22 includes: Step S221: Introduce a deep learning network and train it; Step S222: Standardize the high-frequency detail coefficients, stack them along the channel dimension to form a multidimensional tensor, input it into the trained deep learning network, and output the adjustment coefficient matrix corresponding to the high-frequency detail coefficients; Step S223: Integrate the high-frequency detail coefficients and adjustment coefficient matrix to complete the optimization of the high-frequency detail coefficients.

5. The method for preprocessing multi-source heterogeneous sensor signals of high-voltage cables according to claim 4, characterized in that, In step S221, specifically: The deep learning network is a one-dimensional convolutional neural network with an encoder-bottleneck-decoder architecture, and it is trained by minimizing a composite loss function.

6. The method for preprocessing multi-source heterogeneous sensor signals of high-voltage cables according to claim 1, characterized in that, In step S3, the iteration termination condition for collaborative optimization iteration is reaching the preset maximum number of iterations, and / or the difference between the delay estimates of two adjacent iterations is less than a preset threshold.

7. A preprocessing system for multi-source heterogeneous sensor signals of high-voltage cables, characterized in that, The system includes a processor, a communication interface, a memory, and a communication bus. The processor, the communication interface, and the memory communicate with each other through the communication bus. The processor calls logical instructions in the memory to execute the high-voltage cable multi-source heterogeneous sensing signal preprocessing method according to any one of claims 1 to 6.

Images