Joint detection method of satellite-ground downlink NOMA system

Abstract

The invention discloses a joint detection method for a satellite-ground downlink NOMA (Non-Orthogonal Multiple Access) system, which is characterized by comprising the following steps of: S1, inserting pilot frequencies in a time-frequency grid through pilot frequency signals which are sparsely and uniformly arranged in a sub-channel of a transmitting end; s2, extracting a pilot signal at a receiving end, and performing channel estimation based on a deep residual convolutional network; s3, performing multi-user signal detection based on an iterative retraining method of a deep neural network; s4, optimizing power distribution within a power distribution factor threshold range based on a sparrow search algorithm, and minimizing a system bit error rate; s5, combining the step S2, the step S3 and the step S4 to form a joint detection model based on deep learning; through combination of deep learning and a sparrow search algorithm, the problems of channel estimation precision, signal detection capability, power distribution optimization and the like in the system are solved, the overall performance and reliability of the system are improved, and the method is suitable for large-scale user access and efficient data transmission requirements in a modern satellite communication system.

Description

Technical Field

[0001] The present invention relates to the technical field of joint detection of satellite-ground NOMA systems, and in particular to a joint detection method of a satellite-ground downlink NOMA system. Background Art

[0002] With the rapid development of global communication technology, especially the continuous advancement of satellite communication technology, satellite networks have become an important part of global information transmission. Traditional satellite communication networks mainly rely on ground stations for signal relay, but with the continuous growth of global communication needs, especially the sharp increase in the number of users and data transmission needs, traditional satellite communication methods have gradually exposed problems such as tight spectrum resources, large signal interference and low frequency utilization efficiency. Therefore, how to improve the spectrum utilization efficiency of satellite communications has become a key technical issue in the current satellite communication field. In this context, non-orthogonal multiple access (NOMA) technology, as an emerging wireless communication technology, realizes multi-user multiplexing and improves spectrum utilization efficiency by carrying the signals of multiple users in the same time and frequency resource block. NOMA technology can significantly improve frequency band utilization and provide support for large-scale user access, especially for scenarios with tight spectrum resources. Introducing NOMA technology into satellite communication networks can effectively alleviate the problem of tight spectrum resources and improve the efficiency of satellite communications.

[0003] However, there are some challenges in the practical application of satellite communication systems, especially in the environment of satellite-ground communication links. Due to the complexity and time-varying channels and strong interference between multi-user signals, accurate detection and processing of signals has become an urgent problem to be solved. Existing signal processing methods mostly rely on traditional linear detection technology and channel estimation methods, which are difficult to effectively deal with complex nonlinear channels and multi-user interference. Therefore, how to combine advanced deep learning technology to improve the channel estimation accuracy and multi-user signal detection effect in the satellite-ground downlink NOMA system has become an important research topic in the current satellite communication field. To solve the above problems, we propose a joint detection method for the satellite-ground downlink NOMA system. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a joint detection method for a satellite-ground downlink NOMA system to solve the problems raised in the background technology.

[0005] The above technical objectives of the present invention are achieved through the following technical solutions:

[0006] The joint detection method of the satellite-ground downlink NOMA system comprises the following steps:

[0007] S1. In the subchannel of the transmitting end, the pilot signal is inserted into the time-frequency grid through a sparsely evenly arranged pilot signal;

[0008] S2. Extract the pilot signal at the receiving end, calculate the initial channel frequency response of the pilot position using the least squares algorithm, and perform channel estimation through a deep residual convolutional network to restore the channel frequency response;

[0009] S3. Perform multi-user signal detection based on iterative retraining method of deep neural network, separate user signals and reduce interference between users;

[0010] S4. Optimize power allocation within the power allocation factor threshold range based on the sparrow search algorithm to minimize the system bit error rate;

[0011] S5. Combine steps S2, S3, and S4 to form a deep learning-based joint detection model (SSA-DeepJD) to reliably detect multi-user signals in the satellite-ground downlink NOMA system.

[0012] Preferably, the deep residual convolutional network includes the following structure:

[0013] The first layer is a convolutional layer with 8 convolutional filters and a convolution kernel of 3×3×2, which maps its input to N pf ×N pt ×8 output, used to map the input channel frequency response to a specific feature space;

[0014] The middle four layers are residual convolution blocks, which map their input into N pf ×N pt ×8 output, each block consists of two convolutional layers and a ReLU activation layer, which are connected head to tail to form a residual structure;

[0015] Half-pixel bilinear interpolation layer is used to improve the channel feature resolution and convert the input size to N pf ×N pt ×8 data is mapped into an output size of N f ×N t ×8 data;

[0016] The last one is a convolution layer with 8 convolution filters and a convolution kernel of 3×3×8, which gives N f ×N t ×2 channel estimation data.

[0017] Furthermore, the transmitted superimposed signal is recovered based on the channel frequency response obtained by channel estimation.

[0018] Preferably, the deep neural network for multi-user signal detection comprises the following steps:

[0019] The real and imaginary parts of the complex signal are detected separately and used as input data together with the initial probability of each user signal;

[0020] The time series features of the signal are extracted through the long short-term memory network (LSTM), and the feature mapping is performed in combination with the fully connected layer;

[0021] Generate the soft symbol probability distribution of user signals through the classification layer, and use the cross entropy loss function to optimize the network parameters;

[0022] In iterative retraining, the soft symbol probability weight of each NOMA user signal is adjusted through multiple retraining to improve the signal detection accuracy.

[0023] Preferably, the optimization of the power allocation factor adopts a sparrow search algorithm (SSA), which specifically includes:

[0024] Initialize the sparrow population, each individual represents the power allocation factor;

[0025] The power distribution direction is updated through the discoverer mode, and the discoverer position is updated according to the following formula:

[0026]

[0027] Where t represents the current iteration;

[0028] In the follower mode, the power allocation factor is adjusted by the finder with the highest global fitness;

[0029] In the early warning mode, when the power allocation factor exceeds the safety threshold, the step size control parameter is adjusted to ensure the stability of power allocation.

[0030] Furthermore, It is represented as the nth power allocation factor in the pth group at the tth iteration, P is the maximum number of groups, γ∈[0,1] is a random number, I max represents the maximum number of iterations, Q represents a random number that follows a normal distribution, L represents a 1×N dimensional vector with each element being 1, and R 2 ∈[0, 1] represents the alarm value, S∈[0.5, 1] ​​represents the safety threshold, and the searcher will 2 When it is lower than S, it will continue to search in a larger range. The entire population will be at R 2 When it is higher than S, it quickly moves to a safe area to update its position.

[0031] Preferably, the channel estimation training process is performed by minimizing the L2 loss function, with the goal of minimizing the channel estimation error: The sequence training process of the multi-user detection is performed by minimizing the cross entropy loss function of the 1-of-K coding scheme, with the goal of minimizing the probability decision error:

[0032] Preferably, users in the NOMA cluster are grouped according to channel gain differences, and user signals of each cluster are distinguished by different power allocation factors.

[0033] Preferably, the optimization goal of the power allocation factor is to minimize the bit error rate under the premise of ensuring that the points of user signals in the constellation diagram do not overlap.

[0034] Preferably, the joint detection model (SSA-DeepJD) performs end-to-end training for channel estimation, multi-user signal detection and power allocation optimization, and dynamically adapts to the complexity in different satellite channel environments.

[0035] In summary, the present invention mainly has the following beneficial effects:

[0036] 1. The present invention adopts a deep residual convolutional network (DeepReCE) to effectively solve the accuracy problem of existing channel estimation methods in complex time-varying channel environments, overcome the limitations of traditional methods in dealing with nonlinear channel characteristics, and thus improve the channel frequency response recovery accuracy of the satellite-ground downlink NOMA system. The present invention introduces an iterative deep learning-based multi-user detection model (IDLMuD) combined with a long short-term memory network (LSTM) to extract time series features, which can significantly improve the detection accuracy of multi-user signals, reduce interference between users, and improve system performance.

[0037] 2. The present invention adopts the Sparrow Search Algorithm (SSA) to optimize the power allocation factor, which solves the problems of unreasonable power allocation and high bit error rate in traditional methods, and realizes dynamic optimization of power, thereby minimizing the bit error rate and improving system stability. The present invention combines NOMA technology to carry multiple user signals in the same time and frequency resource block, greatly improving the utilization efficiency of the spectrum, supporting more users to access at the same time, thereby improving the system capacity and data transmission rate.

[0038] 3. The present invention enhances the anti-interference capability of the system through joint channel estimation, multi-user signal detection and power allocation optimization, ensures stability and efficiency in complex environments, can adapt to various communication scenarios and user needs, significantly improves the performance of the satellite-ground downlink NOMA system, and provides an efficient, stable and highly anti-interference capable solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is a flow chart of the combined detection of the present invention;

[0040] Figure 2This is a NOMA receiving system diagram based on the SSA-DeepJD network of the present invention. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings of the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the scope of protection of the present invention.

[0042] The following examples are used to illustrate the present invention, but they cannot be used to limit the scope of protection of the present invention. The conditions in the examples can be further adjusted according to specific conditions. Simple improvements to the method of the present invention under the premise of the concept of the present invention belong to the scope of protection claimed in the present invention.

[0043] Example 1

[0044] refer to Figure 1-Figure 2 , a joint detection method for a satellite-ground downlink NOMA system, comprising the following steps:

[0045] S1. In the subchannel of the transmitting end, the pilot signal is inserted into the time-frequency grid through a sparsely evenly arranged pilot signal;

[0046] S2. Extract the pilot signal at the receiving end, calculate the initial channel frequency response of the pilot position using the least squares algorithm, and perform channel estimation through a deep residual convolutional network to restore the channel frequency response;

[0047] S3. Perform multi-user signal detection based on iterative retraining method of deep neural network, separate user signals and reduce interference between users;

[0048] S4. Optimize power allocation within the power allocation factor threshold range based on the sparrow search algorithm to minimize the system bit error rate;

[0049] S5. Combine steps S2, S3, and S4 to form a deep learning-based joint detection model (SSA-DeepJD) to reliably detect multi-user signals in the satellite-ground downlink NOMA system.

[0050] In the LEO satellite-ground downlink communication network, each satellite communicates with ground users within its single beam coverage area. In the ground receiving network, receiving users are clustered according to the NOMA clustering principle (N users with large channel gain differences are selected in each NOMA user cluster). Each cluster of users occupies a sub-channel, and users in the cluster are distinguished by different powers.

[0051] The cluster set (ie subchannel) occupied by all users covered by a single beam of a LEO satellite is denoted as k = {1, ..., K}, and the user set of the kth subchannel is denoted as U k,n = {U k,1 ,…,U k,N}, where each NOMA user belongs to only one subchannel. On the kth subchannel, each frame data of each user (whose size is N f ×N t ), insert the same grid pilot along the time axis and frequency axis at a given period, and insert N pilots uniformly in the time axis direction. pt The blank pilot position is in the frequency axis direction, at the selected N pt Evenly insert N pf pilot symbols, and other pilot positions are set to 0. On the kth subchannel, the frequency domain superposition transmission signal is expressed as Among them, x k，n To send to U k，n M-QAM modulation symbol; p k is the total transmit power of the kth subchannel; α k，n For U k，n The power allocation factor satisfies

[0052] in U k，n At the receiving end, the received signal in the frequency domain is y k，n =H k，n s k +ω k，n , where H k，n ,ω k，n For LEO satellites to U k，n The frequency response and noise variance of the channel are calculated by extracting the pilot information from each frame of data at the receiving end, and using the LS algorithm to obtain the channel frequency response of the pilot position. In order to extract complex-valued channel features, the channel frequency response is divided into real and imaginary parts to form N pf ×N pt ×2 data input is based on the residual convolutional neural network image super-resolution restoration network (DeepReCE network, whose network system is as follows Figure 2 As shown in Figure 2, channel estimation is performed to restore the true channel frequency response: Among them, F DeepReCE Indicated by θ R The residual convolutional neural network model with parameters, Indicates that the DeepReCE network model is in θ R Channel estimation results under parameters.

[0053] Example 2

[0054] The DeepReCE network model has 8 layers: the first layer is a convolutional layer with 8 convolutional filters and a convolution kernel of 3×3×2, which maps its input to N pf ×N pt ×8 output, followed by four residual convolution blocks to map its input to N pf ×N pt ×8 output, each residual convolution block with residual connection at the beginning and end contains two convolutional layers with 8 convolutional filters and a convolution kernel of 3×3×8. The middle of the two convolutional layers is a ReLU activation layer, because the ReLU activation function can speed up the convergence of the model and overcome the problem of gradient disappearance. In addition, the overall residual convolution structure adds a residual connection so that the input can be directly passed to the last layer of the four residual blocks. The use of residual connections can effectively avoid the problem of gradient disappearance caused by multiple nonlinear operations. Subsequently, half-pixel bilinear interpolation upsampling is performed, and the half-pixel geometric transformation mode is adopted to convert the input size of N pf ×N pt ×8 data is mapped into an output size of N f ×N t ×8 data, and finally a convolution layer with 8 convolution filters and a convolution kernel of 3×3×8, to get N f ×N t ×2 channel estimation data, and finally, the transmitted superimposed signal is restored based on the channel frequency response obtained by channel estimation.

[0055] Next, NOMA multi-user signal detection requires separate detection of the real and imaginary parts of the complex signal. The multi-user detection model based on data-driven iterative deep learning (IDLMuD, whose network system is as follows Figure 2 The introduction of the real part of the signal is shown in Figure 1. First, the training data samples are preprocessed: Among them, Re(·) and Im(·) are the real and imaginary parts of the complex signal. is the sign of the detection decision, p k，n for The initialization probability is set, and the training data set is in, for The IDLMuD network model is divided into seven layers, including a sequence input layer, an LSTM layer, a fully connected layer, a ReLU activation layer, a fully connected layer, a Sigmoid layer, and a classification layer. LSTM combines short-term memory and long-term memory through clever gating, learns the information between the time steps of the sequence data and saves the relevant gradient information, thereby solving the problem of gradient disappearance to a certain extent. The time step is equivalently regarded as a sub-channel in the NOMA-OFDM system. By training the time step module in the LSTM layer, multi-user detection of any sub-channel is realized. Time series feature information is added and removed from the unit state in each time step, and historical valuable information (h t-1 , c t-1 ), which is updated through the operations of input gate, forget gate, candidate unit and output gate: Where q = {1, ..., Q} is the number of iterations, h t is the current hidden state, c t is the current unit state, t = {1, ..., T} is the time step, the LSTM layer contains J hidden units, the state of the historical hidden unit is used as the input of the current hidden unit, and the fully connected layer includes J / 2 neurons, all of which are connected to all neurons in the previous layer, which brings together all the attributes and information collected from the previous layer: fc1 =w fc1 h t +b fc1 ,in, is the weight matrix, is the bias vector, z fc1 After the signal is activated by the ReLU function σ(·), it enters the fully connected layer including 2 neurons to obtain the signal z fc2 =w fc2 σ(z fc1 )b fc2 ,in, Next is the Sigmoid layer, which calculates the probability of the second category as the output through the classification features: The last layer is the classification layer, which infers the probability distribution of soft symbols sent by NOMA users based on the input The proposed algorithm also needs to convert the soft symbol probability Grant Then input into this network for iterative retraining, adjusting the probability weights of different target categories, and obtaining more accurate conditional probability distribution in complex nonlinear settings without the need for prior knowledge of user power allocation. Therefore, when these network modules are retrained to properly implement their classification tasks, the receiver essentially implements iterative multi-user detection in a data-driven manner, based on the final soft symbol probability Maximize the conditional probability to determine U k，n User sends signal Among them, M k，n To send to U k，n The modulation order of the signal.

[0056] Example 3

[0057] In order for the DeepJD receiver to reliably implement signal joint detection, the constructed deep learning network needs to be properly trained. In the DeepReCE network for channel estimation, the network is trained end-to-end through supervised learning of the L2 loss function, with the goal of minimizing the channel estimation error: Then, the restored superimposed signal As the input of the network, pre-train the IDLMuD network model for each user and initialize the probability p k，n After all are set to 1 / 2, the initialized z k Input to the IDLMuD network model for prediction to obtain the pre-trained soft probability distribution. After pre-training, it enters the retraining stage. The soft probability distribution of each retraining is used as the soft input for the next retraining. During the retraining process, as the iteration index grows, this soft probability is iteratively improved. Since each network block is trained separately, only a small number of parameters need to be adapted using sequence training. This shows that the IDLMuD network implements iterative multi-user detection by learning to use a relatively small number of training samples, thereby making its performance better than that of deep receivers with larger training sets. The IDLMuD network model is sequence trained by minimizing the cross entropy loss, where the cost function adopts the cross entropy function of the 1-of-K encoding scheme, that is, each input is assigned to one of K mutually exclusive classes: Among them, the NOMA-OFDM system transmits N s =N f ×N t modulation symbols, represents the IDLMuD model parameters of the nth user at the qth iteration.

[0058] Example 4

[0059] In order to ensure the accuracy of signal detection, it is required that the constellation points between users in each subchannel do not overlap, and a threshold constraint is required for the power allocation factor. Considering the symmetry of the constellation diagram, only the first quadrant of the superimposed signal is analyzed. k，n≠N Any constellation point of user U k，N The first quadrant constellation point superposition requires that all superimposed symbols are mapped into the first quadrant and must meet the following requirements: in,

[0060]

[0061] Since power allocation has a great influence on the system error performance, the optimal power allocation is sought based on the Sparrow Search Algorithm (SSA). The algorithm includes explorer-follower-warner, where each sparrow is a power allocation factor within the power allocation threshold range. The discoverer is responsible for finding new power allocation factors and providing the power allocation range and direction. The follower follows the discoverer with a high adaptability value to obtain the power allocation factor. Under certain conditions, the discoverer and the follower can flexibly use interactive switching, but the ratio between the two remains unchanged. The warner perceives the danger and sends a warning signal in time.

[0062] Example 5

[0063] The finder is responsible for finding new power allocation factors. In each iteration of the optimization search process, the finder is updated as: Where t represents the current iteration, It is represented as the nth power allocation factor in the pth group at the tth iteration, P is the maximum number of groups, γ∈[0,1] is a random number, I max represents the maximum number of iterations, Q represents a random number that follows a normal distribution, L represents a 1×N dimensional vector with each element being 1, and R 2 ∈[0, 1] represents the alarm value, S∈[0.5, 1] ​​represents the safety threshold, and the searcher will 2 When it is lower than S, it will continue to search in a larger range. The entire population will be at R 2 When it is higher than S, it quickly moves to the safe area to update the position. The position of the follower is updated as follows: in, The best explorer at present. represents the current global worst position, A represents a 1×N dimensional vector, in which each element is randomly assigned 1 or -1 and satisfies A + =A T (AA T ) -1 , the early warning is responsible for monitoring the area of ​​the search power allocation factor, and its update formula can be expressed as follows: in, represents the current global optimal power allocation factor, β represents the step size control parameter, which follows a normally distributed random number with a mean of 0 and a variance of 1, η∈[-1,1] represents a random number that controls the direction of movement of the power allocation factor, λ represents a very small value that is not 0 to avoid the denominator from being 0, f t , f worst , f opt They represent the individual fitness value of the current power allocation factor, the current global optimal fitness value, and the current global worst fitness value. After iterative sparrow search heuristic learning, the final Assign to α k，n , accurate and reliable joint signal detection of the entire system can be achieved.

[0064] Working principle: This invention solves a series of key problems in the satellite-ground downlink NOMA system, such as channel estimation accuracy, signal detection capability and power allocation optimization, by combining deep learning with the sparrow search algorithm, thus improving the overall performance and reliability of the system. It is suitable for large-scale user access and efficient data transmission requirements in modern satellite communication systems.

[0065] Although the embodiments of the present invention have been shown and described, it is understood by those skilled in the art that, unless otherwise defined, the technical terms or scientific terms used in the present invention should have the usual meanings understood by persons with ordinary skills in the field to which the present invention belongs, and the words "include" or "comprise" and the like used in the present invention mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents.

[0066] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A joint detection method for a satellite-ground downlink NOMA system, characterized in that: The following steps are involved: S1. In the subchannel of the transmitting end, the pilot signal is inserted into the time-frequency grid through a sparsely evenly arranged pilot signal; S2. Extract the pilot signal at the receiving end, calculate the initial channel frequency response of the pilot position using the least squares algorithm, and perform channel estimation through a deep residual convolutional network to restore the channel frequency response; S3. Perform multi-user signal detection based on iterative retraining method of deep neural network, separate user signals and reduce interference between users; S4. Optimize power allocation within the power allocation factor threshold range based on the sparrow search algorithm to minimize the system bit error rate; S5. Combine steps S2, S3, and S4 to form a deep learning-based joint detection model (SSA-DeepJD) to reliably detect multi-user signals in the satellite-ground downlink NOMA system.

2. The joint detection method of the satellite-ground downlink NOMA system according to claim 1, characterized in that: The deep residual convolutional network includes the following structure: The first layer is a convolutional layer with 8 convolutional filters and a convolution kernel of 3×3×2, which maps its input to N pf ×N pt ×8 output, used to map the input channel frequency response to a specific feature space; The middle four layers are residual convolution blocks, which map their input into N pf ×N pt ×8 output, each block consists of two convolutional layers and a ReLU activation layer, which are connected head to tail to form a residual structure; Half-pixel bilinear interpolation layer is used to improve the channel feature resolution and convert the input size to N pf ×N pt ×8 data is mapped into an output size of N f ×N t ×8 data; The last one is a convolution layer with 8 convolution filters and a convolution kernel of 3×3×8, which gives N f ×N t ×2 channel estimation data.

3. The joint detection method of the satellite-ground downlink NOMA system according to claim 1, characterized in that: The deep neural network for multi-user signal detection comprises the following steps: The real and imaginary parts of the complex signal are detected separately and used as input data together with the initial probability of each user signal; The time series features of the signal are extracted through the long short-term memory network (LSTM), and the feature mapping is performed in combination with the fully connected layer; Generate the soft symbol probability distribution of user signals through the classification layer, and use the cross entropy loss function to optimize the network parameters; In iterative retraining, the soft symbol probability of each NOMA user signal is adjusted through multiple retraining to improve the signal detection accuracy.

4. The joint detection method of the satellite-ground downlink NOMA system according to claim 1, characterized in that: The optimization of the power allocation factor adopts the sparrow search algorithm (SSA), which specifically includes: Initialize the sparrow population, each individual represents the power allocation factor; The power distribution direction is updated through the discoverer mode, and the discoverer position is updated according to the following formula: Where t represents the current iteration; In the follower mode, the power allocation factor is adjusted by the finder with the highest global fitness; In the early warning mode, when the power allocation factor exceeds the safety threshold, the step size control parameter is adjusted to ensure the stability of power allocation.

5. The joint detection method of the satellite-ground downlink NOMA system according to claim 1, characterized in that: The training process of the channel estimation is performed by minimizing the L2 loss function, with the goal of minimizing the channel estimation error: The sequence training process of the multi-user detection is performed by minimizing the cross entropy loss function of the 1-of-K coding scheme, with the goal of minimizing the probability decision error:

6. The joint detection method of the satellite-ground downlink NOMA system according to claim 1, characterized in that: The users in the NOMA cluster are grouped according to the channel gain difference, and the user signals of each cluster are distinguished by different power allocation factors.

7. The joint detection method of the satellite-ground downlink NOMA system according to claim 1, characterized in that: The optimization goal of the power allocation factor is to minimize the bit error rate under the premise of ensuring that the points of user signals in the constellation diagram do not overlap.

8. The joint detection method of the satellite-ground downlink NOMA system according to claim 1, characterized in that: The joint detection model (SSA-DeepJD) performs end-to-end training of channel estimation, multi-user signal detection and power allocation optimization, dynamically adapting to the complexity of different satellite channel environments.

Images