OFDM (Orthogonal Frequency Division Multiplexing) end-to-end communication system with environment adaptive capability

Abstract

The invention belongs to the technical field of communication, and particularly relates to an OFDM (Orthogonal Frequency Division Multiplexing) end-to-end communication system with environment self-adaptive capability. According to the invention, the existing E2E communication system is combined with the CP-OFDM technology, so that the E2E system based on OFDM can effectively resist the multipath effect. In addition, the problem of environmental adaptability of the E2E system in practical application is solved by utilizing a meta-learning technology: firstly, pre-training of a noise simulation network is realized through an MAML technology, so that the noise simulation network can quickly adapt to a changing environment and capture environmental noise; thirdly, widely training the E2E receiving and transmitting network by using noise distribution captured by the channel simulation network, so that the E2E system has relatively high adaptability to channel fading and change of a received signal-to-noise ratio; finally, the OFDM-E2E communication system based on the environment adaptive training strategy can adapt to changes of three dimensions of channel noise, channel fading and a received signal-to-noise ratio, so that the performance close to the optimal performance is always kept in practical application.

Description

Technical Field

[0001] The invention belongs to the field of communication technology, and in particular relates to an OFDM end-to-end communication system with environmental self-adaptation capability. Background Art

[0002] End-to-end communication technology, which has emerged in recent years, is a communication system design method that uses neural networks to build autoencoders and uses deep learning technology to jointly optimize the transmitter and receiver. It is an attempt to break through the limitations of traditional wireless communication physical layer module-by-module optimization and achieve E2E optimization of the entire communication link. In theory, it can obtain the communication waveform with the best global performance (T.O'Shea and J.Hoydis, "An introduction to deep learning for the physical layer," IEEE Trans.Cognit.Commun.Netw., vol.3, no.4, pp.563–575, Dec.2017.). However, the traditional E2E communication system is essentially a single-carrier communication, which will show obvious performance flatness in complex multipath channels. OFDM technology is widely used in wireless communication systems due to its high spectrum efficiency and robustness to channel frequency selectivity. By expanding the existing E2E system into an OFDM architecture with a cyclic prefix (CP), the OFDM-based E2E system can have the same advantages as the traditional OFDM system (A.Felix, S.Cammerer, S. J. Hoydis, and S.ten Brink, “OFDM-autoencoder for end-to-endlearning of communications systems,” in Proc.IEEE 19th Int.Workshop SignalProcess.Adv.Wireless Commun.(SPAWC), Jun.2018, pp.1–5.).

[0003] As a data-driven deep learning technology, the training results of the E2E communication system depend on the samples in the training set. In actual scenarios, the environment encountered by the E2E system during application may follow a different distribution from the environment samples provided during training, and the actual channel environment will also change dynamically. At this time, the performance of the E2E system will deviate from the optimal and may even fail to meet the application requirements. Meta-learning is an idea of ​​training neural networks, the purpose of which is to let the system learn how to learn. In this way, the system can quickly learn new tasks based on existing knowledge. A simple example is to let the neural network learn multiple binary classification tasks, such as cats and dogs, apples and pears, etc. In this way, the network will learn how to distinguish between the two categories of things, and then quickly adapt when encountering other binary classification tasks. Model-Agnostic Meta-Learning (MAML) is a method to achieve meta-learning. Its core idea is to find a very good model initialization parameter through the meta-learning process, and use this initialization parameter as the starting point for training a new model. Only a small number of samples (small samples) are needed to make the model converge quickly when facing tasks similar to those encountered in the meta-learning process (C. Finn, P. Abbeel, and S. Levine, "Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks," arXiv e-prints, p.arXiv:1703.03400, Mar. 2017.). However, the existing E2E system is essentially a single-carrier communication system, which will show obvious performance flattening under complex multipath channels. Summary of the invention

[0004] Aiming at the problem that the existing E2E system has poor performance in complex multipath channels and poor adaptability to channel changes, the present invention combines the existing E2E architecture with OFDM technology, and proposes an environment adaptive training strategy for the E2E communication system based on meta-learning. Finally, an OFDM end-to-end communication system with environment adaptive capability is designed, which can maintain good performance in complex multipath channels and has strong robustness to channel changes.

[0005] The technical solution of the present invention is:

[0006] An OFDM end-to-end communication system with environmental adaptability is constructed by a neural network, including a transmitting network, a noise simulation network, a preprocessing network and a demodulation network. Their parameters are defined as Θ t , Θ s , Θ p and θ de, where the transmitting network is responsible for extracting the features of the information sent by the source and sending it to the channel; the noise simulation network and the preprocessing network are used to match the harsh non-Gaussian noise channel environment. The noise simulation network is only used during system training and is responsible for providing sufficient non-Gaussian noise samples for the joint training of the transceiver network. The preprocessing network is used to suppress the impulse noise component in the received signal and reduce the impact of the fading channel on the receiver demodulation; the demodulation network is responsible for restoring the original information data; the communication process of the system is: define the original transmitted information By FFT Individual messages i Composition, i=1,2,...w FFT , for a single message s i , which consists of a bit sequence of length k, representing one of M possible symbols, where M is the modulation order; the output of s after passing through the transmitting network is a 2-channel real-valued signal, which is regarded as the real and imaginary parts of the complex signal X respectively. FFT messages are assigned to each subcarrier, and the OFDM symbol X is obtained. OFDM After OFDM modulation is completed, add a l-length CP The CP of the transmitted symbol x is obtained. After passing through the channel and noise simulation network, the corresponding inverse process is performed at the receiving end. First, the CP of the received signal y is removed to obtain Y OFDM , then for Y OFDM OFDM demodulation is performed to obtain the frequency domain signal Y. The real and imaginary parts of the complex signal Y are converted into 2-channel real-valued signals, and then passed through the preprocessing network and the demodulation network in turn to restore the original message

[0007] In order to minimize the error between the transmitted data and the data decoded by the receiving end and to adapt to the dynamically changing environment, the system is trained in the following ways:

[0008] 1) Train the noise simulation network. The noise simulation network adopts an adversarial generation network structure that can generate pulses. The loss function of the generator G is:

[0009]

[0010] Among them, E[x] represents the mathematical expectation of the random variable x, θ g With θ d Represent all trainable parameters of the generator G and the discriminator D respectively, z represents a set of random vectors of the generator G input that obey a certain distribution, n represents the environmental noise vector, g represents the simulated noise vector of the generator, α 1 and α 2 Represent the weights of the two regularization terms, ε is a small enough positive number to make 1+ε approach 1, α 1 , α 2and ε are adjustable parameters;

[0011] The loss function of the discriminator D is:

[0012]

[0013] Among them, gp(·) represents the gradient penalty term, λ gp represents the weight of the gradient penalty term, is a convex combination of g and n, The combination coefficient t is sampled from the uniform distribution U[0,1], xp x (x) represents the element x in x i Independent and identically distributed under the given distribution p x (·), Represents the gradient of function f with respect to variable x;

[0014] The noise simulation network is trained using a meta-learning-based training strategy. Specifically, first, T different tasks are sampled from the offline task set. Then build a training set for these tasks With the test set The first stage of training is to iterate over different tasks, using the training set in each iteration. Try to pre-update the parameters of the neural network and then use the test set Calculate the loss of the pre-updated network; when all tasks are traversed and completed, combine all the test losses calculated for the pre-updated network and perform a real update on the original network; through multiple iterations, until the system converges to a fixed indicator, the meta-model of the noise simulation network is obtained;

[0015] The meta-model of the noise simulation network is deployed in the actual channel environment and the noise distribution in the environment is captured to obtain the adapted noise simulation network model;

[0016] 2) Generate a simulated noise training set using the adapted noise simulation network model to train the system's transceiver network:

[0017] Initialize the optimization algorithm Adam, learning rate α, batch size b, number of training rounds e, number of subcarriers w of OFDM symbol FFT , channel distribution set Noise Distribution Set Transmit power set Transmitter network parameters θ t0 , preprocessing network parameters θ p0 , demodulation network parameters θ de0 ;

[0018] Randomly generate a set of 0-1 information bits As the transmission data;

[0019] Make the transmitted data s get the signal through the transmitting network Then perform IFFT on X and add CP to obtain the signal x to be transmitted;

[0020] The channel coefficients and the received signal-to-noise ratio are sampled from the training set, and the transmitted signal x is converted into power according to the signal-to-noise ratio and then passes through the channel fading, and then Gaussian noise or simulated mixed Gaussian pulse noise g generated by the noise simulation network is superimposed to obtain the received signal y;

[0021] After removing the CP and performing FFT on the received signal y, it passes through the preprocessing network and the demodulation network in turn, and finally the recovered information data is obtained.

[0022] Calculate s and E2E loss between

[0023]

[0024] At the same time, the gradient of each network parameter is calculated, and the parameters of the three networks are updated using the Adam optimizer;

[0025] After repeated training, the network parameters are deployed in the system for communication.

[0026] The beneficial effect of the present invention is that the present invention combines the existing E2E communication system with the CP-OFDM technology, so that the proposed OFDM-based E2E system can effectively resist the multipath effect. In addition, the present invention uses meta-learning technology to solve the environmental adaptability problem of the E2E system in practical applications: first, the noise simulation network is pre-trained through the MAML technology, so that it can quickly adapt to the changing environment and capture environmental noise; then, the noise distribution captured by the channel simulation network is used to train the E2E transceiver network extensively, so that the E2E system has strong adaptability to changes in channel fading and received signal-to-noise ratio; finally, this OFDM-E2E communication system based on the environmental adaptive training strategy can achieve adaptation to changes in three dimensions of channel noise, channel fading and received signal-to-noise ratio, so as to always maintain near-optimal performance in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a schematic diagram of the link structure of an OFDM end-to-end communication system provided by an embodiment of the present invention;

[0028] Figure 2 is a schematic diagram of a preprocessing network structure provided by an embodiment of the present invention;

[0029] Figure 3is a flow chart of an environment adaptive training strategy provided by an embodiment of the present invention;

[0030] Figure 4 Schematic diagram of a noise simulation network element training strategy based on MAML provided by an embodiment of the present invention;

[0031] Figure 5 is a schematic diagram comparing the performance of an OFDM-based E2E system provided by an embodiment of the present invention and an existing E2E system;

[0032] Figure 6 is a schematic diagram of an IGAN training curve using MAML as a training strategy provided by an embodiment of the present invention;

[0033] Figure 7 1. It is a schematic diagram of the performance of the OFDM-E2E system adapted to different receiving signal-to-noise ratios provided by an embodiment of the present invention;

[0034] Figure 8 It is a performance diagram of the OFDM-E2E system provided by an embodiment of the present invention adapting to different channel conditions. DETAILED DESCRIPTION

[0036] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments.

[0037] In order to enable the E2E system to effectively resist the multipath effect and always maintain good performance under complex channel conditions, the present invention combines the E2E network architecture with the CP-OFDM technology and designs an OFDM-based E2E communication system. The specific structure of the OFDM-E2E system link is as follows: Figure 1 As shown in Figure 1, the transmitting network, noise simulation network, preprocessing network and demodulation network are all built by neural networks, using Θ t , Θ s , Θ p With Θ de Represent their parameters. The preprocessing network and demodulation network together constitute the receiving network. The traditional E2E communication system based on a single carrier can be composed of these four parts: the transmitting network is responsible for extracting the features of the information sent by the source and sending it to the channel; the noise simulation network and the preprocessing network are used to match the harsh non-Gaussian noise channel environment. The noise simulation network is only used during system training and is responsible for providing sufficient non-Gaussian noise samples for the joint training of the transceiver network. The preprocessing network can suppress the impulse noise component in the received signal and reduce the impact of the fading channel on the receiver demodulation; the demodulation network is responsible for restoring the original information data. The specific network structure will be introduced in detail later.

[0038] Based on the above network structure, the E2E system is combined with CP-OFDM technology: Figure 1 , assuming that the original transmission information By FFT Individual messages i (i=1,2,...w FFT ), for a single message s i , which consists of a bit sequence of length k, representing one of M possible symbols, where M is the modulation order. The OFDM extension of the E2E system is a set of w FFT The output signal of the independent message s after passing through the transmission network is OFDM modulated. The output of the transmission network is a 2-channel real-valued signal, which is regarded as the real and imaginary parts of the complex signal X, creating w FFT equivalent independent sub-channels, and w FFT messages are assigned to each subcarrier, and the OFDM symbol X is obtained. OFDM After OFDM modulation is completed, add a l-length CP After the signal passes through the channel and noise simulation network, the corresponding inverse process is performed at the receiving end. First, the CP of the received signal y is removed to obtain Y OFDM , then for Y OFDM OFDM demodulation is performed to obtain the frequency domain signal Y. The real and imaginary parts of the complex signal Y are converted into 2-channel real-valued signals, and then passed through the preprocessing network and the demodulation network in turn to restore the original message

[0039] The communication model of the OFDM-E2E system is assumed in the present invention to be:

[0040]

[0041] in is the CP-OFDM signal to be sent, s i ∈{0,1},i∈{1,2,…,w FFT} is the information bit data, at this time the modulation order M is 2, f t (·) is the mapping represented by the originator, is the channel matrix, is the channel noise vector, To receive the signal. In order to consider more harsh application environments, the present invention assumes that the distribution of the channel noise n is based on a mixed Gaussian pulse distribution. This mixed noise can be modeled as the sum of impulse noise and Gaussian noise, where the impulse noise is modeled as a symmetric α-stable (SαS) distribution, so the channel noise can be modeled as:

[0042] n=n s +n g ,ns -S α (g,0,0),n g -N(0,s 2 ) (2)

[0043] Where S α (g,0,0) represents a SαS distribution with characteristic parameter α, scale parameter γ, skew parameter β and location parameter δ all being 0; N(0,s 2 ) represents a mean of 0 and a variance of σ 2 Gaussian distribution. At the same time, the pulse intensity λ = 2g α / s 2 , is the ratio of the equivalent power of the pulse noise part to the Gaussian noise part in the mixed noise, representing the pulse strength of the mixed noise in this scenario. The larger the λ, the more pulse components there are in the noise, and vice versa, the closer it is to Gaussian noise. This distribution can also cover the ideal additive Gaussian white noise, that is, λ = 0, and the extremely impulsive impulse noise, that is, λ → ∞, and has a wide range of applicability.

[0044] As a data-driven system design, the E2E architecture requires a large amount of data-assisted training to capture the characteristics of mixed Gaussian and impulsive noise (MGIN) and then design a receiver that is close to the optimal. During the simulation process, noise data can be generated arbitrarily, and the large amount of noise data required by the data-driven system can be met. However, in actual environments, especially complex electromagnetic environment scenarios, only limited noise samples can be obtained when training the system. The use of these limited noise samples is not enough for the E2E transceiver network to fully capture the channel characteristics and design a transceiver that matches the channel, so that the learned E2E system cannot achieve the optimal bit error performance in this environment. To solve this problem, the present invention refers to the idea in the patent (Zhao Chengjie, Wang Jun, Huang Wei, Chen Xiaonan. An end-to-end communication system based on deep learning under non-Gaussian noise [P]. China, invention patent, application number: 2023103497104), and adopts the Impulsive Generative Adversarial Net (IGAN) structure that can generate impulses to construct a noise simulation network, aiming to capture the characteristics of MGIN and simulate its statistical distribution p n , and then use the simulated distribution p g Generates simulated noise g similar to it. The loss function of the generator G is:

[0045]

[0046] In the above formula, E[x] represents the mathematical expectation of the random variable x, θ g With θd Represent all trainable parameters in the generator G and the discriminator D, z represents a set of random vectors of the generator G input that obey a certain distribution (generally using Gaussian distribution), n represents the environmental noise vector, and g represents the simulated noise vector of the generator. The last two terms are the two regularization terms introduced, α 1 and α 2 Represent the weights of the two regularization terms, ε is a small enough positive number to make 1+ε approach 1, α 1 , α 2 Both ε and ε are adjustable parameters that can be adjusted according to different actual environments. The first regularization term can capture the pulse characteristics of MGIN, and the second regularization term can help the generator control the energy distribution of the generated noise to prevent it from focusing too much on the generated pulse value. The two regularization terms work together to enable IGAN to well control the impulse nature of the generated noise distribution.

[0047] The loss function of the discriminator D is:

[0048]

[0049] Among them, gp(·) represents the gradient penalty term, λ gp represents the weight of the gradient penalty term, is a convex combination of g and n, The combination coefficient t is sampled from the uniform distribution U[0,1], xp x (x) represents the element x in x i Independent and identically distributed under the given distribution p x (·), Represents the gradient of function f with respect to variable x.

[0050] The present invention adopts convolutional neural network to design the transmitting network and the receiving network, and uses the multi-level wavelet transform convolutional neural network (MWCNN) architecture to design the preprocessing network in the receiving network. The preprocessing network is composed of stacked discrete wavelet transform (DWT) or inverse discrete wavelet transform (IWT) layers and corresponding convolution blocks, and introduces a channel feature extraction unit to process the multiplicative interference introduced by the channel. The specific structure of the preprocessing network is as follows Figure 2 As shown, Conv1d represents a one-dimensional convolutional layer, BatchNorm1d represents a one-dimensional batch normalization layer, ReLU represents the activation function using ReLU, and Conv Block1 / 2 represents a convolutional network block composed of Conv1d, BatchNorm1d, ReLU, etc.

[0051] The training of the E2E system depends on a pre-provided data set. When the channel environment in actual application is different from that in training, or when the actual channel is constantly changing, the performance of the E2E system will degrade, and thus it will be impossible to maintain near-optimal performance. Therefore, based on the proposed OFDM-E2E system architecture, the present invention designs an environment adaptive training strategy based on meta-learning, so that the E2E system has stronger adaptability to environmental changes, thereby maintaining near-optimal performance.

[0052] In a specific preset scenario, the training task of the E2E system can be expressed as the following optimization problem:

[0053]

[0054] In the formula, f r (·) is the mapping represented by the receiving end. The constraints in the first row represent the information transmission and recovery process, and the constraints in the second row represent the channel distribution preset during training. And the signal power P 0 . But in practice we often encounter problems of the following form:

[0055]

[0056] The distribution of noise and channel fading are respectively in the set and This represents the possible changes in channel conditions. In addition, the received power of the signal also varies in the set In order to solve this problem, the present invention designs an environment adaptive training strategy for the E2E communication system.

[0057] In order to make the system more adaptable to the environment, a simple idea is to add more channel conditions to the channel training set. Let the system be exposed to more environments during training to improve its adaptability to channel changes. However, this method is not completely realistic. Observe model (1), the receiver needs to estimate from y Further recovery This is equivalent to solving the inverse problem of the information transmission process. When we use this idea to fill the channel training set with various distributions of channel fading and noise and change the signal power at the same time, h, x, and n in equation (1) all change together. In this way, to recover from y It will be very difficult, that is, impossible to solve such an inverse problem. Therefore, when two of the components are changing, one component must fix its distribution, so that the system can learn to adapt to at least two dimensions.

[0058] Based on the above analysis, the schematic diagram of the environment adaptive training strategy of the E2E communication system designed by the present invention is shown in Figure 3 The strategy consists of four stages: the first stage is to use a model-independent meta-learning strategy to meta-train the channel noise simulation network (IGAN) of the E2E system, so that it has the property of rapid adaptation under small samples; in the second stage, IGAN begins to quickly capture environmental noise and then provides a noise training set for the training of the transceiver network; the third stage conducts extensive training on the transmitting and receiving networks; after the network has strong environmental adaptability, it can be directly deployed to the environment in the fourth stage. Specifically, the training strategy includes:

[0059] Strategy 1: Meta-learning-based noise simulation network training strategy

[0060] Using the MAML method, a model initialization parameter is found through meta-learning, and this initialization parameter is used as the starting point for training a new model. Only a small number of samples (small samples) are needed to make the model converge quickly when facing tasks similar to those encountered in the meta-learning process. In the MAML process, the input of the system is not just simple data, but multiple tasks. By learning these tasks, the overall characteristics of the tasks are finally captured.

[0061] In the simulation process of non-Gaussian noise, MAML traverses the neural network components through multiple non-Gaussian pulse noises with different distributions to seek a better neural network initialization strategy. For the application scenario of the present invention, different tasks correspond to noises with different pulse intensities λ. When the channel electromagnetic environment fluctuates, the system can quickly track the changes in the noise environment with small samples. Specifically, when training IGAN, first sample T different tasks from the offline task set. Then build a training set for these tasks With the test set The first stage of training is to iterate over different tasks. Each iteration uses the training set Try to pre-update the parameters of the neural network and then use the test set Calculate the loss of the pre-updated network. When all tasks are traversed, combine all the test losses calculated for the pre-updated network and perform a real update on the original network. The whole process can be iterated multiple times until the system converges to a fixed indicator.

[0062] The meta-training process of the noise simulation network is as follows:

[0063] 1. From the distribution p λ T different tasks are sampled, that is, T different noise pulse intensities;

[0064] 2. For each task, Sampling in

[0065] 3. Try to update the discriminator in IGAN through gradient descent:

[0066] 4. Re- Sampling in

[0067] 5. Try to update the generator in IGAN through gradient descent:

[0068] 6. In Repeat steps 2-5 to perform sampling and loss calculation to obtain the test loss of the current task

[0069] L D,j =L D (θ′ d,j ,Θ′ g,j ),L G,j =L G (θ′ g,j );

[0070] 7. Repeat steps 2-6 to traverse all tasks;

[0071] 8. Calculate the total test loss and update the IGAN network parameters using the Adam optimizer

[0072]

[0073] 9. Repeat steps 1-8 for each training round.

[0074] The whole process is shown in the following figure: Figure 4 shown.

[0075] Strategy 2: Extensive training strategy for the sender and receiver networks

[0076] After the IGAN based on the MAML strategy helps the system to quickly capture channel noise, the key to the strong environmental adaptability of the E2E system becomes its adaptability in two dimensions: channel fading and signal power. In this case, different channel fading and signal-to-noise ratio values ​​can be added to the training set of the transceiver network to improve the system's robustness to environmental changes. When using the batch training method, the loss function of the E2E system during training can generally be expressed as:

[0077]

[0078] The present invention rewrites it as:

[0079]

[0080] in is the channel distribution set, is the noise distribution set, is the transmit power set. In this way, the goal of the E2E system becomes to find the general solution for the optimal waveform design under these conditions.

[0081] The present invention adopts the multipath spectrum {τ(t), α(t)} and the normalized maximum Doppler frequency shift ε=f d / △f is used to describe the channel fading. Among them, τ(t) represents the multipath delay, α(t) is the corresponding amplitude, and f d represents the maximum Doppler frequency shift, △f is the subcarrier spacing of OFDM symbols. The present invention adds ε from 10 to the training set. -5 to 10 -1 The channel parameters are set and the training signal-to-noise ratio is selected to vary between 10 dB and 30 dB. In this way, during training, the system's signal will have a varying signal-to-noise ratio and be affected by channel fading randomly selected from the training set.

[0082] By combining strategy 1 and strategy 2, the update process of the E2E system weight finally becomes:

[0083]

[0084] This gives the E2E system adaptability in three dimensions: channel and noise distribution, and received signal-to-noise ratio.

[0085] After completing training using the above training strategy, the OFDM-based E2E communication system designed by the present invention can well adapt to the dynamic changes of the channel and the differences between the channel models during testing and training, and always maintain good performance in complex multipath channels.

[0086] Example

[0087] This embodiment is based on Table 1 and Figure 2 Build an E2E communication system according to the relevant network structure in Figure 1 The process in the previous section expands it into an OFDM architecture with CP. The specific waveform parameters are shown in Table 2. The verification channel is the Extended Vehicular A (EVA) channel. The multipath delay spectrum parameters of the channel are shown in Table 3. In the training set, ε is added from 10 -5 to 10 -1 In addition, two cases, Gaussian noise and mixed Gaussian impulse noise with α=1.5 and λ=1, are considered in the noise set.

[0088] Table 1 Network model parameters used in the embodiment

[0089]

[0090]

[0091] Table 2 Waveform parameters of OFDM-E2E system

[0092] System bandwidth 10MHz Modulation order 2 Number of subcarriers 256 CP length 32

[0093] Table 3 Multipath delay spectrum parameters of EVA channel

[0094] Tap1 Tap2 Tap3 Tap4 Tap5 Tap6 Tap7 Tap8 Tap9 Delay(us) 0 0.03 0.15 0.31 0.37 0.71 1.09 1.73 2.51 Power(dB) 0 -1.5 -1.4 -3.6 -0.6 -9.1 -7.0 -12.0 -16.9

[0095] The specific implementation steps are as follows:

[0096] Step 1: According to Table 1 and Figure 2 The network structure in the E2E communication system is built, and based on Figure 1 The process in is expanded into an OFDM architecture with CP.

[0097] Step 2: Use MAML to meta-train the noise simulation network. The training method flow of the noise simulation network is as follows.

[0098] 1. Initialize the optimization algorithm Adam, learning rate α, batch size b, number of training rounds e, and generator parameter θ g0 , discriminator parameters θ d0 , the number of training tasks T, the preset task collection distribution p λ , Generate corresponding training and test data sets for each task and

[0099] 2. From the distribution p λ T different tasks are sampled, that is, T different pulse intensities;

[0100] 3. In Sampling in

[0101] 4. Try to update the discriminator in IGAN through gradient descent:

[0102] 5. Re- Sampling in

[0103] 6. Try to update the generator in IGAN through gradient descent:

[0104] 7. In Repeat steps 3-6 to perform sampling and loss calculation to obtain the test loss of the current task

[0105] L D,j =L D (θ′ d,j ,Θ′ g,j ),L G,j =L G (θ′ g,j );

[0106] 8. Repeat steps 2-6 to traverse all tasks;

[0107] 9. Calculate the total test loss and update the IGAN network parameters using the Adam optimizer

[0108]

[0109] 10. Repeat steps 2-9 for each training round.

[0110] The metamodel θ of the noise simulation network is obtained MAML .

[0111] Step 3: Metamodel θ of the noise simulation network MAML Deployed to the actual channel environment and quickly capture the noise distribution in the environment, the adapted model θ is obtained * .

[0112] Step 4: Use the adapted noise to simulate the network model θ * Generate simulated noise training set Prepare to train the E2E transceiver network.

[0113] The goal of training the transmitting and receiving networks is to minimize the error between the transmitted data and the data decoded by the receiving end. The present invention uses the binary cross entropy loss function given by formula (7) as the E2E loss. For channel fading, this embodiment adds a 10 -5 to 10 -1 The maximum normalized Doppler shift ε of In addition, this embodiment selects the signal-to-noise ratio to vary between 10dB and 30dB as Under the supervision of the E2E loss of formula (7), the training process of the entire E2E system is as follows:

[0114] 1. Initialize the optimization algorithm Adam, learning rate α, batch size b, number of training rounds e, number of subcarriers w of OFDM symbols FFT , fading and signal-to-noise ratio Transmitter network parameters θ t0 , preprocessing network parameters θ p0 , demodulation network parameters θ de0 ;

[0115] 2. Randomly generate a set of 0-1 information bits As the transmission data;

[0116] 3. Make the transmitted data s get the signal through the transmitting network Then perform IFFT on X and add CP to obtain the signal x to be transmitted;

[0117] 4. Sample channel coefficients and receiving signal-to-noise ratio from the training set, convert the transmitted signal x into power according to the signal-to-noise ratio, and then pass through the channel fading, and then superimpose Gaussian noise or simulated mixed Gaussian pulse noise g generated by the noise simulation network to obtain the received signal y;

[0118] 5. After removing the CP and performing FFT on the received signal y, it passes through the preprocessing network and the demodulation network in turn, and finally the recovered information data is obtained

[0119] 6. Calculate s and The E2E loss between the two networks is calculated, and the gradient of each network parameter is calculated at the same time. The Adam optimizer is used to update the parameters of the three networks.

[0120] 7. Repeat process 2-6 until the training round is completed.

[0121] Step 5: Deploy the extensively trained transmitting and receiving networks to actual scenarios.

[0122] The information bits to be transmitted are grouped according to the predetermined number of subcarriers and sent to the transmission network in sequence. The meta-trained transmission network will design communication waveforms suitable for various scenarios. The transmitted signal passes through the channel and is received by the receiving network. The preprocessing network in the receiving network can suppress the influence of impulse noise and channel fading, and finally the demodulation network restores the original data.

[0123] The present invention adopts Wasserstein-1 distance (WD) to measure the noise simulation effect, and verifies the performance of the OFDM-based E2E communication system in complex multipath channels and its adaptive performance to environmental changes by statistically analyzing the bit error rate (BER) of the system.

[0124] Since there is no finite second-order moment between SαS noise and mixed Gaussian pulse noise, in order to measure the received signal strength in the current environment, the concept of generalized SNR (GSNR) is introduced:

[0125]

[0126] Among them, Ps Represents the average power of the signal. Further introduction of the generalized E b / N 0 The concept (hereinafter referred to as E b / N 0 ):

[0127]

[0128] In the formula, B and R b They represent the system bandwidth and source bit rate of OFDM respectively.

[0129] This embodiment compares the performance of the OFDM-based E2E system and the original E2E system (single carrier) under the EVA channel. Figure 5 It can be seen that the original E2E system has a clear performance plateau under high SNR, and the BER can only reach 10 -3 The OFDM-based E2E system has the same advantages as the traditional OFDM system and can effectively resist the multipath effect. When the signal-to-noise ratio is 30dB, the BER can be reduced to 10 -5 Magnitude.

[0130] In this embodiment, the gain brought by MAML to the noise simulation network is observed under the MGIN channel with α=1.5 and λ=1. Figure 6 The figure depicts the change in WD between the samples generated by IGAN in this embodiment and the real samples after using the MAML-based pre-training strategy as the number of training rounds increases. Figure 6 It can be seen that after using the MAML strategy, IGAN converges significantly faster and achieves more stable results.

[0131] In this embodiment, the effect of the proposed environment adaptive training strategy is observed in scenario 1 (EVA channel, ε=0, Gaussian noise, different received signal-to-noise ratios) and scenario 2 (EVA channel, different ε and noise). Figure 7 The BER performance of the E2E system using the environment adaptation strategy of the present invention in scenario 1 is compared with that of other E2E systems trained under a single signal-to-noise ratio under different signal-to-noise ratios. Figure 7 It can be seen that the E2E system trained with the extended SNR set can achieve near-optimal performance at each received SNR, while the E2E system trained with a single SNR will have a significant performance regression when it is far away from the SNR value for which it was trained. This illustrates the effectiveness of expanding P in equation (7). Figure 8 The performance of the E2E system using the environment adaptation strategy in scenario 2 under different noise distributions and time-selective fading channels. The noise used is Gaussian noise and MGIN with α = 1.5, λ = 1. The channel training set includes ε from 10-1 to 10 -5 Different channel coefficients, ε in the test is 10 -1 , 10 -3 and 10 -5 As can be seen from the figure: the E2E system can always maintain good performance in different channel environments, which proves the effect of ε expansion.

Claims

1. An OFDM end-to-end communication system with environmental adaptability, characterized in that: The system is composed of neural networks, including the transmitting network, the noise simulation network, the preprocessing network and the demodulation network, and their parameters are defined as Θ t , Θ s , Θ p and θ de , where the transmitting network is responsible for extracting the features of the information sent by the source and sending it to the channel; the noise simulation network and the preprocessing network are used to match the harsh non-Gaussian noise channel environment. The noise simulation network is only used during system training and is responsible for providing sufficient non-Gaussian noise samples for the joint training of the transmitting and receiving networks. The preprocessing network is used to suppress the impulse noise component in the received signal and reduce the impact of the fading channel on the receiver demodulation; The demodulation network is responsible for restoring the original information data; the communication process of the system is: define the original transmission information By FFT Individual messages i Composition, i=1,2,...w FFT , for a single message s i , which consists of a bit sequence of length k, representing one of M possible symbols, where M is the modulation order; the output of s after passing through the transmitting network is a 2-channel real-valued signal, which is regarded as the real and imaginary parts of the complex signal X respectively. FFT messages are assigned to each subcarrier, and the OFDM symbol X is obtained. OFDM After OFDM modulation is completed, add a l-length CP The CP of the transmitted symbol x is obtained. After passing through the channel and noise simulation network, the corresponding inverse process is performed at the receiving end. First, the CP of the received signal y is removed to obtain Y OFDM , then for Y OFDM OFDM demodulation is performed to obtain the frequency domain signal Y. The real and imaginary parts of the complex signal Y are converted into 2-channel real-valued signals, and then passed through the preprocessing network and the demodulation network in turn to restore the original message In order to minimize the error between the transmitted data and the data decoded by the receiving end and to adapt to the dynamically changing environment, the system is trained in the following ways: 1) Train the noise simulation network. The noise simulation network adopts an adversarial generation network structure that can generate pulses. The loss function of the generator G is: Among them, E[x] represents the mathematical expectation of the random variable x, θ g With θ d Represent all trainable parameters in the generator G and the discriminator D respectively, z represents a set of random vectors of the generator G input that obey a certain distribution, n represents the environmental noise vector, g represents the simulated noise vector of the generator, α1 and α2 represent the weights of the two regularization terms respectively, ε is a positive number small enough to make 1+ε approach 1, α1, α2 and ε are all adjustable parameters; The loss function of the discriminator D is: Among them, gp(·) represents the gradient penalty term, λ gp represents the weight of the gradient penalty term, is a convex combination of g and n, The combination coefficient t is sampled from the uniform distribution U[0,1], xp x (x) represents the element x in x i Independent and identically distributed under the given distribution p x (·), Represents the gradient of function f with respect to variable x; The noise simulation network is trained using a meta-learning-based training strategy. Specifically, first, T different tasks are sampled from the offline task set. Then build a training set for these tasks With the test set The first stage of training is to iterate over different tasks, using the training set in each iteration. Try to pre-update the parameters of the neural network and then use the test set Calculate the loss of the pre-updated network; when all tasks are traversed and completed, combine all the test losses calculated for the pre-updated network and perform a real update on the original network; through multiple iterations, until the system converges to a fixed indicator, the meta-model of the noise simulation network is obtained; The meta-model of the noise simulation network is deployed in the actual channel environment and the noise distribution in the environment is captured to obtain the adapted noise simulation network model; 2) Generate a simulated noise training set using the adapted noise simulation network model to train the system's transceiver network: Initialize the optimization algorithm Adam, learning rate α, batch size b, number of training rounds e, number of subcarriers w of OFDM symbol FFT , channel distribution set Noise Distribution Set Transmit power set Transmitter network parameters θ t0 , preprocessing network parameters θ p0 , demodulation network parameters θ de0 ; Randomly generate a set of 0-1 information bits As the transmission data; Make the transmitted data s get the signal through the transmitting network Then perform IFFT on X and add CP to obtain the signal x to be transmitted; The channel coefficients and the received signal-to-noise ratio are sampled from the training set, and the transmitted signal x is converted into power according to the signal-to-noise ratio and then passes through the channel fading, and then Gaussian noise or simulated mixed Gaussian pulse noise g generated by the noise simulation network is superimposed to obtain the received signal y; After removing the CP and performing FFT on the received signal y, it passes through the preprocessing network and the demodulation network in turn, and finally the recovered information data is obtained. Calculate s and E2E loss between At the same time, the gradient of each network parameter is calculated, and the parameters of the three networks are updated using the Adam optimizer; After repeated training, the network parameters are deployed in the system for communication.

Images