A joint papr reduction and ini suppression method based on deep unfolding admm network

Abstract

The application relates to a joint PAPR reduction and INI suppression method based on a deep unfolding ADMM network, a multicarrier communication waveform optimization method and device, computer equipment and a storage medium. Through a parameter set interference model and a guard interval injection technology, an optimization problem model is constructed, in which parameter set interference minimization is taken as an optimization objective, a peak-to-average power ratio reaching a threshold value is taken as a constraint condition, and a mixed signal is taken as an unknown quantity to be solved. Data subband input signals of a mixed parameter set multicarrier system are input into the optimization problem model, the optimization problem model is solved by using a parallel alternating direction multiplier method, a waveform-optimized mixed signal is obtained, and the mixed signal is taken as an output signal of the mixed parameter set multicarrier system. The method can generate a multicarrier communication waveform compatible with PAPR reduction and INI suppression.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication networks, in particular to a joint PAPR reduction and INI suppression method based on a deep unfolding ADMM network. BACKGROUND

[0002] Future mobile communication networks will establish a unified connectivity framework to create a connected world. However, in the face of diverse use scenarios and diversified service requirements, the traditional mobile communication air interface is difficult to provide service access at the same time. The mixed parameter set multicarrier system supports the flexible scheduling of heterogeneous resource blocks under the unified air interface by multiplexing different waveforms in the frequency domain, which can adapt to different use scenarios and service requirements, and further expands the flexibility and adaptability of the mobile air interface at the waveform design level. However, the inherent defects of the multicarrier signal affect the system design in a new form. On the one hand, there is a serious peak-to-average power ratio (PAPR) problem, and on the other hand, the out-of-band leakage causes a new form of interference to other users, namely, inter-parameter set interference (INI).

[0003] There is a close relationship between PAPR reduction technology and INI suppression technology. Reducing the PAPR of a signal changes the signal envelope, which may weaken the effect of the INI suppression technology. Suppressing INI in the multi-subband case will affect the effectiveness of the signal PAPR reduction technology, causing the signal peak power to increase again. A too high PAPR will worsen the out-of-band leakage problem after the signal passes through the PA, causing more serious INI. Therefore, the PAPR reduction technology must be compatible with the INI suppression technology.

[0004] However, the existing means of compatibility between PAPR reduction technology and INI suppression technology mostly have problems including high computational complexity, bandwidth efficiency loss, and nonlinearity of the power amplifier. SUMMARY

[0005] Therefore, it is necessary to provide a joint PAPR reduction and INI suppression method based on a deep unfolding ADMM network to solve the above technical problems.

[0006] A multicarrier communication waveform optimization method, the method is implemented in a mixed parameter set multicarrier system, comprising:

[0007] Constructing an inter-parameter set interference model;

[0008] Based on the inter-parameter set interference model and the guard interval injection technology, an optimization problem model is constructed, in which the minimization of the inter-parameter set interference is taken as the optimization objective, the peak-to-average power ratio reaching a threshold value is taken as the constraint condition, and the mixed signal is taken as the unknown quantity to be solved.

[0009] The data subband input signal of the mixed numerology multicarrier system is obtained, the data subband input signal is brought into the optimization problem model, the mixed signal after waveform optimization is obtained by solving the optimization problem model by using a parallelized alternating direction multiplier method, and the mixed signal is taken as a transmission signal of the mixed numerology multicarrier system.

[0010] In one embodiment, the inter-numerology interference model is represented as:

[0011]

[0012] In the above formula, H1 represents a frequency response of a channel, C1 corresponds to an FFT matrix and a CP removal matrix respectively, F2 represents a block diagonal matrix, X2 represents a frequency domain form of the data subband input signal, y INI,1 represents inter-subband interference of subband 1 on subband 2, and subscripts 1 and 2 represent parameters related to two adjacent subbands of subband 1 and subband 2 respectively.

[0013] In one embodiment, before the optimization problem model is solved by using the parallelized alternating direction multiplier method, some variables are replaced to make a constraint relaxation a convex constraint, and an auxiliary variable is added to the optimization problem model to make the model have an equality constraint, which is represented as:

[0014]

[0015] G 1,i = 0, G 2,i = 0, i ∈ G c

[0016] In the above formula, G c represents a complementary set of G, β1 and β2 are penalty factors corresponding to injected signal power, represents an auxiliary variable introduced, which is a final transmission signal to be solved, z represents a mixed signal to be solved, σ represents input signal power of a corresponding subband, F g1 F g2 respectively represent modulation matrices of guard interval injected signals of corresponding numerologies, G1 and G2 respectively represent frequency domain forms of guard interval injected signals of corresponding numerologies, L sys represents a number of time domain sampling points of the mixed signal.

[0017] In one embodiment, when the optimization problem model is solved by using the parallelized alternating direction multiplier method, the optimization problem model is converted into three sub-problems, which are alternately solved in each iteration process, and the three sub-problems are represented as:

[0018]

[0019] In the above formula, G represents an injected signal on a guard interval, represents a variable related to a mixed signal to be solved, and y represents a parameter related to a Lagrange multiplier.

[0020] In one embodiment, when the parallelized ADMM is used to solve the optimization problem model, the parallelized ADMM is unfolded into a deep unfolding network using a deep unfolding network technology, and the optimization problem is solved using the deep unfolding network.

[0021] In one embodiment, the deep unfolding network includes a plurality of iteration layers connected in sequence and having the same structure, each iteration layer corresponding to an iteration step of the parallelized ADMM;

[0022] The iteration layer is divided into a guard band step, a mixed signal step, and a Lagrange multiplier update step according to the solving part of the iteration step, and the model parameters are updated as optimizable learning parameters in each of the steps.

[0023] In one embodiment, a phased training is used when training the deep unfolding network, including a first training phase and a second training phase;

[0024] During the first training phase, the deep unfolding network is trained using a labeled loss function to obtain a trained deep unfolding network, wherein the optimization problem model is solved using the parallelized ADMM according to the data sub-band input signal, and the obtained output signal is used as a label;

[0025] In the application process of using the trained deep unfolding network to solve the mixed signal according to the data sub-band input signal, the trained deep unfolding network is trained in the second phase according to a preset interval time, and an unlabeled index loss function is used for online training in the second phase training.

[0026] The application also provides a joint PAPR reduction and INI suppression device based on a deep unfolding ADMM network, the device comprising:

[0027] A parameter set interference model construction module is configured to construct a parameter set interference model.

[0028] An optimization problem model construction module is configured to construct, based on the parameter set interference model and the guard interval injection technology, an optimization problem model with the minimization of parameter set interference as an optimization objective, the peak-to-average power ratio reaching a threshold value as a constraint condition, and a mixed signal as an unknown quantity to be solved.

[0029] A mixed signal obtaining module is configured to obtain a data subband input signal of the mixed numerology multicarrier system, bring the data subband input signal into the optimization problem model, solve the optimization problem model by using the parallelized alternating direction multiplier method, obtain a waveform-optimized mixed signal, and use the mixed signal as a transmission signal of the mixed numerology multicarrier system.

[0030] A computer device comprises a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program:

[0031] Constructing an inter-numerology interference model;

[0032] Based on the inter-numerology interference model and the guard interval injection technology, constructing an optimization problem model with the minimization of inter-numerology interference as an optimization objective, the peak-to-average power ratio reaching a threshold value as a constraint condition, and a mixed signal as an unknown quantity to be solved;

[0033] Obtaining a data subband input signal of the mixed numerology multicarrier system, bringing the data subband input signal into the optimization problem model, solving the optimization problem model by using the parallelized alternating direction multiplier method, obtaining a waveform-optimized mixed signal, and using the mixed signal as a transmission signal of the mixed numerology multicarrier system.

[0034] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the following steps:

[0035] Constructing an inter-numerology interference model;

[0036] Based on the inter-numerology interference model and the guard interval injection technology, constructing an optimization problem model with the minimization of inter-numerology interference as an optimization objective, the peak-to-average power ratio reaching a threshold value as a constraint condition, and a mixed signal as an unknown quantity to be solved;

[0037] Obtaining a data subband input signal of the mixed numerology multicarrier system, bringing the data subband input signal into the optimization problem model, solving the optimization problem model by using the parallelized alternating direction multiplier method, obtaining a waveform-optimized mixed signal, and using the mixed signal as a transmission signal of the mixed numerology multicarrier system.

[0038] The above-mentioned joint PAPR reduction and INI suppression method based on a deep unfolding ADMM network, by constructing the obtained parameter set interference model and guard interval injection technology, constructs an optimization problem model with the minimization of parameter set interference as the optimization objective, the peak-to-average power ratio reaching the threshold value as the constraint condition, and the mixed signal as the unknown quantity to be solved, inputs the obtained mixed parameter set multicarrier system data subband input signal into the optimization problem model, solves the optimization problem model by using the parallel alternating direction multiplier method, obtains the waveform-optimized mixed signal, and uses the mixed signal as the output signal of the mixed parameter set multicarrier system. This method can generate a PAPR-reduced and INI-suppressed multicarrier communication waveform. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 A flowchart of a joint PAPR reduction and INI suppression method based on a deep unfolding ADMM network in an embodiment;

[0040] Figure 2 A schematic diagram of injecting subcarriers in a guard interval between subbands of a mixed parameter set system in an embodiment;

[0041] Figure 3 A structural diagram of an iteration layer in a deep unfolding network in an embodiment;

[0042] Figure 4 A flowchart of mixed signal generation based on a deep unfolding network in an embodiment;

[0043] Figure 5 A schematic diagram of the theoretical analysis process of INI in an embodiment;

[0044] Figure 6 A structural block diagram of a joint PAPR reduction and INI suppression device based on a deep unfolding ADMM network in an embodiment;

[0045] Figure 7 An internal structure diagram of a computer device in an embodiment. DETAILED DESCRIPTION

[0046] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0047] As shown in Figure 1 The present application provides a joint PAPR reduction and INI suppression method based on a deep unfolding ADMM network, which is implemented in a mixed parameter set multicarrier system and includes the following steps:

[0048] Step S100: Construct an interference model between parameter sets.

[0049] Step S110: Based on the parameter set interference model and the guard interval injection technology, construct an optimization problem model with the parameter set interference minimization as the optimization objective, the peak-to-average power ratio reaching the threshold value as the constraint condition, and the mixed signal as the unknown quantity to be solved.

[0050] Step S120: Obtain the data subband input signal of the hybrid parameter set multicarrier system, input the data subband input signal into the optimization problem model, solve the optimization problem model using the parallelized alternating direction multiplier method, obtain the waveform-optimized hybrid signal, and use the hybrid signal as the transmission signal of the hybrid parameter set multicarrier system.

[0051] In this embodiment, Guard Band Injection (GBI) technology is used, which involves modulating two non-orthogonal symbols superimposed in the guard band. The waveform parameters of these two symbols are the same as those of the adjacent sub-bands, so that each symbol performs its own function to eliminate the parameter set interference caused by its adjacent sub-bands. The mixed signal after frequency domain superposition can also suppress PAPR.

[0052] like Figure 2 As shown, the Guard Band Injection Signal (GBIS) is modulated on the guard band with a relatively small number of subcarriers, thus avoiding the use of valuable data subcarrier resources. Simultaneously, the power of the GBIS is minimized during modeling, saving power. The frequency domain vector representation of the GBIS is as follows: Where, N g To protect the number of subcarriers occupied by the frequency band, the time-domain vector is represented as: Where J is the oversampling factor.

[0053] On the one hand, to avoid introducing signal distortion, the data subcarriers need to be orthogonal to the subcarriers injected by the guard interval. On the other hand, GBIS needs to be able to suppress INI caused by the superposition of different parameter sets. Therefore, in this embodiment, the frequency domain superposition of two non-orthogonal symbols is chosen as GBIS. It can suppress INI caused by the superposition of other subbands within the frequency band of the current subband, and can also reduce the peak-to-average power ratio of the mixed signal without causing distortion.

[0054] Furthermore, in a hybrid parameter set multicarrier system, in subband i, the user's channel impulse response is considered constant over one symbol period, and this constant state is expressed in vector form as follows:

[0055] h i =[hi (0),h i (1),…,h i (L ch,i -1)] (1)

[0056] In formula (1), L ch denotes the length of the channel. In order to focus on the analysis of INI, it is assumed that the ISI (Inter-Symbol Interference) introduced by the channel can be completely eliminated by the CP (Cyclic Prefix), while the interference of the Gaussian noise that the received signal can be subjected to is ignored. Therefore, the received signal of the user in the sub-band i can be expressed as:

[0057]

[0058] In formula (2), F g G g denotes the injected signal of the guard band. T i denotes a Toeplitz matrix, which is equivalent to the convolution operation of the channel, the first column of which is composed of , and the first row is composed of . The first term on the right represents the sub-band signal expected to be obtained, while the second term represents the INI of the current sub-band signal caused by other sub-band signals, and the third term is the injected signal in this chapter for suppressing INI and PAPR reduction.

[0059] In step S100, in order to facilitate the analysis of INI between sub-signals, it is assumed in this embodiment that the SCS used by user 2 is greater than the SCS used by user 1 under the condition of adjacent sub-bands. By analyzing the interference caused by adjacent sub-bands on different sub-carriers, the interference power on each sub-carrier can be derived. When the SCS used by different users satisfies the generalized synchronization assumption, i.e., satisfies f2=2 μ f1. In the decoding process, the symbols of user 1 are superimposed with b s (b s =2) symbols of user 2, as shown in Figure 5 . Because the parameter sets used by different sub-bands are non-orthogonal, the accumulation of inter-parameter set interference is caused. Therefore, the INI of user 1 caused by user 2, i.e., the inter-parameter set interference model, is expressed as:

[0060]

[0061] In formula (3), H1 denotes the frequency response of the channel, C1 corresponds to the FFT matrix and the CP removal matrix respectively, F2 denotes a block diagonal matrix which is the modulation matrix of the input signal, X2 denotes the frequency domain form of the data sub-band input signal, y INI,1represents the inter parameter set interference from subband 2 to subband 1, and the subscripts 1 and 2 represent the parameters related to the two adjacent subbands of subband 1 and subband 2 respectively. In which, since the channel is assumed to be ideal in the analysis of INI, H1 is a unit matrix.

[0062] Further, the element in the mth row and the nth column of G is expressed as:

[0063]

[0064] When the adjacent subbands have the same parameter set, i.e. μ = 0, 2 μ = 1, the subcarriers of each user are orthogonal to each other, i.e. y INI,1 = 0, at this time the out-of-band leakage of the subband does not cause inter parameter set interference. When the system reserved guard interval is large enough, the user 1 is far enough from the frequency band occupied by the user 2, and the interference received by each user can be ignored.

[0065] Similarly, the expression of the INI suffered by each symbol on the mth subcarrier of user 2 can be derived as:

[0066]

[0067] Similarly, the expression of the INI suffered by each symbol on the mth subcarrier of user 2 can be derived as: The expression of the element in the mth row and the nth column of G

[0068]

[0069] Formula (6) not only reflects the interference caused by different subcarrier spacing, but also reflects the influence of the different positions of the sub-signals in a receiving window on y INI,2 .

[0070] In step S110, based on the inter parameter set interference model obtained in step S100, a convex optimization model is constructed with the optimization objective of minimizing the inter parameter set interference, i.e. minimizing the power weighted value of P INI,1 and P INI,2 , and at the same time, the peak-to-average power ratio reaching a threshold value is taken as a constraint condition, then an optimization problem model is obtained with the mixed signal as an unknown quantity to be solved, which is expressed as:

[0071]

[0072] In formula (7), G c is the complementary set of G, β1 and β2 are penalty factors corresponding to the power of the injected signal, the optimization problem model is the sum of two non-convex functions, and the constraint is linear.

[0073] Further, in step S120, the ADMM algorithm is used to solve the convex optimization problem model, i.e., formula (7).

[0074] Considering that the two non-orthogonal symbols in the GIBS work independently, the parameter set interference caused by the respective waveform parameter subbands can be suppressed, and the PAPR of the mixed signal can be reduced after frequency domain superposition. This working mode is highly consistent with the "decomposition-coordination" ADMM algorithm. Before solving the optimization problem model by using the parallelized alternating direction multiplier method, an auxiliary variable is added to the optimization problem model, so that there is an equality constraint in the model, which is represented as:

[0075]

[0076] In formula (8), the separability between the objective function and the constraint variable, and The linear constraint between G1 and G2 is easier to solve in the ADMM framework. The significant advantage of the ADMM algorithm is to avoid solving the original complex optimization problem, but to find the saddle point of the augmented Lagrangian function.

[0077]

[0078] Formula (9) is the augmented Lagrangian function of the optimization problem model formula (8). In formula (9), ρ is the penalty factor of the augmented Lagrangian function, which needs to satisfy ρ>0.

[0079] In order to make the designed algorithm run efficiently on the distributed multi-baseband processing unit of the mixed parameter set system, and at the same time make the data flow mode suitable for another solving mode, i.e., the solving mode based on the deep unfolding network, the ADMM algorithm is also used to solve the optimization problem model. In each iteration calculation, the optimization problem model is converted into three sub-problems for alternating solution, and the three sub-problems are represented as:

[0080]

[0081] In formula (10), G represents the injected signal, represents the variable related to the mixed signal to be solved, and y represents the parameter related to the Lagrange multiplier.

[0082] Specifically, when solving the three sub-problems in formula (10), since it is assumed that INI only has an impact on adjacent subbands, in actual deployment, as long as the frequency domain bandwidth occupied by the user is not too narrow, it can be considered that the frequency distance is far enough, and INI will not have an impact on non-adjacent subbands, so i=1,2. The iteration process involves the solution of three optimization problems, i.e., the sub-problems. Although the form is complex, there is a precise closed-form solution at each stage.

[0083] In updating and , the first sub-problem in formula (10) can be converted into the following equivalent form:

[0084]

[0085] The closed-form solution of can be obtained by taking the partial derivative of formula (11) and setting it equal to zero:

[0086]

[0087] where

[0088]

[0089] Similarly, we have:

[0090]

[0091] where

[0092]

[0093] In updating , first establish the auxiliary variable u l+1 , which is expressed as:

[0094]

[0095] The second sub-problem in formula (10) can be re-expressed as:

[0096]

[0097] Obviously, the problem represented by formula (17) is to solve the closest vector u l+1 to the vector Therefore, can be defined as:

[0098]

[0099] In formula (18), This process can be regarded as the injection process of GBIS, that is, after the superposition of two non-orthogonal symbols in the frequency domain, the PAPR of the mixed signal is suppressed.

[0100] In this embodiment, a pseudo code for solving the optimization problem model by using the above parallelized alternating direction multiplier method is also provided, which is represented by algorithm 1:

[0101]

[0102] In the embodiment, in addition to the above-mentioned direct use of the parallelized alternating direction multiplier method to solve the optimization problem model, the parallelized alternating direction multiplier method is also unfolded into a deep unfolding network (GBI-ADMMnet) by using the deep unfolding network technology, and the optimization problem model is solved by using the deep unfolding network to improve the solving efficiency.

[0103] Optimizing the quadratic penalty coefficient ρ in the ADMM framework can improve the convergence effect of the algorithm and make the algorithm less dependent on the initial selection of ρ. In order to improve the flexibility of the algorithm and maximize the convergence speed of the algorithm, in the embodiment, not only is ρ flexible at each iteration, but also is ρ flexible at each step of each iteration, so as to better adapt to the characteristics of the specific processing object in different steps. In the iteration structure of the ADMM algorithm, the clipping rate γ is a parameter that needs to be set in advance, which represents the degree of adjustment of the mixed signal at each iteration. However, the random characteristics of the mixed signal are not only reflected before the iteration input, but also after each iteration. For some iteration steps, the signal distortion is small, and increasing the adjustment degree can reduce the number of iterations. Therefore, γ is preferably able to determine the adjustment degree of the current mixed signal according to the characteristics of the current mixed signal. Therefore, five learnable parameters that can be optimized are set in the deep unfolding network, including:

[0104] In the embodiment, ADMMnet with almost the same computational complexity as ADMM algorithm is used for optimization. Each iteration process of the above-mentioned parallelized alternating direction multiplier method (GBI-ADMM) for solving the optimization problem model can be divided into three steps, including: operating the guard band, operating the mixed signal, and updating the Lagrange multiplier. The three steps are unfolded into corresponding guard band steps, mixed signal steps, and Lagrange multiplier updating steps in the deep unfolding network, i.e., a single-layer structure of the deep unfolding network as shown in Figure 3 .

[0105] Specifically, the deep unfolding network constructed based on the GBI-ADMM algorithm includes multiple iteration layers connected in sequence and having the same structure. Each iteration layer corresponds to an iteration step of the parallelized alternating direction multiplier method. The iteration layer is divided into a guard band step, a mixed signal step, and a Lagrange multiplier updating step according to the solving part of the iteration step, and the model parameters are updated as learnable learning parameters in each step.

[0106] Further, the training parameters belong to the t-th layer Lagrange multiplier updating step, belong to the t-th layer mixed signal step, belong to the t-th layer guard band step.

[0107] It should be noted that the depth-unfolding of the GBI-ADMM is to fix the number of iterations, because the number of layers in the depth-unfolding network is determined.

[0108] In this embodiment, the forward propagation formula of the GBI-ADMMnet network is represented as:

[0109]

[0110] In the formula (19) to the formula (22), is nonlinear, In the depth-unfolding network, the role of the excitation function is undertaken, and y represents the excitation threshold. When |u t (n)|≤T, When |u t (n)|>T,

[0111] In this embodiment, in the depth-unfolding network described above, the system parameters calculated offline in advance and the original mixed signal are input into each layer in the network. This data connection mode is similar to the "skip connection" in the residual network, which can alleviate the gradient vanishing problem to a certain extent and is conducive to the convergence of the depth-unfolding network.

[0112] In this embodiment, the training of the depth-unfolding network described above is carried out in stages, including a first training stage and a second training stage. In the first training stage, the depth-unfolding network is trained using a labeled loss function to obtain a trained depth-unfolding network. In this process, the parallelized alternating direction multiplier method is used to solve the optimization problem model according to the input signal, and the output signal obtained is used as a label. In the application process of using the trained depth-unfolding network to solve the mixed signal according to the input signal, the trained depth-unfolding network is trained in the second stage according to the preset interval. In the second stage, an unlabeled index loss function is used for online training.

[0113] That is, the first training stage actually trains a usable depth-unfolding network, which has the ability to obtain a waveform-optimized mixed signal according to the input signal. The second training stage is actually an online training stage. After the trained depth-unfolding network is put into practical application, it is trained online every certain period of time to maintain good performance.

[0114] Specifically, in the first training stage, the root mean square error (RMSE) between the network output signal and the expected output signal (label value) is used as the loss function, which is represented by the formula:

[0115]

[0116] In formula (28), represents the output value of the output layer T of the depth unfolding network, M is the number of the training data set, n corresponds to the dimension of the training set, i.e. the number of time domain sampling points, and ||·|| represents the F-norm of a matrix. F represents the F-norm of a matrix.

[0117] Further, in the first training stage, the corresponding mixed signals are obtained by solving a plurality of input signals by using the GBI-ADMM algorithm in this paper, and the input signals are taken as training samples, and the mixed signals obtained by solving are taken as labels, and the depth unfolding network is trained.

[0118] Specifically, in the second training stage, in order to weaken the dependence of the loss function on the label value, an index type loss function is proposed to evaluate the PAPR and INI performance of the output signal, which is used in combination with the RMSE loss function to provide an optimization direction for parameter updating. Wherein, the index type loss function is represented as:

[0119]

[0120] In formula (29), κ INI and κ PAPR respectively represent the weights of INI and PAPR in the loss function, and in this embodiment, the values of the two are preliminarily determined by entropy weight, and then adjusted appropriately according to the training output result and actual needs.

[0121] Further, in formula (29), ||·|| ∞ The infinite norm does not have a gradient, so a sub-gradient is used instead, which is represented as:

[0122]

[0123] Therefore where i is the maximum value index in the vector , so (0,…1 i ,0) T is a sub-gradient of .

[0124] Further, from the calculation of the sub-gradient, it can be seen that in each iteration of the training process, only the value at the maximum value i in the vector is taken as the error and passed back. In the process of reducing PAPR, only the maximum value is concerned in each iteration, but in order to improve the efficiency, the sub-gradient at this place is improved according to the algorithm principle in this chapter. ​​

[0125] The larger the number of backpropagation values ​​*m*, the greater the residual volatility. This is because it deviates from the algorithm's principles, thus convergence cannot be guaranteed, but the CCDF curve becomes steeper. To minimize the number of algorithm iterations while ensuring normal convergence, m = 3 is preferred, meaning the three largest values ​​are considered during backpropagation.

[0126] In this embodiment, due to the complexity of the deep unfolded network structure and the loss function... and With learning parameters The functional relationship between them is difficult to express explicitly, and it is very difficult to directly analyze the gradient expression. The back propagation (BP) algorithm is a fast analytical solution for solving the gradient.

[0127] Furthermore, since the algorithm in this paper involves the processing of complex signals, all gradient calculations must be performed in the complex domain. Therefore, the derivative of a complex matrix is ​​first defined as follows:

[0128]

[0129] In formula (31), Let Re represent any complex matrix, and Re{·} and Im{·} represent the real part matrix and the imaginary part matrix, respectively.

[0130] The derivation of formula (31), that is, the derivation of gradient direction propagation, is as follows:

[0131] In the output layer (t=T):

[0132] Lagrange multiplier update step: The Lagrange multiplier step of the output layer T, according to the algorithm principle, takes effect in layer T+1, therefore the loss function and Y T Irrelevant

[0133]

[0134] Hybridization step: Before deriving the gradient for this layer, the step function needs to be defined as follows:

[0135]

[0136] This step has an explicit relationship with the loss function, so the gradient can be directly calculated as follows:

[0137]

[0138] The previous section presented the subgradient of the mixed signal output by the network using the "multi-valued return subgradient loss function".

[0139]

[0140] The gradients of the two loss functions with respect to γ T are the same:

[0141]

[0142] In formula (36), sum(·) denotes the summation over all elements in the matrix, and the symbol ⊙ denotes element-wise multiplication of matrices.

[0143] The gradients of the two loss functions with respect to the protection band step and are also different. The gradient of the loss function is passed from :

[0144]

[0145] The gradient of the loss function is not only passed from , but also comes from the explicit calculation of the INI part,

[0146]

[0147] The gradient of the training parameter for this step can be further calculated by formula (50).

[0148] At the input layer and the intermediate layers (t = T - 1: -1: 1):

[0149] Lagrange multiplier update step: the step Y t contains the learning parameter From the perspective of backpropagation, the corresponding gradient matrix input to this step is As shown in Figure 3 , the current layer gradient function is expressed as the input gradient matrix:

[0150]

[0151] After the current layer gradient matrix is solved, the derivative of the loss function with respect to the learning parameter of this step can be calculated:

[0152]

[0153] Further, from the perspective of backpropagation, the output step directly connected to this step is The gradient matrix of the current layer can be used to continue to calculate the gradient of the next step according to the chain rule. Then, the gradient matrix of the t-th layer Lagrange multiplier update step obtained by solving will be passed to the t-1-th layer.

[0154] The mixing step: step contains learning parameters The corresponding input gradient matrix of this step in the perspective of back propagation is As shown in formula (43), the relationship between the input gradient and the gradient of the current layer is expressed based on the chain rule, and the gradient of the loss function with respect to the two learning parameters of this step can be obtained: Figure 3

[0155]

[0156] In formula (43), the relationship between the input gradient and the gradient of the current layer is expressed based on the chain rule, and the gradient of the loss function with respect to the two learning parameters of this step can be obtained:

[0157]

[0158] wherein,

[0159]

[0160] Using the gradient calculation results of formula (44) and formula (47), the update optimization of the parameters can be realized. The output gradient matrix of this step is passed to the t-1-th layer.

[0161] The guard band step: step contains learning parameters and The input gradient matrix of this step in the perspective of back propagation is Therefore, The relationship between the input gradient matrix and the gradient of the current layer is:

[0162]

[0163] The relationship between the input gradient matrix and the gradient of the current layer is:

[0164]

[0165] Solving the gradient matrix and of the current step, the gradient of the loss function with respect to the two learning parameters of this step can be obtained:

[0166] ​​

[0167] where, and are defined by equation (51) and equation (52), respectively. So far, the derivation of the gradients of the loss function with respect to the five learning parameters of each layer based on the chain rule is completed, and the error direction propagation process is completed. The pseudo code of the training process of the GBI-ADMMnet deep unfolding network is summarized in Algorithm 2.

[0168]

[0169] In this embodiment, a pseudo code for training the deep unfolding model is also provided, which is represented by Algorithm 2:

[0170]

[0171] In this embodiment, when the trained deep unfolding model is used to solve the mixed signal, it includes: receiving each subband bit stream from the mixed parameter set multicarrier system, generating the corresponding subband signal through constellation mapping, and then inputting the generated subband signal into the deep unfolded GBI-ADMMnet network to generate the mixed signal through the forward propagation process.

[0172] Specifically, before actual deployment, the learnable parameters of the GBI-ADMMnet network need to be trained according to the two-stage training method of the deep unfolding ADMM network described above. In actual deployment, each subband bit stream can be directly converted into a subband signal, and then the subband signal is input into the GBI-ADMMnet network trained in the first stage to generate a mixed signal. Alternatively, the second-stage training method without label values can be adopted to perform online training on the GBI-ADMMnet network to further optimize the network performance.

[0173] As shown in Figure 4 , it is a flowchart of generating a mixed parameter set system generated signal based on the joint PAPR reduction and INI suppression method of the deep unfolding GBI-ADMMnet network.

[0174] In the above joint PAPR reduction and INI suppression method based on a deep unfolded ADMM network, in a multi-subband hybrid parameter set system, the PAPR and INI of a transmission signal are jointly reduced by using a guard interval injection technology, two frequency domain superimposed non-orthogonal symbols are modulated on a guard band, the interference caused by each subband is eliminated, the mixed signal after frequency domain superposition can also achieve peak-to-average ratio reduction, a parameter set interference model is obtained, based on the parameter set interference model, an optimization problem model is constructed, taking the minimization of parameter set interference as the optimization objective, taking the peak-to-average power ratio reaching a threshold value as the constraint condition, and taking the mixed signal as the unknown quantity to be solved, the input signal of the obtained mixed parameter set multicarrier system is brought into the optimization problem model, the parallel alternating direction multiplier method is used to solve the optimization problem model, and the waveform optimized mixed signal is obtained, the decomposition-collaborative injection signal generation mode is combined with the alternating vector multiplier method (ADMM), efficient solving is realized, and the optimality of the algorithm is proved by using the Slater condition and the KKT condition. For the ADMM iteration structure optimization configuration problem, the ADMM iteration structure is unfolded into a layer-by-layer structure of a neural network to obtain a GBI-ADMMnet network, each layer of ADMM contains 5 training parameters, the flexibility of the network is maximized, and a multi-value return subgradient loss function is designed, so that the network training no longer depends on the label value.

[0175] It should be understood that, although Figure 1 the steps in the flowchart of the above method are displayed in sequence according to the arrows, these steps are not necessarily executed in sequence according to the arrows. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other sequences. Moreover, Figure 1 at least part of the steps in the above method can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence of these sub-steps or stages is not necessarily sequential, but can be alternately executed with other steps or at least part of the sub-steps or stages of other steps.

[0176] In one embodiment, as shown in Figure 6 a joint PAPR reduction and INI suppression device based on a deep unfolded ADMM network is provided, comprising: a parameter set interference model construction module 200, an optimization problem model construction module 210, and a mixed signal obtaining module 220, wherein:

[0177] The parameter set interference model construction module 200 is configured to construct a parameter set interference model.

[0178] The optimization problem model construction module 210 is configured to construct an optimization problem model based on the inter-parameter set interference model and the guard interval injection technology, with minimization of inter-parameter set interference as an optimization objective, with the peak-to-average power ratio reaching a threshold value as a constraint condition, and with a mixed signal as an unknown quantity to be solved.

[0179] The mixed signal obtaining module 220 is configured to obtain a data subband input signal of the mixed parameter set multicarrier system, input the data subband input signal into the optimization problem model, solve the optimization problem model by using a parallelized alternating direction multiplier method, obtain a waveform-optimized mixed signal, and use the mixed signal as a transmission signal of the mixed parameter set multicarrier system.

[0180] The specific definitions of the multicarrier communication waveform optimization apparatus can refer to the definitions of the multicarrier communication waveform optimization method in the foregoing, and will not be described herein. Each module in the multicarrier communication waveform optimization apparatus can be realized by software, hardware, or a combination thereof, in whole or in part. Each module can be embedded in or independent of a processor in a computer device in a hardware form, or can be stored in a memory in a computer device in a software form, so as to be called and executed by a processor to perform operations corresponding to each module.

[0181] In an embodiment, a computer device is provided, which can be a terminal, and an internal structure diagram of the computer device can be as shown in FIG. 8. Figure 7 The computer device includes a processor, a memory, a network interface, a display screen, and an input device connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is configured to communicate with an external terminal through a network connection. The computer program is executed by the processor to implement a multicarrier communication waveform optimization method. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or can be a key, trackball, or touchpad arranged on the computer device shell, or can be an external keyboard, touchpad, or mouse, etc.

[0182] Those skilled in the art can understand that Figure 7 The structure shown in FIG. 8 is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. A specific computer device can include more or fewer components than those shown in the figure, or can combine certain components, or have a different component arrangement.

[0183] In one embodiment, a computer device is provided, comprising a memory and a processor, the memory storing a computer program, and the processor implementing the following steps when executing the computer program:

[0184] constructing an inter-parameter set interference model;

[0185] constructing, based on the inter-parameter set interference model and the guard interval injection technique, an optimization problem model with minimization of inter-parameter set interference as an optimization objective, with a peak-to-average power ratio reaching a threshold value as a constraint condition, and with a mixed signal as an unknown quantity to be solved;

[0186] obtaining a data subband input signal of the mixed parameter set multicarrier system, inputting the data subband input signal into the optimization problem model, solving the optimization problem model by using a parallelized alternating direction multiplier method, obtaining a waveform-optimized mixed signal, and taking the mixed signal as a transmission signal of the mixed parameter set multicarrier system.

[0187] In one embodiment, a computer readable storage medium is provided, storing a computer program, and the computer program is executed by a processor to implement the following steps:

[0188] constructing an inter-parameter set interference model;

[0189] constructing, based on the inter-parameter set interference model and the guard interval injection technique, an optimization problem model with minimization of inter-parameter set interference as an optimization objective, with a peak-to-average power ratio reaching a threshold value as a constraint condition, and with a mixed signal as an unknown quantity to be solved;

[0190] obtaining a data subband input signal of the mixed parameter set multicarrier system, inputting the data subband input signal into the optimization problem model, solving the optimization problem model by using a parallelized alternating direction multiplier method, obtaining a waveform-optimized mixed signal, and taking the mixed signal as a transmission signal of the mixed parameter set multicarrier system.

[0191] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to memory, storage, databases, or other media in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0192] The technical features of the above embodiments can be combined in any way. In order to make the description simple, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist, they should be considered as the scope of the present application.

[0193] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, some modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method of multi-carrier communication waveform optimization, characterized by, The method is implemented in a hybrid parameter set multicarrier system, comprising: constructing an inter-parameter set interference model; based on the inter-parameter set interference model and the guard interval injection technology, constructing an optimization problem model with the minimum inter-parameter set interference as the optimization objective, the peak-to-average power ratio reaching the threshold value as the constraint condition, and the hybrid signal as the unknown quantity to be solved; obtaining the data subband input signal of the hybrid parameter set multicarrier system, bringing the data subband input signal into the optimization problem model, solving the optimization problem model by using the parallelized alternating direction multiplier method, obtaining the waveform-optimized hybrid signal, and taking the hybrid signal as the transmission signal of the hybrid parameter set multicarrier system, wherein, before solving the optimization problem model by using the parallelized alternating direction multiplier method, part of the variables are replaced to make the constraint relaxation a convex constraint, and an auxiliary variable is added to the optimization problem model to make the model have an equality constraint, which is expressed as: In the above formula, denotes the complement of and is a penalty factor corresponding to the injected signal power, denotes the introduced auxiliary variable, i.e. the final transmitted signal, denotes the mixed signal to be solved, denotes the input signal power of the corresponding subband, denote the modulation matrix of the guard interval injected signal of the corresponding parameter set, respectively, denote the frequency domain form of the guard interval injected signal of the corresponding parameter set, respectively, denotes the number of time domain sampling points of the mixed signal, denotes the Fourier transform matrix, denotes the operation matrix for removing the cyclic prefix, denotes the frequency domain form of the data subband 1 input signal on the subband 1, correspond to the FFT matrix and the CP removal matrix, respectively, denotes a block diagonal matrix, which is the modulation matrix of the data subband input signal, denotes the frequency domain form of the data subband input signal.

2. The multicarrier communication waveform optimization method of claim 1, wherein, The inter-parameter set interference model is expressed as: In the above equation, denotes the frequency response of the channel, denote the FFT matrix and the de-CP matrix, respectively, denotes a block diagonal matrix is the modulation matrix of the data subband input signal, denotes the frequency domain form of the data subband input signal, denotes the inter-subband interference of subband 1 by subband 2, the subscripts 1 and 2 denote the relevant parameters of the two adjacent subbands of subband 1 and subband 2, respectively.

3. The multicarrier communication waveform optimization method of claim 2, wherein, When the optimization problem model is solved by using the parallelized alternating direction multiplier method, the optimization problem model is converted into three sub-problems which are alternately solved in each iteration process, and the three sub-problems are expressed as: In the above formulae, denotes the injected signal over the guard interval, denotes a variable related to the mixed signal to be solved, denotes a parameter related to the Lagrange multiplier.

4. The multicarrier communication waveform optimization method of any of claims 1-3, wherein, When the optimization problem model is solved by using the parallelized alternating direction multiplier method, the parallelized alternating direction multiplier method is unfolded into a deep unfolding network by using the deep unfolding network technology, and the optimization problem is solved by using the deep unfolding network.

5. The multicarrier communication waveform optimization method of claim 4, wherein, The deep unfolding network comprises a plurality of iteration layers which are connected in sequence and have the same structure, and each iteration layer corresponds to an iteration step of the parallelized alternating direction multiplier method; The iteration layer is divided into a guard band step, a hybrid signal step, and a Lagrange multiplier updating step according to the solving part of the iteration step, and the model parameters are updated as the optimizable learning parameters in each step.

6. The multicarrier communication waveform optimization method of claim 5, wherein, The deep unfolding network is trained in stages, including a first training stage and a second training stage; In the first training stage, the deep unfolding network is trained by using a labeled loss function to obtain a trained deep unfolding network, wherein the output signal obtained by solving the optimization problem model according to the data subband input signal by using the parallelized alternating direction multiplier method is used as the label; In the application process of solving the hybrid signal according to the data subband input signal by using the trained deep unfolding network, the trained deep unfolding network is trained in the second stage according to the preset interval time, and an index loss function without a label value is used for online training in the second stage training process.

7. An apparatus for joint PAPR reduction and INI suppression based on a deep unfolding ADMM network, the apparatus comprising: a processor configured to: obtain a signal; and apply the deep unfolding ADMM network to the signal to generate a modified signal. The device implements the multicarrier communication waveform optimization method according to any one of claims 1-6, comprising: an inter-parameter set interference model construction module for constructing an inter-parameter set interference model; An optimization problem model construction module is configured to construct an optimization problem model based on the inter-parameter set interference model and the guard interval injection technique, with the minimization of the inter-parameter set interference as an optimization objective, the peak-to-average power ratio reaching a threshold value as a constraint condition, and a mixed signal as an unknown quantity to be solved. A mixed signal obtaining module is configured to obtain a data subband input signal of the mixed parameter set multicarrier system, input the data subband input signal into the optimization problem model, solve the optimization problem model by using a parallelized alternating direction multiplier method, obtain a waveform-optimized mixed signal, and use the mixed signal as a transmission signal of the mixed parameter set multicarrier system.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor executes the computer program to implement the steps of the method in any one of claims 1 to 6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method in any one of claims 1 to 6.

Images