Joint beamforming method based on end-to-end learning in multi-ris assisted communication system

Abstract

The application relates to the field of information and communication technology, and relates to a joint beamforming method based on end-to-end learning in a multi-RIS assisted communication system, and aims to improve the bit error rate performance of the communication system.The application provides a joint beamforming method based on end-to-end learning, wherein single reflection and double reflection links are actually considered, and a bit error rate minimization problem is formulated by jointly designing active and passive beamforming.In order to solve the non-convex problem thus generated, a new method is introduced, and the proposed multi-RIS assisted communication system is regarded as an end-to-end optimization task.Specifically, modulation, precoding, passive beamforming, combination and demodulation processes are simultaneously optimized to avoid local optimization.Simulation results show that, compared with a traditional alternating optimization algorithm, the system based on end-to-end learning in the application realizes a competitive performance gain.In addition, under the condition that the total number of reflection elements is the same, the performance enhancement of multiple RISs relative to a single RIS assisted system is verified.

Description

Technical Field

[0001] This invention relates to the field of information and communication technology, and more specifically to a joint beamforming method based on end-to-end learning in a communication scenario within a communication network. Background Technology

[0002] Communication is a field rich in expert knowledge, involving modeling different types of channels, compensating for various hardware deficiencies, and designing optimal signaling and detection schemes to ensure reliable data transmission. Therefore, communication is a complex and mature engineering field with many different research areas, all focusing on performance improvements at the physical layer. Machine learning or deep learning-based methods offer tangible new benefits for optimizing communication systems. Representing real-world images or language with rigorous mathematical models is difficult, but it can now be achieved directly using deep learning algorithms that can learn to perform tasks with accuracy exceeding human levels. However, in communication systems, we need to design transmission signals to support direct analytical algorithms for symbol detection across various channel and system models. Therefore, provided such models fully capture real-world effects, deep learning can produce significant results in optimizing physical layer communication.

[0003] Complex wireless systems can be divided into several simpler, independent modules, such as source encoders, channel encoders, modulators, demodulators, channel decoders, and decoders. Based on this modular design paradigm, the global optimization of the entire communication system can be approximated by optimizing each module individually. However, optimizing each module does not equate to global optimization of the entire system. Therefore, this classic design paradigm becomes a bottleneck limiting the globally optimal performance of wireless communication systems.

[0004] Recently, Reconfigurable Intelligence Surfaces (RIS) have emerged as one of the most promising candidates for enabling intelligent communication systems as an evolving wireless technology. Composed of low-cost passive components, RIS can reflect and incident signals like a scatterer in a signal propagation environment. Each RIS reflecting element can be controlled in a software-defined manner to produce ideal multipath effects, such as signal focusing and interference cancellation, controlling the radio environment to improve wireless communication performance. In particular, combining RIS with multiple-input multiple-output systems can produce very positive effects, such as mitigating propagation blockage through additional paths and reducing the number of antennas required in the transceiver to achieve the desired performance. To leverage these advantages of RIS-assisted communication, implementing high-gain joint beamforming in RIS-assisted communication systems is crucial. A RIS-assisted communication channel consists of two sub-channels: a direct channel between the base station and the user, and an indirect channel generated by the RIS. The end-to-end channel between the base station and the user can be optimized at the link end in a conventional manner. However, due to the limitations of conventional optimization algorithms, alternating optimization of RIS and active beamforming results in performance degradation. Summary of the Invention

[0005] This invention aims to reduce the performance loss of existing beamforming methods and to improve the bit error rate performance of communication systems.

[0006] This provides a joint beamforming method based on end-to-end learning in a multi-RIS assisted communication system.

[0007] The specific solution of the present invention is as follows:

[0008] Step 1: A deep neural network (DNN) is used to represent modulation and precoding at the base station, and demodulation and decoding at the user. The original bitstream b is mapped to x through the network and then restored to x after transmission through the channel. The technical details of the intermediate neural network model in step one can be described as follows:

[0009] Consider an input bit stream b∈{0,1} k At the base station, the encoder maps b to a complex baseband signal. Transmission is handled by N1 antennas. The channel is then defined as a statistical system where the output y follows a conditional probability distribution P(y|x). The decoder will receive the signal. Restore to received bit stream The transmitted signal is from the transmitted bit stream b = [b1, b2, ..., b k ]∈{0,1} k Generated in [the process], after modulation and precoding, it can be represented as [the following].

[0010] x = ps, (1)

[0011] in is the precoding (i.e., active beamforming) vector at the base station, and s is the modulation signal for each bit of signal in b.

[0012] At the user's location, the final signal z can be determined from the received signal y via z = q. T y is calculated, where It is the user's combined vector. Then, we use a hard-discriminate algorithm to recover the signal z, and the recovered bitstream It is obtained through the corresponding demodulation algorithm.

[0013] Step 2: Introduce trainable parameters to represent the passive beamforming of RIS in the embedded layer, that is, the phase shift and amplitude coefficients are represented as trainable weights, and the phase shift and amplitude of the signal change after passing through RIS.

[0014] The technical details of the RIS layer model in step two can be described as follows:

[0015] The RIS is described as an embedding layer. Specifically, the elements Θ1 and Θ2 of the embedding matrix are trainable weights used to characterize the phase reflection and amplitude changes of the signal as it passes through the RIS. The trainable weights of the RIS are jointly trained with the encoder and decoder to ensure optimal performance.

[0016] In a transmission system, multiple RIS transmit signal x to the user. The phase shift and amplitude of x can be independently adjusted by each element of the RIS, where we will use θ n ∈[0,2π) and β n ∈[0,1] represents the phase shift and amplitude reflection coefficient of the nth reflecting element of RIS.

[0017] Step 3: In our proposed end-to-end model, modulation, precoding, passive beamforming, decoding and demodulation processes are designed simultaneously, realizing the joint beamforming method.

[0018] The beneficial effects of this invention are as follows: Addressing the common problem of difficulty in obtaining a globally optimal solution during the optimization of multi-RIS assisted communication systems, this invention studies a joint beamforming method based on end-to-end learning, which practically considers both single-reflection and dual-reflection links. Minimizing the bit error rate is addressed by jointly designing active and passive beamforming. To solve the resulting non-convexity problem, we introduce a novel method that treats the proposed multi-RIS assisted communication system as an end-to-end optimization task. Specifically, modulation, precoding, passive beamforming, combination, and demodulation processes are simultaneously optimized to avoid local optima. Simulation results show that our proposed end-to-end learning-based system achieves competitive performance gains compared to traditional alternating optimization algorithms. Furthermore, with the same total number of reflective elements, the performance enhancement of multiple RIS systems over a single RIS-assisted system is verified. Attached Figure Description

[0019] Figure 1 This is a schematic diagram illustrating the communication principle of the joint beamforming method based on end-to-end learning in the multi-RIS assisted communication system described in this invention.

[0020] Figure 2 This is a schematic diagram comparing the bit error rate performance and signal-to-noise ratio of the proposed method with those of existing technologies, obtained from simulation experiments.

[0021] Figure 3 This is a schematic diagram comparing the bit error rate performance of this method with existing technologies and different RIS elements, obtained from simulation experiments;

[0022] Figure 4 This is a schematic diagram showing the convergence of the proposed method as a function of the number of iterations, obtained from simulation experiments. Detailed Implementation

[0023] Specific Implementation Method 1: This implementation method, illustrated with figures, describes a joint beamforming method based on end-to-end learning in a multi-RIS assisted communication system, comprising the following steps:

[0024] Step 1: A deep neural network (DNN) is used to represent modulation and precoding at the base station, and demodulation and decoding at the user. The original bitstream b is mapped to x through the network and then restored to x after transmission through the channel.

[0025] The technical details of the intermediate neural network model in step one can be described as follows:

[0026] Consider an input bit stream b∈{0,1} k At the base station, the encoder maps b to a complex baseband signal.

[0027] Transmission is handled by N1 antennas. The channel is then defined as a statistical system where the output y follows a conditional probability distribution P(y|x). The decoder will receive the signal. Restore to received bit stream The transmitted signal is from the transmitted bit stream b = [b1, b2, ..., b k ]∈{0,1} k Generated in [the process], after modulation and precoding, it can be represented as [the following].

[0028] x = ps, (1)

[0029] in is the precoding (i.e., active beamforming) vector at the base station, and s is the modulation signal for each bit of signal in b.

[0030] At the user's location, the final signal z can be determined from the received signal y via z = q. T y is calculated, where It is the user's combined vector. Then, we use a hard-discriminate algorithm to recover the signal z, and the recovered bitstream It is obtained through the corresponding demodulation algorithm.

[0031] Step 2: Introduce trainable parameters to represent the passive beamforming of RIS in the embedded layer, that is, the phase shift and amplitude coefficients are represented as trainable weights, and the phase shift and amplitude of the signal change after passing through RIS.

[0032] The technical details of the RIS layer model in step two can be described as follows:

[0033] The RIS is described as an embedding layer. Specifically, the elements Θ1 and Θ2 of the embedding matrix are trainable weights used to characterize the phase reflection and amplitude changes of the signal as it passes through the RIS. The trainable weights of the RIS are jointly trained with the encoder and decoder to ensure optimal performance.

[0034] In a transmission system, multiple RIS transmit signal x to the user. The phase shift and amplitude of x can be independently adjusted by each element of the RIS, where we will use θ n ∈[0,2π) and β n ∈[0,1] represents the phase shift and amplitude reflection coefficient of the nth reflecting element of RIS.

[0035] Step 3: In our proposed end-to-end model, modulation, precoding, passive beamforming, decoding and demodulation processes are designed simultaneously, realizing the joint beamforming method.

[0036] The technical details of the end-to-end model in step three can be described as follows:

[0037] We consider a downlink MIMO communication system assisted by multiple RIS (Reflectors and Reflectors). Since the direct link between the base station and the user is severely obstructed by obstacles, two RIS deployed near the base station and the user are used to enhance the communication link from the base station with N1 antennas to the user with N2 antennas. The RIS are equipped with M1 and M2 reflectors, respectively.

[0038] A multi-RIS-assisted communication system based on end-to-end learning can be implemented as an autoencoder, consisting of four parts: an encoder, a channel, a RIS, and a decoder. The encoder extracts features of the transmitted signal instead of the transmitter, the channel is modeled by a conditional probability density function, the RIS can be viewed as an embedding layer, and the decoder recovers the received signal instead of the receiver. To achieve optimal performance, our goal is to optimize the beamforming process end-to-end by minimizing the difference between the transmitted signal b and the received signal b. The cross-entropy loss function between them can reach the global optimum, and it is expressed as follows:

[0039]

[0040] Mini-batch gradient descent is used to minimize the cross-entropy loss function by simultaneously training the encoder, RIS, and decoder weights. The updated weights are represented as follows:

[0041]

[0042]

[0043]

[0044]

[0045] in and These are the trainable weights of the encoder, RIS1, RIS2, and decoder in the i-th traversal; η represents the learning rate, which affects... Convergence; Let W represent the gradient of the cross-entropy loss function, where W is the gradient of the cross-entropy loss function.

[0046] The encoder and decoder use W E and W D The values ​​of RIS can be obtained from the phase shift and amplitude reflection coefficients. and After the training process, the trained weights can obtain a robust representation of the channel conditions and can be deployed to each module via wired or wireless means.

[0047] The optimization problem of joint beaming in step three can be described as follows:

[0048] This invention focuses on optimizing bit error rate performance, which is achieved by comparing b and The error rate is calculated based on the ratio between these parameters. To minimize the bit error rate during downlink transmission, we consider jointly optimizing the precoder p, passive reflection matrices Θ1 and Θ2 at the base station, and the decoder q at the user. Furthermore, optimization is performed under the constraint of the base station transmit power, i.e. And passive beamforming matrix constraints, i.e., θ n ∈[0,2π) and β n ∈[0,1]. In mathematics, the optimization problem is expressed as:

[0049]

[0050]

[0051]

[0052] Where P e ∈{0,1}, where k is the number of bits transmitted.

[0053] This invention addresses the common problem of difficulty in obtaining a globally optimal solution during the optimization of multi-RIS assisted communication systems. It proposes a joint beamforming method based on end-to-end learning, which practically considers both single-reflection and dual-reflection links. Minimizing the bit error rate is addressed by jointly designing active and passive beamforming. To solve the resulting non-convexity problem, this invention introduces a novel approach that treats the proposed multi-RIS assisted communication system as an end-to-end optimization task. Specifically, it simultaneously optimizes modulation, precoding, passive beamforming, combining, and demodulation processes to avoid local optima.

[0054] Step 4: The multi-RIS assisted communication system proposed in this invention considers both single-reflection links and dual-reflection links, simultaneously receiving user signals from a direct link without RIS, a single-reflection link with RIS, and a dual-reflection link with RIS. Because both single-reflection and dual-reflection links are considered simultaneously, user signals received from the direct link without RIS, the single-reflection link with RIS, and the dual-reflection link with RIS... It can be represented as:

[0055]

[0056] in It is the signal received by the user terminal; This is the transmitted signal on the base station side, denoted by x = ps, where This is the precoding (i.e., active beamforming) stage of the base station, where s∈C is the transmitted signal; It is the received additive white Gaussian noise.

[0057] These are the equivalent channels for the base station → user, base station → RIS 1, base station → RIS 2, RIS 1 → RIS 2, RIS 1 → user, and RIS 2 → user links, respectively.

Claims

1. A joint beamforming method based on end-to-end learning in a multi-RIS assisted communication system, characterized by: The following communication process occurs between the end and the end, where "end-to-end" refers to the communication between the base station and the user. The specific communication process is as follows: Step 1: Use a first-order deep neural network model to represent the modulation and precoding processes of the signal at the base station; use a second-order deep neural network model to represent the demodulation and decoding processes of the signal at the user. Step 2: At the transmitting end, transmit the original bit stream. The complex baseband signal to be emitted was obtained by mapping using a deep neural network model. ; Step 3: Introduce trainable parameters to represent the RIS passive beamforming in the embedded layer, i.e., the phase shift and amplitude coefficients are represented as trainable weights. The RIS is used to modify the phase shift and amplitude of the complex baseband signal to be transmitted, obtained in Step 2. ; The optimization problem for joint beamforming is as follows: in, , It is the number of bits transmitted. For precoded vectors, It is a combination vector of users. and Trainable parameters are used to characterize the phase reflection and amplitude changes of the signal as it passes through the RIS; the original bitstream. The restored bitstream , and Let these be the phase shift and amplitude reflection coefficients of the nth reflecting element of the RIS, respectively; the RIS is equipped with... and One reflective element; Step 4: Deploy M RIS across the end-to-end to enhance the connection from the point of view. A base station with one antenna to a base station with The communication link for each user with one antenna; the M, and All are positive integers; ; Step 5: The transmitting end uses the communication link established in Step 4 to transmit the complex baseband signal to be transmitted, whose phase shift and amplitude have been changed in Step 3. Transmit to the wireless channel; complete a joint beamforming based on end-to-end learning in a multi-RIS assisted communication system; Step Six: At the receiving end, use a receiving antenna to receive the complex baseband signal to be transmitted, whose phase shift and amplitude have been changed in Step Three, in the wireless channel described in Step Five. The received signal is ; , , It transmits signals; It is the received additive white Gaussian noise. , , , , , These are base stations Users, base stations RIS 1, Base Station RIS 2, RIS 1 RIS 2, RIS 1 Users and RIS 2 Equivalent channel of user link; Step 7: Use the second deep neural network model described in Step 1 to process the complex baseband signal to be transmitted, whose phase shift and amplitude have been changed in Step 3 of Step 6. Perform corresponding demodulation and decoding to obtain the restored bitstream. .

2. The joint beamforming method based on end-to-end learning in a multi-RIS assisted communication system according to claim 1, characterized in that... In step one, the first deep neural network model represents the modulation and precoding processes of the signal at the base station; the specific description of the signal processing using the second deep neural network model is as follows: At the base station, consider a set of input bit streams. k is a positive integer, and the encoder will output the bit stream. Mapped to complex baseband signal ,Depend on One antenna transmits; then, the channel is defined as the output. Follows conditional probability distribution The statistical system; the decoder will receive the signal. Restore to received bit stream The transmitted signal is from the transmitted bit stream. Generated in [the process], after modulation and precoding, it can be represented as [the following]. , in It is the precoding vector at the base station. yes The modulated signal for each bit of the signal; At the user's end, the final signal According to the received signal pass Calculation, where It is a combination vector of users; then, a hard discriminant algorithm is used to recover the signal. The recovered bitstream This is achieved through appropriate demodulation methods.

3. The joint beamforming method based on end-to-end learning in a multi-RIS assisted communication system according to claim 1, characterized in that... In step two, the specific technical details of the RIS layer model are as follows: RIS is described as an embedding layer, specifically, the elements of the embedding matrix. and These are trainable weights, used to characterize the phase reflection and amplitude changes of the signal as it passes through the RIS; specifically, the trainable weights of the RIS are jointly trained with the encoder and decoder. In a communication system, M RIS will transmit complex baseband signals. Transmitted to the user; Complex baseband signal The phase shift and amplitude can be independently adjusted by each element of the RIS, where: and These represent the phase shift and the nth amplitude reflection coefficient of the nth reflecting element of the RIS, respectively; n is a positive integer.

4. The joint beamforming method based on end-to-end learning in a multi-RIS assisted communication system according to claim 1, characterized in that... The specific technical details of end-to-end communication within the system are as follows: Consider a downlink MIMO communication system assisted by multiple RIS; since the direct link between the base station and the user is severely obstructed by obstacles, two RIS deployed near the base station and the user are used to enhance the downlink communication. A base station with one antenna to a base station with Communication links for users with one antenna; The RIS are equipped with and One reflective element; the and All are positive integers.

5. The joint beamforming method based on end-to-end learning in a multi-RIS assisted communication system according to claim 4, characterized in that... A multi-RIS assisted communication system can be implemented as an automatic encoder, which consists of four parts: encoder, channel, RIS, and decoder; The encoder is used to extract features of the transmitted signal instead of the transmitter, the channel is modeled by a conditional probability density function, the RIS serves as an embedding layer, and the decoder is used to recover the received signal instead of the receiver.

Images