CNN-BiLSTM-based low-complexity coherent optical communication system damage compensation method

Abstract

The invention relates to the technical field of coherent optical communication signal processing, in particular to a CNN-BiLSTM-based low-complexity coherent optical communication system damage compensation method, which comprises the following steps of: processing a to-be-processed dual-polarization coherent optical signal by adopting a pre-trained CNN-BiLSTM neural network model for low-complexity optical fiber transmission damage compensation, a balanced polarization state signal is obtained; according to the method, dispersion compensation and polarization mode dispersion are mutually constrained and cooperatively optimized in a mode of combining one-dimensional convolution and bidirectional LSTM, so that the model can process dispersion compensation and polarization equilibrium compensation at the same time; one-dimensional convolution is adopted to calculate the phase compensation amount, a dispersion compensation filter of a traditional algorithm is replaced to realize dispersion compensation, dependence on data is reduced, and accuracy is improved; a bidirectional LSTM structure is adopted, a polarization sensing time sequence attention mechanism is combined, and collaborative optimization of dynamic damage balance and low-power-consumption real-time processing is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of coherent optical communication signal processing, and specifically relates to a low-complexity coherent optical communication system impairment compensation method based on CNN-BiLSTM. Background Art

[0002] Chromatic Dispersion (CD) and Polarization Mode Dispersion (PMD) in optical fiber transmission are the main factors restricting signal quality. In a coherent optical communication system, traditional dispersion compensation algorithms rely on frequency domain equalization techniques. For example, the frequency domain dispersion compensation algorithm requires multiple executions of Fourier transform, inverse Fourier transform, and polynomial interpolation, resulting in the computational complexity increasing exponentially with the symbol rate. In addition, the coupling effect of chromatic dispersion and polarization mode dispersion causes traditional algorithms to require step-by-step compensation, with significant residual errors.

[0003] In the prior art, the electrical domain algorithms for dispersion compensation and polarization equalization mainly focus on reducing the number of taps of the filter to reduce the computational complexity. However, due to the separate compensation design of the two, there is still room for complexity reduction. And part of the machine learning fusion compensation technology is dedicated to studying the number of filter taps trained by the neural network model. Its direction of reducing complexity is still limited to step-by-step compensation. Another part of the research selects to model the optical fiber channel transfer function through a neural network. Such methods will bring huge training computational costs due to their large amount of training.

[0004] Therefore, there is an urgent need for an optical fiber signal processing method that can simultaneously process dispersion compensation and polarization equalization compensation with low computational cost. Summary of the Invention

[0005] In view of this, the present invention discloses a low-complexity coherent optical communication system impairment compensation method based on CNN-BiLSTM to solve the above problems, including: obtaining a dual-polarization state coherent optical signal to be processed, and using a pre-trained CNN-BiLSTM neural network model for low-complexity optical fiber transmission impairment compensation to process the dual-polarization state coherent optical signal to be processed to obtain an equalized polarization state signal;

[0006] Further, the CNN-BiLSTM neural network model for low-complexity optical fiber transmission impairment compensation includes, connected in sequence:

[0007] A signal preprocessing module, configured to receive the dual-polarization state coherent optical signal to be processed and preprocess the signal, and decompose the dual-polarization state coherent optical signal into four real number signals;

[0008] One-dimensional convolutional compensation module, based on the CNN neural network, extracts the time-domain features of four signals to obtain the time-domain features of dispersion and the phase compensation amount;

[0009] Bidirectional time-series modeling module, adopting the BiLSTM structure, flattens the time-domain features of dispersion, and further extracts the evolution law of the polarization state based on the time-series attention mechanism to obtain spatio-temporal features;

[0010] Damage compensation output module, maps the spatio-temporal features into compensation signals according to the phase compensation amount, and outputs the equalized polarization state signals.

[0011] The beneficial effects of the present invention include:

[0012] By adopting the bidirectional LSTM structure, the hidden state dynamically captures the dual-polarization coupling features, and combines with the time-series attention mechanism of polarization perception to realize the collaborative optimization of dynamic damage equalization and low-power real-time processing, solving the problems of limited compensation accuracy and high computational complexity caused by module fragmentation in traditional solutions;

[0013] By deeply embedding the optical communication physical model into the network structure and training objectives, using one-dimensional convolution learning to obtain the phase compensation amount to replace the dispersion compensation filter of the traditional algorithm to achieve dispersion compensation, reducing the dependence on data and improving the accuracy;

[0014] By adopting the functional coupling design, combining mechanisms such as parameter sharing and residual connection, the model can simultaneously process dispersion compensation and polarization equalization compensation, and make the dispersion compensation and polarization mode dispersion equalization constrain and optimize each other, solving the problem of increased computational complexity brought by the traditional separate design;

[0015] By introducing the polarization state weight, directly encoding the polarization state correlation law through numerical-driven learning. Compared with the traditional method using the MMA equalizer to update the filter coefficients by iteratively calculating the error signal, it avoids the time delay and memory overhead brought by iterative calculation. Description of the Drawings

[0016] Figure 1 It is a schematic structural diagram of the CNN-BiLSTM neural network model for low-complexity optical fiber transmission damage compensation in the present invention;

[0017] Figure 2 It is a schematic structural diagram of the numerical-driven simulation experiment in the embodiment of the present invention;

[0018] Figure 3 It is a comparison diagram of the effects of pre-training the CNN-BiLSTM neural network model for low-complexity optical fiber transmission damage compensation with different activation functions in the embodiment of the present invention;

[0019] Figure 4It is a comparison chart of the convergence speed of the two-way time series modeling module with different numbers of layers in the embodiments of the present invention;

[0020] Figure 5 It is a comparison chart of the phase of the compensation signal obtained by the method of the present invention and the traditional method in the embodiments of the present invention;

[0021] Figure 6 It is a curve graph showing the change of the bit error rate with the received optical power under different polarization states by the method of the present invention and the traditional method in the embodiments of the present invention. Detailed implementation manners

[0022] In order to make the purpose, technical solutions, features and advantages of the present invention clearer and more understandable, the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0023] This embodiment includes a low-complexity coherent optical communication system impairment compensation method based on CNN-BiLSTM. This embodiment uses a dual-polarization 16QAM modulation coherent optical system. As Figure 2 shown, the dual-polarization 16QAM modulation coherent optical system includes a transmitter, an optical fiber transmission channel, and a receiver.

[0024] In this embodiment, the dual-polarization coherent optical signal to be processed is obtained from the transmitter and the optical fiber transmission channel, and the low-complexity compensation method applied to the DSP at the receiving end of the coherent optical communication system is implemented based on the receiver.

[0025] The transmitter performs the same signal processing on the optical signals of the two polarization states respectively, generates a string of pseudo-random binary bit sequences, and then generates a Barker code and inserts it at the front end of the signal as a preamble for the receiver to perform preamble synchronization. The Barker code is a coding method used to represent complex signals in optical fiber communication, which can effectively reduce the interference in the signal and enhance the stability of the signal. Subsequently, the signal is subjected to 16QAM modulation and shaping filtering. Then the signal is sent into an IQ Mach-Zehnder modulator for optical modulation to obtain an optical signal for transmission in the optical fiber.

[0026] Furthermore, before the signal enters the optical fiber transmission channel, the two optical signals are multiplexed onto two polarization states by a polarization beam splitter PBS, and the multiplexed signal enters the optical fiber for transmission. The optical fiber relay uses an EDFA for amplification to obtain the dual-polarization coherent optical signal to be processed.

[0027] The receiver includes a digital signal processor, which is deployed with the CNN-BiLSTM neural network model for low-complexity optical fiber transmission impairment compensation designed by the present invention. The received signal is successively subjected to IQ imbalance compensation, the CNN-BiLSTM neural network model for low-complexity optical fiber transmission impairment compensation, matched filtering, clock recovery, frequency offset and phase offset recovery, preamble synchronization, phase ambiguity processing in the receiver digital signal processor, and finally 16QAM demodulation to obtain the baseband signal.

[0028] Specifically, as Figure 1 shown, the CNN-BiLSTM neural network model for low-complexity optical fiber transmission impairment compensation includes, connected in sequence: a signal preprocessing module, which is used to receive the dual-polarization state coherent optical signal to be processed and preprocess the signal, and decompose the dual-polarization state coherent optical signal into four real-number signals; a one-dimensional convolution compensation module, based on the CNN neural network, extracts the time-domain features of the four signals to obtain the time-domain features of chromatic dispersion and the phase compensation amount; a bidirectional time series modeling module, adopting the BiLSTM structure, flattens the time-domain features of chromatic dispersion, and further extracts the evolution law of the polarization state based on the time series attention mechanism to obtain spatio-temporal features; an impairment compensation output module, which maps the spatio-temporal features into a compensation signal according to the phase compensation amount and outputs the equalized polarization state signal.

[0029] The processing of the signal preprocessing module for the data includes: using a polarization beam splitter to separate the dual-polarization state coherent optical signal to be processed into two complex signals of I / Q; respectively separating the real part signal and the imaginary part signal of the two complex signals of I / Q to obtain four signals, and processing the four signals by using Min-Max sliding window normalization to obtain four real-number signals. The formula for Min-Max sliding window normalization is:

[0030] x normal =(x - x min ) / (x max - x min )

[0031] where x min represents the minimum value of the sample data, x max represents the maximum value of the sample data, and maps the signal result x normal to between [0 - 1].

[0032] The one-dimensional convolution compensation module includes, connected in sequence: a convolution layer, a 20% dropout layer and a ReLU activation function layer; the convolution layer includes 4 parallel one-dimensional convolution channels, and each channel uses a 128-dimensional convolution kernel. The one-dimensional convolution compensation module is equivalent to a time-domain chromatic dispersion compensation filter, and the frequency-domain response of the equivalent time-domain chromatic dispersion compensation filter satisfies the formula:

[0033]

[0034] Among them, h CD (z, t) represents the amplitude response of the time-domain dispersion compensation filter, j represents the imaginary unit, D represents the fiber dispersion coefficient, λ represents the optical wavelength, c represents the speed of light, and z represents the propagation distance of the signal on the optical fiber.

[0035] Furthermore, the traditional time-domain dispersion compensation filter obtains the filter transfer function through the fast Fourier transform, and the formula is:

[0036]

[0037] Among them, H CD [k] represents the amplitude response of the traditional time-domain dispersion compensation filter, T s represents the sampling period, L represents the fiber length, k represents the wave number, β2 represents the second-order dispersion of the fiber, and the calculation formula of β2 is:

[0038]

[0039] The bidirectional time-series modeling module includes a cascaded forward LSTM layer and a backward LSTM layer, both with a dimension of 32. The polarization state evolution equation of the bidirectional time-series modeling module is:

[0040]

[0041] Among them, j represents the imaginary unit, β2 represents the group velocity dispersion parameter, Δβ represents the birefringence coefficient, and E xI 、E xQ 、E yI 、E yQ represent four-channel real number signals.

[0042] Specifically, the polarization state evolution equation is the physical constitutive equation of the dynamic evolution of the polarization state in the optical fiber, which describes the polarization state coupling law caused by chromatic dispersion (CD, characterized by β2) and birefringence effect (characterized by Δβ) during the transmission of the optical signal. By adding this equation as a physical regularization term to the loss function, it guides the BiLS TM to learn the polarization state evolution mode that conforms to the actual optical fiber transmission law, rather than purely relying on data-driven. The hidden state of the BiLSTM is still calculated and generated by the standard LSTM unit through the input gate, forget gate, and output gate mechanisms, and the equation does not directly participate in the numerical calculation of the hidden state. During pre-training, it is mandatory that the polarization state features (such as the hidden state) output by the BiLSTM satisfy the physical laws described by the equation.

[0043] The processing of data by the bidirectional time-series modeling module includes: propagating the dispersion time-domain features forward and backward based on the time order to obtain the forward hidden state sequence and the backward hidden state sequence, calculating the time-series attention weights at time t through a trainable parameter matrix, and weighted-summing the bidirectional hidden state sequences according to the time-series attention weights to obtain the spatio-temporal features at time t.

[0044] The formula used to calculate the time-series attention weights is:

[0045]

[0046] where α t represents the hidden state weight at time t, and W q and W k represent the trainable parameter matrices, represents the forward hidden state sequence, represents the backward hidden state sequence.

[0047] The formula used to calculate the spatio-temporal features is:

[0048]

[0049] where E BiLSTM represents the spatio-temporal features.

[0050] The hidden state weight α t dynamically adjusts the coupling strength of the X / Y polarization state signals by quantifying the influence of the polarization state evolution at different times on the current compensation, thereby suppressing the polarization crosstalk caused by PMD. For example, in high-speed optical fiber transmission, if the Y-polarization signal strongly interferes with the X-polarization at a certain time, the weight will automatically reduce the contribution of the Y-polarization feature at that time.

[0051] Furthermore, the traditional MMA equalizer needs to update the filter coefficients by iteratively calculating the error signal, while the polarization state weight directly encodes the polarization state correlation law through numerical-driven learning, avoiding the time delay and memory overhead caused by iterative calculation.

[0052] The damage compensation output module uses a complex-domain residual connection to map the spatio-temporal features, and the formula is:

[0053]

[0054] where E comp represents the equalized polarization state signal, E in represents the four-channel real signal, j represents the imaginary unit, θ CNN represents the phase compensation amount, and E BiLSTM represents the spatio-temporal features.

[0055] Furthermore, the CNN-BiLSTM neural network model for low-complexity optical fiber transmission damage compensation uses the root mean square error (RMSE) as the loss function during pre-training. The formula is as follows:

[0056]

[0057] where y i represents the true value of the pre-training sample, represents the predicted value of the pre-training sample, and n represents the number of pre-training samples.

[0058] Table 1. Simulation system parameters

[0059]

[0060] As shown in Table 1, the parameter settings for the simulation test using a dual-polarization 16QAM modulation coherent optical system in this embodiment are based on the parameters in Table 1 for testing. It can be seen from Figure 3 that when the model is trained using the RELU function, the convergence performance of its loss function is the best, taking into account both the convergence speed and the later model fitting ability. Moreover, the model using the RELU function also shows better performance during subsequent generalization testing.

[0061] It can be seen from Figure 4 that the trained model compensates for the signal for joint chromatic dispersion (CD) and polarization mode dispersion (PMD), and compares it with the signal compensated by the traditional algorithm. From the compensation result of the real part of the X-polarized signal shown in the figure, it can be seen that the network model, through its powerful reconstruction ability, compensates the uncompensated signal. This figure mainly interprets the design choices of the network design parameters, and the value settings are derived from the best results in the simulation experiment. Here, the original file can choose this figure to show the signal reconstruction ability of the model to verify the effectiveness of the compensation effect.

[0062] To illustrate the advantages of the present invention in this embodiment, a comparison is made between the traditional time-domain dispersion compensation algorithm and the CMA polarization equalization algorithm. The input signal in the algorithm is a dual-polarization complex signal with a length of 8192. The computational complexity is measured by counting the number of real arithmetic operations. Since complex multiplication requires 4 real multiplications and 2 real additions or subtractions, and complex addition requires 2 real additions. Therefore, complex operations can be represented by the corresponding number of real operations, and the computational complexity is approximately 4 times. Regarding the statistical calculation complexity of the traditional algorithm, the number of real calculations is statistically analyzed according to the above algorithm principle.

[0063] Table 2. Computational complexity comparison

[0064]

[0065] Table 2 shows the comparison of the computational amount between the method of the present invention and the traditional method. It can be seen that the computational amount of the method of the present invention is about 52.122% of that of the traditional compensation algorithm, reducing the computational complexity by about 48%. Moreover, since iterative operations are not required, memory resources are saved, and the hardware and energy requirements are reduced.

[0066] From Figure 5 it can be seen that there is little difference in the effect of the signal phase compensated by the method proposed by the present invention and the signal phase compensated by the traditional algorithm, and subsequent normal demodulation can be achieved.

[0067] From Figure 6 it can be seen that in the case of a large-capacity transmission task with a low noise level, the average bit error rate of the method proposed by the present invention is about 1.41×10 -3 , below the hard decision threshold of 3.6×10 -3 . Compared with the average bit error rate of 1.71×10 -3 of the signal compensated by the traditional algorithm, its performance is better than the average bit error rate of the signal compensated by the traditional algorithm. This shows that the method proposed by the present invention can effectively and accurately compensate for the signal distortion caused by CD and PMD in the optical fiber transmission process in an environment without interference or with a low interference level, achieving a high transmission quality.

[0068] Finally, it should be noted that the above description only presents some embodiments of the present invention. For those skilled in the art, various changes, modifications, substitutions, and variations of these embodiments can be conceived without departing from the principle and spirit of the present invention. The protection scope of the present invention is defined by the appended claims and their equivalents, and the above actions should all be covered within the protection scope of the present invention.

Claims

1. A low-complexity coherent optical communication system impairment compensation method based on CNN-BiLSTM, characterized in that Including: Obtain the dual-polarization coherent optical signal to be processed, and use the pre-trained CNN-BiLSTM neural network model for low-complexity optical fiber transmission damage compensation to process the dual-polarization coherent optical signal to be processed, and obtain an equalized polarization state signal; The CNN-BiLSTM neural network model for low-complexity optical fiber transmission damage compensation includes, connected in sequence: A signal preprocessing module, configured to receive the dual-polarization coherent optical signal to be processed and preprocess the signal, and decompose the dual-polarization coherent optical signal into four real-number signals; A one-dimensional convolutional compensation module, based on the CNN neural network, extracts time-domain features of the four signals to obtain the time-domain features of chromatic dispersion and the phase compensation amount; A bidirectional time-series modeling module, using the BiLSTM structure, flattens the time-domain features of chromatic dispersion, and further extracts the evolution law of the polarization state based on the time-series attention mechanism to obtain spatio-temporal features; A damage compensation output module, maps the spatio-temporal features into a compensation signal according to the phase compensation amount, and outputs an equalized polarization state signal.

2. The method for compensating impairments of a low-complexity coherent optical communication system based on CNN-BiLSTM according to claim 1, wherein, The processing of the data by the signal preprocessing module includes: using a polarization beam splitter to separate the dual-polarization coherent optical signal to be processed into two I / Q complex signals; respectively separating the real part signal and the imaginary part signal of the two I / Q complex signals to obtain four signals, and using Min-Max sliding window normalization to process the four signals to obtain four real-number signals.

3. The method for compensating impairments of a low-complexity coherent optical communication system based on CNN-BiLSTM according to claim 1, wherein The one-dimensional convolutional compensation module includes a convolutional layer, a 20% dropout layer, and a ReLU activation function layer connected in sequence; the convolutional layer includes 4 parallel one-dimensional convolutional channels, and each channel uses a 128-dimensional convolutional kernel.

4. The method for compensating impairments of a low-complexity coherent optical communication system based on CNN-BiLSTM according to claim 1, wherein The one-dimensional convolutional compensation module is equivalent to a time-domain chromatic dispersion compensation filter, and the frequency-domain response of the equivalent time-domain chromatic dispersion compensation filter satisfies: Among them, h CD (z, t) represents the amplitude response of the time-domain dispersion compensation filter, D represents the fiber dispersion coefficient, λ represents the optical wavelength, c represents the speed of light, and z represents the propagation distance of the signal on the optical fiber.

5. The method for compensating impairments of a low-complexity coherent optical communication system based on CNN-BiLSTM according to claim 1, wherein The bidirectional time-series modeling module includes a cascaded forward LSTM layer and a backward LSTM layer, both with a dimension of 32. The polarization state evolution equation used by the bidirectional time-series modeling module is: Among them, β2 represents the group velocity dispersion parameter, Δβ represents the birefringence coefficient, E xI 、E xQ 、E yI 、E yQ represent four-channel real signals.

6. The method for compensating impairments of a low-complexity coherent optical communication system based on CNN-BiLSTM according to claim 5, wherein The processing of the data by the bidirectional time-series modeling module includes: propagating the time-domain features of chromatic dispersion forward and backward based on the time order to obtain a forward hidden state sequence and a backward hidden state sequence, calculating the time-series attention weight at time t through a trainable parameter matrix, and weighted summing the bidirectional hidden state sequences according to the time-series attention weight to obtain the spatio-temporal feature at time t; the formulas for calculating the time-series attention weight and the spatio-temporal feature are: Among them, α t represents the hidden state weight at time t, W q and W k represent trainable parameter matrices, represents the forward hidden state sequence, represents the backward hidden state sequence, E BiLSTM represents spatio-temporal features.

7. The method for compensating impairments of a low-complexity coherent optical communication system based on CNN-BiLSTM according to claim 1, wherein The damage compensation output module maps the spatio-temporal features using complex-domain residual connection, and the formula is: Among them, E comp represents an equalized polarization state signal, and E in represents a four-channel real signal, θ CNN represents the phase compensation amount, and E BiLSTM represents the spatio-temporal characteristics.

8. The method for compensating impairments of a low-complexity coherent optical communication system based on CNN-BiLSTM according to claim 1, wherein The CNN-BiLSTM neural network model for low-complexity optical fiber transmission damage compensation uses the root mean square error RMSE as the loss function during pre-training.

Images