CN117914431A - Combined equipment activity detection and channel estimation method based on deep learning - Google Patents
Combined equipment activity detection and channel estimation method based on deep learning Download PDFInfo
- Publication number
- CN117914431A CN117914431A CN202410063061.6A CN202410063061A CN117914431A CN 117914431 A CN117914431 A CN 117914431A CN 202410063061 A CN202410063061 A CN 202410063061A CN 117914431 A CN117914431 A CN 117914431A
- Authority
- CN
- China
- Prior art keywords
- matrix
- data
- training
- module
- analog
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 43
- 230000000694 effects Effects 0.000 title claims abstract description 26
- 238000001514 detection method Methods 0.000 title claims abstract description 24
- 238000013135 deep learning Methods 0.000 title claims abstract description 21
- 239000011159 matrix material Substances 0.000 claims abstract description 115
- 238000000926 separation method Methods 0.000 claims abstract description 24
- 238000013528 artificial neural network Methods 0.000 claims abstract description 23
- 238000000605 extraction Methods 0.000 claims abstract description 8
- 238000012549 training Methods 0.000 claims description 47
- 230000006870 function Effects 0.000 claims description 18
- 230000004913 activation Effects 0.000 claims description 16
- 238000009826 distribution Methods 0.000 claims description 10
- 230000000306 recurrent effect Effects 0.000 claims description 10
- 238000000354 decomposition reaction Methods 0.000 claims description 7
- 238000004364 calculation method Methods 0.000 claims description 5
- 238000011156 evaluation Methods 0.000 claims description 3
- 238000004458 analytical method Methods 0.000 claims description 2
- 230000005540 biological transmission Effects 0.000 abstract description 11
- 230000009916 joint effect Effects 0.000 abstract description 5
- 238000003062 neural network model Methods 0.000 abstract description 3
- 238000004891 communication Methods 0.000 description 12
- 238000012545 processing Methods 0.000 description 12
- 230000008569 process Effects 0.000 description 10
- 238000013136 deep learning model Methods 0.000 description 6
- 230000008901 benefit Effects 0.000 description 4
- 238000004422 calculation algorithm Methods 0.000 description 4
- 238000013461 design Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 238000012880 independent component analysis Methods 0.000 description 4
- 238000010606 normalization Methods 0.000 description 4
- 230000008054 signal transmission Effects 0.000 description 4
- 238000013475 authorization Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 230000004044 response Effects 0.000 description 3
- 238000013459 approach Methods 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 238000007405 data analysis Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 239000000654 additive Substances 0.000 description 1
- 230000000996 additive effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 238000005562 fading Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000010295 mobile communication Methods 0.000 description 1
- 230000010363 phase shift Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000011946 reduction process Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000001228 spectrum Methods 0.000 description 1
- 238000010998 test method Methods 0.000 description 1
Landscapes
- Mobile Radio Communication Systems (AREA)
Abstract
The invention relates to a joint equipment activity detection and channel estimation method based on deep learning, which comprises the following steps: the SVD module and the CICA module are combined to form the blind source separation module, the pilot frequency matrix S act can be separated from the pilot frequency matrix Y through the blind source separation module, and then the pilot frequency matrix S act and the pilot frequency matrix Y are input into the neural network module to conduct data feature extraction and iterative computation, so that the prediction accuracy of the neural network model is improved, and more accurate prediction data can be generated. The transmission cost can be reduced on the basis of improving the accuracy of joint activity detection and channel estimation, and the trade-off between pilot frequency cost and accuracy is realized.
Description
Technical Field
The invention belongs to the technical field of wireless communication, and particularly relates to the technical field of wireless communication based on deep learning.
Background
The 'everything interconnection' has become an important development direction of the communication industry, and most terminal devices in the future are connected with a network and gradually aggregated into a global internet of things with huge scale and high intelligence. In a conventional mobile communication system, random access is a necessary procedure for establishing a radio link between a terminal and a user equipment, and normal data transmission between the terminal and the user equipment network is performed only after the random access is completed. The random access method without authorization is used as a new technical means, and the random access of the user can be realized under the condition that the user does not need to obtain authorization in advance, thereby improving the user capacity and the access energy efficiency. The breakthrough of this method is that it improves the traditional licensed random access mechanism, making it possible to increase the number of users.
However, there is still a key problem to be solved in the unlicensed random access method: device activity detection and channel estimation (JADCE), i.e., when multiple users are simultaneously accessing, the access point typically needs to perform user activity detection to identify active users and channel estimation to obtain channel state information for the active users. The received data is corrected and recovered by estimating the state information of the channel, such as time domain or frequency domain response, etc., to improve the accuracy of the received signal, thereby allowing the user to effectively access and communicate.
Researchers have proposed solutions to the JADCE problem using deep learning (DEEP LEARNING abbreviated DL). For example, CN 112910806A discloses a joint channel estimation and user activation detection method based on a deep neural network, and the data feature extraction capability of the neural network module is utilized to detect pilot information and user transmission data signals, so as to implement active user identification and channel estimation scheme design. The pilot, which is a special signal used for channel estimation and equalization in a communication system, is typically inserted into a data stream transmitted from a user equipment so that the terminal can estimate characteristics of a channel of the corresponding user equipment, such as fading, time delay, etc., using these known pilot signals.
Increasing the length of the pilot data, and thus making the pilot provide more information, is a common means of improving the accuracy of joint activity detection and channel estimation. The increase of the pilot data length often needs to occupy more precious spectrum resources and transmission time, which has the negative effect of increasing transmission overhead. Therefore, there is an urgent need for a solution that can reduce the transmission overhead, thereby achieving a tradeoff between pilot overhead and accuracy of joint activity detection and channel estimation.
Disclosure of Invention
The invention aims to provide a joint equipment activity detection and channel estimation method based on deep learning, which can reduce transmission overhead and realize the trade-off between pilot frequency overhead and accuracy on the basis of improving the accuracy of equipment activity detection and channel estimation.
The invention is realized by the following technical scheme:
A joint equipment activity detection and channel estimation method based on deep learning, comprising the steps of:
S1, blind source separation
Transmitting a preamble signal matrix Y received by a terminal into a blind source separation module, wherein the blind source separation module comprises an SVD module and a CICA module, the SVD module carries out decomposition calculation on the preamble signal matrix Y so as to obtain an estimated value of the rank of the preamble signal matrix Y, and the CICA module carries out signal separation on the preamble signal matrix Y based on the estimated value, separates out a pilot matrix S act and outputs the pilot matrix S act to a neural network module; the pilot matrix S act is a pilot data set sent by the active user equipment;
S2, data prediction
The neural network module receives the preamble matrix Y and the pilot matrix S act, and generates and outputs prediction data by data feature extraction and iterationThe prediction data/>Including user equipment active data and channel state data.
The SVD module is a singular value decomposition algorithm module (Singular Value Decomposition) for short, is a linear algebraic decomposition method, and is widely applied to the fields of data analysis, signal processing, image compression and the like. The SVD module is used for decomposing one matrix into products of three matrices and estimating the rank of the matrix by utilizing singular values. The user utilizes the SVD module in the prior art to decompose the preamble signal matrix Y into a noise signal subspace and a pure noise space, so as to estimate the rank of the preamble signal matrix Y. The rank of the preamble matrix Y is the active ue number value.
The CICA module is a complex independent component analysis algorithm (Complex Independent component analysis) for short, and is a method for blind source separation (Blind Source Separation, BSS). In the CICA block, it is assumed that the observed signal is linearly mixed from a plurality of independent source signals, and the CICA block aims to separate the original independent source signals from the mixed signals.
After estimating the rank of the preamble matrix Y by using the SVD, the estimated rank and the preamble matrix Y are input into the CICA module, and the user may separate to obtain the pilot matrix S act based on the CICA module in the prior art.
The deep learning model is a machine learning model based on a multi-layer neural network structure, and modeling and prediction of a complex data mode can be realized through multi-layer feature extraction and representation learning. Such as Convolutional Neural Networks (CNNs), recurrent Neural Networks (RNNs), and the like. In the field of joint equipment activity detection and channel estimation, a deep learning model can be utilized to learn complex activity modes and features from multi-user signals, so that efficient detection of multi-user activity is realized. And more accurate channel estimation can be realized by learning the channel characteristics and the data modes through pilot signals in the data. Particularly in complex multipath propagation environments, the deep learning model can learn nonlinear channel characteristics, so that the accuracy of channel estimation is improved.
Compared with directly inputting the preamble matrix Y into the deep learning module for data prediction, the method firstly introduces the blind source separation module formed by combining the SVD module and the CICA module, can separate the pilot matrix S act from the preamble matrix Y through the blind source separation module, then inputs the preamble matrix Y and the pilot matrix S act into the neural network module for data feature extraction and iterative calculation, improves the prediction accuracy of the neural network model, and can generate more accurate predicted dataThe prediction data/>The method comprises the steps of containing user equipment active data and channel state data, wherein the user equipment active data contains state information of user activity or silence, and can be used for distinguishing active users from silent users. The channel state data comprises time domain or frequency domain response of the channel, and aims to correct and recover the received data through time domain or frequency domain response estimation of the channel so as to improve the accuracy of the received signal.
An innovative blind source separation module is introduced in the scheme, and can effectively separate the pilot frequency matrix, so that pilot frequency signals and active user equipment can be in one-to-one correspondence. By the technology, the separated frequency matrixes which can be in one-to-one correspondence with the active user equipment are input into the deep learning module, so that the accuracy of the deep learning model is remarkably improved. Compared with the traditional method for increasing the pilot data length, the method improves the accuracy of joint activity detection and channel estimation and simultaneously effectively reduces the transmission overhead. Notably, this approach not only improves accuracy, but also achieves a tradeoff of pilot overhead and accuracy. In wireless communication, pilot overhead is a key performance indicator, which directly affects the transmission efficiency and reliability of the system. By combining blind source separation and deep learning, the accuracy and the transmission efficiency of the system are improved on the premise of not increasing pilot frequency overhead. In a word, the method combining blind source separation and deep learning not only improves the accuracy of joint activity detection and channel estimation in wireless communication, but also reduces pilot frequency overhead, and opens up new possibilities for the development of wireless communication technology.
As a preferred embodiment of the present invention, the neural network module is a UGAN network model, and the training mode of the UGAN network model is as follows:
SA, generating training data
Generating an analog leading signal matrix Y 'based on matlab, wherein the analog leading signal matrix Y' contains a known analog user data signal matrix x 'and other analog interference signal data, and generating a training data set after the S1 and blind source separation of the analog leading signal matrix Y', wherein the training data set comprises an analog active signal matrix Yact 'and an analog pilot frequency matrix Sact';
SB, training and generating forecast data
The training data set is sent into the UGAN network model for training, and the prediction data is output
We select the UGAN network model as the neural network module in this case. The UGAN network model is a mature deep learning model in the prior art, has strong characteristic learning and classifying capabilities, and is suitable for the fields of signal processing and communication. In the training process of the UGAN network model, a training mode of a countermeasure network (GAN) is generated, and the generated dummy data is more and more similar to the real data through competition and cooperation between the generator and the discriminator. In our case, during the training process of UGAN network model, we make the predicted data more and more approach to the real data by continuously optimizing the network parameters.
Prior to training the UGAN network model, we need to generate the analog preamble matrix Y' based on Matlab platform. This matrix contains the known matrix x' of analog user data signals and the other interference signal data, such as noise, time delay, etc. Through the analog preamble matrix Y ', the blind source separation can be performed in S1, and the analog pilot matrix Sact ' is extracted from the preamble matrix Y ', so that the active user equipment can realize one-to-one correspondence with the pilot matrix. And sending the analog preamble matrix Y 'and the analog pilot matrix Sact' as the training data set to the UGAN network model for training. Thereby generating the prediction data
Through the training process described above, we successfully trained UGAN network models with high accuracy and stability. The model may be generated by generating the predictive dataThereby, the activity of the user equipment and the corresponding channel quality are estimated more accurately.
As a preferred embodiment of the present invention, the expression of the analog preamble matrix Y' is specifically:
Y’=SX’+Z,
wherein, The method comprises the steps of representing a pilot matrix set sent by N devices, wherein L represents pilot length, N and M represent the number of user devices and the number of antennas respectively, and the pilot matrix set can be customized by a trainer; /(I)Representing a noise matrix, subject to a standard Gaussian normal distribution;
in a communication system, noise is an unavoidable factor, which can interfere with and affect signal transmission. The noise matrix Z represents an lxm matrix of these randomly generated noise components whose elements follow a standard gaussian normal distribution. Where L represents the pilot length and M represents the number of antennas. The elements in this matrix are randomly generated and follow a standard gaussian normal distribution. By analyzing and processing the noise matrix Z, the precision and accuracy of signal processing and data analysis can be improved.
S represents a pilot matrix set sent by N devices, and the pilot length L is a key parameter which can be defined by a user to ensure the stability and accuracy of signal transmission. In practical applications, we can optimize the accuracy and stability of signal transmission by adjusting the value of the pilot length L.
As a preferred embodiment of the present invention, the expression of the analog user data signal matrix x' is:
X’=AH
where a represents the user equipment state matrix, AWGM, N, M represent the number of user equipments and the number of antennas, respectively, which can be customized by a trainer;
In a communication system, the analog user data signal matrix X' =ah describes the data signals transmitted by the user equipment and the state of the channel. The user equipment state matrix a not only describes the activity level of the user equipment, but also reflects the communication state of the equipment. And the channel matrix H is used to evaluate the channel quality, which can reflect the variation and loss of the signal during transmission.
In our case, the channel matrix H is selected as an AWGN channel. AWGN channel, also known as additive white gaussian noise channel, is a common noise model in communication systems. This noise affects the signal transmission, so that when constructing the channel matrix H, we construct an nxm matrix using random numbers sampled independently from gaussian distributions with zero mean and N0/2 variance.
As a preferred aspect of the present invention, the performance evaluation method of the UGAN network model is as follows: the prediction data is processedAnd the analog user data signal matrix x' is subjected to normalized error analysis to obtain an NMSE value, wherein the NMSE value is used for judging the UGAN network model to generate the predicted data/>Accuracy of (3).
NMSE is an abbreviation for "Normalized Mean Squared Error", which is an indicator used to measure the difference between model predictions and true values. It is typically used to evaluate the performance of the generated model, including the UGAN network model. In the UGAN network model, the NMSE serves to evaluate the degree of difference between the samples generated by the generator and the real samples. By calculating the prediction dataThe mean square error between the NMSE value and the analog user data signal matrix x' is normalized to obtain an index for measuring the performance of the generator, namely the NMSE value, which is helpful for evaluating the UGAN network model to generate the predicted data/>Accuracy of (3).
As the preferable mode of the invention, the UGAN network model comprises a generator and a discriminator, and the activation functions of the generator and the discriminator adopt ReLU activation functions.
The ReLU activation function is a commonly used activation function, which is collectively referred to as RECTIFIED LINEAR Unit. It can change the negative value of the input to 0, and keep the positive value unchanged, thereby increasing the nonlinear expression capacity of the network. The use of ReLU activation functions in the generator and the arbiter can enhance the representation of the model to better learn and model the distribution of the data. By employing the ReLU activation function in UGAN network model, we can build a generation countermeasure network with strong representation capability and stability, providing powerful support for various machine learning tasks.
As a preferred aspect of the present invention, the generator is a recurrent neural network, and is cascade-connected by a plurality of basic blocks, and the basic blocks include Unet blocks and projection blocks;
The generator in this case is the Recurrent Neural Network (RNN). The recurrent neural network is a relatively mature neural network in the prior art, and is formed by cascading a plurality of basic blocks, wherein each basic block comprises the Unet blocks and the projection blocks. The design concept of the structure is based on the deep understanding of data distribution and the requirement of fine processing, fine extraction of the features is realized through the Unet block, and the normalization processing is carried out on the features through the projection block, so that the structure is better suitable for various data distributions.
In the Recurrent Neural Network (RNN), the Unet blocks and the projection blocks are typically used to process modeling and generation tasks of sequence data. The Unet block is to splice the characteristic diagram of the encoder part and the characteristic diagram of the decoder part by adopting a jump connection mode, so that more space information is reserved in the decoding process. This design allows the generator to better understand the contextual information of the input data and to generate a richer, more accurate output.
The projection block has the function of carrying out normalization processing on the characteristics, so that the attributes such as the scale and the range of the characteristics are more stable, and the input data of different scales can be better processed. In addition, the projection block also has the function of adjusting the feature dimension so that the output result of the generator can be better matched with the dimension of the target data.
By cascading multiple basic blocks, the recurrent neural network can implement gradual abstraction and fine processing of input data. The network structure has remarkable advantages in processing complex data, can effectively extract key information in the data and generate high-quality output results, namely the predicted data
As a preferred aspect of the present invention, the UGAN network model trains the generator by using a block training strategy, where the block training strategy specifically is: setting an initial learning rate eta 0, a learning rate eta 1 and a learning rate II eta 2, firstly training the ith basic block to be converged by using the initial learning rate eta 0, and then training all the basic blocks to be converged by using the learning rate eta 1 and the learning rate II eta 2; the initial learning rate η 0 > the learning rate one η 1 > the learning rate two η 2.
In the training process of UGAN network model, an innovative block training strategy is adopted to achieve better performance and stability. The core idea of this strategy is to break up the whole generator into a number of basic blocks and to train step by step for each basic block.
First, we set one of the initial learning rates η 0, which η 0 determines the step size of the model parameter update. In the training process, we first train the i-th basic block using the initial learning rate η 0 until the basic block converges. This has the advantage that the model can be better adapted to the data distribution and ensures that the generator has a good starting point in the subsequent training.
When the i-th basic block converges, we begin training the remaining basic blocks using two different learning rates, η 1 and two η 2. These two learning rates are typically lower than the initial learning rate α, which helps the model to more finely adjust the parameters later in the training, further improving the performance of the generator.
The advantage of this block training strategy is that it allows us to optimize each basic block in a targeted way while maintaining the overall performance of the model. In addition, by gradually reducing the learning rate, we can better control the update rate of model parameters, thereby avoiding the problems of over-fitting and under-fitting. Overall, by employing a block training strategy, we can train UGAN the network model more efficiently, improving its generator performance and stability.
In summary, the invention has the following beneficial effects:
1. A blind source separation module is introduced, which can effectively separate out the pilot frequency matrix, and realize the one-to-one correspondence between the pilot frequency matrix and the active user equipment. By the aid of the technology, the separated pilot matrix is input into the deep learning module, and accordingly accuracy of the deep learning model is improved remarkably.
2. By calculating the prediction dataThe mean square error between the NMSE value and the analog user data signal matrix x' is normalized to obtain an index for measuring the performance of the generator, namely the NMSE value, which is helpful for evaluating the UGAN network model to generate the predicted data/>Accuracy of (3).
3. By cascading multiple basic blocks, the recurrent neural network can implement gradual abstraction and fine processing of input data. The network structure has remarkable advantages in processing complex data, can effectively extract key information in the data and generate high-quality output results, namely the predicted data
4. By adopting the block training strategy, the network model UGAN can be trained more effectively, and the performance and stability of the generator can be improved.
Drawings
FIG. 1 is a schematic diagram of a UGAN network model training process of the present invention;
FIG. 2 is a schematic diagram of the Unet module configuration of the present invention;
Fig. 3 is a NMSE comparison of the two methods.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings.
The present embodiment is only for explanation of the present invention and is not to be construed as limiting the present invention, and modifications to the present embodiment, which may not creatively contribute to the present invention as required by those skilled in the art after reading the present specification, are all protected by patent laws within the scope of claims of the present invention.
UGAN network model training steps as shown in figure 1,
S1, generating a training data set
S11, each user generates an information bit stream S with the length of L\2, and the information bit stream S is subjected to coded modulation to obtain a pilot signal S to be transmitted, wherein the modulation mode is binary phase shift keying (BPSK for short). The modulated signal matrix S reaches the terminal through the AWGN channel.
S12, as shown in FIG. 1, the analog preamble matrix received by the terminal is:
Y′=SX′+Z (1)
Wherein the method comprises the steps of Representing the pilot matrix,/>Representing AWGN channel matrix,/>Representing device Activity diagonal matrix/>And the noise matrix is represented, and the standard Gaussian normal distribution is obeyed, and L, N and M respectively represent the pilot frequency length, the user quantity and the antenna quantity.
S13, a sparse reconfigurable analog preamble signal matrix of the device active matrix is as follows:
Y′=[H]:,N[S]N,:+Z=HactSact+Z (2)
where N represents the set of users, S act' represents the analog pilot matrix sent by the active device, and H act represents the channel matrix corresponding to the active device.
Since the number of active user equipments is the rank of the H actSact' matrix. The number of active ues can be calculated by calculating the rank K a of H actSact'. As can be seen from equation 2, the rank of the analog preamble matrix Y ' is the rank K a of H actSact ', so the number of active ues can be obtained by calculating the rank of the analog preamble matrix Y '.
Solving the rank of an analog preamble matrix Y' by using a singular value decomposition (Singular Value Decomposition is called SVD for short): decomposing the received analog preamble matrix Y
Y′=U∑VH (3)
Wherein the method comprises the steps ofKmax=min (N.L) > K a, kmax being the singular value of Y act'.
The received analog preamble matrix Y act' is further divided into a noise signal subspace and a pure noise space at a high signal-to-noise ratio.
Wherein the noise signal subspace is expressed as:
The pure noise space can be expressed as:
further, the number of active devices may be estimated as:
S14, based on The analog preamble signal matrix received by the device is shown in formula (2), and the blind source separation of Y act 'is performed by using a complex independent component analysis algorithm (Complex Independent component analysis is abbreviated as CICA) to obtain an analog pilot frequency matrix S act' corresponding to the active device.
S15, S act' separated based on CICA algorithm, wherein complex data are converted into real numbers before the data are sent to UGAN network, and the specific operation is as follows:
Obtaining a data set N represents the number of user equipments
S2, generating prediction data
UGAN network start training data setThe UGAN network model is formed by cascading a generator network and a discriminator network, wherein the generator is composed of a plurality of basic blocks, and the basic blocks comprise two components: unet blocks and projection blocks, a plurality of Unet blocks and projection blocks form a recurrent neural network (Recurrent Neural Network is called RNN for short), and the activation functions of the generator network and the arbiter network all adopt ReLU activation functions.
The Unet module is shown in fig. 2 and comprises a three-layer network: the first layer network is a convolution layer, and the activation function is a ReLU; the second layer network comprises V convolution blocks with the same structure, and each convolution block comprises a convolution layer (Conv layer), a batch normalization (Batch normalization BN layer for short) layer and a ReLU activation layer; the third network comprises a convolutional layer.
During feedforward of input data, data dimension reduction processing is performed first, and then data dimension increase is performed to extract data features. In the data dimension reduction process, 1D convolution with a kernel size of 3 and a step size of 2 and a filling size of 1 is used for sampling, then 1D convolution with a kernel size of 3 and a step size of 2 and a ReLU activation function are used for extracting features, tensors with the same size are overlapped, and the process is repeated twice; 1D conversion with a kernel size of 2 and a step size of 2 is used in a data dimension increasing stage; features are next extracted using a 1D convolution with a kernel of 3 and a step size of 1 kernel ReLU activation function, and this process is repeated twice. In the network design of the arbiter, we use a 1D convolution of kernel size 3, step size 1 and ReLU activation function to extract features, then use a 1D convolution of kernel size 3, step size 2 to downsample, and finally use a 1D convolution of kernel size 3 to obtain the final output.
Will be during trainingAs input to the UGAN network generator, where (-) + represents the pseudo-inverse of the matrix. The input of basic block is/>The projection block projects the output value to/>The outputs of corresponding Unet blocks areThe loss function of the generator is:
The loss function of the arbiter is:
Training a generator by adopting a block training strategy, wherein the initial learning rate is eta 0; the ith basic block is trained to converge at a learning rate η 0, and then all blocks are trained until convergence at learning rates η 1 and η 2.
Output of Unet blocks after K iterations:
S3, performance evaluation
In step S2, generating predictive data, UGAN network model generates predictive dataThe predicted data/>And true data/>And performing comparison and calculation to obtain NMSE, wherein the expression is as follows:
Wherein the method comprises the steps of Is an estimated value,/>Is a true value.
UGAN network model test procedure
The terminal receives the preamble signal matrix Y, and the user activity state and the corresponding channel quality in the preamble signal matrix Y are unknown. The SVD module and CICA module are combined to form the blind source separation module, the pilot frequency matrix S act is firstly separated from the pilot frequency matrix Y received by the terminal through the blind source separation module, and then the pilot frequency matrix S act and the pilot frequency matrix Y are input into a trained UGAN network model to perform data feature extraction and iterative calculation, so that the prediction accuracy of the neural network model is improved, and more accurate prediction data can be generatedThe prediction data/>The method comprises the steps of containing user equipment active data and channel state data, wherein the user equipment active data contains state information of user activity or silence, and can be used for distinguishing active users from silent users. The channel state data is used for estimating the channel quality corresponding to the user equipment so that the user equipment can be effectively accessed and communicated without obtaining authorization in advance.
In this example, the above implementation uses MATLABR2020a and pycharm2020, pytorch3.8 (python 3.8) to simulate the above procedure to perform a simulation experiment, and the data in the simulation are set as follows: pilot length l=32, 40, 48, 58, 64, 72, 80, number of devices n=64, preset active device number ka=8, number of base station antennas m=8, learning rate η 0=5-10-4,η1=0.2η0,η2=0.02η0, standard deviation σ 2 =0.1. A comparison of loss errors under different schemes as pilot length increases is calculated and plotted as shown in fig. 3.
The foregoing description of the preferred embodiments of the invention is not intended to be limiting, but rather is intended to cover all modifications, equivalents, and alternatives falling within the spirit and principles of the invention.
Claims (8)
1. The joint equipment activity detection and channel estimation method based on deep learning is characterized by comprising the following steps:
S1, blind source separation
The method comprises the steps that a terminal receives a preamble signal matrix Y, the preamble signal matrix Y is transmitted to enter a blind source separation module, the blind source separation module comprises an SVD module and a CICA module, the SVD module carries out decomposition calculation on the preamble signal matrix Y to obtain an estimated value of the rank of the preamble signal matrix Y, and the CICA module carries out signal separation on the preamble signal matrix Y based on the estimated value to separate a pilot frequency matrix S act and outputs the pilot frequency matrix S act to a neural network module; the pilot matrix S act is a pilot data set sent by the active user equipment;
S2, data prediction
The neural network module receives the preamble matrix Y and the pilot matrix S act, and generates and outputs prediction data by data feature extraction and iterationThe prediction data/>Including user equipment active data and channel state data.
2. The method for joint equipment activity detection and channel estimation based on deep learning as claimed in claim 1, wherein the neural network module is UGAN network model, and the training mode of the UGAN network model is as follows:
SA, generating training data
Generating an analog leading signal matrix Y 'based on matlab, wherein the analog leading signal matrix Y' contains a known analog user data signal matrix x 'and other analog interference signal data, and generating a training data set after the S1 and blind source separation of the analog leading signal matrix Y', wherein the training data set contains the analog leading signal matrix Y 'and an analog pilot frequency matrix Sact';
SB, training and generating forecast data
The training data set is sent into the UGAN network model for training, and the prediction data is output
3. The method for joint equipment activity detection and channel estimation based on deep learning as claimed in claim 2, wherein the expression of the analog preamble matrix Y' is specifically:
Y’=SX’+Z,
wherein, S= (S 1,......SN) represents the set of pilot matrices sent by N devices,/>The matrix representing the L×N, L, N, M represents the pilot length, the number of the user equipments and the number of antennas respectively, which can be customized by a trainer; representing a noise matrix, subject to a standard Gaussian normal distribution; /(I) Representing an lxm matrix.
4. A joint equipment activity detection and channel estimation method based on deep learning as claimed in claim 3, wherein the expression of the analog user data signal matrix x' is:
X’=AH,
where a represents the user equipment state matrix, H is the channel matrix of AWGM,/>Representing an nxm matrix.
5. The method of deep learning based joint equipment activity detection and channel estimation of claim 2, wherein the UGAN network model training method further comprises generating predictive data at the SB, trainingSubsequent SC, performance evaluation: couple the prediction data/>And the analog user data signal matrix x' is subjected to normalized error analysis to obtain an NMSE value, wherein the NMSE value is used for judging the UGAN network model to generate the predicted data/>Accuracy of (3).
6. A method for joint equipment activity detection and channel estimation based on deep learning as defined in any one of claims 2-5, wherein said UGAN network model includes a generator and a arbiter, and the activation functions of said generator and said arbiter are all ReLU activation functions.
7. The joint equipment activity detection and channel estimation method based on deep learning of claim 6, wherein the generator is a recurrent neural network, and is cascaded by a plurality of basic blocks, and the basic blocks comprise Unet blocks and projection blocks.
8. The method for joint equipment activity detection and channel estimation based on deep learning of claim 7, wherein the UGAN network model trains the generator by using a block training strategy, the block training strategy specifically being: setting an initial learning rate eta 0, a learning rate eta 1 and a learning rate II eta 2, firstly training the ith basic block to be converged by using the initial learning rate eta 0, and then training all the basic blocks to be converged by using the learning rate eta 1 and the learning rate II eta 2; the initial learning rate η 0 > the learning rate one η 1 > the learning rate two η 2.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202410063061.6A CN117914431A (en) | 2024-01-16 | 2024-01-16 | Combined equipment activity detection and channel estimation method based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202410063061.6A CN117914431A (en) | 2024-01-16 | 2024-01-16 | Combined equipment activity detection and channel estimation method based on deep learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN117914431A true CN117914431A (en) | 2024-04-19 |
Family
ID=90690523
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202410063061.6A Pending CN117914431A (en) | 2024-01-16 | 2024-01-16 | Combined equipment activity detection and channel estimation method based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN117914431A (en) |
-
2024
- 2024-01-16 CN CN202410063061.6A patent/CN117914431A/en active Pending
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US7289972B2 (en) | Cognitive radio engine based on genetic algorithms in a network | |
CN108566257B (en) | Signal recovery method based on back propagation neural network | |
Lin et al. | Learning of time-frequency attention mechanism for automatic modulation recognition | |
CN112115821B (en) | Multi-signal intelligent modulation mode identification method based on wavelet approximate coefficient entropy | |
CN110167176B (en) | Wireless network resource allocation method based on distributed machine learning | |
CN111342867A (en) | MIMO iterative detection method based on deep neural network | |
CN112910811B (en) | Blind modulation identification method and device under unknown noise level condition based on joint learning | |
CN114268388B (en) | Channel estimation method based on improved GAN network in large-scale MIMO | |
CN114039870B (en) | Deep learning-based real-time bandwidth prediction method for video stream application in cellular network | |
WO2021088465A1 (en) | Fast modulation recognition method using multilayer perceptron, and employing data fusion of multiple distribution tests | |
CN116841732A (en) | Federal learning resource optimization design method based on single-bit quantization | |
CN110474798B (en) | Method for predicting future signal of wireless communication by using echo state network | |
CN117914431A (en) | Combined equipment activity detection and channel estimation method based on deep learning | |
CN116170066A (en) | Load prediction method for low-orbit satellite Internet of things | |
WO2006011866A2 (en) | Cognitive radio engine application of genetic algorithms | |
CN111541472B (en) | Low-complexity machine learning assisted robust precoding method and device | |
Wang et al. | Blind Recognition Method for Non-Cooperative Communication Signals Based on Aerial Computing | |
Yang et al. | Deep-learning-based bouncing-order prediction for propagation channel characterization using graph modeling | |
CN114764610A (en) | Channel estimation method based on neural network and communication device | |
CN112364845A (en) | Signal-to-noise ratio evaluation method and device, electronic equipment and storage medium | |
US20150103958A1 (en) | Methods, systems, and media for determining whether a signal of interest is present | |
CN116761223B (en) | Method for realizing 4G radio frequency communication by using 5G baseband chip and vehicle-mounted radio frequency system | |
CN116032332B (en) | Large-scale MIMO system detection model construction method suitable for changeable channel state information | |
CN117667749B (en) | Fuzzy test case optimization method and system | |
WO2024057817A1 (en) | Wireless signal processing system and wireless signal processing method |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |