An adaptive time-frequency equalization method and system for submarine cable transmission signals
By using an adaptive time-frequency equalization method, combined with multi-layer wavelet decomposition and neural networks, the nonlinear equalization problem of submarine cable signal transmission systems in complex environments was solved, achieving efficient high and low frequency synchronization and anti-interference capabilities, and improving the nonlinear equalization effect.
Patent Information
- Application Number
- CN202611585291.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-14
- Publication Date
- 2026-11-10
AI Technical Summary
Existing submarine cable signal transmission systems struggle to track channel dynamics in real time under complex submarine cable transmission scenarios, resulting in poor nonlinear equalization performance. In particular, they are unable to simultaneously capture the detailed distortions of high-frequency components and the overall nonlinear trends of low-frequency components in the signal.
An adaptive time-frequency equalization method is adopted. The received optical signal is converted into an electrical signal and digitally processed. Multi-layer wavelet decomposition and neural network are used, combined with the scaling factor of different neurons to perform wavelet transform, capture the low-frequency trend and high-frequency details of the electrical signal, and perform adaptive filtering, noise reduction and time-frequency equalization processing.
It achieves nonlinear equalization with high and low frequency synchronization, improves the nonlinear equalization effect, enhances the system's anti-interference capability against channel fluctuations, and improves the accuracy and efficiency of nonlinear equalization through adaptive threshold matching noise reduction.
Smart Images

Figure CN123074593A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of signal processing, in particular to a method and system for adaptive time-frequency equalization of submarine cable transmission signals. BACKGROUND
[0002] With the rapid development of global cloud computing and big data services, transoceanic data traffic continues to grow, and submarine cables have become an important infrastructure for global information transmission. To meet the growing demand for communication capacity, space division multiplexing (SDM) technology expands system capacity by increasing the number of spatial channels, while reducing the number of relay devices such as optical amplifiers in long-distance submarine cable systems, reducing system construction costs and energy consumption, and has become an important development direction for the next generation of submarine optical cables. Among them, mode division multiplexing (MDM) technology based on orbital angular momentum (OAM) utilizes the orthogonality between different orders of OAM modes to achieve parallel transmission of multiple signals, and is one of the important implementation ways of SDM, which has broad application prospects in the next generation of high-capacity submarine optical communication systems.
[0003] However, the OAM-MDM submarine optical cable system has the characteristics of long transmission distance, large link span and complex operating environment, and the signal is easily affected by fiber nonlinear effects, mode coupling and environmental disturbances during transmission, resulting in random time-varying nonlinear damage, which seriously affects the system transmission performance. Although the existing Volterra equalizer and static neural network equalization method can compensate for part of the nonlinear distortion, the model parameters are fixed and it is difficult to track the dynamic changes of the submarine cable channel in real time, and the equalization performance is limited in complex submarine cable transmission scenarios. Therefore, it is urgent to study nonlinear equalization methods with adaptive adjustment capabilities to realize dynamic perception and real-time compensation of the random time-varying nonlinear damage of the submarine cable channel.
[0004] In recent years, machine learning algorithms have been widely used in nonlinear equalization of OAM-MDM submarine cable transmission signals, but most algorithms are difficult to capture both the detailed distortion of high-frequency components and the overall nonlinear trend of low-frequency components in the signal, resulting in poor nonlinear equalization effect after time-frequency processing. SUMMARY
[0005] The purpose of the embodiments of the present application is to provide a method and system for adaptive time-frequency equalization of submarine cable transmission signals to improve the nonlinear equalization effect of time-frequency processing. The specific technical solutions are as follows:
[0006] In a first aspect, the embodiments of the present application provide a method for adaptive time-frequency equalization of a submarine cable transmission signal, applied to a signal receiving end, comprising:
[0007] Converting the received optical signal into an electrical signal, and performing digital signal processing on the electrical signal to obtain a target signal;
[0008] Sampling the target signal to obtain original feature data;
[0009] De-noising the target signal to obtain a compensation signal, and filtering in the de-noising process using an adaptive threshold calculated based on the target signal;
[0010] Performing sliding window sampling on the compensation signal to obtain compensation feature data;
[0011] Performing multi-layer wavelet decomposition on the compensation feature data to obtain wavelet feature data, the wavelet feature data including feature high-frequency coefficients of each layer and feature low-frequency coefficients of the highest layer, performing wavelet decomposition on the feature low-frequency coefficients of one layer to obtain feature high-frequency coefficients and feature low-frequency coefficients of the next layer;
[0012] Inputting the composite feature data obtained by splicing the original feature data and the wavelet feature data into a neuron;
[0013] Performing wavelet transform on the feature values of each dimension in the composite feature data based on the scale factor and the translation factor of the dimension in the neuron to obtain wavelet transform results, and multiplying the wavelet transform results of each dimension to obtain an output result of the neuron, the scale factor of the preset narrow kernel neuron being less than 1 and greater than 0, and the scale factor of the preset wide kernel neuron being greater than 1;
[0014] Fusing the output results of the neurons and the composite feature data to obtain an electrical signal after time-frequency equalization processing.
[0015] In an embodiment of the present application, the de-noising of the target signal to obtain the compensation signal comprises:
[0016] Performing multi-layer wavelet decomposition calculation on the target signal to obtain de-noising high-frequency coefficients and de-noising low-frequency coefficients of each layer, wherein the de-noising high-frequency coefficients and the de-noising low-frequency coefficients of the first layer are obtained by wavelet decomposition calculation on the target signal, and the de-noising high-frequency coefficients and the de-noising low-frequency coefficients of other layers are obtained by wavelet decomposition calculation on the de-noising low-frequency coefficients of the previous layer;
[0017] Calculating a target standard deviation of the de-noising high-frequency coefficients of the first layer;
[0018] For each layer, based on the standard deviation of the noise-reduced high-frequency coefficients of the layer and the target standard deviation, an adjustment weight of the layer is calculated, and in the noise-reduced high-frequency coefficients of the layer, a target high-frequency coefficient that minimizes the risk value of the unbiased risk estimate is calculated, based on the adjustment weight and the target high-frequency coefficient, an adaptive threshold of the layer is calculated, noise-reduced high-frequency coefficients in the layer that are less than the adaptive threshold in absolute value are set to 0, and noise-reduced high-frequency coefficients in the layer that are greater than or equal to the adaptive threshold in absolute value are multiplied by an adaptive coefficient corresponding to the noise-reduced high-frequency coefficient, wherein the adaptive coefficient is less than 1 and is calculated based on the noise-reduced high-frequency coefficient and the adaptive threshold.
[0019] Based on the noise-reduced low-frequency coefficients and the updated noise-reduced high-frequency coefficients, inverse wavelet decomposition is performed to obtain a compensation signal.
[0020] In an embodiment of the present application, for each dimension, the initial value of the translation factor and / or the initial value of the scale factor of the dimension configured in the neuron are calculated based on the eigenvalue of the dimension.
[0021] In an embodiment of the present application, for each dimension, the initial value of the translation factor of the dimension configured in the neuron is obtained by the following method:
[0022] The eigenvalues of the dimension in the feature data of the input neuron are clustered to obtain a cluster center;
[0023] The cluster center value of the cluster center is determined as the initial value of the translation factor of the dimension.
[0024] In an embodiment of the present application, for each dimension, the initial value of the scale factor of the dimension configured in the neuron is obtained by the following method:
[0025] The standard deviation of the eigenvalue of the dimension is calculated;
[0026] The initial value of the scale factor of the narrow kernel neuron is determined as a first multiple of the standard deviation, and the initial value of the scale factor of the wide kernel neuron is determined as a second multiple of the standard deviation, wherein the first multiple is less than 1 and greater than 0, and the second multiple is greater than 1.
[0027] In an embodiment of the present application, the fusion of the output results of each neuron and the complex feature data to obtain the electric signal after time-frequency equalization processing comprises:
[0028] The output results of each neuron and the complex feature data are weighted and calculated, and a preset network training bias is superimposed to obtain an electric signal after time-frequency equalization processing.
[0029] In an embodiment of the present application, the method further comprises:
[0030] The translation factors and the scale factors of each dimension in each neuron are updated based on a preset loss function by using a back propagation algorithm.
[0031] In a second aspect, the embodiments of the present application provide a system for adaptive time-frequency equalization of a submarine cable transmission signal, comprising: a signal sending end and a signal receiving end, wherein the signal sending end and the signal receiving end are connected by an optical fiber;
[0032] The signal sending end is configured to send an optical signal to the signal receiving end through the optical fiber.
[0033] The signal receiving end is configured to implement the method of any one of the first aspect.
[0034] In a third aspect, the embodiments of the present application provide a device for adaptive time-frequency equalization of a submarine cable transmission signal, which is applied to a signal receiving end and comprises:
[0035] A conversion module is configured to convert a received optical signal into an electrical signal, and perform digital signal processing on the electrical signal to obtain a target signal.
[0036] An original signal sampling module is configured to sample the target signal to obtain original feature data.
[0037] A noise reduction module is configured to reduce noise of the target signal to obtain a compensated signal, and perform filtering on the compensated signal by using an adaptive threshold calculated based on the target signal.
[0038] A sliding window sampling module is configured to perform sliding window sampling on the compensated signal to obtain compensated feature data.
[0039] A decomposition module is configured to perform multi-layer wavelet decomposition on the compensated feature data to obtain wavelet feature data, wherein the wavelet feature data comprises feature high-frequency coefficients of each layer and feature low-frequency coefficients of the highest layer, and the decomposition module is configured to perform wavelet decomposition on the feature low-frequency coefficients of one layer to obtain feature high-frequency coefficients and feature low-frequency coefficients of a next layer.
[0040] An input module is configured to input composite feature data obtained by splicing the original feature data and the wavelet feature data into a neuron.
[0041] A processing module is configured to, for each dimension of feature values in the composite feature data, perform wavelet transform on the feature values of the dimension based on a scale factor and a translation factor of the dimension in the neuron to obtain a wavelet transform result, and multiply the wavelet transform results of all dimensions to obtain an output result of the neuron, wherein the scale factor of a preset narrow kernel neuron is less than 1 and greater than 0, and the scale factor of a preset wide kernel neuron is greater than 1.
[0042] A fusion module is configured to fuse the output results of the neurons and the composite feature data to obtain an electrical signal subjected to time-frequency equalization processing.
[0043] In one embodiment of the present application, the noise reduction module is specifically used for:
[0044] performing multi-layer wavelet decomposition calculation on the target signal to obtain noise reduction high-frequency coefficients and noise reduction low-frequency coefficients of each layer, wherein the noise reduction high-frequency coefficients and noise reduction low-frequency coefficients of the first layer are obtained by performing wavelet decomposition calculation on the target signal, and the noise reduction high-frequency coefficients and noise reduction low-frequency coefficients of other layers are obtained by performing wavelet decomposition calculation on the noise reduction low-frequency coefficients of the previous layer;
[0045] calculating a target standard deviation of the noise reduction high-frequency coefficients of the first layer;
[0046] for each layer, based on the standard deviation of the noise reduction high-frequency coefficients of the layer and the target standard deviation, calculating an adjustment weight of the layer, and in the noise reduction high-frequency coefficients of the layer, calculating a target high-frequency coefficient that minimizes the risk value of the unbiased risk estimate, calculating an adaptive threshold of the layer based on the adjustment weight and the target high-frequency coefficient, setting the noise reduction high-frequency coefficients of the layer whose absolute value is less than the adaptive threshold to 0, and multiplying the noise reduction high-frequency coefficients of the layer whose absolute value is greater than or equal to the adaptive threshold by an adaptive coefficient corresponding to the noise reduction high-frequency coefficient, wherein the adaptive coefficient is less than 1 and is calculated based on the noise reduction high-frequency coefficient and the adaptive threshold;
[0047] performing inverse wavelet decomposition calculation based on the noise reduction low-frequency coefficients and the updated noise reduction high-frequency coefficients to obtain a compensation signal.
[0048] In one embodiment of the present application, for each dimension, the initial value of the translation factor and / or the initial value of the scale factor of the dimension configured in the neuron are calculated based on the feature value of the dimension.
[0049] In one embodiment of the present application, for each dimension, the initial value of the translation factor of the dimension configured in the neuron is obtained by the following module:
[0050] The clustering module is configured to perform clustering processing on the feature value of the dimension in the feature data of the input neuron to obtain a clustering cluster.
[0051] The translation factor calculation module is configured to determine the cluster center value of the clustering cluster as the initial value of the translation factor of the dimension.
[0052] In one embodiment of the present application, for each dimension, the initial value of the scale factor of the dimension configured in the neuron is obtained by the following module:
[0053] The standard deviation calculation module is configured to calculate the standard deviation of the feature value of the dimension.
[0054] The scale factor calculation module is configured to determine an initial value of the scale factor of the narrow-kernel neuron as a first multiple of the standard deviation, and determine an initial value of the scale factor of the wide-kernel neuron as a second multiple of the standard deviation, wherein the first multiple is less than 1 and greater than 0, and the second multiple is greater than 1.
[0055] In an embodiment of the present application, the fusion module is specifically configured to:
[0056] The output results of the neurons and the composite feature data are weighted and calculated, and a preset network training bias is superimposed to obtain the telecommunication signal processed by time-frequency equalization.
[0057] In an embodiment of the present application, the device further comprises:
[0058] The parameter updating module is configured to update the translation factors and the scale factors of the dimensions in the neurons based on a preset loss function by using a back propagation algorithm.
[0059] In a fourth aspect, an embodiment of the present application provides an electronic device, including a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory complete mutual communication through the communication bus;
[0060] The memory is configured to store a computer program.
[0061] The processor is configured to execute the program stored on the memory, and implement the method in any of the first aspect.
[0062] In a fifth aspect, a computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the method in any of the first aspect.
[0063] In a sixth aspect, an embodiment of the present application further provides a computer program product containing instructions, which, when executed on a computer, cause the computer to perform the method in any of the first aspect.
[0064] The embodiment of the present application has the following beneficial effects:
[0065] The adaptive time-frequency equalization method for submarine cable transmission signals provided in the embodiments of the present application has different scale factors configured in different neurons, so that when wavelet transform calculation is performed based on a large scale factor, the low-frequency trend of the electrical signal can be captured, and when wavelet transform calculation is performed based on a small scale factor, the spike, mutation, and high-frequency transient details of the electrical signal can be captured. Thus, the nonlinear trend of the electrical signal can be captured in the whole by combining the high frequency and the low frequency, the nonlinear equalization of the high frequency and the low frequency is realized, and the nonlinear equalization effect is further improved. In addition, the electrical signal is subjected to noise reduction processing before wavelet transform calculation in the present application, so that the noise of the electrical signal is preliminarily corrected, and the frequency energy distribution is reconstructed, and the anti-interference ability of the system to channel fluctuation is further improved. Moreover, the adaptive threshold used in the noise reduction processing is calculated based on the electrical signal, that is, the adaptive threshold matches the condition of the electrical signal itself, and the adaptive filtering and noise reduction based on the adaptive threshold matched with the electrical signal can achieve better noise reduction effect. In addition, the original feature data is retained, the original feature data and the wavelet feature data are spliced and then input into the neuron together, the local time domain information of the original electrical signal can be retained, the wavelet feature data after compensation is complementary, the detail features of the feature data input into the neuron are enriched, and the nonlinear equalization effect of the time-frequency processing after the wavelet transform processing is improved.
[0066] Of course, implementing any product or method of the present application does not necessarily require achieving all the advantages described above at the same time. BRIEF DESCRIPTION OF DRAWINGS
[0067] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other embodiments can also be obtained by those skilled in the art based on these drawings.
[0068] Figure 1 A flowchart of an adaptive time-frequency equalization method for submarine cable transmission signals is provided for the embodiments of the present application;
[0069] Figure 2 A system architecture diagram is provided for the embodiments of the present application;
[0070] Figure 3 A flowchart of an electrical signal compensation process is provided for the embodiments of the present application;
[0071] Figure 4 A flowchart of a multi-layer wavelet decomposition calculation is provided for the embodiments of the present application;
[0072] Figure 5 A flowchart of an inverse wavelet decomposition calculation is provided for the embodiments of the present application;
[0073] Figure 6 This application provides a schematic diagram of a complete process for noise reduction and nonlinear equalization of electrical signals.
[0074] Figure 7 A schematic diagram of the processing flow of a wavelet coefficient adaptive threshold module provided in an embodiment of this application;
[0075] Figure 8 This application provides a schematic diagram of a translation factor initial value calculation process.
[0076] Figure 9 A schematic diagram of the scale factor initial value calculation process provided in this application embodiment;
[0077] Figure 10 A schematic diagram of the processing flow of a multi-core radial basis function-wavelet neural network module provided in an embodiment of this application;
[0078] Figure 11 A schematic diagram of the comparison curves of the first type of bit error rate versus optical power provided for embodiments of this application;
[0079] Figure 12 A schematic diagram of the comparison curves of the second type of bit error rate versus optical power provided for embodiments of this application;
[0080] Figure 13 This is a schematic diagram of power spectral density comparison results provided in an embodiment of this application;
[0081] Figure 14 A schematic diagram of an adaptive time-frequency equalization system for submarine cable transmission signals provided in this application embodiment;
[0082] Figure 15 A schematic diagram of an adaptive time-frequency equalization device for submarine cable transmission signals provided in an embodiment of this application;
[0083] Figure 16 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation
[0084] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art based on this application are within the scope of protection of this application.
[0085] To improve the nonlinear equalization effect of time-frequency processing, this application provides an adaptive time-frequency equalization method and system for submarine cable transmission signals.
[0086] See Figure 1 This application provides an adaptive time-frequency equalization method for submarine cable transmission signals, applied at the signal receiving end. The method includes:
[0087] S101: Convert the received optical signal into an electrical signal, and perform digital signal processing on the electrical signal to obtain the target signal;
[0088] S102: Sample the target signal to obtain the original feature data;
[0089] S103: Noise reduction of the target signal to obtain a compensation signal;
[0090] In noise reduction, an adaptive threshold calculated based on the target signal is used for filtering.
[0091] S104: Obtain compensation feature data by performing sliding window sampling on the compensation signal;
[0092] S105: Perform multi-level wavelet decomposition on the compensated feature data to obtain wavelet feature data;
[0093] Wavelet feature data includes high-frequency coefficients of each layer and low-frequency coefficients of the highest layer. Wavelet decomposition of the compensated feature data yields the high-frequency and low-frequency coefficients of the first layer. Wavelet decomposition of the low-frequency coefficients of the first layer yields the high-frequency and low-frequency coefficients of the next layer.
[0094] S106: Input the composite feature data, formed by concatenating the original feature data and wavelet feature data, into the neuron;
[0095] S107: For the feature value of each dimension in the composite feature data, based on the scaling factor and translation factor of that dimension in the neuron, perform wavelet transform on the feature value of that dimension to obtain the wavelet transform result, and multiply the wavelet transform results of each dimension to obtain the output result of the neuron.
[0096] The scale factor of the preset narrow-nucleus neurons is less than 1 and greater than 0, and the scale factor of the preset wide-nucleus neurons is greater than 1.
[0097] S108: The output results of each neuron are fused with the composite feature data to obtain the time-frequency equalized electrical signal.
[0098] As can be seen from the above, in the adaptive time-frequency equalization method for submarine cable transmission signals provided in this application embodiment, different scale factors are configured in different neurons. Therefore, when performing wavelet transform calculation based on a larger scale factor, the low-frequency trend of the electrical signal can be captured; when performing wavelet transform calculation based on a smaller scale factor, the spikes, abrupt changes, and high-frequency transient details of the electrical signal can be captured. Thus, the nonlinear trend of the electrical signal can be captured holistically by combining high and low frequencies, achieving synchronous nonlinear equalization of high and low frequencies, thereby improving the nonlinear equalization effect. Furthermore, in this application, the electrical signal is first denoised before wavelet transform calculation, thereby initially correcting the noise of the electrical signal and reconstructing the frequency domain energy distribution, further improving the system's anti-interference capability against channel fluctuations. Moreover, the adaptive threshold used in the denoising process is calculated based on the electrical signal, meaning the adaptive threshold matches the electrical signal itself. Adaptive filtering and denoising based on the adaptive threshold matching the electrical signal can achieve better denoising results. In addition, by preserving the original feature data and concatenating it with wavelet feature data before inputting it into the neuron, the local time-domain information of the original electrical signal can be preserved. This complements the compensated wavelet feature data, enriches the detailed features of the feature data input into the neuron, and improves the nonlinear equalization effect of time-frequency processing after wavelet transform.
[0099] See Figure 2 This is a schematic diagram of a system architecture provided in an embodiment of this application.
[0100] The diagram includes an 8-level pulse amplitude modulation (PAM) transmitting module, an orbital angular momentum modulation (OCM) module, an orbital angular momentum demodulation (OCM) module, and an 8-level pulse amplitude modulation (PAM) receiving module. The signal transmission sequence is: 8-level pulse amplitude modulation (PAM) transmitting module - orbital angular momentum modulation (OCM) module - orbital angular momentum demodulation (OCM) module - 8-level pulse amplitude modulation (PAM) receiving module. The 8-level pulse amplitude modulation (PAM) transmitting module and the orbital angular momentum modulation (OCM) module - orbital angular momentum demodulation (OCM) module are the signal transmitting end, and the orbital angular momentum demodulation (OCM) module and the 8-level pulse amplitude modulation (PAM) receiving module are the signal receiving end.
[0101] Specifically, in the 8-level pulse amplitude modulation (PACM) transmitting module, the external-cavity laser (EC laser) outputs a continuous optical carrier. In one example, the EC laser outputs a 1550nm optical carrier, which, after processing by a polarization controller, is fed into a Mach-Zehnder modulator (MZM). A digital signal processor (DSP) generates an 8-level pulse amplitude modulation baseband digital signal, which is output to an arbitrary waveform generator (AWG) to be converted into an analog electrical signal. This signal is then amplified by an electrical amplifier (EA) to drive the MZM. The polarization controller (PC) is used to adjust the polarization state of the input light. The MZM modulator loads the 8-level pulse amplitude modulation electrical signal onto the optical carrier and outputs the modulated optical signal, which is then fed into an optical coupler (OC).
[0102] The optical coupler splits the modulated light into two paths; each path is amplified by an erbium-doped fiber amplifier (EDFA) in the orbital angular momentum modulation module, and beam shaped by a single-mode fiber (SMF); then it passes through a collimator (Col.) and a linear polarizer (LP) in sequence, and the two optical signals are incident on spatial light modulator 1 (SLM1) and spatial light modulator 2 (SLM2) respectively.
[0103] Spatial light modulator 1 and spatial light modulator 2 generate orbital angular momentum vortex beams with different topological charges, respectively. One beam passes through a half-wave plate (HWP), and the other is directly input into a polarization beam combiner (PBC). The two beams are combined by the polarization beam combiner, and then pass through a quarter-wave plate (QWP) for polarization adjustment to obtain a multiplexed orbital angular momentum light signal, which is then sent into a ring-core fiber (RCF).
[0104] Multiplexed orbital angular momentum optical signals introduce noise, nonlinear impairments, and mode crosstalk during transmission in a ring-core fiber. During testing, a 2.3km ring-core fiber can be used to simulate submarine cable transmission.
[0105] When the optical signal arrives at the receiver, it first enters the orbital angular momentum demodulation module; the beam splitter (BS) splits the beam into two paths; the reflector adjusts the optical path; the two paths pass through a quarter-wave plate, vortex phase plate 1 (VPP1) and vortex phase plate 2 (VPP2) respectively to complete the orbital angular momentum mode demultiplexing, separate the beams of different orbital angular momentum modes, and output them to two photodetectors (PDs).
[0106] The photodetector converts the two optical signals into electrical signals; the electrical signals are sent to the oscilloscope (OSC) to complete signal acquisition, obtain sampling data, and send it to the back-end offline digital signal processing link.
[0107] The offline digital signal processing link performs the following processing on the sampled data in sequence: Data Normalization: normalizes the amplitude of the sampled signal; Resampling: completes sampling rate matching; Low-Pass Filtering: filters out out-of-band high-frequency noise; Clock Recovery: restores the clock phase of the received signal; Frame Synchronization: locates the start position of the data frame; Adaptive Nonlinear Equalization, including a wavelet coefficient adaptive threshold module and a multi-kernel radial basis function (RBF)-wavelet neural network (WNN) module; Demapping: restores the equalized sampled values to a bit data stream.
[0108] in, Figure 2 The x[k] shown is the original feature data obtained in step S102, r[k] is the data after noise reduction by the wavelet coefficient adaptive threshold module, the processing performed by the multi-kernel radial basis function-wavelet neural network module corresponds to steps S106-S108, and z[k] is the electrical signal after time-frequency equalization processing.
[0109] In one embodiment, after the received optical signal is converted into an electrical signal, the electrical signal is subjected to at least one of the following digital signal processing methods: oscilloscope sampling processing, data normalization processing, resampling processing, low-pass filtering processing, clock recovery processing, and frame synchronization processing.
[0110] In one embodiment, steps S106-S108 can be implemented by a multi-kernel radial basis function-wavelet neural network. The multi-kernel radial basis function-wavelet neural network includes a feature extraction layer, a wavelet function calculation layer, and an output layer. The feature extraction layer is used to implement step S106, the wavelet function calculation layer is used to implement step S107, and the output layer is used to implement step S108.
[0111] In one embodiment of this application, the target signal is tapped and sampled. The tap window length is 2len. For example, the nth sampled value in the electrical signal is represented as x[n], and the nth data in the original feature data obtained after tapping is X[n]=[x[n-len],…,x[n+len]]. T When the original feature data contains N2 data points, the original feature data is represented in vector form as X=[X[1],…,X[N2]].
[0112] In one embodiment, noise reduction processing of electrical signals can be performed using methods such as Wiener filtering, Kalman filtering, low-pass filtering, memory polynomial equalization, and multilayer perceptron neural networks.
[0113] In one possible embodiment, a sliding window can be used to sample the compensation signal. Adjacent sampling points in the sliding window have a strong correlation. In this application, the subsequent use of wavelet transform calculation can reduce the correlation between the input feature data, so that the variation characteristics of the electrical signal at different scales can be fully represented, thereby enhancing the ability to express nonlinear distortion.
[0114] Let r[n] be the nth data in the compensation signal. In one example, the sliding window length is 2win, and the compensation signal is sampled to obtain the nth data in the compensation feature data: R[n] = [r[n-win], ..., r[n+win]]. T So when the total length of the compensation feature data is N1, the compensation feature data is R=[R[1],…,R[N1]].
[0115] Further, multi-level wavelet decomposition is performed on each data point in the compensation feature data. First, a first-level wavelet decomposition is performed to obtain the first-level high-frequency coefficients D1 and A1. Then, a second-level wavelet decomposition is performed on the first-level low-frequency coefficients A1 to obtain the second-level high-frequency coefficients D2 and A2. Finally, a third-level wavelet decomposition is performed on the second-level low-frequency coefficients A1 to obtain the third-level high-frequency coefficients D3 and A3, and so on. In one example, Discrete Wavelet Transform (DWT) can be used for wavelet decomposition. Other wavelet decomposition algorithms, such as basic wavelet transform decomposition and stationary wavelet transform decomposition, can also be used.
[0116] If a three-level wavelet decomposition is ultimately performed, then for each data point in the compensated feature data, we can obtain C[n]=[D1,D2,D3,A3]. T Finally, the wavelet feature data C=[C[1],…,C[N1]] is obtained, with N dimensions. C .
[0117] The original feature data X is concatenated with the wavelet feature data C to obtain the composite feature data P:
[0118]
[0119] Where P represents composite feature data, C represents wavelet feature data, and X represents original feature data. These are the feature values from the first to the Nth dimension in the original feature data. The number of electrical signals, = +2len+1, 2len represents the dimension of the wavelet feature data, and 2len represents the length of the tap window when the original feature data is obtained through sampling.
[0120] The composite feature data is input into each neuron, and the data input into each neuron is the same.
[0121] The number of narrow-nucleus neurons in a neuron group is the same as the number of wide-nucleus neurons, or the difference between the two is less than a predetermined difference. In other words, neurons are approximately evenly divided into wide-nucleus neuron groups and narrow-nucleus neuron groups based on their number. For example, if the total number of neurons is H, then the index set of wide-nucleus neurons can be represented as... ={1,…,[H / 2]} ={[H / 2]+1,…,H}.
[0122] The scaling factor of wavelet transform controls the temporal support width of the neural nucleus. Narrow-nucleus neurons have a small scaling factor, resulting in a narrow temporal support interval, sharp waveforms, and high temporal resolution. They are sensitive to localized, small-range inputs and are better able to respond to local abrupt changes, transient spikes, and rapid fluctuations in the input signal. They excel at capturing signal details, abrupt changes, noise, and short-term nonlinear distortions. Wide-nucleus neurons have a large scaling factor, resulting in a wide temporal support interval, smooth waveforms, and lower temporal resolution. They can respond to a larger range of input samples, capturing the overall trend of signal changes and excel at capturing global, large-scale nonlinear trends. Therefore, combining the two allows for the capture of both signal details, abrupt changes, noise, and short-term nonlinear distortions, as well as global, large-scale nonlinear trends. This comprehensive capture of both signal details and global nonlinear distortions improves the nonlinear equalization effect in time-frequency processing.
[0123] In one embodiment, the first derivative of the Gaussian wavelet is used as the wavelet function for wavelet transform calculation. Other wavelet functions in related technologies can also be selected for wavelet transform calculation, and this embodiment does not limit this.
[0124] The first derivative of the Gaussian wavelet is expressed as:
[0125]
[0126] in, is the result of calculating the first derivative of the Gaussian wavelet, where x is the independent variable in the input.
[0127] In one embodiment, when performing wavelet transform calculations based on the first derivative of the Gaussian wavelet, the output of the h-th neuron in this application is:
[0128]
[0129] Where P represents the feature data of the input neuron, and h represents the index of the neuron. Let be the output of the h-th neuron, where N is the total number of dimensions of the feature data of the input neurons, and i is the index of the dimension. Let i be the i-th eigenvalue of the n-th feature data. Let i be the translation factor of the i-th dimension configured in the h-th neuron. Let i be the scaling factor configured in the h-th neuron. This indicates that the first derivative of the Gaussian wavelet is being calculated. The wavelet transform result is obtained by performing wavelet transform on the i-th eigenvalue of the n-th feature data from the h-th neuron.
[0130] It should be noted that using the first derivative of the Gaussian wavelet for wavelet transform calculation is only one embodiment. If the first derivative of the Gaussian wavelet is replaced with other wavelet functions, then... After substituting other wavelet functions to obtain the wavelet transform calculation results, the output of the neuron can be obtained by multiplying the results together.
[0131] In one possible embodiment, the outputs of each neuron are summed and then summed with the feature values of each dimension in the original feature data to obtain the time-frequency equalized electrical signal. Since the original feature data is directly fused without any nonlinear activation processing, it can only fit the linear relationship between input and output. This avoids neurons repeatedly learning existing linear features, allowing neurons to focus on modeling the remaining nonlinear distortions.
[0132] In this embodiment, which is implemented using a multi-kernel radial basis function-wavelet neural network, the original feature data is directly input to the output layer through the residual branch and fused with the output of the neuron.
[0133] In one embodiment of this application, see Figure 3 The above-mentioned noise reduction of the target signal to obtain the compensated signal includes:
[0134] S301: Perform multi-level wavelet decomposition on the target signal to obtain the noise reduction high-frequency coefficients and noise reduction low-frequency coefficients of each layer;
[0135] Among them, the noise reduction high-frequency coefficients and noise reduction low-frequency coefficients of the first layer are obtained by wavelet decomposition of the target signal mentioned above, and the noise reduction high-frequency coefficients and noise reduction low-frequency coefficients of the other layers are obtained by wavelet decomposition of the noise reduction low-frequency coefficients of the previous layer.
[0136] S302: Calculate the target standard deviation of the noise reduction high-frequency coefficients of the first layer mentioned above;
[0137] S303: For each layer, based on the standard deviation of the noise reduction high-frequency coefficients of that layer and the target standard deviation mentioned above, calculate the adjustment weight of that layer, and among the noise reduction high-frequency coefficients of that layer, calculate the target high-frequency coefficient that minimizes the risk value of the unbiased risk estimate. Based on the adjustment weights and the target high-frequency coefficients mentioned above, calculate the adaptive threshold of that layer, set the noise reduction high-frequency coefficients in that layer whose absolute value is less than the adaptive threshold to 0, and multiply the noise reduction high-frequency coefficients in that layer whose absolute value is greater than or equal to the adaptive threshold by the adaptive coefficient corresponding to the noise reduction high-frequency coefficient.
[0138] Wherein, the above adaptive coefficient is less than 1, and is calculated based on the noise reduction high-frequency coefficient and the above adaptive threshold;
[0139] S304: The compensation signal is obtained by inverse wavelet decomposition based on the denoised low-frequency coefficients and the updated denoised high-frequency coefficients.
[0140] As can be seen from the above, this embodiment employs wavelet decomposition followed by filtering for noise suppression. The high-frequency coefficients for noise reduction carry the main nonlinear fluctuations and transient characteristics and are more susceptible to frequency-related distortion; therefore, they are the primary focus of filtering. Noise reduction is achieved by modifying the high-frequency coefficients and reconstructing the frequency domain energy distribution, thus improving the system's anti-interference capability against fiber optic, especially submarine cable, channel fluctuations while accurately preserving transient characteristics. Furthermore, this embodiment dynamically adjusts weights and calculates adaptive thresholds based on the actual input electrical signal. Combined with the high-frequency energy distribution of each layer represented by the noise-reduced high-frequency coefficients obtained from wavelet decomposition, the optimal adaptive threshold is adaptively selected, solving the problem of constant deviation caused by traditional fixed filtering thresholds. This allows for dynamic noise suppression while preserving the details of signal abrupt changes, enabling more accurate compensation for nonlinear distortion in subsequent steps, improving the accuracy of signal modeling and recognition, and facilitating subsequent network training.
[0141] In one embodiment, DWT can be used to perform multi-level wavelet decomposition calculations on the electrical signal. Alternatively, other wavelet decomposition algorithms such as basic wavelet transform decomposition and stationary wavelet transform decomposition can be used for wavelet decomposition calculations. There is no limitation on this.
[0142] See Figure 4 The diagram illustrates the process of performing multi-level wavelet decomposition on an electrical signal.
[0143] x[n] is an electrical signal. A first-order Malat decomposition is performed on x[n]. The signal is then processed in parallel using a high-pass filter h[m] and a low-pass filter g[m], and downsampled (represented by ↓2 in the figure) to obtain the high-frequency noise reduction coefficients of the first layer. And the low-frequency coefficient of the first layer of noise reduction After that, A second Marat decomposition is performed, using a high-pass filter h[m] and a low-pass filter g[m] in parallel, and then downsampling is applied to obtain the noise reduction high-frequency coefficients of the second layer. And the low-frequency coefficient of the second layer of noise reduction After that, A third Marat decomposition is performed, with parallel processing through a high-pass filter h[m] and a low-pass filter g[m], followed by downsampling, to obtain the noise reduction high-frequency coefficients of the third layer. And the low-frequency coefficient of the third layer for noise reduction The set of high-frequency coefficients for the three-layer noise reduction is D. j [k].Figure 4 The illustrated embodiment performs three-level wavelet decomposition calculations. This is only an example; in actual scenarios, the number of wavelet decomposition calculation levels can be selected according to requirements.
[0144] The odd function of the filter used for DWT can be the Daubechies 4th order wavelet (db4 wavelet), which has good smoothness, tight support characteristics, and linear phase characteristics. It is suitable for separating signal trends from details. The decomposition process is expressed as follows:
[0145] ;
[0146] ;
[0147] Where L is the length of the filter. Indicates a high-pass filter. This represents a low-pass filter. The formula represents iterating through all the coefficients of a high-pass or low-pass filter from n=0 to n=L-1, and then combining these coefficients with... The displacement samples are multiplied together, where m is the layer index. For the above electrical signal, Let m be the low-frequency noise reduction coefficient of the m-th layer. Let be the high-frequency noise reduction coefficients of the m-th layer, and k be the index of the element in the low-frequency noise reduction coefficients and the high-frequency noise reduction coefficients.
[0148] In one possible implementation, the standard deviation of the high-frequency coefficients of each layer for noise reduction is calculated using the Median Absolute Deviation (MAD) method. Specifically, the standard deviation can be calculated based on the following formula:
[0149]
[0150] in, To calculate the standard deviation, Let be the high-frequency coefficients for noise reduction in the m-th layer, and median() represents the calculation of the median.
[0151] The following formula is used when calculating the target standard deviation:
[0152]
[0153] in, The target standard deviation, This represents the high-frequency coefficient for noise reduction in the first layer.
[0154] The target standard deviation can represent the noise intensity of the second high-frequency coefficient of the first layer.
[0155] Alternatively, other methods in related technologies can be used to calculate the noise standard deviation of the high-frequency coefficients of each layer for noise reduction, which is not limited here.
[0156] To accommodate the differences in energy distribution across different layers, the adjustment weights for each layer can be calculated using the following formula:
[0157]
[0158] Where m is the layer index. Let c be the adjustment weight for the m-th layer, and c be a preset parameter set according to requirements. In one example, it can be 0.2. The target standard deviation, This is a preset, extremely small positive real number, used to prevent the denominator from being 0 without significantly affecting the calculation result. Let m be the high-frequency coefficients for noise reduction in the m-th layer. Let m be the standard deviation of the m-th layer. This indicates that the natural constant e is calculated exponentially.
[0159] In one embodiment, when calculating the target high-frequency coefficient that minimizes the risk value of the unbiased risk estimate, the denoising high-frequency coefficient with the largest value in each layer is first removed, thereby preventing the calculated target high-frequency coefficient from being the maximum value of the denoising high-frequency coefficient in that layer. This results in the calculated adaptive threshold being greater than all denoising high-frequency coefficients in that layer, which in turn causes all denoising high-frequency coefficients in that layer to be set to 0.
[0160] In one possible implementation, for each layer, the elements in the noise reduction high-frequency coefficients of that layer can be sorted by size (e.g., ascending order). The noise reduction high-frequency coefficients of the m-th layer contain... The set of candidate thresholds consists of [number] elements. ,in, ≤ ≤…≤ By iterating through each element in ascending order, the risk value of the unbiased risk estimate is calculated, and the target high-frequency coefficient that minimizes the risk value of the unbiased risk estimate among the noise-reduced high-frequency coefficients of the m-th layer is obtained.
[0161] In another possible implementation, the elements in the noise reduction high-frequency coefficients can be not sorted, and the risk value of the unbiased risk estimate can be calculated directly according to the order of the elements.
[0162] In one embodiment, the risk value of the unbiased risk estimate can be calculated based on the Stan unbiased risk estimation algorithm. In this case, the risk value is calculated based on the following formula:
[0163]
[0164] Where m is the layer index, p is the index of the element in the candidate threshold set, and k is the index of the element in the high-frequency coefficients for denoising. Let m be the risk value of the m-th layer. The target standard deviation, Let p be the p-th element in the candidate threshold set of the m-th layer. The k-th element in the high-frequency coefficients of the m-th layer for noise reduction. Let be the number of elements in the high-frequency coefficients of the m-th layer for noise reduction. express The absolute value is less than or equal to The number of coefficients.
[0165] Other risk estimation methods, such as Monte Carlo-Stein unbiased risk estimation, generalized Stein unbiased risk estimation, and joint Stein unbiased risk estimation, can also be used in other embodiments.
[0166] For each layer, after calculating the target high-frequency coefficient, the adjustment weight is multiplied by the target high-frequency coefficient to obtain the adaptive threshold. Specifically, the adaptive threshold can be calculated using the following formula:
[0167]
[0168] in, The adaptive threshold for the m-th layer. Let m be the adjustment weight of the m-th layer. Let be the target high-frequency coefficient of the m-th layer.
[0169] After obtaining the adaptive thresholds for each layer, the high-frequency coefficients for noise reduction are updated based on the following formula:
[0170]
[0171] in, This refers to the k-th element in the updated high-frequency noise reduction coefficients of the m-th layer. To update the k-th element in the high-frequency coefficients of the previous noise reduction, This is a preset coefficient that can be set according to requirements; in one example, it can be 0.1. The adaptive threshold for the m-th layer. These are adaptive coefficients.
[0172] In this embodiment, when Force it to zero to suppress interference, when At that time, a smoothing and shrinking process is applied to the signal, thereby effectively reducing the constant deviation of the traditional fixed threshold method and improving the accuracy of preserving the amplitude characteristics of the signal edge.
[0173] In one embodiment of this application, an inverse wavelet decomposition method, which is the opposite of the wavelet decomposition method used in wavelet decomposition calculation, is used to perform inverse wavelet decomposition calculation on the denoised low-frequency coefficients and the updated denoised high-frequency coefficients. For example, if DWT is used in wavelet decomposition calculation, then inverse wavelet decomposition calculation is performed using Inverse Discrete Wavelet Transform (IDWT) for reconstruction.
[0174] The specific formula for IDWT reconstruction is as follows:
[0175]
[0176] Where L is the length of the filter. It is the time-reverse version of the high-pass filter in wavelet decomposition. It is the time-reversed version of the low-pass filter in wavelet decomposition. This refers to the k-th element in the (m-1)-th layer of low-frequency coefficients obtained through inverse wavelet decomposition. This is the k-th element in the updated m-th layer of noise reduction high-frequency coefficients.
[0177] See Figure 5 The figure illustrates the inverse wavelet decomposition calculation process in this embodiment. The figure shows the process of performing inverse wavelet decomposition calculation on the denoised high-frequency coefficients after 3-layer update. The number of layers mentioned above is only an example and is not limited in this embodiment.
[0178] It is a collection of three updated noise reduction high-frequency coefficients, including the third layer. The second layer The first layer .
[0179] Firstly adopt (equivalent to the aforementioned) )right After processing, upsampling is performed (represented by ↑2 in the figure), and then... (equivalent to the aforementioned) The low-frequency coefficient of the third layer for noise reduction After processing, upsampling is performed, and then the calculated results are added together to obtain the noise-reduced low-frequency coefficients for the second layer after reconstruction. The above process is equivalent to a Marat reconstruction. Then, the following is adopted: right After processing, upsampling is performed, and then... The low-frequency coefficients of the reconstructed second layer After processing, upsampling is performed, and then the two calculated results are added together to obtain the noise reduction low-frequency coefficients of the first layer after reconstruction. The above process is equivalent to a Marat reconstruction. Then, the following is adopted: right After processing, upsampling is performed, and then... The low-frequency coefficients of the reconstructed first layer After processing, upsampling is performed, and then the two calculated results are added together to obtain the compensation signal r[n]. The above process is equivalent to a Marat reconstruction.
[0180] See Figure 6 This demonstrates a complete process for noise reduction and nonlinear equalization of electrical signals.
[0181] First, regarding D j [k] is decomposed into wavelet values to obtain the first layer D1[k], the second layer D2[k], and the third layer D3[k].
[0182] Figure 6 The diagram shows the detailed process of processing D1[k]. The same operation is performed on D2[k] within the solid box above to obtain T2[k]. The same operation is performed on D3[k] within the solid box above to obtain T3[k]. That is, the same operation as D1[k] is performed.
[0183] The median absolute deviation noise is estimated by using D1[k] only once, and the global noise standard deviation is obtained.
[0184] Furthermore, the first-layer std(D1[k]) is calculated for D1[k], and the weight coefficients (i.e., adjustment weights) are calculated to obtain the first-layer weight coefficients. .
[0185] Additionally, the absolute values of all elements in D1[k] are taken, and then sorted in ascending order. `sort(|D1[k]|)` sorts the absolute values of D1[k] to obtain the candidate threshold set. ,in, ≤ ≤…≤ , =length(D1[k]), where length(D1[k]) is the length of D1[k].
[0186] Take one by one Input all D1[k] to calculate a single Stein's unbiased risk estimation risk, repeated cyclically. -1 time, calculate All of them Stan unbiased risk estimates risk; the smallest Stan unbiased risk estimate is selected. That is, the optimal threshold of the first layer. .
[0187] Calculate the first layer adaptive threshold .
[0188] For each judge ?
[0189] If so, then... If shrinking is not performed, leave it unchanged and output the processed result. for .
[0190] get , , The result of the merger .
[0191] See Figure 7 The processing flow of the wavelet coefficient adaptive threshold module is shown. Figure 7 The x[k] shown is the original feature data obtained in step S102, r[k] is the data after noise reduction and nonlinear equalization by the wavelet coefficient adaptive threshold module, and z[k] is the electrical signal after time-frequency equalization.
[0192] The specific processing flow of the wavelet coefficient adaptive threshold module is as follows:
[0193] First, x[k] is decomposed into W layers using discrete wavelet transform to obtain the set D of noise reduction high-frequency coefficients for each layer. j [k] and A j [k], the weighting coefficients (i.e., adjusted weights) are calculated using the median absolute deviation estimate, and the weighting coefficients and D are then combined. j [k] Input: Stan risk-free estimation - Garrot adaptive shrinking threshold layer; Output: ,Will and A j [k] is the inverse discrete wavelet transform reconstruction layer of layer W, and the output is r[k].
[0194] In one embodiment of this application, for each dimension, the initial values of the translation factor and / or scaling factor configured in the neuron for that dimension are calculated based on the eigenvalues of that dimension.
[0195] As can be seen from the above, the initial values of the translation factor and scaling factor are determined based on the characteristic distribution of the electrical signal, which can capture the dynamic changes of the signal. Through adaptive parameter settings, the parameters calculated by the wavelet transform are matched with the actual characteristic distribution of the electrical signal. Compared with randomly set parameters, this improves computational stability and avoids getting trapped in local optima. Especially for submarine cable signal transmission scenarios with significant nonlinear impairments, the solution provided in this embodiment can achieve better nonlinear equalization, realizing dynamic perception and real-time compensation of random time-varying nonlinear impairments in the submarine cable channel, thereby improving the signal transmission performance of the submarine cable signal.
[0196] See Figure 8 For each dimension, the initial value of the translation factor configured in that dimension in the neuron is obtained through the following steps S801-S802:
[0197] S801: Cluster the feature values of this dimension in the feature data of the input neuron to obtain clusters;
[0198] S802: Determine the cluster center value of the cluster as the initial value of the translation factor for this dimension.
[0199] As can be seen from the above, this embodiment uses clustering to process the feature data of the input neurons and calculates the initial value of the translation factor based on this. This allows the wavelet function to cover the high-density clustering region of the feature data of the input neurons when performing wavelet transform calculation based on the translation factor. This enables the neurons to capture the distribution pattern of the input feature data more accurately and quickly, enhances convergence stability, and improves convergence speed.
[0200] In one embodiment, the K-means clustering algorithm can be used for clustering, or other clustering methods can be used, as detailed in the following formula:
[0201]
[0202] Where h is the index of the neuron, and when the neuron is a narrow-nucleus neuron, This represents the total number of narrow-nucleus neurons; when the neurons are wide-nucleus neurons, Let P be the total number of wide-nucleus neurons, and P be the feature data of the input neurons. Given the feature data set of the h-th neuron as input, Let h be the cluster center vector of the h-th neuron. N is the total number of dimensions of the feature data. Let h be the cluster center value of the h-th neuron in the i-th dimension, and so on. This indicates that the squared Euclidean distance is calculated, where E is the sum of the squared Euclidean distances between the feature data of the input narrow-nucleus neuron or wide-nucleus neuron and the cluster center vector.
[0203] In this embodiment, the calculation yields the result that minimizes E. This involves calculating the cluster centers for each dimension. For narrow-nucleus neurons, the square of the Euclidean distance between the eigenvalues of each dimension and the cluster centers of their respective dimensions is calculated. The sum of the squares of the Euclidean distances for each dimension is then calculated to obtain the Euclidean distance sum. Finally, the sum of the squares of the Euclidean distances for all narrow-nucleus neurons is calculated. The resulting cluster centers minimize the sum of the squares of the Euclidean distances for all narrow-nucleus neurons. For wide-nucleus neurons, the same process is repeated.
[0204] Since the cluster center value is determined as the initial value of the translation factor in this embodiment, the receptive field center of the wavelet transform can be shifted to the cluster center when the initial value of the translation factor is used for wavelet transform. Therefore, the wavelet transform calculation can cover the high-density region of the feature data, enhance the neuron's ability to represent the distribution of feature data, and thus improve the nonlinear equilibrium effect.
[0205] In one embodiment, all wide-nucleus neurons have the same initial translation factor corresponding to the same dimension, and all narrow-nucleus neurons have the same initial translation factor corresponding to the same dimension. The initial translation factors corresponding to the same dimension in wide-nucleus neurons and narrow-nucleus neurons can be the same or different.
[0206] See Figure 9 For each dimension, the initial value of the scaling factor for that dimension configured in the neuron is obtained through the following steps S901-S902:
[0207] S901: Calculate the standard deviation of the eigenvalues for this dimension;
[0208] S902: The initial value of the scaling factor for narrow-nucleus neurons is determined to be a first multiple of the above standard deviation, and the initial value of the scaling factor for wide-nucleus neurons is determined to be a second multiple of the above standard deviation;
[0209] Among them, the first multiple is less than 1 and greater than 0, and the second multiple is greater than 1.
[0210] As can be seen from the above, the scaling factor directly controls the temporal width of the neuron. Setting the initial scaling factor based on the standard deviation of the feature values in each dimension of the input feature data allows the neuron to effectively operate within the actual numerical fluctuation range of the feature values, preventing a large number of feature values from falling outside the neuron's numerical fluctuation range. This, in turn, enables the neuron to effectively extract features from the initial feature data. Furthermore, assigning a smaller scaling factor to narrow-nucleus neurons allows them to accurately capture transient detail distortions in the signal, while assigning a larger scaling factor to wide-nucleus neurons allows them to capture the overall trend of the feature data.
[0211] In one possible embodiment, the first multiple is 0.5 and the second multiple is 3. The values of the first multiple and the second multiple are only one possible value, and this application does not limit them.
[0212] The initial values of the scaling factors for narrow-nucleus neurons and wide-nucleus neurons can be calculated using the following formula:
[0213]
[0214] Where h is the index of the neuron, and i is the index of the dimension. The initial value of the scaling factor for the i-th dimension configured in the h-th neuron. Let be the standard deviation of the i-th eigenvalue. A set of indices for wide-nucleus neurons. Let be the set of indices for narrow-nucleus neurons, where 'a' is a second multiple and 'b' is a first multiple.
[0215] In one embodiment of this application, the above-mentioned fusion of the output results of each neuron and composite feature data to obtain a time-frequency equalized electrical signal includes:
[0216] Step A: The output results of each neuron and the above composite feature data are weighted and calculated, and the preset network training bias is superimposed to obtain the electrical signal after time-frequency equalization processing.
[0217] As can be seen from the above, the weighted calculation method can balance the influence of the calculation results of each channel on the electrical signal after time-frequency equalization and compensate for the bias of network training, so that the final electrical signal after time-frequency equalization is more accurate.
[0218] Step A above can be expressed by the following formula:
[0219]
[0220] in, Let h be the nth element in the electrical signal after time-frequency equalization, h be the index of the neuron, and H be the total number of neurons. Let h be the neuron weight of the h-th neuron. Let be the output of the h-th neuron, where i is the index of the dimension and N is the total number of dimensions. The residual weights for the i-th dimension of the composite feature data are... Let be the feature value of the i-th dimension in the composite feature data, and let bias be the preset network training bias.
[0221] See Figure 10 This is a schematic diagram of the processing flow of the multi-kernel radial basis function-wavelet neural network module.
[0222] Figure 10 The x[k] shown is the original feature data obtained in step S102, r[k] is the data after noise reduction by the wavelet coefficient adaptive threshold module, and z[k] is the electrical signal after time-frequency equalization.
[0223] The multi-kernel radial basis function-wavelet neural network module includes a feature extraction layer, a wavelet function calculation layer (with a total of H neurons), and an output layer.
[0224] In the feature extraction layer, the electrical signal x[n] undergoes local raw tapping to obtain the raw feature data X. The compensation signal r[n] is sampled using a sliding window to obtain the compensation feature data R. The compensation feature data R is decomposed by discrete wavelet transform to obtain wavelet feature data C. X and C are concatenated to obtain the composite feature data P=[C,X]. P is then input into each neuron in the wavelet function calculation layer. The figure contains a wide-nucleus neuron group. and narrow nucleus neuron group Each of the figures contains The circular frame represents a wavelet transform calculation. The initial value of the scaling factor of the i-th dimension in the h-th neuron of the wide-nucleus neuron group is... , The initial value of the scale factor in the i-th dimension of the h-th neuron in the narrow nucleus neuron group is... , .
[0225] The residual branch inputs the composite feature data P into the output layer, and the output layer multiplies the output of the h-th neuron by the neuron weight of the h-th neuron. Multiply the eigenvalue of the i-th dimension of the composite feature data P by the residual weight of the i-th dimension. The weighted results are then summed (∑) to obtain the time-frequency equalized electrical signal z[k].
[0226] In one embodiment of this application, the method further includes:
[0227] Step B: Using the backpropagation algorithm, based on the preset loss function, update the translation factor and scaling factor of each dimension in each neuron.
[0228] As can be seen from the above, this embodiment continuously adjusts the translation and scaling factors of each dimension in the neuron, making the neuron's processing increasingly precise and gradually adapting to the actual characteristic distribution of the input electrical signal. This allows for gradual correction of erroneous responses to signal frequency components and time localization, achieving more accurate training. After training, the received electrical signal can be restored to a near-ideal state after time-frequency equalization, completing the nonlinear equalization of the damaged electrical signal.
[0229] In one embodiment, the aforementioned preset loss function can be a loss function that minimizes the mean squared error (MSE) between the time-frequency equalized electrical signal and the ideal sequence. In one example, the preset loss function can be defined as:
[0230]
[0231] in, Let N be the loss value, N be the number of electrical signals processed, and i be the index of the electrical signal. Let i be the true value of the i-th electrical signal when it is not affected by nonlinear distortion. It is the i-th electrical signal after time-frequency equalization processing.
[0232] The translation factor and scaling factor can be updated using gradient descent. In addition, the neuron weights of each neuron, the residual weights of each dimension, and the network training bias can also be updated.
[0233] The loss function above is just one example and can be replaced with other common loss functions in this field.
[0234] See Figure 11 The figure shows a comparison curve of the first type of bit error rate as a function of optical power.
[0235] Figure 11 The figure shows the comparison curves of bit error rate (BER) of different OAM modes (|l|=2, |l|=3) as a function of received optical power under the following conditions: 2.3 km RCF transmission, 32-GBaud-PAM8 modulation, sliding window length of 101, tap window length of 5, number of neurons of 50, learning rate of 5e-4, number of iterations of 500, training data of 30%, and test data of 60%.
[0236] Figure 11The circular dot curve is a curve obtained by non-linear equalization using the Volterra equalizer, the triangular dot curve is a curve obtained by non-linear equalization using the traditional WNN, and the square dot curve is a curve obtained by using the embodiment of this application.
[0237] Specifically, at the 25% Soft-decision Forward Error Correction (SD-FEC) threshold of 2.4 × 10⁻², for the OAM mode with |l| = 2, this model reduces the required received optical power from 3.4 dBm in Volterra to 3.1 dBm, achieving a sensitivity improvement of 0.3 dB. Since the OAM mode with |l| = 3 can easily meet the 25% SD-FEC threshold at the received optical power (ROP) setting...
[0238] See Figure 12 The figure shows a comparison curve of the second type of bit error rate as a function of optical power.
[0239] Figure 12 The verification environment and Figure 1 The verification environment is the same. Figure 12 The circular dot curve is a curve obtained by non-linear equalization using the Volterra equalizer, the triangular dot curve is a curve obtained by non-linear equalization using the traditional WNN, and the square dot curve is a curve obtained by using the embodiment of this application.
[0240] The 7% hard-decision forward error correction (HD-FEC) threshold is 3.8 × 10⁻⁶. -3 In this embodiment, the required received optical power is reduced from 7.8dBm to 4.5dBm, resulting in a gain of 3.3dB.
[0241] Figure 13 A schematic diagram of power spectral density comparison results is shown.
[0242] Figure 13 The image shows the PSD comparison results when the OAM mode is |l|=3 and the received optical power is 2dBm. Figure 13 The horizontal axis represents the normalized frequency (×π radians / sampling point), and the vertical axis represents the power spectral density (dB / Hz). Figure 13 The gray curve, which slopes from the upper left to the lower right, corresponds to a signal that is affected by nonlinear distortion and has not been equalized. The gray curve, which is centered, corresponds to a signal that is not affected by nonlinear distortion. The black curve corresponds to a signal that has undergone time-frequency equalization processing according to the embodiments of this application. The dashed curve corresponds to a signal that has undergone nonlinear equalization using the Volterra equalizer.
[0243] As can be seen, the unequalized signal exhibits obvious spectral deviation and distortion. Compared with the signal that is nonlinearly equalized using a Volterra equalizer, the signal that is nonlinearly equalized using the scheme provided in this application has a higher curve overlap with the signal that is not affected by nonlinear distortion. This proves that the embodiments of this application can achieve more accurate nonlinear equalization in complex OAM communication environments.
[0244] Corresponding to the aforementioned adaptive time-frequency equalization method for submarine cable transmission signals, this application also provides an adaptive time-frequency equalization system for submarine cable transmission signals.
[0245] See Figure 14 This application provides a schematic diagram of an adaptive time-frequency equalization system for submarine cable signal transmission. The system includes a signal transmitter 1401 and a signal receiver 1402, which are connected by an optical fiber 1403.
[0246] The aforementioned signal transmitting end 1401 is used to transmit optical signals to the signal receiving end through the optical fiber;
[0247] The aforementioned signal receiver 1402 is used to implement the adaptive time-frequency equalization method for submarine cable transmission signals as described above.
[0248] As can be seen from the above, in the adaptive time-frequency equalization method for submarine cable transmission signals provided in this application embodiment, different scale factors are configured in different neurons. Therefore, when performing wavelet transform calculation based on a larger scale factor, the low-frequency trend of the electrical signal can be captured; when performing wavelet transform calculation based on a smaller scale factor, the spikes, abrupt changes, and high-frequency transient details of the electrical signal can be captured. Thus, the nonlinear trend of the electrical signal can be captured holistically by combining high and low frequencies, achieving synchronous nonlinear equalization of high and low frequencies, thereby improving the nonlinear equalization effect. Furthermore, in this application, the electrical signal is first denoised before wavelet transform calculation, thereby initially correcting the noise of the electrical signal and reconstructing the frequency domain energy distribution, further improving the system's anti-interference capability against channel fluctuations. Moreover, the adaptive threshold used in the denoising process is calculated based on the electrical signal, meaning the adaptive threshold matches the electrical signal itself. Adaptive filtering and denoising based on the adaptive threshold matching the electrical signal can achieve better denoising results. In addition, by preserving the original feature data and concatenating it with wavelet feature data before inputting it into the neuron, the local time-domain information of the original electrical signal can be preserved. This complements the compensated wavelet feature data, enriches the detailed features of the feature data input into the neuron, and improves the nonlinear equalization effect of time-frequency processing after wavelet transform.
[0249] Corresponding to the aforementioned adaptive time-frequency equalization method for submarine cable transmission signals, this application also provides an adaptive time-frequency equalization device for submarine cable transmission signals.
[0250] See Figure 15 This is a schematic diagram of an adaptive time-frequency equalization device for submarine cable transmission signals provided in an embodiment of this application. It is applied at the signal receiving end and includes:
[0251] The conversion module 1501 converts the received optical signal into an electrical signal, and performs digital signal processing on the electrical signal to obtain the target signal;
[0252] The original signal sampling module 1502 samples the target signal to obtain the original feature data;
[0253] The noise reduction module 1503 denoises the target signal to obtain a compensation signal. In the noise reduction process, an adaptive threshold calculated based on the target signal is used for filtering.
[0254] The sliding window sampling module 1504 performs sliding window sampling on the compensation signal to obtain compensation feature data;
[0255] The decomposition module 1505 performs multi-level wavelet decomposition on the compensation feature data to obtain wavelet feature data. The wavelet feature data includes the high-frequency coefficients of each layer and the low-frequency coefficients of the highest layer. Wavelet decomposition of the compensation feature data yields the high-frequency coefficients and low-frequency coefficients of the first layer. Wavelet decomposition of the low-frequency coefficients of the first layer yields the high-frequency coefficients and low-frequency coefficients of the next layer.
[0256] Input module 1506 inputs composite feature data, which is formed by concatenating the original feature data and wavelet feature data, into the neuron;
[0257] The processing module 1507 performs wavelet transform on each feature value in the composite feature data, based on the scaling factor and translation factor of that dimension in the neuron, to obtain the wavelet transform result. The wavelet transform results of each dimension are multiplied to obtain the output result of the neuron. The scaling factor of the preset narrow-nucleus neuron is less than 1 and greater than 0, and the scaling factor of the preset wide-nucleus neuron is greater than 1.
[0258] The fusion module 1508 fuses the output results of each neuron with the composite feature data to obtain the time-frequency equalized electrical signal.
[0259] As can be seen from the above, in the adaptive time-frequency equalization method for submarine cable transmission signals provided in this application embodiment, different scale factors are configured in different neurons. Therefore, when performing wavelet transform calculation based on a larger scale factor, the low-frequency trend of the electrical signal can be captured; when performing wavelet transform calculation based on a smaller scale factor, the spikes, abrupt changes, and high-frequency transient details of the electrical signal can be captured. Thus, the nonlinear trend of the electrical signal can be captured holistically by combining high and low frequencies, achieving synchronous nonlinear equalization of high and low frequencies, thereby improving the nonlinear equalization effect. Furthermore, in this application, the electrical signal is first denoised before wavelet transform calculation, thereby initially correcting the noise of the electrical signal and reconstructing the frequency domain energy distribution, further improving the system's anti-interference capability against channel fluctuations. Moreover, the adaptive threshold used in the denoising process is calculated based on the electrical signal, meaning the adaptive threshold matches the electrical signal itself. Adaptive filtering and denoising based on the adaptive threshold matching the electrical signal can achieve better denoising results. In addition, by preserving the original feature data and concatenating it with wavelet feature data before inputting it into the neuron, the local time-domain information of the original electrical signal can be preserved. This complements the compensated wavelet feature data, enriches the detailed features of the feature data input into the neuron, and improves the nonlinear equalization effect of time-frequency processing after wavelet transform.
[0260] In one embodiment of this application, the noise reduction module 1503 is specifically used for:
[0261] The target signal is subjected to multi-level wavelet decomposition calculation to obtain the noise reduction high-frequency coefficients and noise reduction low-frequency coefficients of each layer. The noise reduction high-frequency coefficients and noise reduction low-frequency coefficients of the first layer are obtained by wavelet decomposition calculation of the target signal, and the noise reduction high-frequency coefficients and noise reduction low-frequency coefficients of the other layers are obtained by wavelet decomposition calculation of the noise reduction low-frequency coefficients of the previous layer.
[0262] Calculate the target standard deviation of the noise reduction high-frequency coefficients of the first layer;
[0263] For each layer, an adjustment weight is calculated based on the standard deviation of the noise reduction high-frequency coefficients of that layer and the target standard deviation. Among the noise reduction high-frequency coefficients of that layer, a target high-frequency coefficient that minimizes the risk value of the unbiased risk estimate is calculated. An adaptive threshold for that layer is calculated based on the adjustment weight and the target high-frequency coefficient. Noise reduction high-frequency coefficients in that layer whose absolute value is less than the adaptive threshold are set to 0. Noise reduction high-frequency coefficients in that layer whose absolute value is greater than or equal to the adaptive threshold are multiplied by the adaptive coefficient corresponding to the noise reduction high-frequency coefficient. The adaptive coefficient is less than 1 and is calculated based on the noise reduction high-frequency coefficient and the adaptive threshold.
[0264] The compensation signal is obtained by inverse wavelet decomposition based on the denoised low-frequency coefficients and the updated denoised high-frequency coefficients.
[0265] As can be seen from the above, this embodiment employs wavelet decomposition followed by filtering for noise suppression. The high-frequency coefficients for noise reduction carry the main nonlinear fluctuations and transient characteristics and are more susceptible to frequency-related distortion; therefore, they are the primary focus of filtering. Noise reduction is achieved by modifying the high-frequency coefficients and reconstructing the frequency domain energy distribution, thus improving the system's anti-interference capability against fiber optic, especially submarine cable, channel fluctuations while accurately preserving transient characteristics. Furthermore, this embodiment dynamically adjusts weights and calculates adaptive thresholds based on the actual input electrical signal. Combined with the high-frequency energy distribution of each layer represented by the noise-reduced high-frequency coefficients obtained from wavelet decomposition, the optimal adaptive threshold is adaptively selected, solving the problem of constant deviation caused by traditional fixed filtering thresholds. This allows for dynamic noise suppression while preserving the details of signal abrupt changes, enabling more accurate compensation for nonlinear distortion in subsequent steps, improving the accuracy of signal modeling and recognition, and facilitating subsequent network training.
[0266] In one embodiment of this application, for each dimension, the initial values of the translation factor and / or scaling factor configured in the neuron for that dimension are calculated based on the eigenvalues of that dimension.
[0267] As can be seen from the above, the initial values of the translation factor and scaling factor are determined based on the characteristic distribution of the electrical signal, which can capture the dynamic changes of the signal. Through adaptive parameter settings, the parameters calculated by the wavelet transform are matched with the actual characteristic distribution of the electrical signal. Compared with randomly set parameters, this improves computational stability and avoids getting trapped in local optima. Especially for submarine cable signal transmission scenarios with significant nonlinear impairments, the solution provided in this embodiment can achieve better nonlinear equalization, realizing dynamic perception and real-time compensation of random time-varying nonlinear impairments in the submarine cable channel, thereby improving the signal transmission performance of the submarine cable signal.
[0268] In one embodiment of this application, for each dimension, the initial value of the translation factor configured in the neuron for that dimension is obtained through the following module:
[0269] The clustering module is used to perform clustering processing on the feature values of this dimension in the feature data of the input neurons to obtain clusters;
[0270] The translation factor calculation module is used to determine the initial value of the translation factor for that dimension as the cluster center value of the cluster.
[0271] As can be seen from the above, this embodiment uses clustering to process the feature data of the input neurons and calculates the initial value of the translation factor based on this. This allows the wavelet function to cover the high-density clustering region of the feature data of the input neurons when performing wavelet transform calculation based on the translation factor. This enables the neurons to capture the distribution pattern of the input feature data more accurately and quickly, enhances convergence stability, and improves convergence speed.
[0272] In one embodiment of this application, for each dimension, the initial value of the scaling factor configured in the neuron for that dimension is obtained through the following module:
[0273] The standard deviation calculation module is used to calculate the standard deviation of the eigenvalues for this dimension.
[0274] The scaling factor calculation module is used to determine the initial value of the scaling factor of narrow-nucleus neurons as a first multiple of the standard deviation, and to determine the initial value of the scaling factor of wide-nucleus neurons as a second multiple of the standard deviation, wherein the first multiple is less than 1 and greater than 0, and the second multiple is greater than 1.
[0275] As can be seen from the above, the scaling factor directly controls the temporal width of the neuron. Setting the initial scaling factor based on the standard deviation of the feature values in each dimension of the input feature data allows the neuron to effectively operate within the actual numerical fluctuation range of the feature values, preventing a large number of feature values from falling outside the neuron's numerical fluctuation range. This, in turn, enables the neuron to effectively extract features from the initial feature data. Furthermore, assigning a smaller scaling factor to narrow-nucleus neurons allows them to accurately capture transient detail distortions in the signal, while assigning a larger scaling factor to wide-nucleus neurons allows them to capture the overall trend of the feature data.
[0276] In one embodiment of this application, the fusion module 1508 is specifically used for:
[0277] The output results of each neuron and the composite feature data are weighted and calculated, and a preset network training bias is superimposed to obtain an electrical signal after time-frequency equalization processing.
[0278] As can be seen from the above, the weighted calculation method can balance the influence of the calculation results of each channel on the electrical signal after time-frequency equalization and compensate for the bias of network training, so that the final electrical signal after time-frequency equalization is more accurate.
[0279] In one embodiment of this application, the apparatus further includes:
[0280] The parameter update module is used to update the translation factor and scaling factor of each dimension in each neuron using the backpropagation algorithm and based on a preset loss function.
[0281] As can be seen from the above, this embodiment continuously adjusts the translation and scaling factors of each dimension in the neuron, making the neuron's processing increasingly precise and gradually adapting to the actual characteristic distribution of the input electrical signal. This allows for gradual correction of erroneous responses to signal frequency components and time localization, achieving more accurate training. After training, the received electrical signal can be restored to a near-ideal state after time-frequency equalization, completing the nonlinear equalization of the damaged electrical signal.
[0282] This application also provides an electronic device, such as... Figure 16 As shown, it includes a processor 1601, a communication interface 1602, a memory 1603, and a communication bus 1604, wherein the processor 1601, the communication interface 1602, and the memory 1603 communicate with each other through the communication bus 1604.
[0283] Memory 1603 is used to store computer programs;
[0284] The processor 1601 is used to execute the program stored in the memory 1603 to implement the aforementioned adaptive time-frequency equalization method for submarine cable transmission signals.
[0285] As can be seen from the above, in the adaptive time-frequency equalization method for submarine cable transmission signals provided in this application embodiment, different scale factors are configured in different neurons. Therefore, when performing wavelet transform calculation based on a larger scale factor, the low-frequency trend of the electrical signal can be captured; when performing wavelet transform calculation based on a smaller scale factor, the spikes, abrupt changes, and high-frequency transient details of the electrical signal can be captured. Thus, the nonlinear trend of the electrical signal can be captured holistically by combining high and low frequencies, achieving synchronous nonlinear equalization of high and low frequencies, thereby improving the nonlinear equalization effect. Furthermore, in this application, the electrical signal is first denoised before wavelet transform calculation, thereby initially correcting the noise of the electrical signal and reconstructing the frequency domain energy distribution, further improving the system's anti-interference capability against channel fluctuations. Moreover, the adaptive threshold used in the denoising process is calculated based on the electrical signal, meaning the adaptive threshold matches the electrical signal itself. Adaptive filtering and denoising based on the adaptive threshold matching the electrical signal can achieve better denoising results. In addition, by preserving the original feature data and concatenating it with wavelet feature data before inputting it into the neuron, the local time-domain information of the original electrical signal can be preserved. This complements the compensated wavelet feature data, enriches the detailed features of the feature data input into the neuron, and improves the nonlinear equalization effect of time-frequency processing after wavelet transform.
[0286] The communication bus 1604 mentioned in the above electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus 1604 can be divided into address bus, data bus, control bus, etc. For ease of illustration, it is represented by only one thick line in the figure, but this does not indicate that there is only one bus or one type of bus.
[0287] Communication interface 1602 is used for communication between the above-mentioned electronic device and other devices.
[0288] The memory 1603 may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory 1603 may also be at least one storage device located remotely from the aforementioned processor.
[0289] The processor 1601 mentioned above can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.
[0290] In another embodiment provided in this application, a computer-readable storage medium is also provided, which stores a computer program that, when executed by a processor, implements the steps of any of the above-described adaptive time-frequency equalization methods for submarine cable transmission signals.
[0291] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the adaptive time-frequency equalization methods for submarine cable transmission signals in the above embodiments.
[0292] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).
[0293] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0294] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system, apparatus, electronic device, computer-readable storage medium, and computer program product embodiments are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.
[0295] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.
Claims
1. An adaptive time-frequency equalization method for submarine cable transmission signals, characterized in that, Applied to the signal receiving end, including: The received optical signal is converted into an electrical signal, and the electrical signal is then processed digitally to obtain the target signal. The original feature data is obtained by sampling the target signal; The target signal is denoised to obtain a compensation signal. The denoising process uses an adaptive threshold calculated based on the target signal for filtering. The compensation feature data is obtained by performing sliding window sampling on the compensation signal; Multi-level wavelet decomposition of the compensation feature data yields wavelet feature data, which includes the high-frequency coefficients of each level and the low-frequency coefficients of the highest level. Wavelet decomposition of the compensation feature data yields the high-frequency and low-frequency coefficients of the first level, and wavelet decomposition of the low-frequency coefficients of the first level yields the high-frequency and low-frequency coefficients of the next level. The composite feature data, formed by concatenating the original feature data and wavelet feature data, is input into the neuron; For each feature value in the composite feature data, based on the scaling factor and translation factor of that dimension in the neuron, wavelet transform is performed on the feature value of that dimension to obtain the wavelet transform result. The wavelet transform results of each dimension are multiplied to obtain the output result of the neuron. The scaling factor of the preset narrow-nucleus neuron is less than 1 and greater than 0, and the scaling factor of the preset wide-nucleus neuron is greater than 1. The output results of each neuron are fused with the composite feature data to obtain the time-frequency equalized electrical signal.
2. The method according to claim 1, characterized in that, The noise reduction of the target signal to obtain the compensated signal includes: The target signal is subjected to multi-level wavelet decomposition calculation to obtain the noise reduction high-frequency coefficients and noise reduction low-frequency coefficients of each layer. The noise reduction high-frequency coefficients and noise reduction low-frequency coefficients of the first layer are obtained by wavelet decomposition calculation of the target signal, and the noise reduction high-frequency coefficients and noise reduction low-frequency coefficients of the other layers are obtained by wavelet decomposition calculation of the noise reduction low-frequency coefficients of the previous layer. Calculate the target standard deviation of the noise reduction high-frequency coefficients of the first layer; For each layer, an adjustment weight is calculated based on the standard deviation of the noise reduction high-frequency coefficients of that layer and the target standard deviation. Among the noise reduction high-frequency coefficients of that layer, a target high-frequency coefficient that minimizes the risk value of the unbiased risk estimate is calculated. An adaptive threshold for that layer is calculated based on the adjustment weight and the target high-frequency coefficient. Noise reduction high-frequency coefficients in that layer whose absolute value is less than the adaptive threshold are set to 0. Noise reduction high-frequency coefficients in that layer whose absolute value is greater than or equal to the adaptive threshold are multiplied by the adaptive coefficient corresponding to the noise reduction high-frequency coefficient. The adaptive coefficient is less than 1 and is calculated based on the noise reduction high-frequency coefficient and the adaptive threshold. The compensation signal is obtained by inverse wavelet decomposition based on the denoised low-frequency coefficients and the updated denoised high-frequency coefficients.
3. The method according to claim 1 or 2, characterized in that, For each dimension, the initial values of the translation factor and / or scaling factor configured in the neuron for that dimension are calculated based on the eigenvalues of that dimension.
4. The method according to claim 3, characterized in that, For each dimension, the initial value of the translation factor configured in that dimension within the neuron is obtained as follows: Clustering is performed on the feature values of this dimension in the feature data of the input neurons to obtain clusters; The cluster center value of the cluster is determined as the initial value of the translation factor for this dimension.
5. The method according to claim 3, characterized in that, For each dimension, the initial value of the scaling factor for that dimension configured in the neuron is obtained as follows: Calculate the standard deviation of the eigenvalues for this dimension; The initial value of the scaling factor for narrow-nucleus neurons is determined as a first multiple of the standard deviation, and the initial value of the scaling factor for wide-nucleus neurons is determined as a second multiple of the standard deviation, wherein the first multiple is less than 1 and greater than 0, and the second multiple is greater than 1.
6. The method according to claim 1 or 2, characterized in that, The fusion of the outputs of each neuron and the composite feature data yields a time-frequency equalized electrical signal, including: The output results of each neuron and the composite feature data are weighted and calculated, and a preset network training bias is superimposed to obtain an electrical signal after time-frequency equalization processing.
7. The method according to claim 1 or 2, characterized in that, The method further includes: The backpropagation algorithm is used to update the translation factor and scaling factor of each dimension in each neuron based on a preset loss function.
8. An adaptive time-frequency equalization system for submarine cable transmission signals, characterized in that, The system includes a signal transmitting end and a signal receiving end, which are connected by optical fiber communication. The signal transmitting end is used to transmit optical signals to the signal receiving end through the optical fiber; The signal receiving end is used to implement the method according to any one of claims 1-7.
9. An adaptive time-frequency equalization device for submarine cable transmission signals, characterized in that, Applied to the signal receiving end, including: The conversion module converts the received optical signal into an electrical signal, and performs digital signal processing on the electrical signal to obtain the target signal; The raw signal sampling module samples the target signal to obtain raw feature data; The noise reduction module reduces the noise of the target signal to obtain a compensation signal. The noise reduction process uses an adaptive threshold calculated based on the target signal for filtering. The sliding window sampling module performs sliding window sampling on the compensation signal to obtain compensation feature data; The decomposition module performs multi-level wavelet decomposition on the compensation feature data to obtain wavelet feature data. The wavelet feature data includes the high-frequency coefficients of each layer and the low-frequency coefficients of the highest layer. Wavelet decomposition of the compensation feature data yields the high-frequency and low-frequency coefficients of the first layer. Wavelet decomposition of the low-frequency coefficients of the first layer yields the high-frequency and low-frequency coefficients of the next layer. The input module inputs composite feature data, formed by concatenating the original feature data and wavelet feature data, into the neuron; The processing module performs wavelet transform on each feature value in the composite feature data, based on the scaling factor and translation factor of that dimension in the neuron, to obtain the wavelet transform result. The wavelet transform results of each dimension are multiplied to obtain the output result of the neuron. The scaling factor of the preset narrow-nucleus neuron is less than 1 and greater than 0, and the scaling factor of the preset wide-nucleus neuron is greater than 1. The fusion module combines the output results of each neuron with composite feature data to obtain an electrical signal after time-frequency equalization processing.
10. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the method described in any one of claims 1-7.
