Machine learning-based joint design method for dual-power transceiver network with integrated communication and inductance
Through the joint design method of dual-power transmitter and reception network based on machine learning, the DDPG algorithm and supervised learning are used to solve the joint optimization problem of dual-power transmitting signals and receivers in synesthesia integrated system, and the efficient system in multi-objective scenarios is realized Performance improvement.
Patent Information
- Application Number
- CN202410928804.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-11
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-07-11
AI Technical Summary
There is a lack of joint optimization method for dual-power transmitting signals and dual-power receivers in the existing synesthesia integrated system. Especially in multi-objective scenarios, traditional methods rely on noise and channel statistical characteristics, resulting in high complexity and difficulty in effectively improving the system. performance.
Using a joint design method of dual-power transmitter and reception network based on machine learning, the SO-OCDM signal and dual-power receiver network are designed using deep deterministic policy gradient (DDPG) algorithm and supervised learning, and the dual-power transmitting signal is realized through reinforcement learning and supervised learning. Joint optimization with receiver reduces mathematical model dependence on channels and targets.
The dual-power receiver simultaneously realizes decoding and multi-object detection, which improves the system performance and realizes the joint design of the dual-power transmitting signal and the receiver, which is more flexible and efficient than traditional methods.
Smart Images

Figure CN118944807A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of communication perception integration, and particularly relates to the design of a dual-power transmission signal and a dual-power receiver in a communication perception integration system. Background Art
[0002] As the number of electronic devices increases, spectrum resources become scarce, causing the originally independent communication and perception systems to begin to develop in the direction of synaesthesia integration. Synaesthesia integration has the advantages of reducing system energy consumption and electronic interference, and has become a promising development direction for the next generation of wireless communications.
[0003] In the existing research work on integrated sensing systems, most receivers only consider radar tasks or communication tasks. For coexisting integrated sensing systems, in reference 1 (Chiriyath AR, Paul B, Bliss D. W. Simultaneous radar detection and communications performance with clutter mitigation [C]. 2017 IEEE Radar Conference, Seattle, USA, 2017, 0279-0284), a joint receiver is proposed, which first implements communication decoding, obtains communication transmission information, and then performs continuous interference cancellation to remove communication signals, and finally performs radar detection tasks. In reference 2 (Dong Y, Liu F, Xiong Y. Joint receiver design for integrated sensing and communications [J]. IEEE Communications Letters, 2023, 27 (7): 1854-1858), it is considered that in the presence of communication decoding errors, the target reflection coefficient is estimated by using both the direct communication path and the radar target reflection path. However, there is no related research work on the dual-power receiver of the dual-power integrated sensing system.
[0004] Since communication and perception have different requirements for signals, the transmission signal needs to be designed to balance the performance of communication and perception. At the receiving end, considering that the main task of communication is to decode the transmitted information, and the main task of radar is to detect whether the target exists, the detection is completed by designing the detector, and the decoding is realized by designing the decoder. Furthermore, the signal at the transmitting end and the detector and decoder at the receiving end can be jointly designed to improve the system performance. However, the joint design of transmission and reception under the traditional method depends on the statistical characteristics of noise and channel, and faces the problem of high complexity under the complex statistical characteristics of noise channel. Therefore, in the multiple-input multiple-output (MIMO) dual-function synaesthesia integrated system, it is necessary to consider the joint design method of transmission and reception based on machine learning. Summary of the invention
[0005] The technical problem that the present invention solves in view of the deficiencies of the background technology is to obtain a joint optimization method of a dual-power transmission signal and a dual-power receiver based on machine learning in a synaesthesia integrated system under a multi-target scenario.
[0006] The technical solution of the present invention is a joint design method of a synaesthesia integrated dual-function transceiver network based on machine learning, the method comprising:
[0007] Step 1: Use the SO-OCDM signal as the transmitted duplex signal to construct the mth transmitter at time kT s The transmitted SO-OCDM signal expression is m k, m = 1, ..., M;
[0008]
[0009] Where M is the total number of transmitters, k is the sampling point index, and T s is the sampling time interval, f Δ is the frequency interval between the signals transmitted by adjacent transmitters, E m is the energy of the transmitted signal, P is the number of OCDM symbols, L is the number of chirp subcarriers contained in one OCDM symbol, and b p,l,m The communication symbol modulated on the lth subcarrier in the pth symbol transmitted by the mth transmitter, q p,l,m [k] includes the guard interval and chirp subcarrier;
[0010]
[0011] T GI,m is the duration of the protection interval, T m is the symbol duration, T o,m is the duration of a complete OCDM symbol and satisfies T o,m =T GI,m +T m ;
[0012] Step 2: Express the M transmission signal parameters to be designed as vector form λ;
[0013]
[0014] Among them, λ m =[E m ,T m ] T , M represents the number of transmitters, (·) T represents transpose;
[0015] Step 3: Construct a signal design network, including a DDPG algorithm module and a SO-OCDM signal generator. DDPG stands for Deep Deterministic Policy Gradient. The DDPG algorithm module includes an actor network, a critic network, a target actor network, and a target critic network. During the i-th interaction between the signal design network and the dual-function receiving network, the input of the actor network is the received signal energy vector, and the output is the optimized SO-OCDM signal parameter vector.
[0016]
[0017] Among them, the received signal energy vector γ i =[γ 11 , ..., γ m ,...,γ MN ] T , γ mn is the signal energy received by the nth receiver from the mth transmitter, represents the actor network, θ A are the trainable parameters of the actor network, is the output of the actor network; the input of the critic network is the received signal energy vector γ i and the SO-SO-OCDM signal parameter vector The output is an estimate of the value of the action state θ C are the trainable parameters of the critic network; the target actor network Used to generate the energy vector γ of the received signal at the next moment i+1 The corresponding signal parameter vector are the trainable parameters of the target actor network, γ i+1 is the network input, is the network output; target critic network Used to evaluate the energy vector γ of the received signal at the next moment i+1 and signal parameter vector The corresponding action state value is a trainable parameter of the target critic network; after the dual-function receiving network feeds back the channel information to the signal design network through the wired feedback link, γ i , the signal design network generates the signal parameter vector corresponding to the channel information The SO-OCDM signal generator then generates a corresponding duplex SO-OCDM signal according to the signal parameter vector;
[0018] Step 4: Calculate the value of the nth dual-power receiver at the time kT during the ith interaction s The received signal is r n,i [k];
[0019]
[0020] in, is the sum of the M target reflection path signals reflected by G targets, is the sum of the signals of M direct paths, z n,i [k] is noise; and represent the path fading coefficients of the target reflection path and the direct path, respectively. and represents the delay between the target reflection path and the direct path, s m,i (.) represents the signal transmitted by the mth transmitter during the i-th interaction;
[0022] Step 5: Combine the target information vector and the communication information vector to construct the synaesthesia integrated information vector ψ i ;
[0023]
[0024] in,
[0025] ρ i =[ρ 1,i , ..., ρ j,i , ...ρ j,i ] T
[0026] b i =[b 1,i ,...,b m,i ,...,b M,i ] T
[0027] b m,i =[b 1,1,m,i ,...,b l,p,m,i ,...,b L,P,m,i ]
[0028] b l,p,m,i=[b l,p,m,1,i ,...,b l,p,m,d,i ,...,b l,p,m,D,i ]
[0029] j,i Indicates whether there is a target at the jth grid point, b l,p,m,d,i represents the dth bit on the lth subcarrier in the pth symbol transmitted by the mth transmitter;
[0030] Step 6: Construct a dual-power receiver network based on supervised learning, which includes a signal preprocessing module, a fully connected network and a comparator; in the signal preprocessing module, remove the guard interval, bandpass filter, de-chirp subcarrier, and stack the signal to obtain the signal vector
[0031] in,
[0032]
[0033] K GI,m is the number of sampling points in the guard interval, K o,m is the number of sampling points to complete the symbol, k'T s ∈[pT o,m +T GI,m ,pT o,m +T o,m ), (·) * represents the conjugate operation, k′ represents the sample index after removing the guard interval;
[0034] Step 7: Signal Vector It is input into the fully connected network, and the network output is the estimated synaesthesia integration information vector;
[0035]
[0036] in, The received signal vector at the i-th interaction, θ DFR are the trainable parameters of the dual-power receiver network, is the output of the dual-power receiver network; compare the network output with the threshold and make the final decision ψ i ;
[0037] Step 8: Determine the objective function of the signal design network, the loss function of the dual-power receiver network, and the parameter update method;
[0038] Step 9: Generate a transmit signal based on the SO-OCDM signal model, and then generate a receive signal and the corresponding label ψ i; When training the signal design network, the signal design network interacts with the dual-power receiver network I times, and the signal design network parameters are updated once each interaction; When training the dual-power receiver network, the signal design network interacts with the dual-power receiver network I times, and when all the interaction processes are completed, the dual-power receiver network parameters are updated; Alternately iterate the training of the signal design network and the dual-power receiver network to achieve joint training of dual-power transmission and reception; When training the dual-power receiver network, the signal design network parameters θ are fixed A ,θ C , and For θ DFR Update; when training the signal design network, fix the network parameters θ DFR , for θ A ,θ C , and Update; the two training steps are iterated alternately until the entire network converges; represents the trainable parameters of the target actor network, represents the trainable parameters of the target critic network;
[0039] Step 10: Generate a SO-OCDM signal according to the transmission signal model. After the signal passes through the channel, the received signal is obtained. The received signal is input into the duplex receiver network to obtain the output of the network. Finally, the output of the network ψ i Compare with label ψ i , to evaluate system performance.
[0040] Furthermore, the specific method of step 8 is:
[0041] The objective function L(θ) for training the actor network A )for:
[0042]
[0043] in, represents the critic network, It means to find the expectation about the signal parameter vector and signal energy vector. I represents the total number of interactions, and the signal energy vector γ∈{γ i |i=1,...,I}; at the uth iteration, update θ according to the gradient ascent A ,Right now
[0044]
[0045] Among them, η A is the learning rate of the actor network, ← denotes the assignment operation, θ A,u It is uThe parameters of the actor network at iteration , is L(θ A ) in the parameter θ A =θ A,u-1 The gradient of the target actor network Perform a soft update.
[0046]
[0047] Among them, 0<<β s <1 is the weight factor of soft update, represents the target actor network at the u-th iteration;
[0048] The objective function L(θ) of the critic network C )for,
[0049]
[0050] Among them, rewards represents the reward for the i-th interaction, and the next received signal energy vector γ′∈{γ i+1 |i=1,...,I}, β TD represents the discount factor, represents the target critic network, represents the critic network, It means the expectation of the signal parameter vector, the received signal energy vector, the reward and the received signal energy vector at the next moment; at the uth iteration,
[0051]
[0052] Among them, η C is the learning rate, θ C,u represents the actor network parameters at the u-th iteration, is L(θ C ) in the parameter θ C =θ C,u1 The target critic network and the critic network have the same network architecture and the same initialization network parameters. The target actor network is also soft-updated.
[0053]
[0054] in, represents the target actor network parameters at the u-th iteration;
[0055] The loss function L of the dual-power receiver network DFR (θ DFR )for:
[0056]
[0057] in, represents the expectation of the received signal vector and the synaesthesia integrated information vector, ψ={ψ i |i1,...,I}, ψ q and represents ψ and The qth element in, q=1,...,Q, Q is the total number of elements, l ISAC = {l ISAC,i |i=1,...,I} is the loss of each training data, and the trainable parameters in the network are trained using gradient descent.
[0058]
[0059] Among them, η DFR is the learning rate, is L(θ DFR ) in the parameter θ DFR =θ DFR,u1 The gradient of time.
[0060] The beneficial effects of the present invention are: by designing a dual-power receiver, decoding and multi-target detection tasks are simultaneously realized using the same received signal, thereby improving system performance. Furthermore, by adopting reinforcement learning and supervised learning, the joint design of the dual-power transmission signal and the dual-power receiver is realized. Compared with the traditional method, the method based on machine learning is less dependent on the mathematical models of channels, targets, clutter, etc., and can effectively improve the system performance, providing a feasible method for the joint design problem of transmission and reception in the field of synaesthesia integration. BRIEF DESCRIPTION OF THE DRAWINGS
[0061] Figure 1 Schematic diagram of the deep deterministic policy gradient network architecture.
[0062] Figure 2 Dual-function receiver network based on supervised learning.
[0063] Figure 3 Schematic diagram of the interaction process between the signal design network and the dual-power receiver network.
[0064] Figure 4 The objective function of the signal design network changes with the number of iterations.
[0065] Figure 5 The objective function of the dual-power receiver network varies with the number of iterations.
[0066] Figure 6 Communication decoding performance curves of dual-receiver network and independent receiver network.
[0067] Figure 7 Radar multi-target detection performance curves of dual-power receiver network and independent receiver network.
[0068] Figure 8 Communication decoding performance curves when the transmitter and receiver are jointly designed and when only the receiver is designed.
[0069] Fig. 9 Radar multi-target detection performance curves with jointly designed transmitter and receiver and with only the receiver designed. DETAILED DESCRIPTION
[0070] For the convenience of description, the following definitions are first made:
[0071] (·) T is the transpose, (·) * represents conjugation, It means to find the ergodic average. Represents the gradient operator.
[0072] Consider a MIMO telepresence system with M transmitters and N receivers, which monitors an area with multiple targets while communicating.
[0073] Step 1: Assume that in the SO-OCDM signal transmitted by the mth (m=1,...,M) transmitter, a frame contains P SO-OCDM symbols and the sampling period is T s , the sample number is k (k = 1, ..., K), and the total number of sampling points is K. The SO-OCDM signal of the mth transmitter is
[0074]
[0075] Among them, k is the sampling point index, f Δ is the frequency interval between the signals transmitted by adjacent transmitters, E m is the energy of the transmitted signal, P is the number of OCDM symbols, L is the number of chirp subcarriers contained in one OCDM symbol, and b p,l,m is the communication symbol modulated on the lth subcarrier in the pth symbol transmitted by the mth transmitter and b l,p,m =f Modu ([b l,p,m,1 ,...,b l,p,m,d ,...,b l,p,m,D ]), f Modu (·) is the modulation mode, b l,p,m,d is the bit data transmitted in communication, q p,l,m [k] includes the guard interval and chirp subcarrier,
[0076]
[0077] T GI,m is the duration of the protection interval, T m is the symbol duration, T o,m is the duration of a complete OCDM symbol and satisfies T o,m =T GI,m T m .
[0078] Step 2: From equations (1)-(3), we can see that the SO-OCDM waveform s m k is completely determined by the signal parameters, so the SO-OCDM signal optimization problem is equivalent to the optimization problem of the waveform parameter λ. The signal energy and symbol duration to be designed on the mth transmitter are integrated into a parameter vector, which is expressed as
[0079] λ m =E m ,T m T , (4) The parameter vector composed of the signal parameters of all SO-OCDM at M transmitters is
[0080]
[0081] In order to improve system performance, a dual-power transmission signal is designed.
[0082] Step 3: Use reinforcement learning to solve the design problem of dual-power SO-OCDM signals. In the i-th interaction process, the state is the received signal energy vector γ i , the specific form is
[0083] γ i =[γ 11,i ,...,γ mn,i ,...,γ MN,i ] T , (6)
[0084] Among them, γ mn is the signal energy received by the nth receiver from the mth transmitter. Action is the SO-OCDM signal parameter vector to be designed, and its specific form is
[0085]
[0086] Among them, λ m,i [E m,i ,T m,i ] T . The reward is the synaesthesia integration system reward is
[0087]
[0088] Among them, l ISAC,i It represents the performance index of the receiving end of the synaesthesia integrated system with dual transmit and receive functions. ISAC,i Negatively correlated with signal performance, and are the bonuses for auxiliary energy and duration, β E and T is the weighting factor for energy and symbol duration bonus.
[0089] About energy rewards for,
[0090]
[0091] Among them, a>0 is a constant, is the energy received by all duplex receivers from the mth transmitter. Bonus for symbol duration for
[0092]
[0093] T m,i and T o,m,i are the symbol duration and the complete symbol duration at the i-th interaction, respectively.
[0094] Build a signal design network, including the DDPG algorithm module and the SO-OCDM signal generator. The DDPG algorithm architecture is as follows: Figure 1 As shown in Figure 2. The DDPG algorithm consists of four networks: actor network, critic network, target actor network, and target critic network. The actor network is used to implement the strategy. During the i-th interaction, the input of the actor network is the received signal energy vector γ i , the output of the network is the action
[0095]
[0096] in, represents the actor network, θ A are the trainable parameters of the actor network, is the output of the actor network. The critic network The input is the received signal energy vector γ i and the SO-OCDM signal parameter vector Used to approximate the optimal action-state value function, where θ C is a trainable parameter of the critic network. When the received signal energy vector γ is given i+1 , target actor network Generate optimal actions are the network parameters of the target actor network. Then, the received signal energy vector γ i+1 and actions Input to the target critic network Used to evaluate the corresponding action state value, is the network parameter of the target critic network. For the target actor network and the target critic network, the soft update method is used to update the network parameters. The introduction of the target actor network and the target critic network can promote the convergence of iterations. After the dual-function receiving network feeds back the channel information to the signal design network through the wired feedback link, γ i , the signal design network generates the signal parameter vector corresponding to the channel information The SO-OCDM signal generator then generates a corresponding duplex SO-OCDM signal according to the signal parameter vector.
[0097] Step 4: For the receiving end, the nth (n=1,...,N) duplex receiver at the time kT of the ith interaction s The received signal is
[0098]
[0099] Among them, the target number G<G max , G max is the maximum number of targets that may appear in the monitoring area, is the direct path signal, is the target reflection signal, z n,i [k] is noise. For the mnth direct path, the path loss coefficient is
[0100]
[0101] The nth one is located The receiver and the mth receiver located at The distance between the transmitters is
[0102]
[0103] Here, μ d It means that the energy of the transmitted signal and the energy of the received signal are at a distance of d. mn,i =1, is the propagation delay, c is the speed of light. For the mnth direct path, the target reflection coefficient of the gth target is
[0104]
[0105] Here, g,i is the target reflection coefficient, the mth transmitter and the gth transmitter located at (x g,i ,y g,i) is
[0106]
[0107] The distance between the nth receiver and the gth target is
[0108]
[0109] The delay caused by the target is
[0110]
[0111] The energy of the transmitted signal and the energy of the received signal are at a distance d. mg,i =d gn,i The ratio of 1 is μ t The last term is mixed Gaussian noise,
[0112]
[0113] Among them, CN(0,σ v ) means that the mean is 0 and the variance is σ v The coefficient of the vth Gaussian component is κ v And its satisfaction
[0114] Step 5: All bits carried on l subcarriers in the p-th symbol at the m-th transmitter are expressed as a vector
[0115] b l,p,m,i =[b l,p,m,1,i ,...,b l,p,m,D,i ], (20)
[0116] The bit vector transmitted by the mth transmitter is
[0117] b m,i =[b 1,1,m,i ,...,b l,p,m,i ,,b L,P,m,i ], (twenty one)
[0118] The communication bit vector transmitted by M transmitters is denoted as
[0119] b i =[b 1,i ,...,b m,i ,,b M,i ] T . (twenty two)
[0120] For radar target detection tasks, the monitoring area is divided into J grid points, and the target information vector can be recorded as
[0121] ρi =[ρ 1,i , ..., ρ J,i ] T . (twenty three)
[0122] For the dual-power receiver, its task is to estimate the synaesthesia information vector composed of the communication bit vector and the radar target information vector, that is,
[0123]
[0124] Among them, ψ i ∈0,1 Q , the size of Q is J+MPLD, and the value of Q is related to the number of grid points J, the number of transmitters M, the number of symbols P, the number of subcarriers L and the number of modulated bits D.
[0125] Step 6: Construct a dual-power receiver network based on supervised learning to achieve joint decoding and detection. The dual-power receiver network includes a signal preprocessing module, a fully connected network and a comparator. In the signal preprocessing module, first remove the index kT s ∈[pT o,m ,pT o,m T GI,m ) sampling point to remove the guard interval. Then, the signal from the mth transmitter received by the nth duplex receiver is obtained by bandpass filtering the received signal.
[0126]
[0127] Among them, k′T s ∈[pT o,m +T GI,m ,(p+1)T o,m ,), Then perform a de-chirp operation to obtain the information carried by the lth subcarrier in the pth symbol on the mth transmitter.
[0128]
[0129] Stack the sampling points of the L subcarriers and P symbols of the mnth path, and we have
[0130]
[0131] All sampling points of MN paths are
[0132]
[0133] Step 7: Signal Vector The neural network input into the duplex receiver is used to solve the joint decoding and detection problem. To represent the neural network, the network output is the estimated synaesthesia integration information vector,
[0134]
[0135] Among them, θ DFR are the trainable parameters of the network, The first J elements in the matrix correspond to the probability of whether there is a target at each of the J grid points. The remaining elements are the probability values of the communication bit data being 0 or 1, expressed as Indicates. Each element in ρ The final verdict is
[0136]
[0137] in, and for and The jth element in . If The final verdict is Indicates that the network determines that the jth grid point has a target, and vice versa. Similarly, the element With threshold Λ b For comparison, if Then the corresponding bit data is judged to be 1, otherwise it is 0. Figure 2 In the figure, the network architecture of the dual-power receiver network is shown.
[0138] Step 8: The goal of training the actor network is to adjust the network parameters θ A , the objective function is to approximate the optimal strategy that can maximize the action-state value function, and use gradient ascent to update θ A , and the parameters of the target actor network Perform soft updates. The goal of the critic network is to accurately estimate the optimal action-state value function, minimize the temporal difference error, and use gradient descent to update the critic network parameters θ C , and the network parameters of the target critic network Perform soft update. The task of the dual-power receiver network is a multi-classification task, using cross entropy as its loss function, and gradient descent to update the network parameters θ of the dual-power receiver network DFR .
[0139] Step 9: The tuple used to train the dual-power signal design network is generated by the interaction between the dual-power receiver and the dual-power signal design network. During the ith interaction between the dual-power signal design network and the dual-power receiver, the dual-power signal design network obtains the received signal energy vector γ fed back by the dual-power receiver i According to γ i , the dual-function signal design network generates a signal parameter vector When training the dual-function signal design network, in order to explore the signal parameter space, a random term is added to the action, that is,
[0140]
[0141] in, is a zero-mean Gaussian distribution with variance Σ s The random terms are added only when training the dual-function waveform design network. When training the dual-power receiving network and testing, the signal parameter vector output by the network is directly used Dual-function signal design network using signal parameter vector Generate dual-power transmission signal s m,i [k], the SO-OCDM signal is received by the dual-power receiver after channel attenuation and target reflection. Then, the dual-power receiver calculates the next received signal energy vector γi+1 and the reward And feed it back to the dual-function signal design network. So far, the required tuples for training is stored in the experience replay buffer In the interaction process, Figure 3 shown.
[0142] The task of the dual-power receiver network is to obtain the synaesthesia information vector from the signal vector, so the training data is the signal vector and the label is the synaesthesia information vector. In order to obtain the training data of the dual-power receiver network, the dual-power signal design network is based on the signal parameter vector A dual-power transmission signal is generated, and after the dual-power transmission signal is attenuated by the channel and reflected by the target, it is received by the dual-power receiver. After preprocessing by the dual-power receiver network, a signal vector can be obtained. During the first interaction between the dual-power receiver network and the dual-power signal design network, the data and labels are stored in the buffer. When the interaction is over, use The network of the dual-power receiver network is trained with the samples and labels in . After each training, reset The task of the dual-function receiver network is a multi-classification task, so cross entropy is used as its loss function.
[0143] When training the dual-function signal design network, randomly select the experience buffer B WavExtract a batch of data from the network to update the network parameters θ of the dual-power signal design network A and C , at this time the target actor network parameters Target critic network parameters and the network parameters of the dual-power receiver network θ DFR is fixed. When the parameter θ A and C After updating, the parameters and When training the dual-power receiver network, fix all network parameters in the dual-power signal design network and use the buffer area The training data in updates the network parameters θ of the dual-power receiver network DFR .
[0144] Step 10: When the network converges, the optimized actor network parameters are obtained and dual-power receiving network parameters First, the dual-power SO-OCDM signal with the initialization signal parameters is transmitted to obtain the initialization received signal energy vector γ1 of the environment feedback. The actor network in the dual-power signal design network generates the optimal signal parameter vector in the current scenario according to the received signal energy vector γ1. Then, the SO-OCDM signal generator transmits the signal in duplex mode. After the signal is attenuated by the path and reflected by the target, it is received by the duplex receiver, which outputs the target and communication information vector estimation value ψ2 and the next received signal energy vector γ2, and records ψ2 and its corresponding label. This process is repeated until the test data collection is completed.
[0145] Working principle of the present invention
[0146] In step 6, the dual-power receiver uses the same received signal to simultaneously complete communication decoding and radar multi-target detection. For communication decoding, the bit information is 0 or 1, and the communication decoding task can be regarded as a classification problem. All bits carried on the l subcarriers in the p-th symbol on the m-th transmitter are expressed as a vector form:
[0147] b l,p,m,i =[b l,p,m,1,i ,...,b l,p,m,D,i ], (32) The bit vector transmitted by the mth transmitter is
[0148] b m,i =[b 1,1,m,i ,...,b l,p,m,i ,,b L,P,m,i ], (33) The communication bit vector transmitted by M transmitters is denoted as
[0149] b i =[b 1,i ,...,b m,i ,,b M,i ] T . (34) Assuming that the target exists in a two-dimensional monitoring area,
[0150] G={(x,y)|x∈[x min ,x max ],y∈[y min ,y max ]}, (35)
[0151] Among them, [x min ,x max ] and [y min ,y max ] is the range of the region. Divide the entire region into J=J x J y grid points, among which, represents rounding down, Δx and y are the size of the grid point. g,i ,y g,i ) is the target, and it is considered that at the jth grid point, Since the entire area is divided into J grid points, each grid point has two possible situations, and the target exists ρ j,i =1 or does not exist j,i =0(J=1, ..., J).
[0152] Then the target information vector can be written as
[0153] ρ i =[ρ 1,i ,...,ρ J,i ] T . (36)
[0154] Therefore, the multi-target problem is transformed into the problem of detecting whether there is a target at each grid point, and the task of the radar receiver can also be regarded as a classification task. For the dual-power receiver, its task is to estimate the synaesthesia integrated information vector composed of the communication bit vector and the radar target information vector, that is,
[0155]
[0156] Among them, ψ i ∈0,1 Q , the size of Q is J+MPLD, and the value of Q is related to the number of grid points J, the number of transmitters M, the number of symbols P, the number of subcarriers L and the number of modulated bits D.
[0157] In step 8, the goal of training the actor network is to adjust the network parameters θ A Approximately the optimal policy that can maximize the state value function, the objective function is
[0158]
[0159] Among them, the signal parameter vector The received signal energy vector γ∈{γ i |i1,...,I}. At the uth iteration, update θ according to the gradient ascent A ,Right now
[0160]
[0161] Among them, η A is the learning rate of the actor network, θ A,u are the parameters of the actor network at the u-th iteration, is L(θ A ) in the parameter θ A =θ A,u-1 The gradient of the target actor network Perform a soft update.
[0162]
[0163] Among them, 0<<β s <1 is the weight factor of soft update, represents the target actor network at the u-th iteration.
[0164] The goal of the critic network is to accurately estimate the optimal Q function, and the objective function is the TD error.
[0165]
[0166] Among them, rewards The next received signal energy vector γ′∈{γ i+1 |i=1,...,I}, At the uth iteration, we have
[0167]
[0168] Among them, η C is the learning rate, θ C,u represents the actor network parameters at the u-th iteration, is L(θ C ) in the parameter θ C =θ C,u-1 The target critic network and the critic network have the same network architecture and the same initialization network parameters. The target actor network is also soft-updated.
[0169]
[0170] in, represents the target actor network parameters at the u-th iteration.
[0171] The task of the dual-function receiver network is a multi-classification task, using cross entropy as its loss function.
[0172]
[0173] Among them, ψ∈{ψ i |i=1,...I,, ψ q and represents ψ and The qth (q=1,...,Q)th element in l ISAC = {l ISAC,i |i=1,...,I} is the loss of each training data. Gradient descent is used to train the trainable parameters in the network.
[0174]
[0175] Among them, η DFR is the learning rate, is L(θ DFR ) in the parameter θ DFR =θ DFR,u-1 The gradient of time.
[0176] At the transmitting end, phase shift keying is used to modulate the communication data, that is, D = 1. In order to obtain diversity gain, it is assumed that different transmitters transmit the same information. In order to simplify the simulation difficulty, it is assumed that the number of symbols of the SO-OCDM signal transmitted by all transmitters is the same P = 1, the number of subcarriers is L = 8, and the duration of the guard interval is fixed to T GI,m =1.6ms, K o,m =K o =14. Initialize the transmission signal parameter to λ0, where the signal energy distribution ratio is E1E m 1 / M, symbol duration is T1=...=T M =3.2×10 -3 ms. Assume that there are M = 2 transmitters in the system, located at (1.8, 1) km and (1.8, 4) km, and N = 2 receivers, located at (4.4, 1.4) km and (4.4, 3.7) km. The range of the monitored area is x min =3.3km,x max =3.7km,y min =1.7km and y max=3.2km. The transmit and receive energy ratios of the direct path and the target reflection path are μ=8×10 6 and t =4×10 12 The size of the grid is Δx = 400m and Δy = 500m, that is, J = 2. The number of interactions is set to I = 5.
[0177] When constructing the network, the actor network contains 2 hidden layers, 1 output layer and 1 normalization layer. The number of nodes in the two hidden layers is 128 and 64 respectively, and the number of nodes in the output layer is 2M. Finally, the normalization layer is introduced to make the signal parameters meet the constraints. The critic network contains 2 hidden layers, with 128 and 64 nodes respectively, and 1 output layer with 1 node. For the actor network and the critic network, the activation function of the middle layer is ReLU. The learning rate of the actor network and the critic network in the dual-function signal design network is η A =η C = 0.001. The discount factor in the action value function is βd = 0.9, and the weight of the target network soft update is β s =0.99. The weight factors of energy reward and time efficiency reward in the reward are β E = 0.3 and β T =0.2. The number of nodes in the two hidden layers of the dual-power receiver network is 500 and 100 respectively, and the number of nodes in the output layer is J+L×P. ReLU is selected as the activation function of the hidden layer, sigmoid is selected as the activation function of the output layer, and the learning rate is η DFR =0.004.
[0178] During training, the experience replay buffer The capacity is 5000. During each training, the dual-function signal design network is 32 tuples are randomly selected from the network to update the parameters of the dual-power signal design network. The number of alternating iterations of the dual-power signal design network and the dual-power receiver network is 128. During training, the signal-to-noise ratio is set to 20 dB. During testing, the signal-to-noise ratio range is SNR∈[6.4,13] dB, and 10240 scenarios are randomly generated at each signal-to-noise ratio to test the performance.
[0179] exist Figure 4 In Figure 2, we show how the cumulative rewards in ε scenarios change as the number of alternating iterations of the dual-power signal design network and the dual-power receiver network increases. Figure 5 The loss value of the dual-power receiver network changes with the increase of the number of iterations. It can be seen that after 80 alternations between the dual-power signal design network and the dual-power receiver network, the cumulative reward and loss value changes tend to be stable, indicating that the iteration of the entire network has converged.
[0180] For communication decoding tasks, SER is used as the performance evaluation index of communication. For radar multi-target detection tasks, the detection probability P is used. D As an evaluation indicator. In the multi-target detection problem, the detection probability is defined as the probability that the number and location of multiple targets are correctly judged. Figure 6 and Figure 7 In the figure, the differences in communication SER and radar detection performance between an independent radar communication receiver and a dual-power receiver are shown. It can be seen that the performance of the dual-power receiver is better than that of the independent radar communication receiver in terms of both communication and perception performance, because the network realizes the sharing of radar communication information through parameter sharing.
[0181] exist Figure 8 and Fig. 9 In the figure, the communication SER and radar detection performance of the case of joint optimization of the transceiver network and optimization of the dual-power receiver network are shown respectively. Whether in terms of communication or perception performance, the performance of the joint design of the transceiver network is better than that of the dual-power receiver network design using the initialization signal, and the performance gain brought by the joint design of the transceiver is very obvious, which shows the effectiveness of the dual-power transmission signal design.
Claims
1. A joint design method of a synaesthesia integrated dual-function transceiver network based on machine learning, the method comprising: Step 1: Use the SO-OCDM signal as the transmitted duplex signal to construct the mth transmitter at time kT s The transmitted SO-OCDM signal expression is m k, m = 1, ..., M; Where M is the total number of transmitters, k is the sampling point index, and T s is the sampling time interval, f Δ is the frequency interval between the signals transmitted by adjacent transmitters, E m is the energy of the transmitted signal, P is the number of OCDM symbols, L is the number of chirp subcarriers contained in one OCDM symbol, and b p,l,m The communication symbol modulated on the lth subcarrier in the pth symbol transmitted by the mth transmitter, q p,l,m [k] includes the guard interval and chirp subcarrier; T GI,m is the duration of the protection interval, T m is the symbol duration, T o,m is the duration of a complete OCDM symbol and satisfies T o,m =T GI,m +T m ; Step 2: Express the M transmission signal parameters to be designed as vector form λ; Among them, λ m [E m ,T m ] T , M represents the number of transmitters, (·) T represents transpose; Step 3: Construct a signal design network, including a DDPG algorithm module and a SO-OCDM signal generator. DDPG stands for Deep Deterministic Policy Gradient. The DDPG algorithm module includes an actor network, a critic network, a target actor network, and a target critic network. During the i-th interaction between the signal design network and the dual-function receiving network, the input of the actor network is the received signal energy vector, and the output is the optimized SO-OCDM signal parameter vector. Among them, the received signal energy vector γ i =[γ 11 ,...,γ mn ,...γ mn ] T , γ mn is the signal energy received by the nth receiver from the mth transmitter, represents the actor network, θ A are the trainable parameters of the actor network, is the output of the actor network; the input of the critic network is the received signal energy vector γ i and the SO-OCDM signal parameter vector The output is an estimate of the value of the action state θ C are the trainable parameters of the critic network; the target actor network Used to generate the energy vector γ of the received signal at the next moment i+1 The corresponding signal parameter vector are the trainable parameters of the target actor network, γ i+1 is the network input, is the network output; target critic network Used to evaluate the energy vector γ of the received signal at the next moment i+1 and signal parameter vector The corresponding action state value is a trainable parameter of the target critic network; after the dual-function receiving network feeds back the channel information to the signal design network through the wired feedback link, γ i , the signal design network generates the signal parameter vector corresponding to the channel information The SO-OCDM signal generator then generates a corresponding duplex SO-OCDM signal according to the signal parameter vector; Step 4: Calculate the value of the nth dual-power receiver at the time kT during the ith interaction s The received signal is r n,i [k]; in, is the sum of the M target reflection path signals reflected by G targets, is the sum of the signals of M direct paths, z n,i [k] is noise; and represent the path fading coefficients of the target reflection path and the direct path, respectively. and represents the delay between the target reflection path and the direct path, s m,i (.) represents the signal transmitted by the mth transmitter during the i-th interaction; Step 5: Combine the target information vector and the communication information vector to construct the synaesthesia integrated information vector ψ i ; in, r i =[ρ 1,i ,...,r j,i ,...,r J,i ] T b i =[b 1,i ,...,b m,i ,...,b M,i ] T b m,i =[b 1,1,m,i ,...,b l,p,m,i ,...,b L,P,m,i ] b l,p,m,i =[b l,p,m,1,i ,...,b l,p,m,d,i ,...,b l,p,m,D,i ] o j,i Indicates whether there is a target at the jth grid point, b l,p,m,d,i represents the dth bit on the lth subcarrier in the pth symbol transmitted by the mth transmitter; Step 6: Construct a dual-power receiver network based on supervised learning, which includes a signal preprocessing module, a fully connected network and a comparator; in the signal preprocessing module, remove the guard interval, bandpass filter, de-chirp subcarrier, and stack the signal to obtain the signal vector in, K GI,m is the number of sampling points in the guard interval, K o,m is the number of sampling points to complete the symbol, k'T s ∈[pT o,m +T GI,m ,pT o,m +T o,m ), (·) * represents the conjugate operation, k′ represents the sample index after removing the guard interval; Step 7: Signal Vector It is input into the fully connected network, and the network output is the estimated synaesthesia integration information vector; in, The received signal vector at the i-th interaction, θ DFR are the trainable parameters of the dual-power receiver network, is the output of the dual-power receiver network; compare the network output with the threshold and make the final decision ψ i ; Step 8: Determine the objective function of the signal design network, the loss function of the dual-power receiver network, and the parameter update method; Step 9: Generate a transmit signal based on the SO-OCDM signal model, and then generate a receive signal and the corresponding label ψ i ; When training the signal design network, the signal design network interacts with the dual-power receiver network I times, and the signal design network parameters are updated once each interaction; When training the dual-power receiver network, the signal design network interacts with the dual-power receiver network I times, and when all the interaction processes are completed, the dual-power receiver network parameters are updated; Alternately iterate the training of the signal design network and the dual-power receiver network to achieve joint training of dual-power transmission and reception; When training the dual-power receiver network, the signal design network parameters θ are fixed A ,θ C , and For θ DFR Update; when training the signal design network, fix the network parameters θ DFR , for θ A ,θ C , and Update; the two training steps are iterated alternately until the entire network converges; represents the trainable parameters of the target actor network, represents the trainable parameters of the target critic network; Step 10: Generate a SO-OCDM signal according to the transmission signal model. After the signal passes through the channel, the received signal is obtained. The received signal is input into the duplex receiver network to obtain the output of the network. Finally, the output ψi of the network is compared with the label ψi to evaluate the system performance.
2. The method for designing a synaesthesia integrated dual-function transceiver network based on machine learning as claimed in claim 1, characterized in that: The specific method of step 8 is: The objective function L(θ) for training the actor network A )for: in, represents the critic network, It means to find the expectation about the signal parameter vector and signal energy vector. I represents the total number of interactions, and the signal energy vector γ∈{γ i |i=1,...,I}; At the uth iteration, update θ according to the gradient ascent A ,Right now Among them, η A is the learning rate of the actor network, ← denotes the assignment operation, θ A,u are the parameters of the actor network at the u-th iteration, is L(θ A ) in the parameter θ A =θ A,u-1 The gradient of the target actor network Perform a soft update. Among them, 0<<β s <1 is the weight factor of soft update, represents the target actor network at the u-th iteration; The objective function L(θ) of the critic network C )for, Among them, rewards represents the reward for the i-th interaction, and the next received signal energy vector γ′∈{γ i+1 |i=1,...,I}, β TD represents the discount factor, represents the target critic network, represents the critic network, It means the expectation of the signal parameter vector, the received signal energy vector, the reward and the received signal energy vector at the next moment; at the uth iteration, Among them, η C is the learning rate, θ C,u represents the actor network parameters at the u-th iteration, is L(θ C ) in the parameter θ C =θ C,u-1 The target critic network and the critic network have the same network architecture and the same initialization network parameters. The target actor network is also soft-updated. in, represents the target actor network parameters at the u-th iteration; The loss function L of the dual-power receiver network DFR (θ DFR )for: in, represents the expectation of the received signal vector and the synaesthesia integrated information vector, ψ={ψ i |i=1,...,I}, ψ q and represents ψ and The qth element in, q = 1, ..., Q, Q is the total number of elements, l ISAC = {l ISAC,i |i=1,...,I} is the loss of each training data, and the trainable parameters in the network are trained using gradient descent. Among them, η DFR is the learning rate, is L(θ DFR ) in the parameter θ DFR =θ DFR,u-1 The gradient of time.
Citation Information
Patent Citations
Radar communication integrated signal design method based on OCDM
CN112737998A
MIMO radar communication double-work integrated joint transmitting and receiving method based on machine learning
CN117254999A
Cognitive interference integrated beam forming method based on deep reinforcement learning
CN117614501A
Cited By
Pseudo-satellite signal design method and device based on intelligent framework
CN120085327A
Pseudo-satellite signal design method and device based on intelligent framework
CN120085327B
