An intelligent auxiliary multi-satellite combined activation user detection method
By utilizing a multi-satellite joint deep neural network detection method in low-Earth orbit satellite communication systems, the problem of insufficient accuracy in detecting active users on a single satellite was solved, achieving efficient detection of active user sets and improving user access performance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- THE 54TH RESEARCH INSTITUTE OF CHINA ELECTRONICS TECHNOLOGY GROUP CORPORATION
- Filing Date
- 2024-01-09
- Publication Date
- 2026-07-28
AI Technical Summary
In low-Earth orbit satellite communication systems, the unlicensed access technology for single satellites suffers from insufficient accuracy in detecting active users, resulting in long user access delays and wasted spectrum resources.
An active user detection method based on intelligent assisted multi-satellite collaboration is adopted. This method processes signals by mounting deep neural networks on multiple satellites and then sending the results back to the cloud for merging. The detection accuracy is improved by utilizing the channel differences of multiple satellites, and the user activation probability is estimated by using a fully connected deep neural network and the Softmax function.
It improves the detection accuracy of active user sets, reduces computational complexity and backhaul signaling overhead, and enhances user access performance in unauthorized access scenarios.
Smart Images
Figure CN117880015B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of low-Earth orbit satellite communication, and in particular to an active user detection method based on intelligent assisted multi-satellite collaboration. Background Technology
[0002] In low-Earth orbit (LEO) satellite communication systems, due to the long transmission links and rapid relative movement between the satellite and the ground, authorized access based on multiple handshakes suffers from significant latency and severe signaling overhead. Unlicensed access technology simplifies the access process, reducing the number of interactions between users and the base station. By allowing users to send pilot and data signals in a "send whenever they want" mode, it significantly reduces signaling exchange and lowers user access latency, thus gaining widespread discussion. In unlicensed access systems, users are in a randomly active state. Since the base station lacks prior information about the active user set, it must first detect the active user set after receiving pilot or data signals from users. Considering the large user load of the system, user pilots are typically non-orthogonal to reduce the signaling overhead required for user identification, leading to interference between pilot signals sent by multiple users. Therefore, accurately detecting the active user set at the base station is challenging. For curved-tube satellites, active user detection is completed at the ground station, with the satellite only responsible for signal relay. In future systems, assuming the satellite possesses base station functionality or simple signal processing capabilities, it can process received signals before relaying them to the ground station.
[0003] Currently, compressed sensing algorithms are commonly used in unlicensed access systems for user activity detection. This approach is computationally complex and limited by the pilot set length, making accurate active user detection difficult when the pilot set is short. Furthermore, some researchers have explored using deep learning techniques to assist in active user set detection in unlicensed access systems. These methods construct deep neural networks as detection receivers to detect received signals and estimate the active user set. However, due to strong interference in user transmission pilots, the detection accuracy for the active user set is significantly insufficient. Moreover, considering the limited coverage time of a single satellite for ground users and the significant differences in large-scale fading from users to satellites, the accuracy of activation detection varies for different users. Failure to detect the active user set leads to longer user access delays, placing a significant burden on users and wasting spectrum resources. Summary of the Invention
[0004] In view of this, this invention proposes an active user detection method based on intelligent assisted multi-satellite collaboration to solve the problem of insufficient accuracy of single-satellite detection of active users with unlicensed access. This invention effectively utilizes the channel differences between users and multiple satellites, and employs backhaul data from low-Earth orbit satellites and intelligent merging of multi-satellite data at the ground base station. Compared to single-satellite detection schemes based on compressed sensing and schemes deploying neural networks on a single satellite, this method can improve the accuracy of active user set detection.
[0005] The technical solution adopted in this invention is as follows:
[0006] An active user detection method based on intelligent assisted multi-satellite collaboration includes the following steps:
[0007] Step 1: The user randomly activates and sends an uplink signal. The user uses a pre-configured extension sequence to extend the transmitted symbol. The extension sequence varies from user to user.
[0008] Step 2: K satellites respectively receive the signal from the activated user;
[0009] Step 3: The satellite uses its onboard neural network to process the received signals and outputs a vector consisting of the activation probabilities of each user. The satellite then transmits the output back to the cloud.
[0010] Step 4: Receive the return signals from K satellites in the cloud, merge them, and input them into the cloud-based neural network. The output is an N-dimensional vector. in, The probability of activation for user N;
[0011] Step 5: Set the total loss function J(Θ) to evaluate the difference between the estimated set of active users and the actual set of active users;
[0012] Step 6: Input the training samples into the satellite neural network, and use the gradient descent method to train the satellite neural network and the cloud neural network together to obtain the network parameter set Θ that minimizes the total loss function J(Θ);
[0013] Step 7: After training is completed, the corresponding neural network is mounted on satellites and the cloud to perform intelligent unauthorized access user detection through multi-satellite collaboration.
[0014] Furthermore, the neural network onboard the satellite is a fully connected deep neural network, which uses the sigmoid function or the ReLU function as the activation function to process the input signal layer by layer, and the final output is a vector of length N. The probability of activating for user N.
[0015] Furthermore, the cloud-based neural network contains L fully connected layers and one softmax function; after receiving the return signals from K satellites, the cloud merges them into P = [p1, ..., p2]. K ], as the input signal for the cloud neural network.
[0016] Furthermore, the loss function of the cloud-based neural network is p e Compared with the actual user activation indicator vector B e = [β1, ... β N Cross-entropy J e (Θ), the loss function of the satellite neural network is p k Compared with the actual user activation indicator vector B e = [β1, ..., β] N Cross-entropy J k The total loss function J(Θ) is the loss function of the K satellite neural networks. k (Θ) and the loss function J of the cloud neural network e Weighted sum of (Θ):
[0017]
[0018] Where Θ is the neural network parameter set, η e η represents the weights of the loss function in a cloud-based neural network. k The weights are the weights of the loss function for the satellite neural network.
[0019] The beneficial effects of this invention are as follows:
[0020] 1. This invention utilizes the channel fading differences caused by the location differences between users and multiple satellites to propose a multi-satellite joint active user detection method. Considering that the ground area may be covered by multiple satellites simultaneously, the same neural network structure is deployed on each satellite to process the received signals of that satellite, estimate whether each user is active, and obtain the corresponding activation probability.
[0021] 2. The satellite of this invention will transmit its detection results back to the ground base station. The ground base station will merge the estimation results of multiple satellites through a cloud network to obtain a summary result, which will be used as the final estimation result.
[0022] 3. In the neural network training stage, the present invention adopts a combined training method of multi-satellite neural network and cloud network. The cost function of the neural network is a weighted value of the performance evaluation of the output results of each satellite network and the performance evaluation of the cloud network results. Attached Figure Description
[0023] Figure 1This is a schematic diagram illustrating a scenario of random user activation and access under multi-satellite coverage in an embodiment of the present invention;
[0024] Figure 2 This is a schematic diagram of the structure of the multi-satellite joint user activation detection network in an embodiment of the present invention;
[0025] Figure 3 This diagram illustrates a performance comparison between the single-satellite activation detection algorithm and the method of this invention. Detailed Implementation
[0026] To better illustrate the purpose and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.
[0027] An active user detection method based on intelligent assisted multi-satellite joint detection is proposed. In an unlicensed access system covered by low-Earth orbit satellites, active user detection is performed by equipping multiple satellites with deep neural networks. The detection results are then transmitted back to the ground base station for merging detection by the neural network. This method improves the accuracy of active user set detection under low-Earth orbit satellite coverage, addresses the problem of insufficient accuracy of single satellite in active user detection, and ensures that a single low-Earth orbit satellite can complete the joint detection of multiple satellites with low computational complexity and low backhaul signaling overhead.
[0028] This method specifically includes the following steps:
[0029] Step 1: There are N users in total. Users are randomly activated and send uplink signals. The symbol of user i is s. i Using a specific extended sequence d i =[d i，1 , ..., d i，M ] T The symbol is extended, and the extension sequence is configured individually for each user before the signal transmission process. Each user's extension sequence is different, and the extended signal is q. i =d i s i Different spreading sequences can be used to distinguish different users;
[0030] Step 2: K satellites receive the signal from the activated user. The signal received by the k-th satellite is...
[0031]
[0032] Where, β i =1 indicates that the i-th device is active, β i =0 indicates that the i-th device is inactive, h i，k =[h i，k，1 , ..., h i，k，M ] TLet i be the channel vector between the i-th device and base station k. It is a complex Gaussian noise vector;
[0033] Step 3: At the satellite k end, the onboard neural network will be used to process the received signal Y. k The neural network, consisting of multiple fully connected layers and hidden layers, outputs a vector of length N. The probability of activation for user N is determined and transmitted back to the cloud, which can be a ground base station or the core network.
[0034] Step 4: The cloud receives the return signals from K satellites, merges them as the input signal for the cloud network, and then outputs an N-dimensional vector through the cloud network. The probability of activation for user n;
[0035] Step 5: Set the overall loss function J(Θ) to the intermediate loss function J of the K single-satellite subnetworks passing through the Softmax function. k (Θ) and the loss function J of the detection enhancement network deployed in the cloud. e Weighted sum of (Θ):
[0036]
[0037] Wherein, the loss function is the difference between the estimated set of activated users and the actual set of activated users, and can be a function such as mean squared error or cross-entropy; Θ is the set of neural network parameters, η e The weights η of the loss function for the detection augmentation network deployed in the cloud. k The weights of the intermediate loss function for the output of a single satellite subnetwork;
[0038] Step 6: Use the training set and gradient descent to train the network parameter set Θ that minimizes the loss function;
[0039] Step 7: After training is complete, the neural network is actually deployed on multiple satellites and in the cloud.
[0040] Step 8: When testing and applying the trained neural network, sort the output results in the cloud and select N with the highest probability. a Use elements to obtain an estimate of the target active user set:
[0041]
[0042] This method can jointly utilize the received signal information from multiple satellites and combine it with the ability of deep neural networks to mine the received spread spectrum signals. Data fusion processing is performed in the cloud to improve the accuracy of active user detection and ensure user access performance in unlicensed scenarios.
[0043] The principle of this method is as follows:
[0044] Considering the long link distance and large transmission delay between low-Earth orbit satellites and the ground, unlicensed access can reduce latency and signaling overhead. This method optimizes the user activation detection method under unlicensed access. This method utilizes the received signals from multiple satellites covering the same area for separate detection and then merges them in the cloud. It leverages the channel gain difference caused by the relative distance difference between the user and the multiple satellites to enhance the detection performance of active users under unlicensed access.
[0045] This method employs a deep neural network to design an active user detection scheme. Sub-networks are mounted on each satellite to estimate the status of active users, and the estimation results are sent back to the cloud for merging processing using a cloud-based neural network. This approach can improve the information mining capabilities of deep neural networks in data signals received from multiple satellites, thereby enhancing detection accuracy.
[0046] Here is a more specific example:
[0047] A method for detecting active users based on intelligent assisted multi-satellite collaboration is proposed. This method improves user activation detection performance by using a multi-satellite collaborative intelligent unauthorized access user detection approach. The structure of the intelligent detection network is as follows: Figure 2 As shown, it includes the following steps:
[0048] Step 1, for example Figure 1 The system model shown illustrates multi-satellite coverage, where K satellites can simultaneously receive signals from N ground users. Users are randomly activated and transmit uplink signals. The transmission symbol for user n is s. i Then, a specific extended sequence d is used. i =[d i，1 , ...,d i，M ] T The symbol is extended, and the extended signal is q. i =d i s i ;
[0049] Step 2: K satellites receive the signal from the activated user. The signal received by the k-th satellite is...
[0050]
[0051] Where, β i =1 indicates that the i-th device is active, β i=1 indicates that the i-th device is inactive, h i，k =[h i，k，1 , ..., h i，k，M ] T Let i be the channel vector between the i-th device and base station k. It is a complex Gaussian noise vector;
[0052] Step 3: At satellite k, the onboard neural network, i.e., sub-network k, will be used to process the received signal Y. k In this embodiment, subnetwork k can be a regular fully connected deep neural network. Its activation function can be either the sigmoid function or the ReLU function to process the input signal layer by layer, and the final output is a vector of length N. The probability of activation for user n. Satellite k transmits this information back to the cloud, which can be deployed at ground base stations or the core network;
[0053] Step 4: A cloud network is deployed in the cloud. One possible network structure includes L fully connected layers and one Softmax function. After receiving the return signals from K satellites, the cloud merges them as the input signal for the cloud network. One possible merging method is P = [p1, ..., p2]. K Then, an N-dimensional vector is output via the cloud network. The probability of activation for user n;
[0054] Step 5: Set the loss function of the cloud network to p. e Compared with the actual user activation indicator vector B e = [β1, ..., β] N Cross-entropy J e (Θ). Set the loss function of the sub-network to p. k Compared with the actual user activation indicator vector B e = [β1, ..., β] N Cross-entropy J k (Θ). Set the total loss function J(Θ) to the intermediate loss functions J of the K single-satellite subnetworks. k (Θ) and the loss function J of the detection enhancement network deployed in the cloud. e Weighted sum of (Θ):
[0055]
[0056] Where Θ is the neural network parameter set, η e The weights η of the loss function for the detection augmentation network deployed in the cloud. kThis refers to the weights of the intermediate loss function in the output of the single-satellite subnetwork. The weighting factor can be set empirically; one feasible setting is η. e η k =K:1.
[0057] Step 6: Use the training set and gradient descent to train the network parameter set Θ that minimizes the loss function;
[0058] Step 7: After training is complete, the neural network is actually deployed on multiple satellites and in the cloud.
[0059] Step 8: When testing and applying the trained neural network, sort the output results in the cloud and select N with the highest probability. a Use elements to obtain an estimate of the target active user set:
[0060]
[0061] This invention uses multiple satellites in a low-Earth orbit (LEO) satellite communication network to carry deep neural networks to detect active users, and then transmits the detection results back to the ground for merging detection by the neural network. This can improve the accuracy of active user detection when applying unlicensed access in LEO satellite communication scenarios, and ensure the access performance of users in unlicensed scenarios.
[0062] In summary, compared with traditional active user set detection schemes, this invention adopts a multi-satellite joint detection method, which can realize unauthorized access user set detection through multi-satellite joint detection. In scenarios where multiple satellites cover the same area, this invention can detect randomly activated users by using multiple satellites individually equipped with intelligent receivers, and then transmit the detection results back to the ground base station for merging and unified processing, thereby achieving active user detection.
Claims
1. A method for detecting active users based on intelligent assisted multi-satellite collaboration, characterized in that, Includes the following steps: Step 1: The user randomly activates and sends an uplink signal. The user uses a pre-configured extension sequence to extend the transmitted symbol. The extension sequence varies from user to user. Step 2: K satellites respectively receive the signal from the activated user; Step 3: The satellite uses its onboard neural network to process the received signals and outputs a vector consisting of the activation probabilities of each user. The satellite then transmits the output back to the cloud. Step 4: Receive the return signals from K satellites in the cloud, merge them, and input them into the cloud-based neural network. The output is an N-dimensional vector. in, The probability of activation for user N; Step 5: Set the total loss function J(Θ) to evaluate the difference between the estimated set of active users and the actual set of active users; Step 6: Input the training samples into the satellite neural network, and use the gradient descent method to train the satellite neural network and the cloud neural network together to obtain the network parameter set Θ that minimizes the total loss function J(Θ); Step 7: After training is completed, the corresponding neural network is mounted on satellites and the cloud to perform intelligent unauthorized access user detection through multi-satellite collaboration.
2. The method for detecting active users based on intelligent assisted multi-satellite collaboration according to claim 1, characterized in that, The satellite carries a fully connected deep neural network, which uses either the sigmoid or ReLU function as activation functions to process the input signal layer by layer, ultimately outputting a vector of length N. The probability of activating for user N.
3. The method for detecting active users based on intelligent assisted multi-satellite collaboration according to claim 2, characterized in that, The cloud-based neural network contains L fully connected layers and one softmax function; after receiving back signals from K satellites, the cloud merges them into P = [p1, ..., p2]. K ], as the input signal for the cloud neural network.
4. The method for detecting active users based on intelligent assisted multi-satellite collaboration according to claim 3, characterized in that, The loss function of a cloud-based neural network is p e Compared with the actual user activation indicator vector B e = [β1, ..., β] N Cross-entropy J e (Θ), the loss function of the satellite neural network is p k Compared with the actual user activation indicator vector B e = [β1, ..., β] N Cross-entropy J k The total loss function J(Θ) is the loss function of the K satellite neural networks. k (Θ) and the loss function J of the cloud neural network e Weighted sum of (Θ): Where Θ is the neural network parameter set, η e η represents the weights of the loss function in a cloud-based neural network. k The weights are the weights of the loss function for the satellite neural network.