A power data acquisition method for a submarine optical cable

Abstract

The application belongs to the field of data acquisition, and particularly relates to a power data acquisition method for a submarine optical cable, which aims to solve the technical problem that a traditional filtering method is difficult to deal with complex noise and has poor data purification effect, and comprises the following steps: S1, obtaining a frequency domain input signal; S2, obtaining a prediction error; calculating a frequency domain gain by using an enhanced frequency variable forgetting factor matrix; S3, when performing update correction, performing cross-band correction on filter coefficients of all frequency points, and applying an attenuation weight to a correction amount in a known interference harmonic frequency band; and S4, obtaining purified data in a frequency domain by using the corrected filter coefficients, and obtaining time domain purified data through inverse fast Fourier transform. The application can better extract effective power data of the submarine optical cable in a strong noise background, and improves the signal-to-noise ratio and reliability of the acquired data.

Description

Technical Field

[0001] This invention belongs to the field of data acquisition, specifically a method for acquiring power data for submarine optical cables. Background Technology

[0002] Submarine optical cables are the lifeblood of global information and communication. By collecting and analyzing the power data of submarine optical cables in real time, the operating status of key equipment such as repeaters can be effectively monitored, and potential faults can be warned in a timely manner. The seabed environment is relatively complex, and power data is often affected by strong electromagnetic interference and marine environmental noise during transmission, resulting in an extremely low signal-to-noise ratio and severely suppressing useful signals.

[0003] In order to extract effective power characteristic signals from a background of strong noise, traditional methods such as least mean square (LMS) and recursive least squares (RLS) in the time domain, as well as adaptive filtering in the frequency domain, are used. Although they can suppress noise to some extent, their effects are limited.

[0004] Specifically, traditional methods use fixed window functions for frame segmentation, which cannot adapt to changes in signal spectrum characteristics and are prone to severe spectral leakage and reduced frequency resolution. Traditional frequency-domain RLS (FRLS) algorithms cannot match the uneven energy distribution of noise across different frequency bands, causing unnecessary computational overhead and coefficient misalignment risks at high signal-to-noise ratios, and also ignoring the inherent correlation between broadband or harmonic interference across frequency bands. For known strong periodic interference such as power frequency harmonics, there is a lack of targeted suppression strategies that match the interference characteristics, leaving significant interference components in the cleaned data and affecting the accuracy of subsequent state assessments. Summary of the Invention

[0005] This invention provides a power data acquisition method for submarine optical cables to solve the technical problems of traditional filtering methods being unable to cope with complex noise and having poor data purification effects.

[0006] A method for acquiring power data for submarine optical cables includes the following steps:

[0007] S1. Acquire raw power data and divide it into frames. Obtain an asymmetric hybrid window function based on the spectral characteristic parameters of each data frame. Apply the asymmetric hybrid window function to each data frame and perform windowing and fast Fourier transform to obtain the frequency domain input signal.

[0008] S2, establish a state-space prediction model for the filter coefficients at each frequency point, obtain the prior prediction value for each frequency point, calculate the prediction value of the frequency domain input signal based on the prior prediction value, and thus obtain the prediction error; calculate the frequency-varying forgetting factor matrix based on the signal-to-noise ratio of each frequency point and its adjacent frequency points, enhance the frequency-varying forgetting factor matrix, and use the enhanced frequency-varying forgetting factor matrix to calculate the frequency domain gain.

[0009] S3, when the statistical distance between the statistical distribution of the prediction error and the preset baseline noise model exceeds a preset threshold, the filter coefficient update correction is triggered; when performing the update correction, the filter coefficients at all frequency points are cross-band corrected using the frequency domain gain and prediction error, and an attenuation weight related to the interference characteristics is applied to the correction amount within the known interference harmonic frequency band.

[0010] S4 uses the corrected filter coefficients to obtain the frequency domain of the purified data, and then performs an inverse fast Fourier transform to obtain the time domain purified data.

[0011] Furthermore, the power supply voltage or current signal of the submarine optical cable is collected to obtain a discrete digital signal sequence. The digital signal sequence is then divided into multiple data frames of a set length, with several overlapping sampling points between adjacent data frames.

[0012] Further, in S1, an asymmetric mixing window function is obtained based on the spectral characteristic parameters of each data frame. This asymmetric mixing window function is then used to perform windowing and fast Fourier transform on each data frame to obtain the frequency domain input signal, including:

[0013] Calculate the spectral kurtosis of the data frame, use the spectral kurtosis as a transient characteristic index to calculate the total harmonic distortion, and use the total harmonic distortion as a harmonic contamination index.

[0014] The standard deviation parameter σ of the Gaussian window function is obtained based on the transient characteristic index, and the parameters of the Blackman-Harris window function are determined based on the harmonic contamination index.

[0015] By concatenating the first half of the Gaussian window function with the second half of the Blackman-Harris window function, an asymmetric hybrid window function is obtained.

[0016] Furthermore, adopt Calculate the standard deviation of the Gaussian window function ,in, for The maximum value, for The minimum value, The reference midpoint for spectral kurtosis. It is a constant. At maximum value and minimum value middle, This is the transient characteristic index of the data frame.

[0017] Furthermore, in S2, based on the signal-to-noise ratio of each frequency point and its adjacent frequency points, the frequency-varying forgetting factor matrix is ​​calculated, including:

[0018] The signal-to-noise ratio at frequency k is calculated using the current signal power and the estimated noise power. ;

[0019] Through S-shaped function Calculate the initial forgetting factor for each frequency point, where, Let k be the forgetting factor at frequency k. This represents the minimum value of the forgetting factor. This represents the maximum value of the forgetting factor. This is the reference threshold for the signal-to-noise ratio. This is the slope control parameter for the sigmoid function;

[0020] The sequence of initial forgetting factors at all frequencies is subjected to moving average filtering along the frequency axis to obtain the diagonal elements of the frequency-varying forgetting factor matrix.

[0021] Furthermore, in S3, data is collected in advance during periods with no signal or only background noise, and a Gaussian mixture model representing the amplitude distribution of the noise signal is established. This Gaussian mixture model is used as the baseline noise model.

[0022] Furthermore, in S3, the set of prediction errors e(k,n) for all frequency points of the current data frame is collected, and the probability density function of the set is calculated to form a real-time error distribution; KL divergence is used to calculate the distance between the real-time error distribution and the preset baseline noise model.

[0023] Furthermore, in S3, when the statistical distance between the statistical distribution of the prediction error and the preset baseline noise model exceeds a preset threshold, the filter coefficients are updated and corrected, including:

[0024] The baseline noise model is a Gaussian distribution model. ,variance It is determined by the average noise power of several consecutive data frames collected by the system in the initial state without power load;

[0025] At each time step, the prediction error of the most recent sample points is extracted, and its probability density function P(e) is calculated.

[0026] P(e) and the baseline noise model are calculated using KL divergence. Statistical distance between ;

[0027] when When the value exceeds the preset threshold, the filter coefficients are updated and corrected.

[0028] Furthermore, during the update correction, cross-band correction is performed on the filter coefficients at all frequencies using frequency domain gain and prediction error, and an attenuation weight related to the interference characteristics is applied to the correction amount within the known interference harmonic frequency band, including:

[0029] Obtain the tridiagonal coupling matrix representing the correlation between frequency points. The main diagonal elements are 1, and the secondary diagonal elements are the energy correlation coefficients of adjacent frequency points obtained from historical data statistics.

[0030] The initial correction vector calculated using frequency domain gain and prediction error will be used. With the tridiagonal coupling matrix Multiply to obtain the coupling correction amount. ;

[0031] For known power frequency harmonic interference, construct an attenuation weight vector. Set a weight value less than 1 at the corresponding harmonic frequency point and several frequency points to its left and right, and set a weight value of 1 at the other frequency points;

[0032] Coupling correction amount With decay weight vector Element-wise multiplication yields the correction amount applied to the filter coefficients.

[0033] Furthermore, in S4, the corrected filter coefficients are used to perform a multiplication operation on the frequency domain input signal to obtain the posterior prediction error, which is the frequency domain of the cleaned signal.

[0034] The beneficial effects are as follows: Compared with existing technologies, this invention constructs a window function based on the spectral characteristic parameters of each data frame, reducing spectral leakage. The filtering process uses an update triggering mechanism based on prediction error, avoiding unnecessary coefficient adjustments when the signal is stable. Simultaneously, a forgetting factor matrix associated with the signal-to-noise ratio of each frequency point is constructed, enabling a better balance between tracking speed and steady-state error in complex environments with uneven noise distribution. Cross-band correction is used to eliminate strong periodic interference. This invention can better extract effective power data from submarine optical cables in strong noise backgrounds, improving the signal-to-noise ratio and reliability of the acquired data. Attached Figure Description

[0035] Figure 1 This is a flowchart of a power data acquisition method for submarine optical cables;

[0036] Figure 2 This is a schematic diagram of an asymmetric mixed window function;

[0037] Figure 3 This is a schematic diagram of harmonic band attenuation weights. Detailed Implementation

[0038] An embodiment of the power data acquisition method for submarine optical cables provided by this invention:

[0039] like Figure 1 As shown, a method for acquiring power data for submarine optical cables includes the following steps:

[0040] S1. Acquire raw power data and divide it into frames. Obtain an asymmetric hybrid window function based on the spectral characteristic parameters of each data frame. Apply the asymmetric hybrid window function to each data frame and perform windowing and fast Fourier transform to obtain the frequency domain input signal.

[0041] The implementation method of step S1 is as follows:

[0042] The power supply voltage or current signal of the submarine optical cable is acquired to obtain a discrete digital signal sequence. This digital signal sequence is then divided into multiple data frames of a predetermined length, with several overlapping sampling points between adjacent data frames. For example, the predetermined length is 2048 sampling points, and adjacent data frames overlap by 1024 sampling points, i.e., a 50% overlap rate. A Fourier transform is performed on the original data frames to calculate the spectral centroid and spectral flatness. Based on the height of the spectral centroid and the magnitude of the spectral flatness, the composition ratio of the mixing window is determined.

[0043] In a preferred embodiment, an asymmetric mixing window function is obtained based on the spectral feature parameters of each data frame. The asymmetric mixing window function is then used to perform windowing and fast Fourier transform on each data frame to obtain the frequency domain input signal, including:

[0044] Calculate the spectral kurtosis of the data frame, use the spectral kurtosis as a transient characteristic index to calculate the total harmonic distortion, and use the total harmonic distortion as a harmonic contamination index.

[0045] The standard deviation parameter of the Gaussian window function is obtained based on the transient characteristic index. The parameters of the Blackman-Harris window function are determined based on the harmonic pollution index.

[0046] By concatenating the first half of the Gaussian window function with the second half of the Blackman-Harris window function, an asymmetric hybrid window function is obtained.

[0047] Voltage sags or inrush currents generate transient signals. These signals are brief and abrupt, requiring high time resolution to pinpoint their occurrence. Gaussian window functions provide good temporal focus. Steady-state harmonics, on the other hand, need frequency separation to avoid spectral leakage. This necessitates window functions with extremely low sidelobes; therefore, the Blackman-Harris window function is employed. By dynamically constructing an asymmetric hybrid window, the first half of the asymmetric window function can be adapted to transient characteristics, while the second half adapts to steady-state characteristics. For example, when a data frame contains a transient overvoltage caused by a lightning strike, its spectral kurtosis is calculated to be as high as 9.5, a typical indicator of a strong transient signal. The standard deviation parameter of the Gaussian window function... A small value (e.g., 0.12) is set to form a very sharp first half of the window. Meanwhile, the total harmonic distortion of this data frame may only be 3%, indicating that harmonic pollution is not severe, and standard Blackman-Harris window parameters are used. The sharp Gaussian window's first half is then spliced ​​with the Blackman-Harris window's second half, which has good sidelobe suppression, to obtain a hybrid window.

[0048] Optionally, adopt Calculate the standard deviation of the Gaussian window function ,in, for The maximum value, for The minimum value, The reference midpoint for spectral kurtosis. The transient characteristic exponent of the data frame is a constant. At that time, the value of the sigmoid function was 0.5. At maximum value and minimum value middle.

[0049] The preferred Blackman-Harris window function ,in, For position Window function value at that point, The length of the window function. , , , Let [x] be the coefficients of the window function; obtain the basic vector of coefficients for the Hanning window [0.5, 0.5, 0, 0], denoted as [x]. The fundamental vector of coefficients of the Blackman-Harris window [ The expression [0.35875, 0.48829, 0.14128, 0.01168] is denoted as... ,use Calculate the mixing factor β(H), where, The harmonic pollution index of the signal. As the reference midpoint for harmonic pollution, These are parameters that control the steepness of the sigmoid function. According to... Calculated .

[0050] In another embodiment, the hybrid window is composed of a weighted combination of a rectangular window in the left half and a Caesar window in the right half. The weighting of this combination and the smoothness of the transition region are determined by spectral feature parameters, such as... Figure 2As shown. When a transient impact is detected at the beginning of the signal, the weight of the rectangular window is increased to reduce distortion. The generated asymmetric mixed window function is multiplied point-by-point with the current data frame, and then a fast Fourier transform is performed on the windowed data frame to obtain the complex frequency domain input signal of the data frame.

[0051] S2, establish a state-space prediction model for the filter coefficients at each frequency point, obtain the prior prediction value for each frequency point, calculate the prediction value of the frequency domain input signal based on the prior prediction value, and thus obtain the prediction error; calculate the frequency-varying forgetting factor matrix based on the signal-to-noise ratio of each frequency point and its adjacent frequency points, enhance the frequency-varying forgetting factor matrix, and use the enhanced frequency-varying forgetting factor matrix to calculate the frequency domain gain.

[0052] The implementation method of step S2 is as follows:

[0053] Using the filter coefficient vector W(k, n) at each frequency point k as a state variable, and employing a first-order autoregressive model as the state-space prediction model, the filter coefficients for the next time step are predicted by multiplying the current filter coefficients by a constant close to 1 (e.g., 0.999), thus obtaining the prior prediction weight coefficients. The predicted value of the frequency domain signal is obtained by multiplying the prior prediction weights with the conjugate transpose of the current frequency domain input signal. The actual frequency domain input signal is used as the desired signal, and the desired signal and the predicted value are calculated. The difference between the values ​​is used to obtain the prior prediction error e(k, n) for each frequency point. In another embodiment, a linear or nonlinear Kalman filter can be used as the state-space prediction model; in yet another embodiment, a neural network model including multiple fully connected layers can also be used as the state-space prediction model.

[0054] The noise power at each frequency point is continuously estimated using the least statistical method, and the current signal power is calculated to obtain the real-time signal-to-noise ratio (SNR)(k) for each frequency point k. A diagonal matrix is ​​constructed as the frequency-varying forgetting factor matrix, where the value of the k-th element λ(k) on the diagonal is determined by a nonlinear function. The input to the nonlinear function is the weighted average SNR of five frequency points (frequency point k and its two adjacent frequencies on each side). The λ(k) value corresponding to the frequency band with high SNR is close to 1, while the λ(k) value corresponding to the frequency band with low SNR is smaller (e.g., 0.98). In the covariance matrix update step of the recursive least squares algorithm, the frequency-varying forgetting factor matrix is ​​used instead of the traditional scalar forgetting factor to calculate the gain vector g(k, n) for each frequency point k.

[0055] In a preferred embodiment, the frequency-varying forgetting factor matrix is ​​calculated based on the signal-to-noise ratio of each frequency point and its adjacent frequency points, including:

[0056] The signal-to-noise ratio at frequency k is calculated using the current signal power and the estimated noise power. ;

[0057] Through S-shaped function Calculate the initial forgetting factor for each frequency point, where, Let k be the forgetting factor at frequency k. This represents the minimum value of the forgetting factor. This represents the maximum value of the forgetting factor. This is the reference threshold for the signal-to-noise ratio. Here, represents the slope control parameter of the sigmoid function; =0.98, =1.0, =15;

[0058] The sequence of initial forgetting factors at all frequencies is subjected to moving average filtering along the frequency axis to obtain the diagonal elements of the frequency-varying forgetting factor matrix.

[0059] In power signals, the signal-to-noise ratio (SNR) of major components such as the fundamental and harmonics is usually high, and the signal characteristics are stable. Therefore, frequent adjustments to the filter coefficients are not recommended, and a forgetting factor close to 1 is needed to retain historical information. Conversely, certain high-frequency bands may be filled with random noise, resulting in a low SNR. The filter needs to quickly track these changes, requiring a smaller forgetting factor to give greater weight to the current data. The sigmoid function provides a smooth transition from 0.98 to close to 1.0, with its transition center determined by a 15dB SNR reference threshold.

[0060] Suppose that at a certain moment, the signal-to-noise ratio (SNR) at the 50Hz fundamental frequency reaches 40dB, exceeding the reference value of 15dB. The initial forgetting factor calculated using the sigmoid function will be very close to 1.0, for example, 0.9998, indicating a high degree of confidence in the historical state of that frequency. However, at the 2kHz frequency, due to noise interference, the SNR is only 5dB, lower than the reference value. The calculated forgetting factor will be close to 0.98, such as 0.9803, allowing the filter to adapt to noise fluctuations more quickly. A moving average of five frequencies is applied to the initial forgetting factor sequence across all frequencies. For example, the final forgetting factor at the 100th frequency is obtained by averaging the initial forgetting factors from the 98th to the 102nd frequencies. This ensures that the filter's adaptability changes smoothly across the entire spectrum, avoiding filter coefficient oscillations caused by sudden changes in SNR at a single frequency.

[0061] The adaptive forgetting rate at each frequency point is independent, but in actual signals, there is usually a certain correlation between adjacent frequency points. Furthermore, the frequency-varying forgetting factor matrix is ​​enhanced by taking the calculated diagonal elements as the main diagonal.

[0062] The normalized spectral correlation coefficient between adjacent frequency points is calculated based on the frequency domain amplitude statistics of recent data frames. ;

[0063] The second diagonal elements of the frequency-varying forgetting factor matrix Set as ,in and These are the main diagonal elements corresponding to the frequency points. In an optional embodiment, the frequency domain gain can also be calculated directly using the factor matrix without enhancement.

[0064] S3, when the statistical distance between the statistical distribution of the prediction error and the preset baseline noise model exceeds a preset threshold, the filter coefficient update correction is triggered; when performing the update correction, the filter coefficients at all frequency points are cross-band corrected using the frequency domain gain and prediction error, and an attenuation weight related to the interference characteristics is applied to the correction amount within the known interference harmonic frequency band.

[0065] The implementation method of step S3 is as follows:

[0066] Data is pre-collected during periods with no signal or only background noise to establish a Gaussian mixture model representing the amplitude distribution of the noise signal. This Gaussian mixture model is used as the baseline noise model. During real-time processing, the set of prediction errors e(k,n) for all frequency points in the current data frame is collected, and the probability density function of this set is calculated to form a real-time error distribution. The distance between this real-time error distribution and the preset baseline noise model is calculated using KL divergence. An empirical threshold (e.g., 0.05) is set. When the calculated KL divergence value is greater than this threshold, it indicates a significant change in signal characteristics, and the filter coefficient update process for this frame is initiated; if it is less than or equal to the threshold, the update is skipped, and the filter coefficients from the previous time step are used directly.

[0067] If an update is triggered, the standard correction for each frequency point k is calculated, which is the product of the gain vector g(k,n) and the prediction error e(k,n). Then, through coupling correction, the final correction for frequency point k is adjusted to a weighted sum of its own standard correction and the standard corrections of its adjacent frequencies k-1 and k+1, with the weighting coefficients set according to the correlation between frequency bands. Based on the known power supply system frequencies, the specific set of frequency points containing power frequency harmonics (e.g., 50Hz, 150Hz, 250Hz) is determined. For these frequency points, the calculated correction is multiplied by a preset attenuation weight to enhance the suppression of these known strong interferences. The correction with the attenuation weight is then added to the prior prediction weight coefficients, completing the update of all frequency point filter coefficients.

[0068] In a preferred embodiment, when the statistical distance between the statistical distribution of the prediction error and the preset baseline noise model exceeds a preset threshold, the filter coefficients are updated and corrected, including:

[0069] The baseline noise model is a Gaussian distribution model. ,variance It is determined by the average noise power of several consecutive data frames collected by the system in the initial state without power load;

[0070] At each time step, the prediction error of the most recent sample points is extracted, and its probability density function P(e) is calculated.

[0071] P(e) and the baseline noise model are calculated using KL divergence. Statistical distance between ;

[0072] when When the value exceeds the preset threshold, the filter coefficients are updated and corrected.

[0073] For example, the inherent background noise is characterized by collecting 100 data frames under no-load conditions. For instance, if the average noise power is measured to be 0.0025, a zero-mean Gaussian distribution with a variance of 0.0025 is established as the baseline noise model, representing the statistical form of the prediction error that the filter should have under ideal operating conditions. During normal operation, the filter continuously operates and generates prediction errors. The most recent 256 error samples are cached, and their distribution characteristics are analyzed in real time. If the state is stable, the prediction error mainly consists of background noise, and its probability density function will be very close to the preset baseline Gaussian model, resulting in a small KL divergence value, such as 0.01. When a large inductive load is suddenly connected, it introduces new harmonics and transient disturbances. The adaptive filter cannot immediately and completely predict these new components, causing the distribution of the prediction error to become distorted and no longer a simple Gaussian distribution. At this time, the calculated KL divergence value may rise to 0.12. If 0.12 is greater than the set threshold of 0.05, the current filter model fails, triggering a costly update and correction process.

[0074] In a preferred embodiment, during the update correction, cross-band correction is performed on the filter coefficients at all frequencies using the frequency domain gain and prediction error, and an attenuation weight related to the interference characteristics is applied to the correction amount within the known interference harmonic frequency band, including:

[0075] Obtain the tridiagonal coupling matrix representing the correlation between frequency points. The main diagonal elements are 1, and the secondary diagonal elements are the energy correlation coefficients of adjacent frequency points obtained from historical data statistics.

[0076] The initial correction vector calculated using frequency domain gain and prediction error will be used. With the tridiagonal coupling matrix Multiply to obtain the coupling correction amount. ;

[0077] For known power frequency harmonic interference, construct an attenuation weight vector. Set a weight value less than 1 at the corresponding harmonic frequency point and several frequency points to its left and right, and set a weight value of 1 at the other frequency points;

[0078] Coupling correction amount With decay weight vector Element-wise multiplication yields the correction amount applied to the filter coefficients.

[0079] By using a coupling matrix, adjusting a coefficient at a certain frequency point can affect its adjacent frequencies. For example, when a large correction is needed at 125Hz, this can be achieved through a tridiagonal coupling matrix. Multiplying these values, 30% of the correction is applied to each of its adjacent frequency points, while the frequency point itself is also affected by the corrections from its neighboring frequencies. In one embodiment, for each frequency point, the correction of its filter coefficients is equal to the Kalman gain (frequency domain gain) of that frequency point multiplied by the prediction error of that frequency point; the corrections of all frequency points are combined into an attenuation weight vector. This constitutes the initial correction vector covering the entire frequency band. In the attenuation weight vector... In this process, the element value at a specified frequency position is set to a weight value less than 1 (such as 0.2), while the element values ​​at other frequency positions are set to 1 to obtain the attenuation weight vector. .

[0080] For the 50Hz fundamental frequency and known power frequency harmonics such as 150Hz and 250Hz, generated by common loads, these are inherent components of the signal rather than noise that needs to be eliminated. To prevent erroneous suppression of these important components, the update amount for these specific frequency bands is limited. By constructing an attenuation weight vector, the correction is attenuated to 20% of its original value within a five-frequency width centered on these harmonic frequencies, such as... Figure 3 As shown. For example, even if a correction of 1 unit is calculated for 150Hz, after attenuation weighting, the applied correction is only 0.2. In other non-harmonic frequency bands, the weight is 1. This correction method, which combines cross-band coupling and harmonic protection, ensures that the filter can adapt to unknown disturbances without damaging the known key harmonic structures in the signal.

[0081] S4 uses the corrected filter coefficients to obtain the frequency domain of the purified data, and then performs an inverse fast Fourier transform to obtain the time domain purified data.

[0082] The implementation method of step S4 is as follows:

[0083] Using the corrected filter coefficients Multiplication is performed on the frequency domain input signal X(k,n) to obtain the posterior prediction error. This posterior prediction error is the frequency domain representation of the purified signal. The purified frequency domain data of all frequency points These are combined into a complete frequency domain sequence. A 2048-point inverse fast Fourier transform is performed on the sequence to obtain a time domain data block. The continuously generated time domain data blocks are then spliced ​​together using an overlap-addition method according to a previously set overlap of 1024 points, thereby reconstructing the noise-suppressed time domain purified power data stream.

[0084] In addition, in the description of this specification, "multiple" means at least two, such as two, three or more, etc., unless otherwise expressly and specifically defined.

Claims

1. A method for acquiring power data for submarine optical cables, characterized in that, Includes the following steps: S1. Acquire raw power data and divide it into frames. Obtain an asymmetric hybrid window function based on the spectral characteristic parameters of each data frame. Apply the asymmetric hybrid window function to each data frame and perform windowing and fast Fourier transform to obtain the frequency domain input signal. S2, establish a state-space prediction model for the filter coefficients at each frequency point, obtain the prior prediction value for each frequency point, calculate the prediction value of the frequency domain input signal based on the prior prediction value, and thus obtain the prediction error; calculate the frequency-varying forgetting factor matrix based on the signal-to-noise ratio of each frequency point and its adjacent frequency points, enhance the frequency-varying forgetting factor matrix, and use the enhanced frequency-varying forgetting factor matrix to calculate the frequency domain gain. S3, when the statistical distance between the statistical distribution of the prediction error and the preset baseline noise model exceeds a preset threshold, the filter coefficient update correction is triggered; when performing the update correction, the filter coefficients at all frequency points are cross-band corrected using the frequency domain gain and prediction error, and an attenuation weight related to the interference characteristics is applied to the correction amount within the known interference harmonic frequency band. S4 uses the corrected filter coefficients to obtain the frequency domain of the purified data, and then performs an inverse fast Fourier transform to obtain the time domain purified data.

2. The power data acquisition method for submarine optical cables according to claim 1, characterized in that, The power supply voltage or current signal of the submarine optical cable is collected to obtain a discrete digital signal sequence. The digital signal sequence is then divided into multiple data frames of a set length, with several overlapping sampling points between adjacent data frames.

3. The power data acquisition method for submarine optical cables according to claim 1, characterized in that, In S1, an asymmetric mixing window function is obtained based on the spectral characteristic parameters of each data frame. This asymmetric mixing window function is then used to perform windowing and Fast Fourier Transform on each data frame to obtain the frequency domain input signal, including: Calculate the spectral kurtosis of the data frame, use the spectral kurtosis as a transient characteristic index to calculate the total harmonic distortion, and use the total harmonic distortion as a harmonic contamination index. The standard deviation parameter σ of the Gaussian window function is obtained based on the transient characteristic index, and the parameters of the Blackman-Harris window function are determined based on the harmonic contamination index. By concatenating the first half of the Gaussian window function with the second half of the Blackman-Harris window function, an asymmetric hybrid window function is obtained.

4. The power data acquisition method for submarine optical cables according to claim 3, characterized in that, use Calculate the standard deviation of the Gaussian window function ,in, for The maximum value, for The minimum value, The reference midpoint for spectral kurtosis. It is a constant. At maximum value and minimum value middle, This is the transient characteristic index of the data frame.

5. The power data acquisition method for submarine optical cables according to claim 1, characterized in that, In S2, the frequency-varying forgetting factor matrix is ​​calculated based on the signal-to-noise ratio of each frequency point and its adjacent frequency points, including: The signal-to-noise ratio at frequency k is calculated using the current signal power and the estimated noise power. ; Through S-shaped function Calculate the initial forgetting factor for each frequency point, where, Let k be the forgetting factor at frequency point k. This represents the minimum value of the forgetting factor. This represents the maximum value of the forgetting factor. This is the reference threshold for the signal-to-noise ratio. This is the slope control parameter for the sigmoid function; The sequence of initial forgetting factors at all frequencies is subjected to moving average filtering along the frequency axis to obtain the diagonal elements of the frequency-varying forgetting factor matrix.

6. The method for acquiring power data for submarine optical cables according to claim 1, characterized in that, In S3, data is collected in advance during periods with no signal or only background noise, and a Gaussian mixture model representing the amplitude distribution of the noise signal is established. This Gaussian mixture model is used as the baseline noise model.

7. The power data acquisition method for submarine optical cables according to claim 6, characterized in that, In S3, the set of prediction errors e(k,n) for all frequency points of the current data frame is collected, and the probability density function of the set is calculated to form a real-time error distribution. KL divergence is used to calculate the distance between the real-time error distribution and the preset baseline noise model.

8. The method for acquiring power data for submarine optical cables according to any one of claims 1-7, characterized in that, In S3, when the statistical distance between the statistical distribution of the prediction error and the preset baseline noise model exceeds a preset threshold, the filter coefficients are updated and corrected, including: The baseline noise model is a Gaussian distribution model. ,variance It is determined by the average noise power of several consecutive data frames collected by the system in the initial state without power load; At each time step, the prediction error of the most recent sample points is extracted, and its probability density function P(e) is calculated. P(e) and the baseline noise model are calculated using KL divergence. Statistical distance between ; when When the value exceeds the preset threshold, the filter coefficients are updated and corrected.

9. The method for acquiring power data for submarine optical cables according to claim 8, characterized in that, During the update correction, the filter coefficients at all frequencies are cross-band corrected using frequency domain gain and prediction error. An attenuation weight related to the interference characteristics is applied to the correction amount within the known interference harmonic frequency band, including: Obtain the tridiagonal coupling matrix representing the correlation between frequency points. The main diagonal elements are 1, and the secondary diagonal elements are the energy correlation coefficients of adjacent frequency points obtained from historical data statistics. The initial correction vector calculated using frequency domain gain and prediction error will be used. With the tridiagonal coupling matrix Multiply to obtain the coupling correction amount. ; For known power frequency harmonic interference, construct an attenuation weight vector. Set a weight value less than 1 at the corresponding harmonic frequency point and several frequency points to its left and right, and set a weight value of 1 at the other frequency points; Coupling correction amount With decay weight vector Element-wise multiplication yields the correction amount applied to the filter coefficients.

10. The method for acquiring power data for submarine optical cables according to claim 1, characterized in that, In S4, the corrected filter coefficients are used to perform a multiplication operation on the frequency domain input signal to obtain the posterior prediction error, which is the frequency domain of the cleaned signal.

Images