CN-122028181-A - Downlink NOMA system power distribution method based on deep learning

Abstract

The invention belongs to the technical field of wireless communication, and particularly relates to a power distribution method of a downlink NOMA system based on deep learning; the invention adopts a two-stage decoupling architecture of off-line high-quality data generation and on-line deep learning rapid reasoning, adopts an improved particle swarm optimization algorithm to generate a globally optimal or nearly optimal power distribution data set aiming at maximizing the total effective capacity of NOMA users in an off-line stage, and utilizes the data set to train a deep neural network model in an on-line stage to realize rapid mapping from a channel state to an optimal power distribution coefficient, thereby realizing millisecond-level real-time power distribution while guaranteeing statistical service quality.

Inventors

• WANG YIFAN
• JI FENGLEI
• CHI XUEFEN

Assignees

• 吉林大学

Dates

Publication Date

20260512

Application Date

20260205

Claims (3)

1. 1. The power distribution method of the downstream NOMA system based on deep learning is characterized by comprising the following steps of: Step one, obtaining The data are specifically as follows: Step 1.1, generating a channel gain sample pair of a user at a time slot t , ) Wherein For the channel gain of user 1 at time slot t, The channel gain sample pairs corresponding to M time slots are recorded as channel gain samples, and the channel original components of the channel gain samples are extracted, namely the complex channel gain from the base station to the user 1 under each time slot And base station to user 2 complex channel gain Wherein t=1, 2 a. M; Step 1.2, the channel gain samples are processed as follows to obtain channel gain samples Data: the sum of all complex channel gain squares of the channel gain samples are spliced to form an input eigenvector of the channel gain samples : , Wherein the first M dimension is the channel gain square sum sequence of the user 1, and the second M dimension is the channel gain square sum sequence of the user 2; input eigenvector to channel gain samples Z-score normalization was performed: , Wherein the method comprises the steps of , Statistical feature vector of channel state information of channel gain samples respectively Mean and standard deviation of (a); for inputting feature vectors Z-score normalization results of (C); step two, constructing a cascade deep neural network model, wherein the model is formed by sequentially connecting a trained first-stage neural network model and a trained second-stage neural network model; The first-stage neural network model has the structure that: an input layer comprising 200 neurons; A full connection layer; an output layer comprising 2 neurons; the second-level neural network model has the structure as follows: An input layer comprising 2 neurons; three hidden layers, each layer uses a ReLU activation function, dropout regularization is applied, and the discarding rate is 0.2; An output layer comprising 1 neuron, using linear activation; Step three, the channel gain samples are processed Inputting the data into a trained cascade deep neural network model, outputting optimal power distribution coefficient for channel gain sample prediction by the model ; Step four, the base station is according to Calculating the actual power allocated to user 1 in the M time slots Actual power allocated to user 2 。
2. 2. A method for power allocation of a deep learning based downstream NOMA system according to claim 1, wherein said step 1.1 is performed by combining 4-dimensional raw components of channel gain samples [ , , , And (c) recombining, namely calculating the square sum of complex channel gains under each time slot in the channel gain sample: Wherein , Complex channel gains for base station to user k in corresponding time slots Real and imaginary parts of (a) are provided.
3. 3. The method for power allocation of a downstream NOMA system based on deep learning as claimed in claim 1, wherein the training process of the cascaded deep neural network model is as follows: step one, data preparation, which specifically comprises the following steps: s1.1 defining a downstream NOMA System model Received signal model at time slot t, received signal for user k Expressed as: , Wherein: for a coded signal sent by a base station to user k, k=1 or 2, representing the kth user; for the complex channel gain from the base station to user k, obeying independent rayleigh fading; the base station is allocated transmit power for user k in M time slots, The total transmit power allocated to all users in M time slots for the base station satisfies the constraint: ; is additive white gaussian noise; channel state information definition According to the rayleigh fading model, Obeying mean value of Is an exponential distribution of (a), the probability density function of which is: , Assuming that the noise power of the receiving end of the user is normalized, wherein strong and weak users are distinguished by the channel gain strength between the base station and the user, the instantaneous service rate of the user k The method comprises the following steps: instantaneous service rate for strong user, user 1 The method comprises the following steps: , instantaneous service rate for weak user, user 2 The method comprises the following steps: , Wherein: for the channel gain of user 1 at time slot t, Channel gain for user 2 at time slot t; S1.2 defining statistical QoS measurement and optimization target based on effective capacity Effective capacity formula for user k with independent co-distributed block fading channels and specific transmission mode The method comprises the following steps: , Wherein the method comprises the steps of A delay QoS index for user k; probability of successful transmission of information for user k, wherein: probability of successful transmission of information by user 1 : , Probability of successful transmission of information by user 2 : , Wherein the threshold value Joint channel requirements that must be met when using SIC decoding for user 1; for a preset transmission rate of user 1, A preset transmission rate for user 2; Is a probability density function; System optimization objective: Will be And The expression of (2) is substituted into the effective capacity respectively The calculation formula of (2) to obtain the total effective capacity of the system : , S1.3, optimizing inter-partition particle swarm to generate optimal labels S1.31 the optimization interval 0, Dividing into four sub-intervals, wherein: , , , S1.32, generating channel sample batch and extracting characteristics: generating channel gain sample pair of user at time slot t , ) The channel gain sample pair corresponding to each M time slots is recorded as a group of channel gain samples, and the average value of the channel gain squares of the group of channel gain samples is calculated And Statistical feature vector as the set of sample channel state information ; S1.34 performing the following operations to find the optimum : S1.341 first interval Optimizing by initializing a PSO particle group in the interval to Performing PSO iteration as a fitness function, and recording the optimal solution found in the interval And Wherein Is the optimal power distribution value in the interval, namely the interval is optimal And (3) the preparation method Is the maximum system total effective capacity value for that interval; Second interval Optimizing by initializing a PSO particle group in the interval to Performing PSO iteration to obtain optimal solution for fitness function And Wherein Is the optimum power distribution value of the interval, namely the optimum interval And (3) the preparation method Is the maximum system total effective capacity value for that interval; third interval Optimizing by initializing a PSO particle group in the interval to Performing PSO iteration to obtain optimal solution for fitness function And Wherein Is the optimum power distribution value of the interval, namely the optimum interval And (3) the preparation method Is the maximum system total effective capacity value for that interval; S1.342, performing physical feasibility verification: If it is Thought to be Invalidating; If it is Thought to be Invalidating; removing invalid data and maximum system total effective capacity value corresponding to the invalid data from the three subintervals, and selecting the maximum system total effective capacity value corresponding to the maximum system total effective capacity value from the effective data The value is regarded as globally optimal in the state of the set of channel gain samples I.e. ; S1.343 calculating an optimal power allocation coefficient for the set of channel gain samples: ; S1.4 construction of the final dataset Corresponding the set of channel gain samples Statistical feature vector of channel state information And an optimal power distribution coefficient Pairing to form a data sample Obtaining The group data samples form a large-scale data set D; s1.5, preparing tag data: Reading the statistical feature vector of the channel state information corresponding to each group of channel gain samples from the data set D And an optimal power distribution coefficient Will be in each group of data And Together as tag data; S1.6, calculating the corresponding channel gain samples according to the step 1.2 Data; s1.7, training process: training of a first-level neural network model: Will be Taking z data corresponding to the group of channel gain samples as input information of a model, and taking statistical feature vectors of channel state information in tag data corresponding to each group of channel gain samples The output data is used as the output data of the model, and the output data is input into the first-stage neural network model for training until the loss function of the first-stage neural network model converges and the training is finished; the first-stage neural network model adopts an Adam optimizer to update parameters, wherein the initial learning rate is the same as the initial learning rate The learning rate adopts exponential decay, and the learning rate decay of each 50 training steps is 0.95 times of the original learning rate decay; The first level neural network model uses the average absolute error as a loss function: , Wherein the method comprises the steps of Is the loss value of the first-level neural network model, To train the number of sets of channel gain samples used, Is the first The base station corresponding to the group channel gain samples sets a target set value of a statistical feature vector of channel state information of the user 1, Is the first The base station corresponding to the group channel gain sample sets a target set value of a statistical feature vector of channel state information of the user 2; Training of a second-level neural network model: Will be Corresponding to group channel gain samples As input information of the model, optimal power distribution coefficients in the tag data corresponding to each group of channel gain samples are obtained The output data is used as the output data of the model, and the output data is input into the second-level neural network model for training until the loss function of the second-level neural network model converges and the training is finished; the second-level neural network model adopts an Adam optimizer to update parameters, wherein the initial learning rate is the same as the initial learning rate The learning rate is a fixed learning rate; the second-level neural network model adopts standard mean square error as a loss function: , Wherein the method comprises the steps of The loss value for the second level neural network model, To train the number of sets of channel gain samples used, A target set value for an optimal power allocation coefficient; output for second level neural network model A power allocation coefficient predictor for the group channel gain samples; Training of cascade network: the original of the first-level neural network model is trained Optimal power distribution coefficients in data and tag data for training a second level neural network model Respectively inputting the input information and the output information of the cascade network into the cascade network for training until the loss function of the cascade network converges and the training is finished; the cascade network adopts an Adam optimizer to update parameters, wherein the initial learning rate is the same as the initial learning rate The learning rate adopts exponential decay, and the learning rate decay of every 30 training steps is 0.95 times of the original learning rate decay; the cascade network adopts a mixed loss function : , Wherein: mean square error; Is Huber loss; regularizing the term for L2.

Description

Downlink NOMA system power distribution method based on deep learning Technical Field The invention belongs to the technical field of wireless communication, and particularly relates to a power distribution method of a downlink NOMA system based on deep learning. Background In order to meet the stringent requirements of the fifth generation and future mobile communication systems for mass connection, ultra-high capacity and low latency, the Non-orthogonal multiple access (Non-Orthogonal Multiple Access, NOMA) technology has become a research hotspot due to its excellent spectral efficiency. Unlike conventional orthogonal multiple access, NOMA simultaneously serves multiple users on the same time-frequency resource block by performing non-orthogonal superposition of power domain or code domain on multiple users at the transmitting end and decoding by using successive interference cancellation technique at the receiving end. The downlink power domain NOMA is a simple and efficient scheme, and the core of the scheme is how to allocate appropriate power for coexisting users so as to balance fairness among users, total throughput of the system and individual service quality. Despite the great potential of NOMA, its performance advantages are highly dependent on accurate, adaptive power allocation. Existing power allocation schemes face the following significant challenges when dealing with dynamic wireless environments and stringent QoS guarantee requirements: (1) The traditional optimization method has insufficient real-time performance and is difficult to adapt to dynamic environments, and most of the existing researches adopt traditional mathematical tools based on convex optimization, partial planning or Lagrange multiplier method and the like to solve the power distribution problem. Such methods, while theoretically capable of finding (sub) optimal solutions, typically involve complex iterative computations and a large number of floating point operations. In the actual scene of rapid change of channel state information (such as high-speed movement and complex fading), the real-time calculation complexity of the algorithms is extremely high, and the solution and the distribution are difficult to complete within the channel coherence time. This results in a lag in system response, and cannot meet the requirements of scenes such as ultra-reliable low-delay communication on millisecond-level or microsecond-level decisions. (2) The effective capacity optimization problem solving for the statistical QoS guarantee is difficult, and with the appearance of novel services such as URLLC, the traditional indexes such as sum rate maximization and the like are insufficient for representing the time delay performance of the system. The effective capacity theory is used as a tool for describing the maximum sustainable arrival rate under the given time delay constraint, and becomes a key index for evaluating the statistical QoS guarantee performance. However, the NOMA power allocation problem, which aims to maximize the total effective capacity of the system, is more pronounced with non-convex, nonlinear characteristics of the objective function, and there is typically no closed-form solution. The existing method is severely dependent on genetic algorithm, particle swarm optimization and other meta-heuristic algorithms for searching, has low convergence speed and huge calculation cost, is only suitable for offline simulation research, and cannot be deployed in an online system requiring quick response. (3) The data-driven intelligent method faces the bottleneck of high-quality training data, and in recent years, a machine learning and deep learning method is introduced to solve the problem of poor real-time performance of traditional optimization. The method learns the mapping relation from the channel state to the power distribution coefficient by training a neural network, thereby realizing millisecond-level rapid decision in the reasoning stage. However, the performance of such methods is severely limited by the quality of the training data. The method is faced with the 'label dilemma' in that if a traditional suboptimal algorithm (such as a fixed proportion-based method) is adopted to generate the training label, the upper limit of the performance of the trained model is locked and cannot exceed the 'teacher' algorithm, and breakthrough gain is difficult to obtain, and if a meta heuristic algorithm which can approach to a global optimal solution is adopted to generate the training label, although high-quality data can be obtained, the data generation process (such as multiple iterative evaluation of PSO) takes very long time, has high calculation cost, and basically does not solve the fundamental contradiction of 'slow online calculation'. This results in the construction of large-scale, high quality training data sets, which are themselves an efficiency bottleneck. Disclosure of Invention In order to overcome the probl