SCMA intelligent detection method and device based on full-connection neural network
By using a fully connected neural network for feature splicing and training, the problems of delayed detection and high bit error rate at the SCMA receiver are solved, achieving efficient and accurate signal detection in complex channel environments, and exhibiting good adaptability and robustness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-13
- Publication Date
- 2026-04-03
AI Technical Summary
Existing multi-user detection methods for SCMA receivers suffer from delayed detection and high bit error rate in complex channel environments, making it difficult to meet real-time communication requirements and lacking adaptability to channel changes.
Signal detection is performed using a fully connected neural network. The network is input into the concatenated vector generated by feature splicing and training, and directly learns the mapping relationship from the received signal features to the original encoding of the user at the transmitting end. The network is trained by combining Smooth L1 Loss and AdamW optimizer, and a Dropout mechanism is added to improve generalization ability.
It achieves accurate signal detection with lower computational complexity in complex channel environments, possesses high robustness and good generalization ability, and can maintain good detection performance in unseen channel environments.
Smart Images
Figure CN121793045A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of communication technology, and in particular to an SCMA intelligent detection method and device based on a fully connected neural network. Background Technology
[0002] With the rapid development of integrated space-ground communication networks, the scale of equipment such as low-Earth orbit satellite constellations, UAV swarms, and ground communication nodes continues to expand. The network needs to simultaneously meet the demands of massive concurrent node access and highly reliable data transmission. Non-Orthogonal Multiple Access (NOMA) technology, due to its ability to improve spectrum utilization, has become one of the core technologies suitable for this scenario. Among them, Sparse Code Multiple Access (SCMA) technology, by mapping user bits to sparse multidimensional codewords and reusing finite resource blocks, can achieve "overloaded access," further enhancing system capacity and showing broad application prospects in integrated space-ground scenarios.
[0003] However, traditional multi-user detection at the SCMA receiver relies on the Message Passing Algorithm (MPA), which has significant limitations: it requires multiple iterations to converge, leading to increased detection latency and making it difficult to meet real-time communication requirements; it is also sensitive to rapidly time-varying channels such as Doppler shift and atmospheric fading in satellite-to-ground links, and the bit error rate (BER) increases significantly under low signal-to-noise ratio (SNR), affecting transmission reliability. Existing SCMA intelligent detection schemes mainly fall into two technical paths. One is the end-to-end autoencoder approach, which replaces the traditional MPA by jointly learning the codebook and the detection process. While this achieves single-pass forward inference, it requires redesigning the codebook. The other is the algorithm acceleration approach, which reduces latency by performing network unfolding or knowledge distillation on the MPA iteration process. While this approach can approach MPA detection performance, it lacks adaptability to changing scenarios.
[0004] Therefore, there is an urgent need for an SCMA intelligent detection method that does not require codebook redesign, is non-iterative, and is highly robust, and can achieve more accurate signal detection than traditional methods with lower computational complexity in complex channel environments. Summary of the Invention
[0005] This application provides an SCMA intelligent detection method and device based on a fully connected neural network, which can achieve more accurate signal detection than traditional methods with lower computational complexity in complex channel environments.
[0006] This application provides an SCMA intelligent detection method based on a fully connected neural network. The method includes: acquiring the received signal at the receiver, the complex channel coefficients during signal propagation, the initial prior log-likelihood ratio of each transmitting user, and the channel noise power; the received signal is a signal that is superimposed by the transmitted signals of multiple transmitting users using SCMA technology; the initial prior log-likelihood ratio of each transmitting user is obtained by running a message passing algorithm once. Based on the number of users at the transmitting end and the number of resource blocks, the received signal at the receiving end, the complex channel coefficients during signal propagation, the initial prior log-likelihood ratio of each user at the transmitting end, and the channel noise power are feature-concatenated to obtain a concatenated vector. The concatenated vector is input into a pre-trained fully connected neural network to obtain the original encoding of the transmitted signal of each transmitting user corresponding to the received signal.
[0007] According to the present invention, an SCMA intelligent detection method based on a fully connected neural network is provided. This method involves concatenating features from the received signal at the receiver, complex channel coefficients during signal propagation, the initial prior log-likelihood ratio of each transmitter user, and the channel noise power based on the number of transmitter users and resource blocks to obtain a concatenated vector. The method includes: The received signal at the receiving end is split into real and imaginary parts and normalized. Then, it is expanded with the number of resource blocks as the dimension, and the resulting continuous real-valued features are used as the features of the received signal. The initial prior log-likelihood ratio of each sending end user is expanded according to the dimension of "user × bit" to obtain the prior soft information features; The complex channel coefficients are expanded according to the "user × resource" dimension and then split into real and imaginary parts. The resulting continuous real-valued channel state information vector is used as the channel state feature. The channel noise power is used as a scalar and concatenated with the received signal features, the prior soft information features, and the channel state features, and then normalized to the same scale to obtain a concatenated vector.
[0008] According to the present invention, an SCMA intelligent detection method based on a fully connected neural network is provided. The fully connected neural network is a six-layer fully connected structure, including an input layer, a first hidden layer, a second hidden layer, a third hidden layer, a fourth hidden layer, and an output layer. The input layer consists of 63 neurons, a GELU activation function module, a normalization module, and a random deactivation module. The first hidden layer consists of 32 neurons, a GELU activation function module, and a normalization module; The second hidden layer consists of 64 neurons, a GELU activation function module, and a normalization module; The third hidden layer consists of 128 neurons, a GELU activation function module, and a normalization module; The fourth hidden layer consists of 64 neurons, a GELU activation function module, a normalization module, and a random inactivation module; The output layer consists of 32 neurons and a Tanh activation function module.
[0009] According to the present invention, an SCMA intelligent detection method based on a fully connected neural network is provided, the method further comprising: The fully connected neural network is trained in the following manner: Within the applicable signal-to-noise ratio range for the space-ground integrated scenario, a set of communication datasets is generated per decibel; the communication datasets include data sent and received by multiple sending end users using multiple resource blocks; The communication dataset is divided according to a set ratio and used as the training set, validation set, and test set in the training process of the fully connected neural network.
[0010] According to the present invention, an SCMA intelligent detection method based on a fully connected neural network is provided, the method further comprising: During training, Smooth L1 Loss is used as the loss function, in conjunction with the ReduceLROnPlateau learning rate decay scheduler and the AdamW optimizer for weight updates. We employ an early stopping mechanism and an optimal model saving strategy to determine the parameters at which the fully connected neural network reaches its peak performance.
[0011] According to the SCMA intelligent detection method based on a fully connected neural network provided by the present invention, after obtaining the original encoding of the transmitted signals of each transmitting end user corresponding to the received signal, the method further includes: Determine the confidence level of the original encoding of the transmitted signal of each transmitting end user. If the confidence level is less than a set threshold, then re-execute the SCMA intelligent detection method based on a fully connected neural network.
[0012] This application also provides an SCMA intelligent detection device based on a fully connected neural network, the device comprising: The acquisition module is used to acquire the received signal from the receiver, the complex channel coefficients during signal propagation, the initial prior log-likelihood ratio of each transmitting user, and the channel noise power; the received signal is a signal that is superimposed by the transmitted signals of multiple transmitting users using SCMA technology; the initial prior log-likelihood ratio of each transmitting user is obtained by running a message passing algorithm once. The processing module is used to perform feature concatenation on the received signal of the receiving end, the complex channel coefficients during signal propagation, the initial prior log-likelihood ratio of each sending end user, and the channel noise power based on the number of sending end users and the number of resource blocks to obtain a concatenation vector; The processing module is further configured to input the spliced vector into a pre-trained fully connected neural network to obtain the original encoding of the transmitted signal of each transmitting end user corresponding to the received signal.
[0013] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the SCMA intelligent detection method based on a fully connected neural network as described above.
[0014] This application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the SCMA intelligent detection method based on a fully connected neural network as described above.
[0015] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the SCMA intelligent detection method based on a fully connected neural network as described above.
[0016] This application provides an SCMA intelligent detection method based on a fully connected neural network. The fully connected neural network, trained on a large amount of data, can directly learn the mapping relationship from the signal features at the receiving end to the original codes of each user at the transmitting end. Compared to traditional MPA algorithms, it does not require redesigning the codebook or multiple iterative calculations, and can achieve more accurate signal detection with lower computational complexity in complex channel environments. Furthermore, the fully connected neural network covers various channel conditions, signal-to-noise ratios, and user configurations during training, giving it strong generalization capabilities. It can maintain good detection performance even when facing new and unseen channel environments without requiring algorithm redesign or adjustment. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0018] Figure 1 A flowchart illustrating an SCMA intelligent detection method based on a fully connected neural network, provided in an embodiment of this application; Figure 2A schematic diagram illustrating the feature splicing process provided in an embodiment of this application; Figure 3 This is a schematic diagram of the network structure of a fully connected neural network provided in an embodiment of this application; Figure 4 A schematic diagram of the loss curve for training a Rayleigh channel model; Figure 5 A schematic diagram of training the loss curve for the Rice channel model; Figure 6 This is a schematic diagram of the SNR-BER comparison curves under the Rayleigh channel. Figure 7 This is a schematic diagram of the SNR-BER comparison curves under the Ricean channel; Figure 8 A schematic diagram of the structure of an SCMA intelligent detection device based on a fully connected neural network provided in an embodiment of this application; Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0020] In the embodiments of this application, "multiple" refers to two or more. Terms such as "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or order.
[0021] Figure 1 A flowchart illustrating an SCMA intelligent detection method based on a fully connected neural network, as provided in this application embodiment, is shown below. Figure 1 As shown, the method includes the following steps: Step 101: Obtain the received signal from the receiver, the complex channel coefficients during signal propagation, the initial prior log-likelihood ratio of each transmitting user, and the channel noise power.
[0022] The embodiments of this application are applicable to SCMA technology in the space-ground integrated scenario. The basic idea of SCMA technology is to closely combine the modulation and spread spectrum processes of the signal, and to map the data streams of multiple users into multi-dimensional sparse codewords and then multiplex them on a limited number of resource nodes. In the embodiments of this application, the received signal is the signal that is superimposed by the transmitted signals of multiple transmitting users using SCMA technology.
[0023] The initial prior log-likelihood ratio for each sending user is obtained by running the message passing MPA algorithm once. Specifically, the calculation method for the initial prior log-likelihood ratio for each sending user is as follows: The MPA decoding process is based on a factor graph model. The factor graph structure includes variable nodes (VN) and function nodes (FN), where variable nodes (VN) correspond to the sending user and function nodes (FN) correspond to resource blocks. It is assumed that the codewords are stored... probability distribution Then receive the signal With superimposed user code The likelihood function is 1. Initialization steps Assuming each user sends codewords with equal probability, the prior probability of the initial VN is given.
[0024] in For users j The codebook collection.
[0025] 2. Update external information of variable nodes (VN) VN updates codeword probabilities based on messages from associated FNs and sends external information to all connected FNs. Its input is the likelihood message from the FN, i.e. User nodes update their external information as follows The message indicates that the user j Exclude resource blocks k The confidence level of the codeword probability after the update. All VNs update messages independently, with no data dependency. Indicates user j The set of resource blocks occupied.
[0026] 3. Function Node (FN) External Information Update FN calculates the conditional probability based on the received signal and the VN message, and feeds back the likelihood information to the associated VN. The input is the received signal. Channel parameters And VN messages: Resource nodes update their external information as follows The message indicates that in the resource block k Above, when given other users to type words, the user j The code-like information.
[0027] 4. Output Log-Likelihood Ratio (LLR) By combining information from all associated resource blocks, the posterior probability of the edge is calculated. T After the next iteration, the user j The code x j The posterior probability is in Indicates the first T Resource blocks after the next iteration k Passed to users j The likelihood information is then obtained. Next, the log-likelihood ratio (LLR) of the original encoded bits is calculated based on the posterior probability of the codeword. For codewords... x j The corresponding number l bits Its LLR definition is as follows: Among them, the numerator summation traversal codebook The Middle For all codewords with a bit value of 1, sum the denominators of the first codeword. Codewords with a bit value of 0.
[0028] MPA decoding requires T iterations to obtain the log-likelihood ratio (LLR), while the initial prior log-likelihood ratio of the sending user in this application only requires one execution of the message passing algorithm, i.e., T=1.
[0029] Step 102: Based on the number of users at the transmitting end and the number of resource blocks, the received signal at the receiving end, the complex channel coefficients during signal propagation, the initial prior log-likelihood ratio of each user at the transmitting end, and the channel noise power are feature-concatenated to obtain a concatenated vector.
[0030] In one possible implementation, refer to Figure 2 Step 102 involves concatenating the received signal from the receiver, the complex channel coefficients during signal propagation, the initial prior log-likelihood ratio of each transmitter user, and the channel noise power based on the number of transmitter users and the number of resource blocks to obtain a concatenated vector. This includes the following steps: Step 201: Decompose the received signal at the receiving end into real and imaginary parts and normalize them. Then expand it with the number of resource blocks as the dimension and use the resulting continuous real-valued features as the features of the received signal.
[0031] Step 202: Expand the initial prior log-likelihood ratio of each sending end user according to the dimension of "user × bit" to obtain the prior soft information features.
[0032] Step 203: Expand the complex channel coefficients according to the "user × resource" dimension and then split them into real and imaginary parts. Use the resulting continuous real-valued channel state information vector as the channel state feature.
[0033] Step 204: Take the channel noise power as a scalar, concatenate it with the received signal features, prior soft information features and channel state features, and normalize it on the same scale to obtain the concatenated vector.
[0034] Assume the number of users at the sending end is The number of resource blocks is The LLR dimension per user per symbol is The noise characteristic length is The total input dimension D of the resulting concatenated vector is: For example, suppose there are 6 transmitting users and 4 resource blocks. Each user has a unique codebook, which is a 4×4 complex matrix where rows represent resource blocks and columns represent codewords. Each user transmits 2 bits of information at a time, containing four possible values, each corresponding to one of the four codewords. The received signal at the receiving end is split into real and imaginary parts and normalized to form an 8-dimensional received signal feature. One MPA is performed to obtain the initial prior log-likelihood ratio for each transmitting user. Then, it is expanded according to the "user × bit" dimension to obtain a 6-dimensional prior soft information feature. The complex channel coefficients are expanded according to the "user × resource" dimension and then split into real and imaginary parts to obtain a 48-dimensional channel state feature. Each SNR level is uniformly set as a scalar, and concatenated with the 8-dimensional received signal feature, the 6-dimensional prior soft information feature, and the 48-dimensional channel state feature, and normalized to the same scale to obtain a 63-dimensional concatenated vector.
[0035] This application effectively integrates and normalizes different types of features, such as the received signal at the receiver, the complex channel coefficients during signal propagation, the initial prior log-likelihood ratio of each transmitter user, and the channel noise power. The resulting concatenated vector is more suitable as the input to a fully connected neural network, thereby improving the detection performance and generalization ability of the fully connected neural network.
[0036] It should be noted that this application does not specify the execution order of steps 201 to 203 above.
[0037] Step 103: Input the concatenated vector into a pre-trained fully connected neural network to obtain the original encoding of the transmitted signals of each transmitting user corresponding to the received signal.
[0038] The fully connected neural network provided in this application is designed based on an existing codebook. With a configuration of 6 users J=6, 4 resource blocks K=4, LLR dimension B=1 per user per symbol, and noise feature length E=1, the input dimension D=63, where 62 dimensions represent the standardized signal features, and the additional dimension represents the normalized noise power N0 under the current channel conditions. The hidden layers in the fully connected neural network are configured with 32, 64, 128, 64, and 32 neurons respectively. A BatchNorm normalization module and a GELU activation function are embedded between layers to enhance the model's expressive power and convergence stability. Simultaneously, to reduce overfitting, a random deactivation module Dropout (0.1) is added between layers, randomly masking approximately 10% of neurons during training to break co-adaptation and improve generalization. The output layer uses the Tanh activation function, restricting the network output to the range [-1, 1], consistent with the distribution of the processed target bit labels.
[0039] Figure 3 The diagram below shows the network structure of the fully connected neural network provided in the embodiments of this application; Table 1 is a description of the network structure of the fully connected neural network provided in the embodiments of this application.
[0040] Table 1. Description of the network structure of fully connected neural networks
[0041] like Figure 3 As shown in Table 1, the fully connected neural network has a six-layer fully connected structure, including an input layer, a first hidden layer, a second hidden layer, a third hidden layer, a fourth hidden layer, and an output layer.
[0042] The input layer consists of 63 neurons, a GELU activation function module, a normalization module, and a random deactivation module. The first hidden layer consists of 32 neurons, a GELU activation function module, and a normalization module; The second hidden layer consists of 64 neurons, a GELU activation function module, and a normalization module; The third hidden layer consists of 128 neurons, a GELU activation function module, and a normalization module; The fourth hidden layer consists of 64 neurons, a GELU activation function module, a normalization module, and a random inactivation module; The output layer consists of 32 neurons and a Tanh activation function module.
[0043] Furthermore, the fully connected neural network is trained in the following way: Within the applicable signal-to-noise ratio (SNR) range for integrated satellite-ground scenarios, a communication dataset is generated per decibel. This dataset includes data transmitted and received by multiple transmitting users using multiple resource blocks. The communication dataset is then divided according to a predetermined ratio to serve as the training, validation, and test sets for the fully connected neural network training process. For example, a communication dataset for an integrated satellite-ground scenario is generated within the SNR range of 1 to 15 dB (15 complete data packets are generated at SNRs of 1 dB to 15 dB, each packet containing transmit and receive data from 6 parallel users, with each user having 64,800 original encoded bits). This dataset is then divided into training, validation, and test sets in a ratio of 8:1.5:0.5.
[0044] During training, Smooth L1 Loss is used as the loss function, along with ReduceLROnPlateau learning rate decay scheduler and AdamW optimizer for weight updates. This ensures stable convergence of the fully connected neural network under different sample distributions, improves the convergence speed of the fully connected neural network, and enhances its adaptability and generalization ability under different SNR conditions.
[0045] An early stopping mechanism and an optimal model saving strategy are adopted to determine the parameters at which the performance of the fully connected neural network reaches its peak. This ensures that training is terminated and the parameters are recorded in a timely manner when the performance of the fully connected neural network reaches its peak, thus preventing overfitting.
[0046] This application employs a cross-signal-to-noise ratio joint training strategy to construct the dataset, which enhances the robustness of the fully connected neural network under different channel quality conditions and reduces the retraining frequency during scene switching. The pre-trained fully connected neural network fully considers the characteristics of the SCMA system detection task, and can maintain a compact structure while taking into account nonlinear fitting ability and convergence efficiency.
[0047] In one possible implementation, after obtaining the original encoding of the transmitted signal of each transmitting end user corresponding to the received signal in step 103, the method further includes: determining the confidence level of the original encoding of the transmitted signal of each transmitting end user; if the confidence level is less than a set threshold, then re-executing the SCMA intelligent detection method based on a fully connected neural network until the confidence level requirement is met, so as to ensure the reliability of the detection results.
[0048] To verify the accuracy of the SCMA intelligent detection algorithm based on a fully connected neural network provided in this application, simulation comparison experiments were conducted in a hybrid environment of MATLAB and Python. The dataset was generated by an SCMA uplink transceiver link built in MATLAB, and the training, validation, and testing of the SCMA intelligent detection model based on the fully connected neural network were completed in the Python programming environment. The specific network parameters and model parameter configurations are shown in the table below.
[0049] 1. Verify convergence: Figure 4 Train the loss curve for the Rayleigh channel model; Figure 5 Loss curves were trained for the Ricean channel model. As shown in Figures 4 and 5, in 100 training rounds under both the Rayleigh and Ricean channels, the loss values of both the training and validation sets decreased rapidly and tended to stabilize, with no obvious overfitting. The Dropout mechanism effectively suppressed overfitting, ensuring that the fully connected neural network operates stably in different channel environments in the satellite-ground integrated scenario.
[0050] 2. Verify BER performance: Furthermore, the BER performance of the proposed SCMA intelligent detection algorithm based on a fully connected neural network is examined, and compared with MPA as the benchmark algorithm. The BER simulation results for Rayleigh fading and Ricean fading channel environments are shown below. Figure 6 and 7 As shown in the figure, the SCMA intelligent detection algorithm based on a fully connected neural network consistently outperforms MPA in terms of reliability, with a more pronounced performance advantage under Rayleigh channels. Under Rayleigh channels, the intelligent detection algorithm exhibits approximately a 2dB performance gain compared to MPA. Under Ricean channels, the performance gap between the two algorithms narrows due to the enhanced communication quality provided by the direct path; however, in the low to medium SNR region, the intelligent detection algorithm still maintains a performance gain of around 1dB.
[0051] Overall, the SCMA intelligent detection algorithm based on fully connected neural networks shows a more significant performance improvement over the traditional MPA algorithm at lower SNR. This indicates that the fully connected neural network proposed in this application has stronger robustness under poor channel quality conditions, and can fully utilize sparse structure and prior knowledge to make effective decisions, thereby improving the overall signal detection performance of the system.
[0052] This application provides an SCMA intelligent detection method based on a fully connected neural network. The fully connected neural network, trained on a large amount of data, can directly learn the mapping relationship from the signal features at the receiving end to the original codes of each user at the transmitting end. Compared to traditional MPA algorithms, it does not require redesigning the codebook or multiple iterative calculations, and can achieve more accurate signal detection with lower computational complexity in complex channel environments. Furthermore, the fully connected neural network covers various channel conditions, signal-to-noise ratios, and user configurations during training, giving it strong generalization capabilities. Even when facing new and unseen channel environments, the model can still maintain good detection performance without the need to redesign or adjust the algorithm.
[0053] The SCMA intelligent detection device based on a fully connected neural network provided in this application is described below. The SCMA intelligent detection device based on a fully connected neural network described below can be referred to in correspondence with the SCMA intelligent detection method based on a fully connected neural network described above.
[0054] Figure 8 A schematic diagram of the SCMA intelligent detection device based on a fully connected neural network provided in this application embodiment is shown below. Figure 8 As shown, the device 800 includes: The acquisition module 810 is used to acquire the received signal from the receiver, the complex channel coefficients during signal propagation, the initial prior log-likelihood ratio of each transmitting user, and the channel noise power; the received signal is a signal that is superimposed by the transmitted signals of multiple transmitting users using SCMA technology; the initial prior log-likelihood ratio of each transmitting user is obtained by running a message passing algorithm once. The processing module 820 is used to perform feature concatenation on the received signal of the receiving end, the complex channel coefficients during signal propagation, the initial prior log-likelihood ratio of each transmitting end user, and the channel noise power based on the number of transmitting end users and the number of resource blocks to obtain a concatenation vector; The processing module 820 is further configured to input the spliced vector into a pre-trained fully connected neural network to obtain the original encoding of the transmitted signal of each transmitting end user corresponding to the received signal.
[0055] It should be noted that the SCMA intelligent detection device based on fully connected neural networks provided in this embodiment of the invention can implement all the method steps implemented in the above-mentioned SCMA intelligent detection method embodiment based on fully connected neural networks, and can achieve the same technical effect. Here, the parts that are the same as those in the method embodiment and the beneficial effects will not be described in detail.
[0056] In one possible design, the processing module 820 is further configured to: split the received signal at the receiving end into real and imaginary parts and normalize them, then expand them with the number of resource blocks as the dimension, and use the resulting continuous real-valued features as received signal features; expand the initial prior log-likelihood ratio of each transmitting user according to the dimension of "user × bit" to obtain prior soft information features; expand the complex channel coefficients according to the dimension of "user × resource" and then split them into real and imaginary parts, using the resulting continuous real-valued channel state information vector as channel state features; and use the channel noise power as a scalar, concatenate it with the received signal features, the prior soft information features, and the channel state features, and perform same-scale normalization to obtain a concatenated vector.
[0057] In one possible design, the fully connected neural network is a six-layer fully connected structure, including an input layer, a first hidden layer, a second hidden layer, a third hidden layer, a fourth hidden layer, and an output layer. The input layer consists of 63 neurons, a GELU activation function module, a normalization module, and a random deactivation module. The first hidden layer consists of 32 neurons, a GELU activation function module, and a normalization module; The second hidden layer consists of 64 neurons, a GELU activation function module, and a normalization module; The third hidden layer consists of 128 neurons, a GELU activation function module, and a normalization module; The fourth hidden layer consists of 64 neurons, a GELU activation function module, a normalization module, and a random inactivation module; The output layer consists of 32 neurons and a Tanh activation function module.
[0058] In one possible design, the device further includes a training module 830 for training the fully connected neural network in the following manner: Within the applicable signal-to-noise ratio range for the space-ground integrated scenario, a set of communication datasets is generated per decibel; the communication datasets include data sent and received by multiple sending end users using multiple resource blocks; The communication dataset is divided according to a set ratio and used as the training set, validation set, and test set in the training process of the fully connected neural network.
[0059] In one possible design, the training module 830 is also used to update weights during training by using Smooth L1 Loss as the loss function, in conjunction with the ReduceLROnPlateau learning rate decay scheduler and the AdamW optimizer. We employ an early stopping mechanism and an optimal model saving strategy to determine the parameters at which the fully connected neural network reaches its peak performance.
[0060] In one possible design, the processing module 820 is further configured to determine the confidence level of the original encoding of the transmitted signal of each transmitting end user, and if the confidence level is less than a set threshold, then the SCMA intelligent detection method based on a fully connected neural network is re-executed.
[0061] Figure 9 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 9As shown, the electronic device may include a processor 910, a communications interface 920, a memory 930, and a communication bus 940. The processor 910, communications interface 920, and memory 930 communicate with each other via the communication bus 940. The processor 910 can call logical instructions from the memory 930 to execute the SCMA intelligent detection method based on a fully connected neural network.
[0062] Furthermore, the logical instructions in the aforementioned memory 930 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0063] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the SCMA intelligent detection method based on a fully connected neural network provided by the above methods.
[0064] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the SCMA intelligent detection method based on a fully connected neural network provided by the above methods.
[0065] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0066] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. An SCMA intelligent detection method based on a fully connected neural network, characterized in that, The method includes: The receiver obtains the received signal, the complex channel coefficients during signal propagation, the initial prior log-likelihood ratio of each transmitting user, and the channel noise power; the received signal is a signal that is superimposed by the transmitted signals of multiple transmitting users using SCMA technology; the initial prior log-likelihood ratio of each transmitting user is obtained by running a message passing algorithm once. Based on the number of users at the transmitting end and the number of resource blocks, the received signal at the receiving end, the complex channel coefficients during signal propagation, the initial prior log-likelihood ratio of each user at the transmitting end, and the channel noise power are feature-concatenated to obtain a concatenated vector. The concatenated vector is input into a pre-trained fully connected neural network to obtain the original encoding of the transmitted signal of each transmitting user corresponding to the received signal.
2. The method according to claim 1, characterized in that, Based on the number of transmitting users and the number of resource blocks, the received signal at the receiving end, the complex channel coefficients during signal propagation, the initial prior log-likelihood ratio of each transmitting user, and the channel noise power are feature-concatenated to obtain a concatenated vector, including: The received signal at the receiving end is split into real and imaginary parts and normalized. Then, it is expanded with the number of resource blocks as the dimension, and the resulting continuous real-valued features are used as the features of the received signal. The initial prior log-likelihood ratio of each sending user is expanded according to the dimension of "user × bit" to obtain the prior soft information features; The complex channel coefficients are expanded according to the "user × resource" dimension and then split into real and imaginary parts. The resulting continuous real-valued channel state information vector is used as the channel state feature. The channel noise power is used as a scalar and concatenated with the received signal features, the prior soft information features, and the channel state features, and then normalized to the same scale to obtain a concatenated vector.
3. The method according to claim 1, characterized in that, The fully connected neural network is a six-layer fully connected structure, including an input layer, a first hidden layer, a second hidden layer, a third hidden layer, a fourth hidden layer, and an output layer; The input layer consists of 63 neurons, a GELU activation function module, a normalization module, and a random deactivation module. The first hidden layer consists of 32 neurons, a GELU activation function module, and a normalization module; The second hidden layer consists of 64 neurons, a GELU activation function module, and a normalization module; The third hidden layer consists of 128 neurons, a GELU activation function module, and a normalization module; The fourth hidden layer consists of 64 neurons, a GELU activation function module, a normalization module, and a random inactivation module; The output layer consists of 32 neurons and a Tanh activation function module.
4. The method according to claim 1, characterized in that, The method further includes: The fully connected neural network is trained in the following manner: Within the applicable signal-to-noise ratio range for the space-ground integrated scenario, a set of communication datasets is generated per decibel; the communication datasets include data sent and received by multiple sending end users using multiple resource blocks; The communication dataset is divided according to a set ratio and used as the training set, validation set, and test set in the training process of the fully connected neural network.
5. The method according to claim 4, characterized in that, The method further includes: During training, Smooth L1 Loss is used as the loss function, in conjunction with the ReduceLROnPlateau learning rate decay scheduler and the AdamW optimizer for weight updates. We employ an early stopping mechanism and an optimal model saving strategy to determine the parameters at which the fully connected neural network reaches its peak performance.
6. The method according to claim 1, characterized in that, After obtaining the original encoding of the transmitted signals of each transmitting user corresponding to the received signal, the process also includes: Determine the confidence level of the original encoding of the transmitted signal of each transmitting end user. If the confidence level is less than a set threshold, then re-execute the SCMA intelligent detection method based on a fully connected neural network.
7. An SCMA intelligent detection device based on a fully connected neural network, characterized in that, The device includes: The acquisition module is used to acquire the received signal from the receiver, the complex channel coefficients during signal propagation, the initial prior log-likelihood ratio of each transmitting user, and the channel noise power; the received signal is a signal that is superimposed by the transmitted signals of multiple transmitting users using SCMA technology; the initial prior log-likelihood ratio of each transmitting user is obtained by running a message passing algorithm once. The processing module is used to perform feature concatenation on the received signal of the receiver, the complex channel coefficients during signal propagation, the initial prior log-likelihood ratio of each transmitter user, and the channel noise power based on the number of transmitter users and the number of resource blocks to obtain a concatenation vector. The processing module is further configured to input the spliced vector into a pre-trained fully connected neural network to obtain the original encoding of the transmitted signal of each transmitting end user corresponding to the received signal.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the SCMA intelligent detection method based on a fully connected neural network as described in any one of claims 1 to 6.
9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the SCMA intelligent detection method based on a fully connected neural network as described in any one of claims 1 to 6.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the SCMA intelligent detection method based on a fully connected neural network as described in any one of claims 1 to 6.