Nonlinear equalization method and device based on neural network

Abstract

The invention discloses a nonlinear equalization method and device based on a neural network, and relates to the field of digital signal processing, and the method comprises the steps: constructing a first neural network model based on a full connection layer and a convolutional layer, the first neural network model at least having a one-dimensional convolutional layer branch, and extracting a nonlinear feature from an upper layer output through the convolutional layer branch; determining a training set according to the sampled signals, and sampling and expanding the training set based on a preset sliding window to obtain a first data set; training a first neural network model based on the first data set; and inputting a to-be-tested signal into the trained first neural network model to obtain a nonlinear equalization result of the to-be-tested signal. According to the method, the neural network model is constructed through the full connection layer and the convolutional layer, the neural network model has at least one one-dimensional convolutional layer branch, the complexity is lower than that of the traditional NNLE while nonlinearity is provided for the neural network model, and due to the fact that the convolutional layer branch is of a positive feed structure, the NNLE can have higher robustness.

Description

Technical Field

[0001] The present application relates to the field of digital signal processing, and in particular to a nonlinear equalization method and device based on a neural network. Background Art

[0002] With the development of optical communication systems, the conversion of PAM (Pulse Amplitude Modulation) signals needs to be performed on high-bandwidth optical communication transceiver systems. When the baud rate of the transmission signal reaches the limit of the modulator, performance degradation will occur. At the same time, after a certain distance of optical fiber transmission, the transmission signal will also produce various nonlinear losses due to the dispersion effect, which requires nonlinear compensation.

[0003] There are three main solutions to the nonlinear compensation problem of optical communication transceiver systems: the traditional MLSE (Maximum Likelihood Sequence Estimation) algorithm, the VNLE (Volterra Series-based Nonlinear Equalizer) algorithm, and the NNLE (Neural Network based Nonlinear Equalization) algorithm. Compared with the other two algorithms, the NNLE algorithm uses activation functions instead of high-order terms to simulate nonlinear models, has lower complexity, and shows great potential in improving system performance.

[0004] However, since NNLE is data-driven, its performance is heavily dependent on high-quality signals. In practical applications, limited bandwidth, excessive noise, or severe nonlinearity will affect the quality of the signal, which in turn affects the training of NNLE, resulting in low robustness of NNLE. Summary of the invention

[0005] The present application provides a nonlinear equalization method and device based on a neural network, which can improve the robustness of the NNLE in the prior art.

[0006] In a first aspect, an embodiment of the present application provides a nonlinear equalization method based on a neural network, the method comprising:

[0007] Building a first neural network model based on a fully connected layer and a convolutional layer, wherein the first neural network model has at least one one-dimensional convolutional layer branch, and nonlinear features are extracted from an upper layer output through the convolutional layer branch;

[0008] Determine a training set according to the sampled signal, sample and expand the training set based on a preset sliding window, and obtain a first data set;

[0009] Based on the first data set, training the first neural network model;

[0010] The signal to be tested is input into the trained first neural network model to obtain a nonlinear equalization result of the signal to be tested.

[0011] In combination with the first aspect, in one implementation, constructing a first neural network model based on a fully connected layer and a convolutional layer includes:

[0012] Constructing a first neural network model based on at least one fully connected layer and at least one one-dimensional convolutional layer;

[0013] After the first fully connected layer of the first neural network model, there is at least one one-dimensional convolutional layer.

[0014] In combination with the first aspect, in one implementation, constructing a first neural network model based on at least one fully connected layer and at least one one-dimensional convolutional layer includes:

[0015] Construct the first neural network model based on two fully connected layers and two convolutional layers;

[0016] In the first neural network model, the first fully connected layer is connected to the first convolutional layer, the first convolutional layer is connected to the second convolutional layer and the second fully connected layer, and the input of the second fully connected layer is the output of the first convolutional layer and the second convolutional layer;

[0017] The activation function of the first neural network model is a swish function.

[0018] In combination with the first aspect, in one implementation, the sampled signal includes: a four-level pulse amplitude modulation signal input into a preset optical communication system, and an output signal of a preset optical communication system;

[0019] The output signal of the preset optical communication system is subjected to clock recovery, phase synchronization, and oversampling processing.

[0020] In combination with the first aspect, in one implementation, determining a training set according to the sampled signal includes:

[0021] All four-level pulse amplitude modulation signals are used as feature value arrays of training sets, and all output signals of preset optical communication systems are used as target value arrays of training sets.

[0022] In combination with the first aspect, in one implementation, sampling based on a preset sliding window and expanding the training set to obtain a first data set includes:

[0023] Moving a preset sliding window based on a preset distance, obtaining a set of values ​​of a feature value array and a value of a target value array within the preset sliding window;

[0024] Establishing a mapping between the set of values ​​of the eigenvalue array and the values ​​of the target value array until the tail end of the sliding window coincides with the tail end of the eigenvalue array;

[0025] A first data set is generated based on all mappings.

[0026] In combination with the first aspect, in one implementation, training the first neural network model based on the first data set includes:

[0027] Setting multiple rounds of training, and dividing the first data set into multiple batches based on the amount of data used in each round of training;

[0028] When each round of training is completed, the parameters of the first neural network model obtained from the previous round of training are passed to the next round;

[0029] After all training rounds are completed, the trained first neural network model is obtained.

[0030] In combination with the first aspect, in one implementation, the optimizer used to train the first neural network model includes an Adam optimizer.

[0031] In combination with the first aspect, in one implementation, the loss function used to train the first neural network model includes a mean square error function and a cross entropy loss function.

[0032] In a second aspect, an embodiment of the present application provides a nonlinear equalization device based on a neural network, characterized in that the device comprises:

[0033] A construction module, configured to construct a first neural network model based on a fully connected layer and a convolutional layer, wherein the first neural network model has at least one one-dimensional convolutional layer branch, and nonlinear features are extracted from an upper layer output through the convolutional layer branch;

[0034] A data processing module, used to determine a training set according to the sampled signal, sample and expand the training set based on a sliding window, and obtain a first data set;

[0035] A training module, used for training the first neural network model based on the first data set;

[0036] The equalization module is used to input the signal to be tested into the trained first neural network model to obtain a nonlinear equalization result of the signal to be tested.

[0037] The beneficial effects brought by the technical solution provided in the embodiments of the present application include:

[0038] The present application constructs a neural network model through a fully connected layer and a convolutional layer, and the neural network model has at least one one-dimensional convolutional layer branch, which provides nonlinearity to the neural network model while being less complex than traditional NNLE. Moreover, since the convolutional layer branch is a feedforward structure, the NNLE can be made more robust. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flowchart of a nonlinear equalization method based on a neural network according to an embodiment of the present application;

[0040] Figure 2 This is a structural diagram of a neural network model of a nonlinear equalization method based on a neural network in an embodiment of the present application;

[0041] Figure 3 This is a structural diagram of an optical communication system based on a nonlinear equalization method of a neural network according to an embodiment of the present application;

[0042] Figure 4 This is a sliding window sampling flow chart of a nonlinear equalization method based on a neural network according to an embodiment of the present application. DETAILED DESCRIPTION

[0043] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0044] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0045] First, please refer to Figure 1 , Figure 1 The flowchart of the nonlinear equalization method based on a neural network according to the embodiment of the present application is shown in FIG. The nonlinear equalization method based on a neural network according to the embodiment of the present application comprises the following steps:

[0046] Step S1: constructing a first neural network model based on a fully connected layer and a convolutional layer, wherein the first neural network model has at least one one-dimensional convolutional layer branch, through which nonlinear features are extracted from an upper layer output.

[0047] Step S2: Determine a training set according to the sampled signal, sample and expand the training set based on a preset sliding window, and obtain a first data set.

[0048] Step S3: Based on the first data set, train a first neural network model.

[0049] Step S4: input the signal to be tested into the trained first neural network model to obtain a nonlinear equalization result of the signal to be tested.

[0050] A fully connected neural network is a basic artificial neural network. Its neurons, as the basic primitive units, are often connected front to back. At the same time, the transmission of data between the layers of the neural network is also one-way transmission, without feedback process, so it is also called a feedforward neural network. Each connection of a fully connected neural network will generate a weight, which is learned through the input training data. Each neuron performs weighted operations on these weights and output data, and then obtains the final output through activation functions such as ReLU, sigmoid or tanh.

[0051] Convolutional neural networks are built by imitating the "perception domain" of organisms. They can perform supervised learning and unsupervised learning. They are also a feedforward neural network, and their main function is to complete the convolution operation on data. Compared with fully connected neural networks, convolutional neural networks have less computational complexity and their system complexity is relatively low.

[0052] The present application combines the above two neural network structures, constructs a neural network model through a fully connected layer and a convolutional layer, and the neural network model has at least one one-dimensional convolutional layer branch, which provides nonlinearity to the neural network model while being less complex than traditional NNLE. Moreover, since the convolutional layer branch is a feedforward structure, the NNLE can be made more robust.

[0053] In some embodiments, in the above step S1, a first neural network model is constructed based on a fully connected layer and a convolutional layer, specifically:

[0054] A first neural network model is constructed based on at least one fully connected layer and at least one one-dimensional convolutional layer, and at least one one-dimensional convolutional layer is connected after the first fully connected layer of the first neural network model.

[0055] In some embodiments, the first neural network model is constructed based on at least one fully connected layer and at least one one-dimensional convolutional layer, specifically:

[0056] A first neural network model is constructed based on two fully connected layers and two convolutional layers. In the first neural network model, the first fully connected layer is connected to the first convolutional layer, the first convolutional layer is connected to the second convolutional layer and the second fully connected layer, the input of the second fully connected layer is the output of the first convolutional layer and the second convolutional layer, and the activation function of the first neural network model is the swish function.

[0057] Specifically, the formula of the swish function is as follows:

[0058] Swish(x)=x·sigmoid(βx)(1)

[0059] In formula (1), β is a learnable parameter, x is the input value of the swish activation function, Swish(x) is the output value of the swish activation function, and sigmoid refers to the sigmoid activation function. When β = 0, the function is equivalent to a linear function When β = 1, swish(x) = x·sigmoid(x); when β approaches infinity, the swish function approaches the ReLU function.

[0060] It should be noted that the swish function can often provide a larger gradient during forward transmission, alleviating the problem of gradient vanishing, thereby improving the training efficiency of the neural network model. At the same time, due to the non-monotonic nature of the swish function, it can improve the expressiveness of the neural network model in some intervals, which helps to improve the performance of the model. In addition, due to the learnability of the β parameter, the swish function can smoothly transition from a linear function to a ReLU function, improving the generalization ability of the model.

[0061] In this embodiment, please refer to the specific structure of the first neural network model. Figure 2 , Figure 2 This is a structural diagram of a neural network model of a nonlinear equalization method based on a neural network in an embodiment of the present application.

[0062] exist Figure 2 In the figure, the fully connected layer 1 receives the input data of the neural network model, and the output of the fully connected layer 1 is processed by the swish activation function and used as the input of the convolution layer 1. The convolution layer 1 is connected to the convolution layer 2 and the fully connected layer 2. The convolution layer 2 is a one-dimensional convolution layer branch, and the convolution layer 1 and the convolution layer 2 form a feedforward structure. The output of the convolution layer 1 is processed by the swish activation function and used as the input of the fully connected layer 2 and the convolution layer 2. The output of the convolution layer 2 is processed by the swish activation function and used as the input of the fully connected layer 2. The output of the fully connected layer 2 is the final output result of the neural network model.

[0063] Specifically, the output size of fully connected layer 1 is 40, and the output size of fully connected layer 2 is 1. The convolution kernel size of convolution layer 1 is 1×3, the number of filters is 32, and the convolution kernel size of convolution layer 2 is 1×5, and the number of filters is 64.

[0064] In some embodiments, in the above step S2, the sampled signal includes: a four-level pulse amplitude modulation signal input into a preset optical communication system, and an output signal of a preset optical communication system, and the output signal of the preset optical communication system is subjected to clock recovery, phase synchronization, and oversampling processing.

[0065] In this embodiment, the preset optical communication system is shown in FIG. Figure 3 , Figure 3 This is a structural diagram of an optical communication system based on a nonlinear equalization method of a neural network according to an embodiment of the present application.

[0066] exist Figure 3 In the optical communication system, the optical path and the circuit are two parts. The optical path emits laser light through a laser, and after the polarization of the optical signal is adjusted by a PC (Polarization Controller), it is input into an IM-DD (Intensity Modulation Direct Detection) modulator. The circuit part maps the digital signal of the custom sequence to four different levels through PAM-4 (Pulse Amplitude Modulation with 4levels), and then converts the digital signal into an analog signal and inputs it into a driver. The driver amplifies the analog signal to a level sufficient to drive the IM-DD modulator.

[0067] Afterwards, the IM-DD modulator adjusts the incident laser by sensing the high and low levels in the analog signal, modulates the laser power, and modulates the laser without signal into the laser carrying signal. After the laser carrying the signal passes through the optical fiber and EDFA (Erbium-Doped Fiber Amplifier), it is input into the optical filter, which filters out the noise and unnecessary frequencies in the laser to improve the quality of the optical signal.

[0068] The photodetector receives the processed laser light, and senses the strength of the photocurrent through the PN junction and other structures in the photodetector, forming electrical signals with different voltages. The electrical signal is transmitted to the oscilloscope through the RF transmission line, and the electrical signal is collected by an external computer. After clock calibration, digital carrier recovery, transverse filter equalization, judgment, bit error rate calculation and other operations are performed on the collected signal, the final output signal is obtained.

[0069] In this embodiment, the number of four-level pulse amplitude modulation signals input into the optical communication system is 32678, and the number of output signals collected from the optical communication system is 262144.

[0070] In some embodiments, in the above step S2, determining the training set according to the sampled signal includes the following steps:

[0071] All four-level pulse amplitude modulation signals are used as feature value arrays of training sets, and all output signals of preset optical communication systems are used as target value arrays of training sets.

[0072] In some embodiments, in the above step S2, sampling and expanding the training set based on a preset sliding window to obtain a first data set includes the following steps:

[0073] S201: moving a preset sliding window based on a preset distance, and obtaining a set of values ​​of a feature value array and a value of a target value array in the preset sliding window.

[0074] S202: Establishing a mapping between the set of values ​​of the eigenvalue array and the values ​​of the target value array until the tail end of the sliding window coincides with the tail end of the eigenvalue array, and generating a first data set according to all the mappings.

[0075] In this embodiment, please refer to Figure 4 , Figure 4 This is a flow chart of the sliding window sampling of the nonlinear equalization method based on a neural network in the embodiment of the present application. The specific process of obtaining the first data set based on the preset sliding window sampling and expanding the training set is as follows:

[0076] Step A01: Input the feature value array X and the target value array Y.

[0077] Step A02: Define the sliding window size as w and the input data length as L, where the input data length is determined based on the length of the feature value array X and the target value array Y.

[0078] Step A03: Create empty cell arrays XTrain and YTrain.

[0079] Step A04: i=1+w.

[0080] Step A05: Assign the i-th element to the i+2w-1-th element of X to the i-th element of XTrain, and assign the i+w-1-th element of Y to the i-th element of YTrain.

[0081] Step A06: i=i+1.

[0082] Step A07: Determine whether i is less than Lw. If so, return to step A05; if not, go to step A08.

[0083] Step A08: Output XTrain and YTrain.

[0084] Step A09: End.

[0085] In some embodiments, in the above step S3, training the first neural network model based on the first data set includes the following steps:

[0086] First, multiple rounds of training are set up, and the first data set is divided into multiple batches based on the amount of data used in each round of training.

[0087] When each round of training is completed, the parameters of the first neural network model obtained from the previous round of training are passed to the next round.

[0088] When all training rounds are completed, the trained first neural network model is obtained.

[0089] Specifically, during the training process, the optimizer used is the Adam optimizer. The Adam optimizer adjusts the learning rate of each parameter by calculating the first-order and second-order momentum of the gradient. Specifically, it combines the advantages of the momentum optimization algorithm and the RMSProp optimization algorithm, and can adaptively adjust the learning rate of each parameter while reducing the variance of the gradient update, thereby accelerating the convergence of the model.

[0090] At the same time, during the training process, the number of training rounds is set to 50, the amount of data used in each round of training is 512, the initial learning rate is 1e-3, the learning rate reduction factor is 0.5, the learning rate reduction cycle is 50, and the L2 regularization factor is 1e-5. The above training parameters can increase the robustness of the model while ensuring the performance of the model.

[0091] In some embodiments, the loss function used to train the first neural network model includes a mean square error function and a cross entropy loss function.

[0092] In a second aspect, the present application provides a nonlinear equalization device based on a neural network, which includes a construction module, a data processing module, a training module and an equalization module.

[0093] The building module is used to build a first neural network model based on a fully connected layer and a convolutional layer. The first neural network model has at least one one-dimensional convolutional layer branch, and nonlinear features are extracted from an upper layer output through the convolutional layer branch.

[0094] The data processing module is used to determine a training set according to the sampled signal, sample and expand the training set based on the sliding window, and obtain a first data set.

[0095] The training module is used to train a first neural network model based on a first data set.

[0096] The equalization module is used to input the signal to be tested into the trained first neural network model to obtain a nonlinear equalization result of the signal to be tested.

[0097] It should be noted that the serial numbers of the above-mentioned embodiments of the present application are only for description and do not represent the advantages or disadvantages of the embodiments.

[0098] The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units that are not listed, or optionally includes other steps or units inherent to these processes, methods, products or devices. The terms "first", "second" and "third" are used to distinguish different objects, etc., and do not represent a sequence, nor do they limit "first", "second" and "third" to different types.

[0099] In the description of the embodiments of the present application, "exemplary", "for example" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary", "for example" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary", "for example" or "for example" is intended to present related concepts in a specific way.

[0100] In the description of the embodiments of the present application, unless otherwise specified, “ / ” means or, for example, A / B can mean A or B; the “and / or” in the text is merely a description of the association relationship of associated objects, indicating that three relationships may exist, for example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, in the description of the embodiments of the present application, “multiple” refers to two or more than two.

[0101] In some processes described in the embodiments of the present application, multiple operations or steps that appear in a specific order are included, but it should be understood that these operations or steps may not be executed in the order in which they appear in the embodiments of the present application or in parallel, and the sequence number of the operation is only used to distinguish the different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed in sequence or in parallel, and these operations or steps may be combined.

[0102] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, disk, CD) as described above, and includes a number of instructions for a terminal device to execute the methods described in each embodiment of the present application.

[0103] The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A nonlinear equalization method based on a neural network, characterized in that: The method comprises: Building a first neural network model based on a fully connected layer and a convolutional layer, wherein the first neural network model has at least one one-dimensional convolutional layer branch, and nonlinear features are extracted from an upper layer output through the convolutional layer branch; Determine a training set according to the sampled signal, sample and expand the training set based on a preset sliding window, and obtain a first data set; Based on the first data set, training the first neural network model; The signal to be tested is input into the trained first neural network model to obtain a nonlinear equalization result of the signal to be tested.

2. A nonlinear equalization method based on a neural network as claimed in claim 1, characterized in that: The first neural network model is constructed based on the fully connected layer and the convolutional layer, including: Constructing a first neural network model based on at least one fully connected layer and at least one one-dimensional convolutional layer; After the first fully connected layer of the first neural network model, there is at least one one-dimensional convolutional layer.

3. A nonlinear equalization method based on a neural network as claimed in claim 2, characterized in that: The first neural network model is constructed based on at least one fully connected layer and at least one one-dimensional convolutional layer, comprising: Construct the first neural network model based on two fully connected layers and two convolutional layers; In the first neural network model, the first fully connected layer is connected to the first convolutional layer, the first convolutional layer is connected to the second convolutional layer and the second fully connected layer, and the input of the second fully connected layer is the output of the first convolutional layer and the second convolutional layer; The activation function of the first neural network model is a swish function.

4. A nonlinear equalization method based on a neural network as claimed in claim 1, characterized in that: The sampled signals include: a four-level pulse amplitude modulation signal input into a preset optical communication system, and an output signal of a preset optical communication system; The output signal of the preset optical communication system is subjected to clock recovery, phase synchronization, and oversampling processing.

5. A nonlinear equalization method based on a neural network as claimed in claim 4, characterized in that: The step of determining a training set according to the sampled signal comprises: All four-level pulse amplitude modulation signals are used as feature value arrays of training sets, and all output signals of preset optical communication systems are used as target value arrays of training sets.

6. A nonlinear equalization method based on a neural network as claimed in claim 5, characterized in that: The sampling based on the preset sliding window and expanding the training set to obtain the first data set includes: Moving a preset sliding window based on a preset distance, obtaining a set of values ​​of a feature value array and a value of a target value array within the preset sliding window; Establishing a mapping between the set of values ​​of the eigenvalue array and the values ​​of the target value array until the tail end of the sliding window coincides with the tail end of the eigenvalue array; A first data set is generated based on all mappings.

7. The nonlinear equalization method based on a neural network as claimed in claim 1, characterized in that: The step of training the first neural network model based on the first data set includes: Setting multiple rounds of training, and dividing the first data set into multiple batches based on the amount of data used in each round of training; When each round of training is completed, the parameters of the first neural network model obtained from the previous round of training are passed to the next round; After all training rounds are completed, the trained first neural network model is obtained.

8. A nonlinear equalization method based on a neural network as claimed in claim 7, characterized in that: The optimizer used to train the first neural network model includes an Adam optimizer.

9. A nonlinear equalization method based on a neural network as claimed in claim 7, characterized in that: The loss functions used to train the first neural network model include a mean square error function and a cross entropy loss function.

10. A nonlinear equalization device based on a neural network according to the method of any one of claims 1 to 9, characterized in that: The device comprises: A construction module, configured to construct a first neural network model based on a fully connected layer and a convolutional layer, wherein the first neural network model has at least one one-dimensional convolutional layer branch, and nonlinear features are extracted from an upper layer output through the convolutional layer branch; A data processing module, used to determine a training set according to the sampled signal, sample and expand the training set based on a sliding window, and obtain a first data set; A training module, used for training the first neural network model based on the first data set; The equalization module is used to input the signal to be tested into the trained first neural network model to obtain a nonlinear equalization result of the signal to be tested.

Images