Signal transceiving processing method and system based on deep neural network

Abstract

The invention discloses a signal transceiving processing method and system based on a deep neural network, and relates to the technical field of wireless communication. Signals of different modulation systems can be generated, and the bit error rate performance is remarkably improved through receiving and transmitting joint optimization; on the basis of meeting SEM constraints, transition band and stop band resources of SEM are fully utilized, and remarkable SE gain is realized; in a frequency selective fading channel, compared with the conventional modulation scheme, the maximum signal-to-noise ratio gain of 20dB is obtained; an introduced data detector based on large kernel attention not only effectively reduces large-span interference information, but also greatly reduces calculation complexity; the method does not depend on a channel prior model, and channel characteristics are implicitly embedded through data driving, so that the method can be switched as required in diversified wireless environments, and stable performance is kept.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication, and particularly relates to a signal transceiving processing method and system based on a deep neural network. BACKGROUND

[0002] In the development history of wireless communication technology, single carrier (SC) modulation and multi-carrier (MC) modulation are two important modulation methods. SC waveform has the advantages of short symbol period, low peak-to-average power ratio (PAPR), and low processing complexity, so the SC modulation technology plays a key role in early communication systems due to its simple implementation. With the increasing demand for communication capacity, the SC time-domain waveform pulse period is reduced, resulting in an increase in signal bandwidth, and the influence of frequency-selective fading becomes more and more significant. In order to solve the above problems, when the mobile communication develops to 4G, the modulation waveform has experienced a transition from SC to MC. Among them, orthogonal frequency division multiplexing (OFDM) has become the most common MC scheme due to its low implementation complexity, flexible resource allocation, strong anti-multipath fading ability, and other advantages, and has been included in IEEE 802.11, IEEE 802.16, long-term evolution technology (LTE) and fifth generation new radio technology (5G NR) and other widely used wireless standards. However, the traditional orthogonal modulation method has advantages in terms of technical maturity and stability, but is limited by the strict orthogonal condition constraint, and its spectral efficiency has been difficult to meet the demand for high spectral utilization of future communication, so the non-orthogonal modulation method of SC / MC is also being explored. Among them, non-orthogonal SC modulation is realized by compressing the symbol time interval, and non-orthogonal MC modulation such as high spectral efficiency frequency division multiplexing (SEFDM) is realized by compressing the subcarrier frequency domain interval. According to the super-Nyquist transmission theory proposed by Mazo, as long as the compression ratio is controlled within a certain range, the transmission rate can be improved while maintaining good detection performance. This method of introducing artificial interference to obtain capacity gain is called Faster than Nyquist (FTN), which provides a new idea for spectrum resource utilization. The above traditional modulation technologies have promoted the continuous development of communication systems.

[0003] With the development of mobile communication to 5G, the application standards of mobile communication put forward higher performance requirements for wireless communication, such as ultra-high transmission rate and ultra-low latency. At the same time, the increasingly complex and diverse communication scenarios pose new challenges to the traditional fixed waveform generation architecture, and the traditional modulation technology gradually exposes its limitations. The traditional waveform generation architecture cannot adapt to complex application scenarios that are difficult to model mathematically, because its effectiveness depends on the correctness of the pre-modeling of fixed scenarios. Complex communication scenarios often have a large number of optimization variables, and the traditional waveform design scheme still uses fixed optimization criteria, so it is difficult to design a waveform that is optimal for the scenario. In addition, when the communication scenario needs to be dynamically switched between SC / MC, the existing solution relies on physical switching of the hardware architecture, which limits the flexibility and generalization ability of the system. Establishing a unified architecture that adapts to SC / MC system waveforms can solve the flexibility problem, but designing modules based on the traditional architecture cannot build the required architecture. Therefore, the traditional modulation technology cannot meet the performance requirements of 5G and future communication systems in complex scenarios, and innovative breakthroughs are needed in optimization technology and communication architecture.

[0004] In recent years, artificial intelligence (AI) technologies represented by deep learning (DL) have developed rapidly and have been successfully applied in many fields. Deep neural networks (DNNs) have strong abstract representation capabilities and can learn features from large amounts of data without relying on prior models. They can solve problems that are difficult to model mathematically or have high complexity and many optimization variables. DNNs have made significant achievements in communication physical layer waveform design. It has been proven that DNNs can capture interference characteristics and utilize spectrum template (SEM) transition band and stop band resources through end-to-end learning, so they perform well in suppressing inter-carrier interference and improving spectral efficiency (SE). In 6G communication research, the deep integration of AI and communication has become a trend. To meet the adaptability requirements of future communication networks to diverse business scenarios and complex electromagnetic environments, existing DL-based research still has obvious limitations: although DL technology has made some improvements in optimizing the performance of SC or MC systems, these improvements are still limited to the framework of traditional waveform design and have not broken through the inherent boundaries between SC and MC. SUMMARY

[0005] The present application is directed to the deficiencies of the prior art, and proposes a signal transceiving processing method and system based on a deep neural network, which is a unified architecture communication framework suitable for SC / MC and orthogonal / non-orthogonal modulation systems. The framework maximizes the system transmission rate as the optimization objective, and jointly optimizes the transceiver neural network parameters under the selected constraints to improve communication performance.

[0006] In a first aspect, the present application provides a signal transceiving processing method based on a deep neural network, which includes the following steps:

[0007] Transmitter processing steps:

[0008] Symbol grouping: The input encoded bit stream b is symbolically grouped to obtain several grouped symbols b. j Where j is the group index, J represents the number of groups, each group represents N subcarriers or time intervals, and N is a preset value;

[0009] Symbol mapping: J symbol modulation networks are used to map each group symbol b. j Symbol modulation processing is performed, with each symbol modulation network processing a group of block symbols b. j The symbol modulation network maps the N-dimensional real-valued vector of the current group to a 2N-dimensional real-valued vector v. j ;

[0010] Shaping and normalization: For each real-valued vector v j Perform plastic surgery to reduce the size of the v-shaped area. j Reshape into an N-dimensional complex vector Then, regarding the complex value symbol Perform center normalization, and then concatenate the symbol vectors of all groups to obtain the symbol sequence a;

[0011] Signal generation: A signal generation network is used to map the symbol sequence a to a time-domain signal s; where the time-domain signal s is sampled at a rate κN. sym / T is generated, and the signal dimension of the time-domain signal s is N. up N sym ×N sym κ is the sampling rate, N up N is the upsampling factor. sym Where N is the number of symbols and T is the symbol period; adjust the value of the sampling rate κ, when N up When / κ=1, it represents an orthogonal waveform; when N up When / κ<1, it is characterized as a non-orthogonal waveform;

[0012] Cyclic prefix addition: Add an integer length L≥l after the time-domain signal s. max The cyclic prefix forms the baseband transmitted signal. Among them l nax This is the sampled value representing the maximum delay spread of the wireless channel;

[0013] Receiver processing steps:

[0014] Channel reception: After down-conversion of the received signal by the RF front-end and removal of the cyclic prefix, the received signal sequence x is obtained, which has a dimension of N. up N sym dimension;

[0015] Receiving demodulation processing: a receiving network is adopted to map the received signal sequence x to N sym a symbol complex vector r;

[0016] Data detection processing: a data detection network based on large kernel attention (LKA) is adopted to perform data detection processing on the symbol complex vector r, and an estimated bit stream

[0017] Further, the symbol modulation network adopts a single hidden layer full connection, and the activation function of the hidden layer is a linear rectifier function ReLu.

[0018] Further, the complex-valued symbol is centered and normalized as:

[0019]

[0020] where M represents the number of bits carried by each symbol, represents the mth element in the set corresponding to the matrix, where the set represents the complex-valued vector all possible vector sets.

[0021] Further, the signal generation network adopts a complex linear fully connected neural network.

[0022] Further, the receiving network adopts a single complex linear fully connected layer.

[0023] Further, the data detection network sequentially includes a first full connection layer, a plurality of LKA layers, and a second full connection layer.

[0024] Each LKA layer sequentially includes a depth convolution layer, a depth dilation convolution layer, and a channel convolution layer with a convolution kernel of 1x1; the input of the first LKA layer is the output of the first full connection layer, the input of the remaining LKA layers is the product of the input and output of the previous LKA layer, and the input of the second full connection layer is the product of the output and input of the last LKA layer.

[0025] Further, the networks of the transmitting end and the receiving end are jointly optimized, and the optimization objective function is set as:

[0026]

[0027] where W is the trainable weight of the transmitting and receiving end networks, and θ is the trainable bias of the transmitting and receiving end networks, represents the batch size of the training data, b i respectively represent the encoded bit stream b and the estimated bit stream

[0028] Further, when jointly optimizing each network of the transmitting end and the receiving end, and setting the optimization objective function as

[0029]

[0030] wherein W t1 and θ t1 are the weight and bias of the symbol modulation network respectively; W r2 and θ r2 are the weight and bias parameters of the data detection network respectively, W T is the weight of the signal generation network, W R is the weight of the receiving network, BW is the bandwidth allocated by the regulatory authority, S mask is the SEM template, S(f i ) represents the power spectral density of the transmitted signal at the discrete frequency point f i , and S mask (f i ) represents the power upper limit of the SEM at the frequency point f i .

[0031] In a second aspect, the present application further provides a signal transmitting and receiving processing system based on a deep neural network, wherein the transmitting end and the receiving end respectively comprise:

[0032] The transmitting end comprises a symbol grouping module, a symbol mapping module, a shaping and normalization module, a splicing module, a signal generation module and a cyclic prefix adding module.

[0033] The receiving end comprises a channel receiving module, a receiving demodulation processing module and a data detection processing module.

[0034] Wherein,

[0035] The symbol grouping module is configured to group the input encoded bit stream b into a plurality of grouped symbols b j and send them to the symbol mapping module; wherein j is the grouping index, J is defined as the number of groups, each group represents N subcarriers or time intervals, and N is a preset value.

[0036] The symbol mapping module is configured to perform symbol modulation processing on each grouped symbol b j using J symbol modulation networks, each symbol modulation network processes one group of grouped symbols b j , maps the N-dimensional real value vector of the current group to a 2N-dimensional real value vector v j through the symbol modulation network, and sends it to a shaping and normalization module.

[0037] The shaping and normalization module is configured to perform shaping and normalization on the input real value vector vj Shaping is performed to transform v j into an N-dimensional complex-valued vector The complex-valued symbols are reshaped and normalized to obtain the complex-valued symbols a j output by the shaping and normalization module, and are sent to the splicing module.

[0038] The splicing module is configured to splice the symbol vectors of all groups to obtain the symbol sequence a and send it to the signal generation module.

[0039] The signal generation module is configured to map the symbol sequence a to the time-domain signal s using a signal generation network and send it to the cyclic prefix adding module; wherein the time-domain signal s is generated at a sampling rate κN sym , and the signal dimension of the time-domain signal s is N up . sym ×N sym , κ is a sampling rate, N up is an up-sampling factor, N sym is the number of symbols, and T is a symbol period; the value of the sampling rate κ is adjusted, when N up / κ = 1, it represents an orthogonal waveform; when N up / κ < 1, it represents a non-orthogonal waveform.

[0040] The cyclic prefix adding module is configured to add a cyclic prefix of an integer length L ≥ l max to the time-domain signal s to form a baseband transmission signal , wherein l max is the sampling value of the maximum delay spread of the wireless channel.

[0041] The channel receiving module is configured to perform radio frequency front-end down-conversion on the received signal, remove the cyclic prefix, and obtain the received signal sequence x and send it to the receive demodulation processing module; wherein the dimension of the received signal sequence x is N up . sym

[0042] The receive demodulation processing module is configured to map the received signal sequence x to an N sym -dimensional symbol complex vector r using a receive network and send it to the data detection processing module.

[0043] The data detection processing module is configured to perform data detection processing on the symbol complex vector r using a data detection network based on large kernel attention to output an estimated bit stream

[0044] The technical solution provided by the present application at least brings the following beneficial effects:

[0045] ​The application can generate signals of different modulation systems; after joint optimization of transceiving, the bit error rate (BER) performance is significantly improved; on the basis of meeting the SEM constraint, the transition band and stop band resources of the SEM are fully utilized, and significant SE gain is achieved; in a frequency selective fading channel, a signal-to-noise ratio gain of up to 20 dB is obtained compared with the current traditional modulation scheme; the data detector based on large kernel attention (LKA) introduced not only effectively reduces the large-span interference information, but also greatly reduces the computational complexity; since it does not depend on the channel prior model, the channel characteristics are implicitly embedded through data-driven, so that it can switch on demand in a diversified wireless environment and maintain stable performance. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0047] Figure 1 The generation process of the traditional SC modulation signal.

[0048] Figure 2 The generation process of the traditional MC modulation signal.

[0049] Figure 3 The overall block diagram of the end-to-end unified architecture communication system based on DNN.

[0050] Figure 4 The data detection network structure diagram of the receiving end.

[0051] Figure 5 The comparison of the single carrier baseband signals generated by the proposed architecture and the traditional scheme.

[0052] Figure 6 The comparison of the multi-carrier baseband signals generated by the proposed architecture and the traditional scheme.

[0053] Figure 7 The BER performance diagram when the number of bits carried by each symbol M = 1 under quasi-static channel conditions and equivalent SE.

[0054] Figure 8 The BER performance diagram when the number of bits carried by each symbol M = 2 under quasi-static channel conditions and equivalent SE.

[0055] Figure 9 The BER performance diagram when the number of bits carried by each symbol M = 2 and the non-orthogonal modulation method is used under quasi-static channel conditions.

[0056] Figure 10Signal power spectral density (PSD) diagrams for comparing the proposed scheme with other schemes under quasi-static channels. DETAILED DESCRIPTION

[0057] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will describe the technical solutions in the embodiments of the present application in a detailed and complete manner with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Generally, the components of the embodiments of the present application described and shown in the accompanying drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application.

[0058] The embodiments of the present application provide a signal transceiving processing method and system based on a deep neural network, which breaks the inherent boundaries between SC / MC and orthogonal / non-orthogonal modulation systems, realizes unified scheduling of multiple modulation systems and adaptive matching of transceivers and channels through a full neural network module design combined with an end-to-end joint optimization strategy, thereby maximizing the flexibility of waveform design and improving SE.

[0059] The specific design idea of the signal transceiving processing method based on a deep neural network provided by the embodiments of the present application is as follows:

[0060] As shown in Figure 1 and Figure 2 , they are the generation processes of traditional SC / MC modulation signals respectively. It can be seen that under the SC / MC system, the generation processes of orthogonal / non-orthogonal waveforms are similar. First, both of them perform symbol mapping on the input binary bit sequence b, and then obtain the digital baseband signal through rate matching and waveform synthesis. Rate matching fills zeros to make the length of the symbol sequence a meet the needs of subsequent waveform synthesis, so as to ensure that the transmission rate matches the channel and the modulation mode. Specifically, for SC modulation, N1 (orthogonal) or N2 (non-orthogonal) zeros are inserted between symbols to obtain a higher rate After shaping filtering, linear convolution is performed to obtain the baseband signal where * represents convolution, and H represents the shaping filter response. For MC modulation, zeros are filled before and after the symbol to N points to obtain N is the number of inverse fast Fourier transform (IFFT) points. At this time, single IFFT or grouped IFFT+phase compensation / accumulation operations are performed, and the orthogonal / non-orthogonal baseband signal

[0061] The differences among several waveforms mainly come from the symbol sequence a to the baseband signal rate matching and waveform synthesis process, and both of them can be represented as linear transformations. For the rate matching process, it is represented by a linear transformation . For the waveform synthesis process, it can be regarded as a linear transformation on the sequence after rate matching to generate a baseband signal Therefore, rate matching and waveform synthesis can be integrated into a linear transformation, i.e., signal generation g(·):

[0062]

[0063] where G = A2A1, A1 and A2 are both transformation matrices, i.e., A1 is a rate matching matrix and A2 is a waveform synthesis matrix. Thus, the SC / MC system is unified in the form of rate matching and waveform synthesis calculation.

[0064] Based on the above idea and using the duality of transceiver signal processing, the final redesigned complete unified architecture of the transceiver is as shown in Figure 3 .

[0065] At the transmitter, the binary information bit sequence to be transmitted is channel coded to obtain a coded bit sequence b, i.e., a coded bit stream. The sequence b is first sent through a grouped symbol modulation network (ModNet) to complete the nonlinear operation of symbol mapping, then shaped and normalized to obtain symbol vectors of each group, and then the concatenated symbol vectors are sent to a signal generation network (GenNet), and finally a cyclic prefix (CP) is added to map the coded bit sequence b to a digital baseband signal

[0066] At the receiver, the received signal x after removing the CP is first sent to a receiving demodulation network (RecNet) to map back to a symbol vector r, and then an estimated bit sequence , i.e., an estimated bit stream, is obtained through a data detection network (DDNet). In order to handle large-span self-interference FTN-ISI / ICI when non-orthogonal waveform modulation is used, a data detector based on large kernel attention (LKA) is used in the data detection step.

[0067] Then a transceiver joint optimization function is constructed, and the batch size of training is set as The mean square error of the transmitted bits and the estimated bits is minimized, and the optimization parameters are the trainable weights W and the bias θ of the network at the transmitting and receiving ends. Parameter training can be performed for different transmission mode scenarios respectively, so that the unified architecture is applicable to signal transmission of different waveforms. The optimization function is as follows:

[0068]

[0069] where b irespectively represent the encoding bit sequence and the estimated bit sequence of the i th data sample.

[0070] The constraints that can be considered in the optimization solution include: SEM constraint; PAPR constraint; transmission power constraint; and the like.

[0071] The method aims to maximize the system transmission rate, improve SE, and minimize BER under a channel with specific statistical characteristics by means of the powerful learning and abstraction ability of the neural network under the premise of the end-to-end unified architecture based on the deep neural network. Meanwhile, through end-to-end training, the transceiver parameters can be matched with the channel and adapted to different modulation systems.

[0072] To achieve the above-mentioned target, the embodiments of the present application use a trainable NN to complete the bit-to-symbol sequence mapping and the generation of the time-domain signal. Compared with the design mode of the traditional SC / MC system using a fixed modulation scheme or relying on a rectangular pulse as a prototype filter, the unified architecture dynamically generates mapping rules and waveform parameters that adapt to different modulation systems through the neural network, and can achieve flexible waveform design and spectrum resource optimization in the cross-system scene of SC / MC, orthogonal / non-orthogonal. The method of grouping symbol mapping effectively solves the problem of high computational complexity caused by only a single modulation network when the number of subcarriers or the number of time intervals is large. The present application introduces the LKA architecture to deconstruct the traditional convolution method, greatly reducing the computational complexity caused by the large convolution kernel when processing large-span interference information. The SEM constraint and the BER optimization target can also be combined through the augmented Lagrange method, fully utilizing the available degrees of freedom of the link, iteratively updating the transceiver neural network parameters, and ensuring that the transmission signal spectrum meets the specific communication protocol while improving the transmission rate.

[0073] Therefore, when communicating by using the deep neural network-based end-to-end unified architecture communication system of the present application, the specific processing procedures of the transmitting end and the receiving end are as follows:

[0074] Transmitting end processing steps:

[0075] Grouping symbol mapping: multiple groups of parallel ModNet are used to map bits to symbols to reduce the computational complexity caused by a single fully connected network and obtain diversity gain. Therefore, before modulation, the encoding bit sequence needs to be grouped. Specifically, the total number of symbols N sym is divided into J groups, and each group represents N subcarriers or time intervals. Since the number of bits carried by each symbol is M, each group contains MN bits. Correspondingly, the input encoding bit sequence b is allocated into J groups {b1, b2…b J} and the total transmission bit sequence can be represented as:

[0076]

[0077] where, represent column concatenation.

[0078] ModNet mapping: Each group of ModNet network is a single hidden layer fully connected trainable network, and the activation function of the hidden layer is linear rectifier function ReLu. The modulation network can be represented as f t1 (·), and is parameterized by weights W t1 and bias θ t1 . Taking the jth group as an example, ModNet maps the N-dimensional real-valued vector b j of the jth group to a 2N-dimensional real-valued vector v j :

[0079] v j =f t1 (b j )=W t1 b j +θ t1 (4)

[0080] Reshaping and normalization: First, the 2N-dimensional real-valued vector v j is reshaped into an N-dimensional complex-valued vector The real part and the imaginary part correspond to the first N dimensions and the last N dimensions of v j respectively, and the relationship can be represented as

[0081]

[0082] Then, the complex-valued symbol is center-normalized, that is, the mean is eliminated and the energy is normalized. Since there are 2 MN possible, the set representing all possible vector sets is constructed. Therefore, the normalization formula is:

[0083]

[0084] Symbol concatenation: After normalization, the symbol vectors of all groups are concatenated to obtain the entire symbol sequence

[0085]

[0086] GenNet processing: This network can control the orthogonality of single / multi-carrier. In this invention, the time domain signal is generated using a trainable complex linear fully connected neural network GenNet, which is represented by f t2 (·). GenNet maps a to time domain signal s:

[0087] s=f t2 (a)=W Ta (8)

[0088] Among them, W T It is a dimension of N up N sym ×N sym A trainable complex-valued matrix. A continuous-time signal is generated by s at a sampling rate κN. sym / T is generated. This is due to the introduction of the sampling multiplier κ and the upsampling factor N. up The symbol period T at this time τ The relationship between the symbol period T when it is orthogonal and the symbol period T is:

[0089]

[0090] in, The compression ratio is called the symbol period. When τ = 1, it represents an orthogonal waveform. When τ < 1, the symbol period is shortened, representing a non-orthogonal waveform, which improves SE at the cost of introducing additional ICI / ISI. Therefore, the compression ratio τ can be adjusted to modify the duration of each symbol, thus providing flexibility in SE tuning.

[0091] Cyclic prefix (CP) addition: Adding an integer length L≥l to the end of the time-domain signal. max (where l) max A cyclic prefix (of the sampled value representing the maximum delay spread of the wireless channel) is used to form the baseband transmitted signal.

[0092] Receiver processing steps:

[0093] Channel reception model: Considering a quasi-static channel, after the received signal is down-converted by the RF front-end and the CP is removed, the relationship between the resulting sequences x and s can be expressed as:

[0094] x=Hs+ω (10)

[0095] Among them, noise Let I be the noise variance, and I be N. sym N up ×N sym N up A 3D identity matrix, the channel matrix H includes multipath effects:

[0096]

[0097] Where P is the number of paths, h i For complex gain, Π is the cyclic shift matrix, and l i This is for path delay.

[0098] RecNet processing: The receiver is a RecNet receiving network consisting of a single complex linear fully connected layer, which will process N...up N sym N sym N r1 dimensional vector r. The network can be denoted as f r1 (·).

[0099] r = f R (x) = W R x (12)

[0100] where W COM is the MMSE equalization matrix driven by the model and the data-driven learnable matrix W R are concatenated.

[0101] W COM = W MMSE W COM (13)

[0102] where W sym is of dimension N sym N up composed of learnable parameters optimized during the training process to adapt to the transmitter and effectively downsample the signal, thus RecNet possesses stronger robustness and flexibility under various channel conditions.

[0103] LKA layer input preprocessing: the symbol vector r is processed by the LKA-based data detection network (DDNet) to estimate the transmitted bits

[0104] In an embodiment, the framework of DDNet is shown in Figure 4 , which replaces the traditional demodulation process, which can be represented as Since the neural network is processing real numbers, the real and imaginary parts of the N sym dimensional complex vector r are first separated and reshaped into an N sym x 2 matrix M. As shown in Figure 4 , the matrix M is first input into an FC network, and the output is an L h adjustable N sym x L h matrix

[0105] LKA layer processing: the LKA architecture decomposes the standard convolution operation into deep convolution (DW-Conv), deep dilated convolution (DW-D-Conv), and channel convolution (1x1 convolution). This design effectively captures long-range dependencies with minimal computational overhead and parameter complexity, and integrates an attention mechanism, thus mitigating large-span ICI / ISI.

[0106] AsFigure 4 As shown in FIG. 1, the LKA detection network has L D Layers. The input-output relationship of the i-th layer can be expressed as:

[0107]

[0108] where, The attention matrix as the refined feature, and represent matrix point multiplication. The calculation formula of the deep convolution DW-Conv layer is:

[0109]

[0110] where, is the convolution kernel of the k-th channel, H DW represents the size of the convolution kernel, j is the symbol index, and k is the feature channel index. Similarly, the deep dilated convolution DW-D-Conv layer introduces a dilation rate d to expand the receptive field, and the calculation formula is:

[0111]

[0112] Bit estimation: as shown in FIG. 2, the output of the last LKA layer Figure 4 is fed into an FC network with hidden layers to generate an estimated bit sequence The estimation formula is as follows:

[0113]

[0114] where, out,1 and W out,2 are weight matrices, and γ out,1 is a bias vector.

[0115] Transmitter-receiver joint optimization:

[0116] Based on the above unified architecture, the present application proposes an end-to-end learning optimization framework, which fully utilizes the degrees of freedom that can be optimized in the transmitter-receiver link. Specifically, the present application takes minimizing the mean square error between the transmitted bit sequence and the estimated bit sequence at the receiving end as the core optimization objective, while considering spectrum compliance and power constraints, and constructs a multi-objective optimization model. For training data with a batch size of , the objective function is expressed as:

[0117]

[0118] where, r2 and θ r2 are the trainable weights and bias parameters in the entire DDNet.

[0119] ​To meet the needs of practical applications, the PSD must satisfy the SEM constraint to ensure that out-of-band radiation does not interfere with adjacent frequency bands.

[0120]

[0121] Of which, BW is the bandwidth allocated by the regulatory agency, and S... mask This is a SEM template.

[0122] Furthermore, a closed-form expression for the PSD of the transmitted signal in a communication system based on the framework proposed in this invention is derived. In this system, symbols within each packet are correlated, while symbols between packets are independent; the PSD of the transmitted signal can be expressed as the sum of the PSDs of all packets:

[0123]

[0124] Since all possible symbol sequences within a group are equally likely, the PSD of each group can be represented as:

[0125]

[0126] Among them, among them, A represents the discrete-time Fourier transform. m [n] represents the element of the m-th normalized symbol vector in the n-th dimension.

[0127] In addition, the average transmission power is also subject to constraints:

[0128]

[0129] In summary, the entire optimization problem can be modeled as follows:

[0130]

[0131] Specifically, the augmented Lagrangian method can be used to solve this problem, adding a penalty term to the loss function. For power constraints, the time-domain signal output by GenNet... Perform normalization directly:

[0132]

[0133] This directly ensures that the average power meets the constraints, eliminating the constraint term. For the SEM constraint, the sum of the excess PSD across all frequency points is defined to quantify the degree of SEM violation.

[0134]

[0135] In summary, the final loss function can be expressed as:

[0136]

[0137] where λ [u] and η [u] are the Lagrange multipliers updated during training. The hard constraints are converted into differentiable loss function terms, enabling joint optimization in end-to-end training.

[0138] The training process is shown in Algorithm 1:

[0139]

[0140]

[0141] According to experience, generally set κ=1.001, λ [0] =0, η [0] =0.01.

[0142] Embodiment:

[0143] To verify the ability of the unified architecture to generate SC / MC signals, Figure 5 and Figure 6 are the time-domain SC / MC signals generated by the proposed unified architecture and the traditional method before joint optimization of transceiver. The single carrier uses QPSK modulation, and the shaping filter uses a root-raised cosine filter with a roll-off factor of 0.35 and a symbol length of 6, with 8 sampling points per symbol. The multi-carrier uses OFDM system, with 64 subcarriers and QPSK modulation for each subcarrier. It can be seen that the waveforms generated by the two methods are almost identical, which proves the effectiveness of the unified architecture.

[0144] This embodiment is based on the unified architecture proposed in Figure 3 , and describes the implementation method and performance analysis of the communication system by taking the generation of non-orthogonal multi-carrier as an example. The training and testing parameters used in this embodiment are shown in Table 1:

[0145] Table 1 Training and testing parameters of the system

[0146]

[0147]

[0148] This embodiment evaluates the impact of the optimized waveform under quasi-static channels, focusing on the performance of BER and PSD. Assume that the complex channel gain h i obeys the distribution The channel state information (CSI) at the receiving end is perfectly estimated, the number of multipaths P is set to 64, the CP length is set to 63, and the delay l i of the i-th path is set to i-1.

[0149] Set K = 4 and use the general NR spectrum emission template for FR 2 as the constraint. Figure 7 and Figure 8 The BER performance of the proposed scheme and other schemes is shown for each symbol carrying bits M = 1 and M = 2, in quasi-static channels and equivalent SE conditions, respectively. The results are combined with Figure 7 and Figure 8 It can be seen that, compared with OFDM and multi-carrier automatic encoder (MC-AE) algorithms, the proposed scheme has a signal-to-noise ratio gain of at least 5-8 dB at a BER of 10 -4 -4 and the same N. Figure 7 It is also shown that, in the case of M = 2, the proposed scheme achieves almost the same BER performance when the number of symbols per group N is 1 and 4, respectively. This indicates that, when M = 2, the use of a memoryless modulation method with N = 1 can achieve most of the performance gain, greatly reducing the computational complexity.

[0150] Figure 9 The BER performance of the proposed scheme and other schemes is shown for each symbol carrying bits M = 1 and M = 2, in quasi-static channels and equivalent SE conditions, respectively. The results are combined with up = 4, when the channel is quasi-static, the proposed system and other schemes have the same BER performance. The proposed scheme is used for multi-carrier modulation and N up = 4, the proposed scheme is set to have the same SE as OFDM and single-carrier modulation schemes using quadrature phase shift keying (QPSK) modulation, and at a BER of 10 -4 -4, the proposed scheme achieves a signal-to-noise ratio gain of about 19 dB and 6 dB, respectively; similarly, when N up = 4, the proposed scheme is set to have the same SE as OFDM and single-carrier modulation schemes using 8-quadrature amplitude modulation (8-QAM) modulation, and at a BER of 10 -4 -4, the proposed scheme achieves a signal-to-noise ratio gain of about 20 dB and 7 dB, respectively.

[0151] Figure 10 The PSD of the proposed scheme is shown under quasi-static channels. It can be seen that the MC-AE scheme under multi-carrier modulation, the traditional OFDM scheme, and the traditional single-carrier modulation scheme all produce waveforms with low power in the stopband and transition band, which limits their ability to obtain diversity gain using these regions. In contrast, the proposed scheme makes more efficient use of the SEM stopband and transition band, effectively improving the diversity gain and reducing the BER.

[0152] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the same; although the present application has been described in detail with reference to the foregoing examples, those of ordinary skill in the art should understand that the technical solutions recorded in the foregoing examples can still be modified, or some technical features therein can be replaced by equivalents; and such modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

[0153] The above only describes some embodiments of the present application. For those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can be made, which are all within the protection scope of the present application.

Claims

1. A signal transceiver processing method based on a deep neural network, characterized in that, Includes the following steps: Transmitter processing steps: Symbol grouping: The input encoded bit stream b is symbolically grouped to obtain several grouped symbols b. j Where j is the group index, J represents the number of groups, each group represents N subcarriers or time intervals, and N is a preset value; Symbol mapping: J symbol modulation networks are used to map each group symbol b. j Symbol modulation processing is performed, with each symbol modulation network processing a group of block symbols b. j The symbol modulation network maps the N-dimensional real-valued vector of the current group to a 2N-dimensional real-valued vector v. j ; Shaping and normalization: For each real-valued vector v j Perform plastic surgery to reduce the size of the v-shaped area. j Reshape into an N-dimensional complex vector Then, regarding the complex value symbol Perform center normalization, and then concatenate the symbol vectors of all groups to obtain the symbol sequence a; Signal generation: A signal generation network is used to map the symbol sequence a to a time-domain signal s; where the time-domain signal s is sampled at a rate κN. sym / T is generated, and the signal dimension of the time-domain signal s is N. up N sym ×N sym κ is the sampling rate, N up N is the upsampling factor. sym Where N is the number of symbols and T is the symbol period; adjust the value of the sampling rate κ, when N up When / κ=1, it represents an orthogonal waveform; when N up When / κ<1, it is characterized as a non-orthogonal waveform; Cyclic prefix addition: Add an integer length L≥l after the time-domain signal s. max The cyclic prefix forms the baseband transmitted signal. Among them l max This is the sampled value representing the maximum delay spread of the wireless channel; Receiver processing steps: Channel reception: After down-conversion of the received signal by the RF front-end and removal of the cyclic prefix, the received signal sequence x is obtained, which has a dimension of N. up N sym dimension; Receiver demodulation processing: A receiving network is used to map the received signal sequence x to N. sym A signified complex vector r of dimension 1; Data detection and processing: A data detection network based on large kernel attention (LKA) is used to perform data detection processing on the symbol complex vector r, and the estimated bit stream is output.

2. The method as described in claim 1, characterized in that, The symbolic modulation network uses a single fully connected hidden layer, and the activation function of the hidden layer is the ReLU linear rectified function.

3. The method as described in claim 1, characterized in that, For complex value symbols Central normalization is performed as follows: Where M represents the number of bits carried by each symbol. Represents a set The nth dimension value of the m-th symbol vector in the set, where Represents a complex-valued vector The set of all possible vectors.

4. The method as described in claim 1, characterized in that, The signal generation network employs a complex linear fully connected neural network.

5. The method as described in claim 1, characterized in that, The receiving network employs a single complex linear fully connected layer.

6. The method as described in claim 1, characterized in that, The data detection network consists of: a first fully connected layer, several LKA layers, and a second fully connected layer. In the algorithm, each LKA layer consists of a depthwise convolutional layer, a depthwise dilated convolutional layer, and a channel convolutional layer with a 1×1 kernel. The input of the first LKA layer is the output of the first fully connected layer, and the input of the remaining LKA layers is the product of the input and output of the previous LKA layer. The input of the second fully connected layer is the product of the output and input of the last LKA layer.

7. The method as described in claim 1, characterized in that, The networks at both the transmitting and receiving ends are jointly optimized, and the objective function is set as follows: Where W represents the trainable weights of the transmitting and receiving networks, and θ represents the trainable biases of the transmitting and receiving networks. Indicates the batch size of the training data. b i Let b represent the encoded bitstream and the estimated bitstream of the i-th sample, respectively.

8. The method as described in claim 7, characterized in that, When jointly optimizing the networks at both the transmitting and receiving ends, the objective function is set as follows: Among them, W t1 ,θ t1 These represent the weights and biases of the symbol modulation network, respectively; W r2 ,θ r2 These are the weights and bias parameters of the data detection network, W. T W represents the weights of the signal generation network. R To receive the network weights, BW allocates bandwidth to regulatory agencies, S mask For SEM template, S(f i ) indicates that the transmitted signal is at a discrete frequency point f i Power spectral density at S mask (f i ) indicates that the SEM is at frequency point f i The power limit at that location.

9. A signal transceiver processing system based on a deep neural network, characterized in that, Its transmitter and receiver respectively include: The transmitter includes: a symbol grouping module, a symbol mapping module, a shaping and normalization module, a splicing module, a signal generation module, and a cyclic prefix addition module; The receiving end includes: a channel receiving module, a receiving demodulation processing module, and a data detection processing module; in, The symbol grouping module is used to group the input encoded bitstream b into several symbol groups b. j And send it to the symbol mapping module; where j is the group index, J is defined as the number of groups, each group represents N subcarriers or time intervals, and N is a preset value; The symbol mapping module uses J symbol modulation networks to map each group symbol b. j Symbol modulation processing is performed, with each symbol modulation network processing a group of block symbols b. j The symbol modulation network maps the N-dimensional real-valued vector of the current group to a 2N-dimensional real-valued vector v. j And send it into the shaping and normalization module; The shaping and normalization module is used to shape the input real-valued vector v. j Perform plastic surgery to reduce the size of the v-shaped area. j Reshape into an N-dimensional complex vector Then, regarding the complex value symbol Perform center normalization to obtain the complex value symbol 'a' output from each shaping and normalization module. j And send it into the splicing module; The splicing module is used to splice the symbol vectors of all groups to obtain the symbol sequence a and send it to the signal generation module; The signal generation module uses a signal generation network to map the symbol sequence 'a' to a time-domain signal 's' and then feeds it into the cyclic prefix addition module; wherein, the time-domain signal 's' has a sampling rate κN. sym / T is generated, and the signal dimension of the time-domain signal s is N. up N sym ×N sym κ is the sampling rate, N up N is the upsampling factor. sym Where N is the number of symbols and T is the symbol period; adjust the value of the sampling rate κ, when N up When / κ=1, it represents an orthogonal waveform; when N up When / κ<1, it is characterized as a non-orthogonal waveform; The cyclic prefix addition module is used to add an integer length L≥l after the time-domain signal s. max The cyclic prefix forms the baseband transmitted signal. Among them l max This is the sampled value representing the maximum delay spread of the wireless channel; The channel receiving module performs RF front-end down-conversion on the received signal, removes the cyclic prefix, and obtains the received signal sequence x, which is then sent to the receiving demodulation processing module; wherein the dimension of the received signal sequence x is N. up N sym dimension; The receiver demodulation module uses a receiving network to map the received signal sequence x to N. syn A signified complex vector r of dimension is generated and fed into the data detection and processing module; The data detection and processing module uses a large kernel attention-based data detection network to perform data detection processing on the symbolic complex vector r, and outputs an estimated bitstream.

Images