A Single-Base-Station Backscatter Channel Estimation Method Based on Deep Unfolding

Abstract

The present invention discloses a single-base station backscatter channel estimation method based on deep unfolding, which belongs to the field of channel estimation of wireless communications. The present invention proposes a deep unfolding neural network designed specifically for channel estimation of a single-base station multi-antenna multi-label backscatter communication network. The network expands the iterative algorithm based on gradient descent to solve the LS problem, replaces each round of iterative calculation of the channel parameters with a single-layer neural network with a fixed structure, and uses the activation function in the neural network to replace the nonlinear mapping of the output, thereby constructing a model-driven channel estimation network. At the same time, some learnable parameters are introduced into the network, and the parameters are adjusted based on the back propagation of the error gradient. Then, the required channel estimation parameters are obtained by simulating the iterative calculation process through the connection of multi-layer neural networks. The present invention has a more compromise advantage than the traditional LS method and the black box neural network in terms of channel estimation accuracy and computational cost.

Description

Technical Field

[0001] The present invention belongs to the field of channel estimation of wireless communications, and in particular relates to a single base station backscatter channel estimation method based on depth expansion. Background Art

[0002] As a new communication paradigm in large-scale Internet of Things and wireless sensor networks, backscattering has unique advantages in facilitating energy acquisition and reducing circuit costs. Backscattering communication networks usually consist of three parts: a carrier generator, a backscattering device, and a signal reader. The integration of the carrier generator and the signal reader is called a single-base station backscattering architecture. Most existing backscattering studies are based on the assumption of complete channel state information (CSI). However, unlike traditional wireless communications, the CSI of backscattering communication is difficult to obtain accurately, and its channel estimation faces two major challenges: limited modulation circuit energy and channel parameters that change with tag status. Therefore, how to perform accurate and effective channel estimation has become an important research direction.

[0003] Most of the existing channel estimation methods for backscatter systems are traditional numerical methods. For example, the prior art document [1] proposed a blind channel estimation algorithm based on expectation maximization. The prior art document [2] derived a new least squares estimator based on the forward and backward links between a full-duplex multi-input multi-output reader and a single tag, and obtained the corresponding linear minimum mean square error estimation of the backscatter channel. The prior art document [3] proposed a method based on the least squares (LS) algorithm to obtain the initial estimated parameters of the tag in different states. With the continuous deepening of research work, some deep learning-based methods that have been effective in traditional wireless communications, such as supervised learning, reinforcement learning, neural networks, and transfer learning, have also been applied to backscatter communications. In the prior art document [4], the authors transformed the signal detection problem into a clustering problem, and then restored the unlabeled signal based on the learned parameters. The prior art document [5] proposed a detection method based on machine learning, which transformed the problem of using energy detectors or minimum mean square error detectors to detect tag signals with high bit error rates (BER) into a classification problem. Prior art literature [6] proposed a fast and flexible convolutional neural network (FFDNet) to approximate the optimal solution, while using a deep neural network (DNN) with a customized loss function to estimate the forward channel coefficients directly from the backscattered signal. Prior art literature [7] proposed a CNN-based deep residual network (CDRN) to implicitly learn residual noise to recover the channel coefficients from noisy pilot-based observations.

[0004] The traditional numerical methods described above often have high computational complexity and are generally unsuitable for channel estimation in backscatter systems composed of multi-antenna readers and multiple tags. Deep learning-based channel estimation schemes, on the other hand, avoid the complex solution of the channel matrix and instead simply use neural networks to learn from the dataset to obtain channel parameters. This allows for efficient solutions even for multi-antenna, multi-tag systems. However, these data-driven deep learning networks treat the communication system as a black box and train it using large amounts of data. This lacks theoretical foundations and interpretability in the network design process, and they require a high amount of data, resulting in poor generalization capabilities.

[0005] [1] S.Ma, G.Wang, R.Fan, and C.Tellambura, "Blind Channel Estimation for Ambient Backscatter Communication Systems," IEEE Commun.Lett., vol.22, no.6, pp.1296–1299, Jun.2018, doi:10.1109 / LCOMM.2018.2817555.

[0006] [2] D. Mishra and EG Larsson, "Optimal Channel Estimation for Reciprocity-Based Backscattering With a Full-Duplex MIMO Reader," IEEETrans.Signal Process., vol.67, no.6, pp.1662–1677, Mar.2019, doi:10.1109 / TSP.2019.2893859.

[0007] [3] Y.Zhu, G.Wang, H.Tang, R.He, and Y.Zou, "Channel Estimation for AmbientBackscatter Systems over Frequency-Selective Channels," in 2018 IEEE / CICInternational Conference on Communications in China(ICCC),2018,pp.384–388.doi:10.1109 / ICCChina.2018.8641250.

[0008] [4]Q. Zhang, H. Guo, Y.-C. Liang, and X. Yuan, “Constellation Learning-Based Signal Detection for Ambient Backscatter Communication Systems,” IEEE J. Sel. Areas Commun., vol. 37, no. 2, pp. 452–463, Feb. 2019, doi: 10.1109 / JSAC.2018.2872382.

[0009] [5]Y. Hu, P. Wang, Z. Lin, M. Ding, and Y.-C. Liang, “Machine Learning Based Signal Detection for Ambient Backscatter Communications,” in ICC 2019 - 2019 IEEE International Conference on Communications (ICC), May 2019, pp. 1–6. doi: 10.1109 / ICC.2019.8761796.

[0010] [6]M. Yerzhanova and Y.H. Kim, “Channel Estimation via Model and Learning for Monostatic Multiantenna Backscatter Communication,” IEEE Access, vol. 9, pp. 165341–165350, 2021, doi: 10.1109 / ACCESS.2021.3'134961.

[0011] [7]C. Liu, X. Liu, D.W.Kwan Ng, and J. Yuan, “Deep Residual Network Empowered Channel Estimation for IRS-Assisted Multi-User Communication Systems,” in ICC 2021 - IEEE International Conference on Communications. Jun. 2021, pp. 1 - 7. doi: 10.1109 / ICC42927.2021.9500708. Summary of the Invention

[0012] The purpose of the present invention is to solve the problems existing in the prior art and provide a single base station backscatter channel estimation method based on depth expansion.

[0013] The specific technical solutions adopted in the present invention are as follows:

[0014] A single base station backscatter channel estimation method based on deep expansion includes the following steps:

[0015] S1. In the environmental interference channel estimation stage, all backscatter devices BD in the multi-antenna single base station backscatter communication system are set to silent state. After the reader R sends the pilot S0, the received signal is obtained and the environmental interference channel parameters are estimated.

[0016] S2. In the backscatter channel estimation phase, the reader R sends a pilot signal S0 to each backscatter device BD in turn and obtains the received signal Y CE , and in the process of sending the pilot signal, only the backscattering device BD that transmits the pilot signal S0 is activated, but other backscattering devices BD remain silent; using the estimated environmental interference channel parameters From the received signal Y CE Remove the environmental interference signal and obtain the interference-free signal Y;

[0017] S3. Vectorize the backscatter channel parameter matrix H, pilot signal S0, and interference-removing signal Y of all backscatter devices BD, converting them from matrix form to vector form. Then, convert the environmental interference channel parameters, pilot signal, and interference-removing signal in vector form from the complex domain to the real domain, respectively, to obtain the backscatter channel parameter h′, pilot signal S′, and interference-removing signal y′ in real form.

[0018] S4. The environmental interference channel parameter h′, the pilot signal S′, and the interference-removed signal y′ are used as inputs of the deep unfolding channel estimation network. The deep unfolding channel estimation network obtains the real domain estimate of the backscatter channel through multiple layers of iteration. The forward reasoning calculation process of any k-th layer iteration in the network is as follows:

[0019]

[0020] V k =f(W 2k z k +B 2k )

[0021] h′ k =g(W 3k z k +B 3k )

[0022] Where: ReLU(·) is the RELU activation function; W 1k , W 2k , W 3k , B 1k , B 2k , B 3k , δ k ,λ k are all learnable parameters of the kth layer during the network iteration process, which are optimized in advance through network training; h′ k-1 、V k-1 are the two outputs of the k-1th layer during the network iteration, and for the first layer during the network iteration, h′0 and V0 are both initialized to all-zero vectors; f(·) and g(·) both represent nonlinear activation functions; z k is the intermediate quantity of the calculation process; k = 1, 2, ..., L, where L is the total number of iterative layers of the depth-expanded channel estimation network;

[0023] Finally, the real domain estimation value of the backscatter channel is converted back to the complex domain to obtain the final estimation result of the backscatter channel.

[0024] Preferably, the multi-antenna single-base station backscatter communication system is a communication system consisting of a full-duplex reader R with multiple antennas and M backscatter devices BD, and the reader R has N antennas, each backscatter device BD is configured with a single antenna, and M and N are both integers greater than 1.

[0025] Preferably, in the multi-antenna single-base station backscatter communication system, within a coherence time, except for the sample time used for signal reception and decoding, the remaining sample time is evenly divided into M+1 segments for performing environmental interference channel estimation and backscatter channel estimation.

[0026] As a preference, in said S1, the environmental interference channel parameter It is estimated by the least squares algorithm.

[0027] As a preference, in said S2, the estimated environmental interference channel parameters are used From the received signal Y CE The method to remove environmental interference signals is as follows:

[0028]

[0029] Where: 1 M×1 Represents an M×1 vector of all ones.

[0030] Preferably, in S4, the deep unfolding channel estimation network needs to be pre-trained before being used for actual reasoning, and the loss function of the network training is in the form of:

[0031]

[0032] Where: h′ l h′ is the backscatter channel parameter estimate output by any iterative output of layer l in the network. LS h′ l The corresponding true value, log(l) is the logarithmic weighted weight.

[0033] Preferably, in S4, during the training process of the deep unfolded channel estimation network, an Adam optimizer is used to perform error back propagation and parameter adjustment based on the loss function until the network converges.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] (1) Low computational-space complexity: This paper uses neural network layer learning parameters to avoid the tedious iterative calculation processes such as matrix inversion and derivation in traditional parameter estimation methods, thereby improving computational efficiency. At the same time, it constructs an interpretable network based on the channel model, circumventing the "black box" problem existing in deep learning, greatly reducing the scale of parameters that need to be learned and reducing space complexity.

[0036] (2) Applicable to large-scale antennas and tags: In traditional methods, increasing the number of antennas and tags will undoubtedly lead to huge computational difficulties, thus affecting communication quality. The method proposed in this paper only needs a single efficient training to be applicable to repeated channel estimation of the same scale, and can adaptively adjust the network structure according to the number of antennas and tags, showing good generalization ability.

[0037] (3) Fewer training samples are required: Previous data-driven deep learning networks require a large amount of data to effectively learn network parameters. The model-driven method proposed in this invention only requires a small number of samples to achieve performance comparable to that of traditional black-box networks trained with a large number of samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Schematic diagram of a multi-antenna single base station backscatter communication system;

[0039] Figure 2 Interference and channel estimation protocols corresponding to deep unfolding channel estimation networks;

[0040] Figure 3 Schematic diagram of the network structure of the deep unfolding channel estimation network;

[0041] Figure 4 Schematic diagram of a single-layer network structure for the iterative calculation process of a deep unfolding channel estimation network;

[0042] Figure 5 A comparison of the normalized mean square error averages of the three networks after training under various system settings.

[0043] Figure 6 This is a comparison chart of the computational time costs of the three networks under various system settings. DETAILED DESCRIPTION

[0044] The present invention will be further described and illustrated below with reference to the accompanying drawings and specific embodiments. The technical features of each embodiment of the present invention may be combined accordingly, provided that there is no conflict between them.

[0045] In order to make the above-mentioned objects, features and advantages of the present invention more clearly understood, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. The technical features in the various embodiments of the present invention can be combined accordingly without conflicting with each other.

[0046] In a preferred implementation of the present invention, a single-base station backscatter channel estimation method based on deep expansion is provided. It should be noted that the multi-antenna single-base station backscatter communication system in the present invention is a communication system composed of a full-duplex reader R with multiple antennas and M backscatter devices BD, and the reader R has N antennas, and each backscatter device BD is configured with a single antenna, and M and N are both integers greater than 1. Moreover, in this multi-antenna single-base station backscatter communication system, within a coherent time, in addition to the sample time used for signal reception and decoding, the remaining sample time is evenly divided into M+1 segments for performing environmental interference channel estimation and backscatter channel estimation. The single-base station backscatter channel estimation method based on deep expansion specifically includes the following steps:

[0047] S1. In the environmental interference channel estimation stage, all backscatter devices BD in the multi-antenna single base station backscatter communication system are set to silent state. After the reader R sends the pilot S0, the received signal is obtained and the environmental interference channel parameters are estimated.

[0048] In the above step S1, the environmental interference channel parameters The channel is estimated by the least squares (LS) algorithm. Channel estimation by the LS algorithm belongs to the prior art and will not be described in detail.

[0049] S2. In the backscatter channel estimation phase, the reader R sends a pilot signal S0 to each backscatter device BD in turn and obtains the received signal Y CE , and in the process of sending the pilot signal, only the backscattering device BD that transmits the pilot signal S0 is activated, but other backscattering devices BD remain silent; using the estimated environmental interference channel parameters From the received signal Y CE Remove the environmental interference signal and obtain the interference-free signal Y.

[0050] In the above step S2 of the present invention, the estimated environmental interference channel parameters are used From the received signal Y CE The method to remove environmental interference signals is as follows:

[0051]

[0052] Where: 1 M×1 Represents an M×1 vector of all ones.

[0053] S3. Vectorize the backscatter channel parameter matrix H, pilot S0, and interference removal signal Y of all backscatter devices BD respectively, convert them from matrix form to vector form, and then convert the environmental interference channel parameters, pilot, and interference removal signal in vector form from the complex domain to the real domain, and obtain the backscatter channel parameters h′, pilot S′, and interference removal signal y′ in real form.

[0054] It should be noted that matrix vectorization is designed to meet the input requirements of neural networks. The specific vectorization method can refer to conventional matrix vectorization methods. Similarly, the conversion of signal data between the complex domain and the real domain is also a prior art.

[0055] S4. The environmental interference channel parameter h′, the pilot signal S′, and the interference-removed signal y′ are used as inputs of the deep unfolding channel estimation network. The deep unfolding channel estimation network obtains the real domain estimate of the backscatter channel through multiple layers of iteration. The forward reasoning calculation process of any k-th layer iteration in the network is as follows:

[0056]

[0057] V k =f(W 2k z k +B 2k )

[0058] h′ k =g(W 3k z k +B 3k )

[0059] Where: ReLU(·) is the RELU activation function; W 1k , W 2k , W 3k , B 1k , B 2k , B 3k , δ k ,λ k are all learnable parameters of the kth layer during the network iteration process, which are optimized in advance through network training; h′ k-1 、V k-1 are the two outputs of the k-1th layer in the network iteration process, and for the first layer in the network iteration process, h′0 and V0 are both initialized to all-zero vectors; f(·) and g(·) represent nonlinear activation functions; zk is an intermediate quantity in the calculation process; k = 1, 2, ..., L, where L is the total number of iterative layers of the deep unfolded channel estimation network;

[0060] Finally, the real domain estimation value of the backscatter channel is converted back to the complex domain to obtain the final estimation result of the backscatter channel.

[0061] In the above step S4 of the present invention, the deep unfolding channel estimation network needs to be pre-trained before being used for actual reasoning, and the loss function of the network training is in the form of:

[0062]

[0063] Where: h′ l h′ is the backscatter channel parameter estimate output by any iterative output of layer l in the network. LS h′ l The corresponding true value, log(l) is the logarithmic weighted weight.

[0064] Furthermore, during the training process of the aforementioned deep unrolled channel estimation network, the Adam optimizer is used to perform error backpropagation and parameter adjustment based on the aforementioned loss function until the network converges. The specific neural network training method is known in the art and will not be described in detail.

[0065] The principle and specific implementation process of the single-base station backscatter channel estimation method based on depth expansion described in S1 to S4 will be demonstrated below through a specific embodiment.

[0066] Example

[0067] In this embodiment, for the single-base-station multi-antenna multi-tag backscatter channel estimation problem, consider the following Figure 1The communication system shown in the figure consists of a full-duplex reader with multiple antennas and M backscatter devices (BDs). The reader R has N antennas, and each backscatter device BD is configured with a single antenna. We propose a sequential estimation strategy similar to the preamble design in traditional communications. That is, the M BDs are estimated sequentially, and when in the estimation phase of the i-th BD (denoted as BDi), only the BDi is in the active state, while the other BDs are in the silent state. For the sake of convenience, the active backscatter device will be denoted as BD_on, and the inactive device will be denoted as BD_off.

[0068] During the communication between the reader R and the activated backscatter device BD_on, the signal received by the reader R includes not only the backscatter signal modulated by BD_on on the downlink signal from the reader R, but also the reflection from the inactive device BD_off, the reflection from other objects near R and the interference caused by random noise (collectively referred to as ambient interference). The existence of these interferences is very detrimental to the correct demodulation of the backscatter signal by R. Therefore, before the backscatter channel estimation is performed, the interference channel is estimated first. Therefore, within the samples durations (a coherence time), in addition to the sample time used for signal reception and decoding, the remaining sample time is evenly divided into M+1 segments for performing ambient interference channel estimation and backscatter channel estimation. The entire interference and channel estimation protocol process is as follows: Figure 2 The following describes these two stages respectively.

[0069] 1. Environmental interference channel estimation stage

[0070] For the multi-antenna single-base station backscatter communication system considered in the present invention, the reader R transmits orthogonal pilots of length l0 samples through N antennas to achieve channel estimation. In order to estimate environmental interference, all tags are set to a silent state at this stage, that is, the modulation parameter is a0. For convenience of representation, l0=N is set here without loss of generality, and the transmitted pilot matrix is represented by S0. Therefore, the duration of an environmental interference channel estimation stage is τ CE0 The received signal of the inner reader R can be expressed as:

[0071]

[0072] Where S0 and Y AI is an N×N complex matrix, H AI Represents the environmental interference channel parameters, which is an N×N complex matrix. kRepresents the backscatter channel parameters of BDi, which is an N×N complex matrix, W0 is a zero-mean squared matrix The size of the complex additive white Gaussian noise (AWGN) is N×N. According to the least squares algorithm (LS algorithm), let the pseudo inverse matrix of the pilot signal be S0 + =S0 H (S0S0 H ) -1 , then the LS estimation of the interference channel can be expressed as:

[0073]

[0074] in Represents the estimation error of the interference channel, p is the power of the transmitted pilot, and satisfies SS H =pI N ,I N is the N×N identity matrix.

[0075] 2. Backscatter channel estimation stage

[0076] The next step is to model and estimate the backscatter channels of M BDs. During the whole process, the reader R sends a pilot signal S0 to each BD in the estimation phase. The duration of the whole backscatter channel estimation phase is The received signal of the inner reader can be expressed as:

[0077]

[0078] Where A is the combined modulation matrix of BD, size M×M, to satisfy:

[0079]

[0080] Y CE The size is MN×N, W1 is the additive white Gaussian noise interference of size MN×N. H is the joint estimation parameter of the MN×N backscatter channel, satisfying:

[0081] H=[h1 h2…h M ] T (5)

[0082] In order to achieve an accurate estimate of H, it is necessary to first use the previously estimated First remove the environmental interference from the received signal, that is:

[0083]

[0084] in: Therefore, the present invention proposes the following LS estimation problem for multi-tag multi-antenna backscatter channels:

[0085]

[0086] Based on the introduction of the above two stages, the present invention is a deep unfolding neural network MBCE-unfolding designed specifically for channel estimation in a single-base station multi-antenna multi-label backscatter communication network. By unfolding the iterative algorithm based on gradient descent to solve the LS problem, replacing each round of iterative calculation of the channel parameters with a single-layer neural network with a fixed structure, and using the activation function in the neural network to replace the nonlinear mapping of the output, a model-driven channel estimation network is constructed. At the same time, some learnable parameters are introduced into the network, and the parameters are adjusted based on the back propagation of the error gradient. Then, the required channel estimation parameters are obtained by simulating the iterative calculation process through the connection of multi-layer neural networks. The specific approach is as follows:

[0087] 1. Data Preprocessing-Vectorization

[0088] The channel estimation problem model for a single-base station full-duplex backscatter communication system was previously obtained. Since directly solving the LS algorithm involves matrix inversion, which is computationally very complex, we use an iterative gradient descent algorithm instead. First, equation (6) can be simplified as:

[0089] Y=S′0H+W (8)

[0090] in:

[0091]

[0092]

[0093]

[0094]

[0095] Consider using gradient descent to solve the LS problem. Since h k are unrelated, so there is no need to use complex multivariate gradient descent. Instead, we can convert the matrix into a unary gradient descent by vectorizing it. To facilitate the representation in the vectorization process, the following definition is proposed: for a matrix B, B(i) represents the i-th row of the matrix, and B(j) represents the j-th column of the matrix. Therefore, Y is vectorized as:

[0096] y vec =Sh vec +w vec (9)

[0097] y vec =vec{Y}=[y 1 y 2 …y M ]T MN 2 ×1 column vector, where Is y i N is obtained by horizontal splicing and transposing the rows 2 ×1 column vector, that is:

[0098] y i =[y i (1;), y i (2;),…,y i (N;)] T (10)

[0099] The same applies to the vectorization of W.

[0100] h vec =vec{H}=[h 1 h 2 …h M ] T , for a MN 2 ×1 column vector. Each channel parameter matrix h i The N columns (i=1, 2, ..., M) are spliced vertically in sequence 2 ×1 column vector, that is:

[0101] h j =[h j (;1) h j (;2)…h j (;N)] T (11)

[0102] S is a new pilot matrix containing BD modulation parameters obtained according to the previous operation process to adapt to the vector operation. but:

[0103]

[0104] 2. Data Preprocessing - Real Domain Conversion

[0105] The above steps complete the vectorization process of each input signal. It is worth noting that the above transmitted signal, received signal and channel parameters are all in the complex domain. In order to facilitate the processing using gradient descent, this embodiment first performs an equivalent transformation in the real domain:

[0106]

[0107] in and Respectively represent the real and imaginary parts of a complex number.

[0108] 3. Construction and training of channel estimation network based on deep expansion

[0109] Deep unfolding can construct a neural network based on prior knowledge of wireless communication systems to solve the required parameters, which can not only avoid the tedious iterative calculation process in traditional parameter estimation methods, but also circumvent the "black box" problem in deep learning. Therefore, the present invention proposes a channel estimation network MBCE-unfolding that combines deep unfolding technology with LS-based channel estimation solution. The overall structure is as follows Figure 3 shown.

[0110] The MBCE-unfolding network introduces some learnable parameters and uses the Adam optimizer to perform error backpropagation and parameter adjustment based on the loss function. The required channel estimation parameters are obtained through L-layer iteration. The total number of layers, L, in the MBCE-unfolding network is an optimizable hyperparameter that can be optimized based on actual results. This network structure design eliminates the complex matrix inversion operation in the LS algorithm, reducing computational complexity and improving the efficiency of channel estimation, thus enabling it to adapt to changing channels. Even when the channel state changes, the network can estimate the channel parameters based on the current transmitted and received signals. The network's loss function uses a logarithmic weighting method to consider the output of each network layer. Errors closer to the output layer contribute more to the loss function. The loss function is as follows:

[0111]

[0112] Each layer is a single-layer neural network constructed based on the iterative calculation process of parameter solution, such as Figure 4 shown.

[0113] The forward reasoning calculation process of each layer of the network is as follows:

[0114]

[0115] V k =f(W 2k z k +B 2k )

[0116] h′ k =g(W 3k z k +B 3k ) (15)

[0117] Where: ReLU(x) = max{0, x}, f(·), g(·) are nonlinear activation functions, h′ k-1 、V k-1are the two outputs of the k-1th layer during the network iteration, and for the first layer during the network iteration, h′0 and V0 are both initialized to all zero vectors. S′ T y′, S′ T S′ and the output of the previous layer As the input of the current k-th layer. Add trainable parameters to the network Optimized in advance through network training. k is an intermediate variable in the computational process; k = 1, 2, ..., L, where L is the total number of iterative layers in the deep unrolled channel estimation network. Furthermore, an auxiliary variable V is added to increase the width of the network. The width of a deep neural network determines the amount of feature information extracted at each layer. The more feature information, the easier it is to train the network.

[0118] The ambient interference channel parameters h′, pilot signal S′, and interference-removing signal y′, after vectorization and real-domain conversion, serve as inputs to the MBCE-unfolding network. The MBCE-unfolding network iterates through L layers to obtain a real-domain estimate of the backscatter channel. Finally, the real-domain estimate of the backscatter channel is converted back to the complex domain to obtain the final backscatter channel estimate.

[0119] In order to verify the estimation performance of the MBCE-unfolding in the present invention, this embodiment uses the normalized mean square error shown in formula (16) to measure the accuracy of channel estimation, where h m represents the real channel parameters between BD and R, Represents the channel parameters estimated by MBCE-unfolding.

[0120]

[0121] Figure 5 This is a comparison chart of the average normalized mean square error of channel estimation for 6000 training samples using MBCE-unfolding and black-box neural networks under different M and N settings, and then using the two trained neural networks and the traditional LS method to estimate the channel for another 2000 samples. Note that the bar chart in the figure uses the left scale as the y-axis, and the line chart uses the right scale as the y-axis. Figure 6 This figure compares the computational time costs of the three methods under various system settings. Note that the bar graph uses the left scale as the y-axis, while the line graph uses the right scale as the y-axis. It can be seen that MBCE-unfolding offers a significant trade-off advantage over traditional LS methods and black-box networks in both channel estimation accuracy and computational cost.

[0122] The embodiment described above is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Persons skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent substitution or equivalent transformation falls within the scope of protection of the present invention.

Claims

1. A single-base station backscatter channel estimation method based on deep expansion, characterized in that: include: S1. In the environmental interference channel estimation stage, all backscatter devices BD in the multi-antenna single base station backscatter communication system are set to silent state. After the reader R sends the pilot S0, the received signal is obtained and the environmental interference channel parameters are estimated. S2. In the backscatter channel estimation phase, the reader R sends a pilot signal S0 to each backscatter device BD in turn and obtains the received signal Y CE , and in the process of sending the pilot signal, only the backscattering device BD that transmits the pilot signal S0 is activated, but other backscattering devices BD remain silent; using the estimated environmental interference channel parameters From the received signal Y CE Remove the environmental interference signal and obtain the interference-free signal Y; S3. Vectorize the backscatter channel parameter matrix H, pilot signal S0, and interference-removing signal Y of all backscatter devices BD, converting them from matrix form to vector form. Then, convert the environmental interference channel parameters, pilot signal, and interference-removing signal in vector form from the complex domain to the real domain, respectively, to obtain the backscatter channel parameter h′, pilot signal S′, and interference-removing signal y′ in real form. S4. The environmental interference channel parameter h′, the pilot signal S′, and the interference-removed signal y′ are used as inputs of the deep unfolding channel estimation network. The deep unfolding channel estimation network obtains the real domain estimate of the backscatter channel through multiple layers of iteration. The forward reasoning calculation process of any k-th layer iteration in the network is as follows: V k =f(W 2k z k +B 2k ) h' k =g(W 3k z k +B 3k ) Where: ReLU(·) is the RELU activation function; W 1k ,W 2k ,W 3k ,B 1k ,B 2k ,B 3k ,δ k ,λ k are all learnable parameters of the kth layer during the network iteration process, which are optimized in advance through network training; h' k-1 、V k-1 are the two outputs of the k-1th layer during the network iteration, and for the first layer during the network iteration, h'0 and V0 are both initialized to all-zero vectors; f(·) and g(·) both represent nonlinear activation functions; z k is the intermediate quantity of the calculation process; k = 1, 2, ..., L, L is the total number of iterative layers of the depth-expanded channel estimation network; Finally, the real domain estimation value of the backscatter channel is converted back to the complex domain to obtain the final estimation result of the backscatter channel.

2. The single-base station backscatter channel estimation method based on deep expansion according to claim 1, characterized in that The multi-antenna single-base station backscatter communication system is a communication system consisting of a full-duplex reader R with multiple antennas and M backscatter devices BD, wherein the reader R has N antennas, each backscatter device BD is configured with a single antenna, and M and N are both integers greater than 1.

3. The single-base station backscatter channel estimation method based on deep expansion according to claim 2, characterized in that In the multi-antenna single-base station backscatter communication system, within a coherence time, except for the sample time used for signal reception and decoding, the remaining sample time is evenly divided into M+1 segments for performing environmental interference channel estimation and backscatter channel estimation.

4. The single-base station backscatter channel estimation method based on deep expansion according to claim 1, characterized in that In S1, the environmental interference channel parameters It is estimated by the least squares algorithm.

5. The single-base station backscatter channel estimation method based on deep expansion according to claim 1, characterized in that: In S2, the estimated environmental interference channel parameters are used From the received signal Y CE The method to remove environmental interference signals is as follows: Where: 1 M×1 Represents an M×1 vector of all ones.

6. The single-base station backscatter channel estimation method based on deep expansion according to claim 1, characterized in that In S4, the deep unfolding channel estimation network needs to be pre-trained before being used for actual reasoning, and the loss function of the network training is: Where: h' l h' is the backscatter channel parameter estimate output by any iterative output of layer l in the network. LS h' l The corresponding true value, log(l) is the logarithmic weighted weight.

7. The single-base station backscatter channel estimation method based on deep expansion according to claim 5, characterized in that: In S4, during the training process of the deep unfolded channel estimation network, the Adam optimizer is used to perform error backpropagation and parameter adjustment based on the loss function until the network converges.

Images