A communication method for remote reasoning tasks

Abstract

The present invention relates to a communication method for remote reasoning tasks, including: collecting pictures of multiple perspectives of an object by an integrated device, and extracting feature information of the picture by a feature encoder; then using a channel precoder to convert it into a data format that conforms to wireless channel transmission; the transmitted signal is reflected by a RIS reflection unit; while increasing the transmission distance, the communication performance is increased, and the edge server receives signals from multiple devices and fuses them into multi-perspective features of the object; and the classifier is used to calculate the classification result. The present invention can effectively improve the task relevance of RIS and improve the potential of RIS to improve communication performance.

Description

Technical Field

[0001] The invention relates to a communication method oriented to remote reasoning tasks, and belongs to the technical field of wireless communications. Background Art

[0002] As we enter the era of digital transformation driven by the Internet of Things, smart cities, and connected devices, the demand for communication networks is also growing rapidly [Laghari AA, Wu K, Laghari RA, et al. A review and state of art of Internet of Things (IoT) [J]. Archives of Computational Methods in Engineering, 2021: 1-19.]. These advanced applications require not only high data transmission rates, but also priority processing of specific tasks that rely on timely and accurate information delivery. To address these challenges, task-oriented communication, which optimizes network resources by focusing on the specific needs of task execution, is considered a key strategy. Traditional communication systems focus on the complete transmission of data without considering its direct utility, while task-oriented communication is committed to identifying and transmitting critical data required to perform specific tasks, thereby significantly reducing the load on network resources. With the tremendous progress of artificial intelligence (AI), most existing works utilize deep neural networks (DNNs) to extract task-related features and serve various downstream tasks. However, practical applications of task-oriented communication often face challenges such as signal blocking, multipath fading, and interference, which can significantly degrade communication performance. Reconfigurable smart surfaces (RIS) consist of electronic components that can control electromagnetic waves. By adjusting the phase and amplitude of the input signal, the wireless communication environment can be effectively controlled to transform potential obstacles into favorable reflection surfaces, thereby improving communication performance. In many existing works, RIS has been deployed in wireless communications and has achieved remarkable results. However, RIS does not distinguish between the type of data being transmitted or the communication task.Most of the existing RIS designs focus on improving communication performance by enhancing channel conditions, solving problems such as minimizing power [Wu Q, Zhang R. Beamforming optimization for wireless network aided by intelligent reflecting surface with discrete phase shifts [J]. IEEE Transactions on Communications, 2019, 68 (3): 1838-1851.], maximizing achievable rate [Kumar V, Flanagan MF, Zhang R, et al. Achievable rate maximization for underlay spectrum sharing MIMO system with intelligent reflecting surface [J]. IEEE Wireless Communications Letters, 2022, 11 (8): 1758-1762.] or overall improved data throughput [Zhao J, Ye J, Guo S, et al. Reconfigurable intelligent surface enabled joint backscattering and communication [J]. IEEE Transactions on Vehicular Technology, 2023.] Although these general methods can effectively improve the overall network performance, they often ignore the requirements of specific communication tasks. The misalignment between the communication mission and the design goals of RIS may hinder the potential of RIS in improving communication performance, leading to unnecessary signal amplification and waste of bandwidth and energy. Summary of the invention

[0003] In view of the shortcomings of the prior art, the present invention provides a task-oriented joint RIS and communication design method to ensure that the enhancement provided by RIS and the transmitted data are highly relevant to the task. Specifically, the present invention proposes a unified method to design feature encoders, precoders and RIS, aiming to maximize the coding rate reduction (MCR) 2), while considering the transmission power limitation and the amplitude constraint of the RIS reflection unit, the system design problem is converted into an alternating optimization problem through problem transformation, and the ABGP and RGP methods are proposed respectively for solving the design schemes of the precoder and RIS. Under the same range of SNR (signal-to-noise ratio), the classification effect of the present invention on the ModelNet10 dataset has achieved higher classification accuracy than the traditional data recovery-based method (MMSE). In addition, compared with the general RIS design method ACM (maximum achievable rate), this method can achieve similar or even better performance as ACM with fewer RIS units.

[0004] Terminology explanation:

[0005] 1. Rayleigh channel is a wireless communication channel model used to describe the fading characteristics of signals in a multipath propagation environment. In this environment, electromagnetic waves propagate through multiple paths such as reflection, refraction and scattering to reach the receiver. Due to the different delay times of each path, the intensity and phase of the received signal fluctuate, resulting in Rayleigh fading. The present invention uses Rayleigh channel as a carrier for wireless signal transmission, with the purpose of simulating the performance of signals in actual environments.

[0006] 2. Channel transmission matrix, usually refers to a complex matrix that describes the multipath effect experienced by a signal when it propagates in a wireless channel. In a multiple-input multiple-output (MIMO) system, the channel transmission matrix represents the transmission characteristics from each transmitting antenna to each receiving antenna, including path loss, phase offset, and delay.

[0007] The technical solution of the present invention is:

[0008] A communication method for remote reasoning tasks, the method is applied to a multi-device collaborative edge reasoning system; the multi-device collaborative edge reasoning system includes multiple integrated devices equipped with cameras and multiple antennas, a feature encoder, a channel precoder, an edge server equipped with multiple antennas, and a classifier; including:

[0009] The integrated device collects images of objects from multiple perspectives, and extracts the feature information of the images through a feature encoder; the channel precoder is then used to convert the information into a data format that conforms to wireless channel transmission; the transmitted signal is reflected by the RIS reflection unit; the transmission distance is increased while the communication performance is increased; the edge server receives signals from multiple devices and fuses them into multi-perspective features of the object; the classifier is used to calculate the classification result.

[0010] According to the preferred embodiment of the present invention, the feature vector, i.e., the feature information of the image, is Adjust to plural form As shown in formula (I):

[0011]

[0012] In formula (I), represents the imaginary unit, assuming D k =2F k ;

[0013] is a vector, D k refers to The dimension, F k refers to Half the dimension of k can be divided by integers, and Refers to the notation of the fields of complex and real numbers.

[0014] According to the preferred embodiment of the present invention, the device is defined Channel precoder The output signal is N k is the number of transmit antennas of device k, x k Satisfy the transmit power constraint shown in formula (II):

[0015]

[0016] In formula (II), E(·) represents the mean operator, ‖·‖ 2 represents the two-norm operator, express The covariance matrix, P k represents the maximum transmission power of device k, and the superscript H represents the conjugate transpose.

[0017] Preferably, according to the present invention, the communication channel established is a Rayleigh channel, and the established communication signal transmission model is expressed as formula (III):

[0018]

[0019] In formula (III), y represents the edge server fusing the received reflection signal into the multi-view feature vector of the target; represents the transmission channel matrix of device k, where represents the linear channel matrix between device k and the edge server, represents the reflection channel matrix between the RIS reflection unit and the edge server about device k, represents the transmission channel matrix between device k and the RIS reflection unit; Represented by RIS phase shift vector is the matrix of the diagonal, And θi ∈(0,2π] represents the phase shift caused by the i-th RIS reflection unit; N r is the number of receiving antennas of the edge server; x k represents the transmission signal of device k, represents the feature vector extracted by the feature encoder of device k; V k represents the channel precoding matrix of device k; represents a Gaussian noise vector, where δ represents the noise power and I represents the identity matrix.

[0020] Preferably, according to the present invention, the RIS phase shift vector satisfies the constraint condition of formula (IV):

[0021]

[0022] In formula (IV), N is the dimension of the RIS phase shift vector φ, and i=1, 2, ..., N represents the i-th element.

[0023] Further preferably, the channel transmission matrices of K devices are spliced: in, Indicates the total number of transmitting antennas of the transmitting device; definition in, is the dimension after concatenating the feature vectors of K devices, rewriting formula (III) as:

[0024]

[0025] The fused multi-view feature vector y is further used as the input of the classifier, and the inference label is obtained after calculation Complete remote reasoning tasks.

[0026] According to the preferred embodiment of the present invention, the design of the unified feature encoder, channel precoder and RIS reflection unit is expressed as: maximizing the MCR of the edge server receiving the signal 2 , while satisfying the power constraints of multiple edge devices and the amplitude constraints of the RIS reflection unit, the optimization problem is established as formula (VI):

[0027]

[0028] In formula (VI), f(·,ψ) represents a MCR 2 The neural network pre-trained for the objective function, ψ represents the learnable parameters of the neural network; is the phase shift vector of RIS; in, is the lossy decoding rate represents the transmission channel matrix; and Represent the covariance matrix of the eigenvector and the covariance matrix of the eigenvector of category j, p j is the prior probability of the feature vector received by the edge server and associated with category j.

[0029] Preferably, according to the present invention, the multivariable optimization problem in formula (VI) is transformed into an alternating optimization problem with single variable constraints, specifically comprising:

[0030] 1) Fix the RIS phase shift vector φ and transform equation (VI) into equation (VII):

[0031]

[0032] 2) Fix the channel precoding matrix V of K devices and transform equation (VI) into equation (VIII):

[0033]

[0034] Preferably, according to the present invention, an ABGP algorithm is proposed to solve the precoding matrix problem, that is, formula (VII); comprising:

[0035] Convert equation (VII) to alternately solve the precoding matrix V of K devices k Specifically, fix the precoding matrix of all devices except device k and optimize V k , repeat the optimization process until convergence, and the optimal precoding matrix V is obtained.

[0036] Further preferably, according to the law tr(B H C)=vec(B) H vec(C) and Where b represents the number of columns of matrix B, the constraint in formula (VII) is transformed into:

[0037]

[0038] In formula (IX) vec represents vectorized operation. represents the Kronecker product, T represents the matrix transpose;

[0039] Calculation formula (VII) The objective function is about V k The gradient is:

[0040]

[0041] In formula (X), and Represent Σ and Σ respectively j The k-th column matrix block is specifically Σ=[Σ (1),…,Σ (K) ];

[0042] When the transmission power of each device meets When Ψ(v k ), so The projection is:

[0043]

[0044] In formula (XI),

[0045] By formula (XII) k To update:

[0046]

[0047] In formula (XII),

[0048] By finding a θ∈[0,π / 2) that satisfies θ=argmaxΨ(V k ), based on formula (XII) the optimal v k and V k ;

[0049] By repeating the above optimization process for each device, the optimal precoding matrix V with a fixed RIS phase shift vector φ can be obtained.

[0050] Preferably, according to the present invention, an RGP algorithm is proposed to solve the RIS optimization problem, that is, formula (VIII); the constraints in formula (VIII) are converted into convex constraints:

[0051] tr(φφ H )=N and|φ| ∞ ≤1(XIII);

[0052] In formula (XIII), tr represents the trace operation, and the sum of the diagonal elements of the matrix is ​​calculated, ||.|| ∞ represents the infinity norm, and represents the AND operator.

[0053] More preferably, using Approximate infinite norm l ∞ ; Using the barrier method, using the logarithmic barrier function Integrate the non-negative constraints to approximate the penalty for violating the constraints, where t is a constant used to adjust the penalty;

[0054] The objective function in formula (VIII) is rewritten as:

[0055] max φΥ(φ,h)=Ψ(φ)+I(1-‖φ‖ h ) (XIV)

[0056] Calculate the gradient of the objective function of formula (XIV) with respect to φ:

[0057]

[0058] In formula (XV), at the same time:

[0059]

[0060] Project the search direction into formula (XVII):

[0061]

[0062] Among them, <,> represents the inner product, by finding a satisfy And update φ using formula (XVIII):

[0063]

[0064] Project φ using formula (XIX) to satisfy the constraints:

[0065]

[0066] The above optimization process is repeated until convergence, and the optimal RIS phase shift vector φ that meets the constraints under the condition of a given precoding matrix V is obtained.

[0067] Preferably, according to the present invention, an alternating iterative algorithm is used to solve the optimization problem of a multi-device collaborative edge reasoning system; comprising:

[0068] Initialize the precoding matrix V to satisfy Initialize the RIS phase shift vector φ to satisfy |φ i |=1;

[0069] Calculated according to formula (X) Calculated according to formula (XI) And according to Find a To update V according to formula (XII) k and v k , repeat the above steps to the precoding matrix of K devices;

[0070] Calculated according to formula (XV) Project it according to formula (XVII), and according to Find a Update φ by equation (XVIII) and project φ according to equation (XIX) to satisfy the constraint;

[0071] Repeat the above steps until convergence; finally, the maximum MCR is obtained under the premise of meeting constraints (II) and (IV). 2 The precoding matrix V and RIS phase shift vector φ of the value.

[0072] The beneficial effects of the present invention are:

[0073] In view of the current demand for task-related data in wireless communications and the fact that general RIS design ignores specific task requirements, which hinders RIS from improving communication performance, the present invention proposes a design method for jointly designing RIS and communication strategy. The design target of the feature encoder, channel precoder and RIS phase shift vector in the unified model is MCR. 2 , the maximization problem is solved with the transmit power and the phase shift vector amplitude as constraints; the ABGP and RGP methods are proposed to solve the channel precoder and the RIS phase shift vector respectively. Under the strategy of alternating optimization, the final designed scheme achieves higher inference accuracy in a low channel environment (SNR is -20dB to 0dB) than the traditional data recovery-based method; in addition, compared with the general RIS design method (such as ACM), the designed scheme can achieve similar performance with fewer RIS reflection units, which shows that the present invention can effectively improve the task relevance of RIS and enhance the potential of RIS to improve communication performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 It is an architecture diagram of a multi-device collaborative edge inference system constructed by the present invention;

[0075] FIG2(a) is a schematic diagram showing a comparison of the classification accuracy of the present invention and the traditional data recovery-based communication (maximum a posteriori probability classifier, MAP) at different receiving ends;

[0076] FIG2( b ) is a schematic diagram showing a comparison of the classification accuracy of the present invention and the traditional data recovery-based communication (K nearest neighbor classifier, KNN) at different receiving ends;

[0077] FIG2(c) is a schematic diagram showing a comparison of the classification accuracy of the present invention and the traditional data recovery-based communication (neural network classifier, NN) at different receiving ends;

[0078] Figure 3 This is a comparison chart of the classification accuracy of the present invention and the general RIS design method (ACM) when NN is used as the receiving end and SNR = -20dB;

[0079] Figure 4is the number of RIS reflection units N and edge server receiving antennas N in the method of the present invention r Plot of the relationship with classification accuracy. DETAILED DESCRIPTION

[0080] The present invention will be further described below with reference to the accompanying drawings and embodiments, but is not limited thereto.

[0081] Example 1

[0082] A communication method for remote reasoning tasks, the method is applied to a multi-device collaborative edge reasoning system; the multi-device collaborative edge reasoning system includes multiple integrated devices equipped with high-precision cameras and multiple antennas, a feature encoder, a channel precoder, an edge server equipped with multiple antennas, and a classifier; including:

[0083] The integrated device collects images of objects from multiple perspectives and extracts the feature information of the images through the feature encoder; feature information refers to key data extracted from the image that helps describe and identify the image content, such as the color, shape, and texture of the object. The feature encoder is a model composed of a multi-layer neural network. By training the neural network parameters of the feature encoder, the feature encoder can learn the feature information in the image (saved in the form of a feature vector). The channel precoder is then used to convert it into a data format that conforms to wireless channel transmission; the implementation process is as follows: V k is the channel precoder; the transmitted signal is reflected by the RIS reflection unit; the implementation process is φ is a diagonal matrix with RIS phase shift vector as diagonal element. Under the action of RIS, the channel transmission matrix is ​​composed of H dk Became H φk In order to improve the transmission distance and communication performance, the edge server receives signals from multiple devices and fuses them into multi-view features of the object. The implementation process is formula (III), which adds the transmission signals received from K devices. The classifier is used to calculate the classification result.

[0084] Establish Figure 1The multi-device collaborative edge inference system assisted by RIS is shown in the figure. K devices equipped with cameras and multiple antennas cooperate with the edge server and complete a remote inference task with the help of N RIS reflection units. The camera is responsible for taking pictures of the target from multiple perspectives, and the feature encoder extracts the feature vector. Before sending the transmission signal, the channel precoder needs to convert the feature vector into a signal form that can be transmitted by the channel. RIS is embedded on the surface of the surrounding buildings to reflect, modulate and phase shift the sent signal. The edge server deploys multiple receiving antennas. After receiving the signal reflected by RIS, it fuses it into a multi-perspective feature vector of the target, and calculates the category through the classifier to complete the edge inference task.

[0085] Example 2

[0086] The communication method for remote reasoning tasks described in Example 1 is different in that:

[0087] The feature vector is the feature information of the image. Adjust to plural form For channel transmission; the feature vector is to save the feature information in the form of a one-dimensional vector; the feature encoder receives the picture as input and outputs the feature vector. As shown in formula (I):

[0088]

[0089] In formula (I), represents the imaginary unit, assuming D k =2F k ;

[0090] is a vector, D k refers to The dimension, F k refers to Half the dimension of k can be divided by integers, and Refers to the symbols of the complex and real number fields. For example represent The dimension of the negative domain is F k Vector.

[0091] Defining devices Channel precoder The output signal is N k is the number of transmit antennas of device k, x k Satisfy the transmit power constraint shown in formula (II):

[0092]

[0093] In formula (II), E(·) represents the mean operator, ‖·‖ 2 represents the two-norm operator, express The covariance matrix, P k represents the maximum transmission power of device k, and the superscript H represents the conjugate transpose.

[0094] The communication channel is established as a Rayleigh channel, and the Rayleigh channel is used as the carrier of wireless signal transmission to simulate the performance of the signal in the actual environment. The established communication signal transmission model is expressed as formula (III):

[0095]

[0096] In formula (III), y represents the edge server fusing the received reflection signal into the multi-view feature vector of the target; represents the transmission channel matrix of device k, where represents the linear channel matrix between device k and the edge server, represents the reflection channel matrix between the RIS reflection unit and the edge server about device k, represents the transmission channel matrix between device k and the RIS reflection unit; Represented by RIS phase shift vector is the matrix of the diagonal, And θ i ∈(0,2π] represents the phase shift caused by the i-th RIS reflection unit; N r is the number of receiving antennas of the edge server; x k represents the transmission signal of device k, represents the feature vector extracted by the feature encoder of device k; V k represents the channel precoding matrix of device k; its function is to convert the feature vector output by the feature encoder into a signal that can be transmitted through the channel; represents a Gaussian noise vector, where δ represents the noise power and I represents the identity matrix.

[0097] Without considering the amplitude modulation effect of RIS, the RIS phase shift vector satisfies the constraint condition of formula (IV):

[0098]

[0099] In formula (IV), N is the dimension of the RIS phase shift vector φ, and i=1, 2, ..., N represents the i-th element.

[0100] Splice the channel transmission matrices of K devices: in, Indicates the total number of transmitting antennas of the transmitting device; definition in, is the dimension after concatenating the feature vectors of K devices, rewriting formula (III) as:

[0101]

[0102] The fused multi-view feature vector y is further used as the input of the classifier, and the inference label is obtained after calculation Complete remote reasoning tasks.

[0103] The design of unified feature encoder, channel precoder and RIS reflection unit is expressed as: Maximize the MCR of the signal received by the edge server 2 , while satisfying the power constraints of multiple edge devices and the amplitude constraints of the RIS reflection unit, the optimization problem is established as formula (VI):

[0104]

[0105] In formula (VI), f(·,ψ) represents a MCR 2 The neural network pre-trained for the objective function, ψ represents the learnable parameters of the neural network; is the phase shift vector of RIS; in, ε is the lossy decoding rate; represents the transmission channel matrix; and Represent the covariance matrix of the eigenvector and the covariance matrix of the eigenvector of category j, p j is the prior probability of the feature vector received by the edge server and associated with category j.

[0106] The multivariable optimization problem in formula (VI) is transformed into an alternating optimization problem with single variable constraints, specifically including:

[0107] 1) Fix the RIS phase shift vector φ and transform equation (VI) into equation (VII):

[0108]

[0109] 2) Fix the channel precoding matrix V of K devices and transform equation (VI) into equation (VIII):

[0110]

[0111] The ABGP algorithm is proposed to solve the precoding matrix problem, namely, formula (VII); in this problem, only the precoder V is the optimization variable. It includes:

[0112] Convert equation (VII) to alternately solve the precoding matrix V of K devices k Specifically, fix the precoding matrix of all devices except device k and optimize V k , repeat the optimization process until convergence, and the optimal precoding matrix V is obtained.

[0113] According to the law tr(B H C)=vec(B) H vec(C) and Where b represents the number of columns of matrix B, the constraint in formula (VII) is transformed into:

[0114]

[0115] In formula (IX) vec represents vectorized operation. represents the Kronecker product, T represents the matrix transpose;

[0116] Calculation formula (VII) The objective function is about V k The gradient is:

[0117]

[0118] In formula (X), and Represent Σ and Σ respectively j The k-th column matrix block is specifically Σ=[Σ (1) ,…,Σ (K) ];

[0119] When the transmission power of each device meets When Ψ(v k ), so The projection is:

[0120]

[0121] In formula (XI),

[0122] By formula (XII) k To update:

[0123]

[0124] In formula (XII),

[0125] By finding a θ∈[0,π / 2) that satisfies θ=argmaxΨ(V k), based on formula (XII) the optimal v k and V k ;

[0126] By repeating the above optimization process for each device, the optimal precoding matrix V with a fixed RIS phase shift vector φ can be obtained.

[0127] The RGP algorithm is proposed to solve the RIS optimization problem, namely, formula (VIII); in formula (VIII), only the RIS phase shift vector φ is the optimization variable. The constraints in formula (VIII) are converted into convex constraints:

[0128] tr(φφ H )=N and|φ ∞ ≤1(XIII);

[0129] In formula (XIII), tr represents the trace operation, and the sum of the diagonal elements of the matrix is ​​calculated, ||.|| ∞ represents the infinity norm, and represents the AND operator.

[0130] use Approximate infinite norm l ∞ ; Using the barrier method, using the logarithmic barrier function Integrate the non-negative constraints to approximate the penalty for violating the constraints, where t is a constant used to adjust the penalty;

[0131] The objective function in formula (VIII) is rewritten as:

[0132] max φ Υ(φ,h)=Ψ(φ)+I(1-‖φ‖ h )(XIV)

[0133] Calculate the gradient of the objective function of formula (XIV) with respect to φ:

[0134]

[0135] In formula (XV), at the same time:

[0136]

[0137] Project the search direction into formula (XVII):

[0138]

[0139] Among them, <,> represents the inner product, by finding a satisfy And update φ using formula (XVIII):

[0140]

[0141] Project φ using formula (XIX) to satisfy the constraints:

[0142]

[0143] The above optimization process is repeated until convergence, and the optimal RIS phase shift vector φ that meets the constraints under the condition of a given precoding matrix V is obtained.

[0144] An alternating iterative algorithm is used to solve the optimization problem of a multi-device collaborative edge inference system; including:

[0145] Initialize the precoding matrix V to satisfy Initialize the RIS phase shift vector φ to satisfy |φ i |=1;

[0146] Calculated according to formula (X) Calculated according to formula (XI) And according to Find a To update V according to formula (XII) k and v k , repeat the above steps to the precoding matrix of K devices;

[0147] Calculated according to formula (XV) Project it according to formula (XVII), and according to Find a Update φ by equation (XVIII) and project φ according to equation (XIX) to satisfy the constraint;

[0148] Repeat the above steps until convergence; finally, the maximum MCR is obtained under the premise of meeting constraints (II) and (IV). 2 The precoding matrix V and RIS phase shift vector φ of the value.

[0149] FIG2(a) is a schematic diagram showing a comparison of the classification accuracy of the present invention and the traditional communication based on data recovery (maximum a posteriori probability classifier, MAP) at different receiving ends; FIG2(b) is a schematic diagram showing a comparison of the classification accuracy of the present invention and the traditional communication based on data recovery (K nearest neighbor classifier, KNN) at different receiving ends; FIG2(c) is a schematic diagram showing a comparison of the classification accuracy of the present invention and the traditional communication based on data recovery (neural network classifier, NN) at different receiving ends; Figure 3 This is a comparison chart of the classification accuracy of the present invention and the general RIS design method (ACM) when NN is used as the receiving end and SNR = -20dB; Figure 4 is the number of RIS reflection units N and edge server receiving antennas N in the method of the present invention rThe relationship diagram with classification accuracy. Under the same range of SNR (signal-to-noise ratio), the classification effect of the present invention on the ModelNet10 dataset has achieved higher classification accuracy than the traditional data recovery-based method (MMSE). In addition, compared with the general RIS design method ACM (maximum achievable rate), this method can achieve similar or even better performance than ACM with fewer RIS units.

Claims

1. A communication method for remote reasoning tasks, characterized in that: The method is applied to a multi-device collaborative edge inference system; the multi-device collaborative edge inference system includes multiple integrated devices equipped with cameras and multiple antennas, a feature encoder, a channel precoder, an edge server equipped with multiple antennas, and a classifier; including: The integrated device collects pictures of objects from multiple perspectives, and extracts the feature information of the pictures through the feature encoder; then uses the channel precoder to convert it into a data format that conforms to wireless channel transmission; the transmitted signal is reflected by the RIS reflection unit; the transmission distance is increased while the communication performance is increased; the edge server receives signals from multiple devices and fuses them into multi-perspective features of the object; the classifier is used to calculate the classification result; The design of unified feature encoder, channel precoder and RIS reflection unit is expressed as: maximizing the MCR of the signal received by the edge server 2 , while satisfying the power constraints of multiple edge devices and the amplitude constraints of the RIS reflection unit, the optimization problem is established as formula (VI): In formula (VI), f(·,ψ) represents a MCR 2 The neural network pre-trained for the objective function, ψ represents the learnable parameters of the neural network; is the phase shift vector of RIS; in, ε is the lossy decoding rate; represents the transmission channel matrix; and Represent the covariance matrix of the eigenvector and the covariance matrix of the eigenvector of category j, p j is the prior probability of the feature vector received by the edge server and related to category j; The multivariable optimization problem in formula (VI) is transformed into an alternating optimization problem with single variable constraints, specifically including: 1) Fix the RIS phase shift vector φ and transform equation (VI) into equation (VII): 2) Fix the channel precoding matrix V of K devices and transform equation (VI) into equation (VIII): The ABGP algorithm is proposed to solve the precoding matrix problem, namely, formula (VII); including: Convert equation (VII) to alternately solve the precoding matrix V of K devices k Specifically, fix the precoding matrix of all devices except device k and optimize V k , repeat the optimization process until convergence, and obtain the optimal precoding matrix V; According to the law tr(B H C)=vec(B) H vec(C) and Where b represents the number of columns of matrix B, the constraint in formula (VII) is transformed into: In formula (IX) vec represents vectorized operation. represents the Kronecker product, T represents the matrix transpose; Calculation formula (VII) The objective function is about V k The gradient is: In formula (X), and Represent Σ and Σ respectively j The k-th column matrix block is specifically Σ=[Σ (1) ,…,Σ (K) ]; When the transmission power of each device meets When Ψ(v k ), so The projection is: In formula (XI), By formula (XII) k To update: In formula (XII), By finding a θ∈[0,π / 2) that satisfies θ=argmaxΨ(V k ), based on formula (XII) the optimal v k and V k ; Repeating the above optimization process for each device, the optimal precoding matrix V for a fixed RIS phase shift vector φ can be obtained; The RGP algorithm is proposed to solve the RIS optimization problem, namely formula (VIII); the constraints in formula (VIII) are converted into convex constraints: tr(φφ H )=N and‖φ ∞ ≤1(XIII); In formula (XIII), tr represents the trace operation, and the sum of the diagonal elements of the matrix is ​​calculated, ||.|| ∞ represents the infinite norm, and represents the AND operator; use Approximate infinite norm l ∞ ; Using the barrier method, using the logarithmic barrier function Integrate the non-negative constraints to approximate the penalty for violating the constraints, where t is a constant used to adjust the penalty; The objective function in formula (VIII) is rewritten as: max φ Y(φ,h)=Ψ(φ)+I(1-‖φ‖ h ) (XIV) Calculate the gradient of the objective function of formula (XIV) with respect to φ: In formula (XV), at the same time: Project the search direction into formula (XVII): Among them, <,> represents the inner product, by finding a satisfy And update φ using formula (XVIII): Project φ using formula (XIX) to satisfy the constraints: Repeat the above optimization process until convergence, and obtain the optimal RIS phase shift vector φ that meets the constraints under the condition of a given precoding matrix V; An alternating iterative algorithm is used to solve the optimization problem of a multi-device collaborative edge inference system; including: Initialize the precoding matrix V to satisfy Initialize the RIS phase shift vector φ to satisfy |φi|=1; Calculated according to formula (X) Calculated according to formula (XI) And according to Find a To update V according to formula (XII) k and v k , repeat the above steps to the precoding matrix of K devices; Calculated according to formula (XV) Project it according to formula (XVII), and according to Find a Update φ by equation (XVIII) and project φ according to equation (XIX) to satisfy the constraint; Repeat the above steps until convergence; finally, the maximum MCR is obtained under the premise of meeting constraints (II) and (IV). 2 The precoding matrix V and RIS phase shift vector φ of the value; Adjust the feature vector, i.e., the feature information of the image, to a plural form; The communication channel is established as a Rayleigh channel, and the established communication signal transmission model is expressed as formula (III): In formula (III), y represents the edge server fusing the received reflection signal into the multi-view feature vector of the target; represents the transmission channel matrix of device k, where represents the linear channel matrix between device k and the edge server, represents the reflection channel matrix between the RIS reflection unit and the edge server about device k, represents the transmission channel matrix between device k and the RIS reflection unit; Represented by RIS phase shift vector is the matrix of the diagonal, And θ i ∈(0,2π] represents the phase shift caused by the i-th RIS reflection unit; N r is the number of receiving antennas of the edge server; x k represents the transmission signal of device k, represents the feature vector extracted by the feature encoder of device k; V k represents the channel precoding matrix of device k; represents the Gaussian noise vector, where δ represents the noise power and I represents the identity matrix; The RIS phase shift vector satisfies the constraint condition of formula (IV): In formula (IV), N is the dimension of the RIS phase shift vector φ, and i=1, 2, ..., N represents the i-th element.

2. A communication method for remote reasoning tasks according to claim 1, characterized in that: The feature vector is the feature information of the image. Adjust to plural form As shown in formula (I): In formula (I), represents the imaginary unit, assuming D k =2F k ; is a vector, D k refers to The dimension, F k refers to Half the dimension of k can be divided by integers, and Refers to the notation of the fields of complex and real numbers.

3. A communication method for remote reasoning tasks according to claim 1, characterized in that: Defining devices Channel precoder The output signal is N k is the number of transmit antennas of device k, x k Satisfy the transmit power constraint shown in formula (II): In formula (II), E(·) represents the mean operator, ‖·‖2 represents the two-norm operator, express The covariance matrix, P k represents the maximum transmission power of device k, and the superscript H represents the conjugate transpose.

4. A communication method for remote reasoning tasks according to claim 1, characterized in that: Splice the channel transmission matrices of K devices: in, Indicates the total number of transmitting antennas of the transmitting device; definition in, is the dimension after concatenating the feature vectors of K devices, rewriting formula (III) as: The fused multi-view feature vector y is further used as the input of the classifier, and the inference label is obtained after calculation Complete remote reasoning tasks.

Images