Underlay mode D2D cooperative multi-hop relay communication power control method based on deep neural network

Abstract

The invention discloses an Underlay mode D2D cooperative multi-hop relay communication power control method based on a deep neural network, and the method achieves the control purpose of maximizing the energy efficiency of an arbitrary-hop D2D cooperative multi-hop relay communication system, employs the deep neural network, builds a universal solution, can quickly solve an approximately optimal power control scheme, and improves the power control efficiency of the D2D cooperative multi-hop relay communication system. And a traditional high-complexity iterative solution mode for coping with a non-convex optimization problem is effectively avoided.

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a high-energy-efficiency Underlay mode D2D collaborative multi-hop relay communication power control method based on a deep neural network. Background Art

[0002] D2D (device-to-device) communication refers to a technology for direct communication between two peer user nodes, which does not rely on network infrastructure, such as base stations (BS) or access points (AP). The 3GPP organization has been discussing this technology since 2013. This technology has been developed since 4G and is now widely used in 5G communication systems. D2D communication reuse in cellular networks comes from the authorized frequency bands of cellular users. There are two mainstream reuse methods, namely Overlay mode and Underlay mode. The former opportunistically accesses idle cellular authorized frequency bands, and the latter can access frequency bands being used by cellular users. Although the latter will cause interference to cellular users, it can also improve the spectrum utilization of the authorized frequency bands, so it has attracted extensive research.

[0003] In the D2D communication in the Underlay mode, the channel condition between the D2D transmitter and the D2D receiver is sometimes not good. In this case, other D2D terminals can be used as relays to use cooperative relay technology to ensure the performance between the D2D transmitter and the D2D receiver. Since most mobile terminals are limited by battery capacity, energy efficiency has become an important indicator for measuring the performance of mobile communication systems and achieving green communications.

[0004] In the prior art, based on matching theory, the relay selection and wireless resource allocation methods are studied to maximize the energy efficiency of the D2D communication system. For example, for two hops, that is, when a D2D terminal is used as a relay, the transmission power of D2D s (D2D source) and the forwarding power of D2D r (D2D relay) are optimized. Figure 1 shown.

[0005] Defects and shortcomings of the existing technology:

[0006] 1. D2D cooperative multi-hop relay communication in cellular networks with the goal of maximizing energy efficiency. For example, for the case of only two hops mentioned above, the power control problem is generally non-convex, and the optimal solution is difficult to express analytically. Generally, only suboptimal solutions can be obtained, and the suboptimal solutions need to be solved iteratively, which is relatively complex.

[0007] 2. The solution to the power control problem of D2D relay communication with more hops requires clever design and complex derivation, and there is no universal solution.

[0008] In recent years, deep learning technology has been widely used in various research fields of wireless communications, including wireless resource allocation. It is based on deep neural networks and can theoretically approximate any function. However, most resource allocation optimization problems are nonlinear or even non-convex. Using deep neural networks to solve these problems can bypass the complex derivation process. At the same time, thanks to the development of GPU technology and the increase in computing power in recent years, the convergence speed of deep neural networks has been greatly improved, which can cope with the real-time allocation of wireless resources. Summary of the invention

[0009] The present invention provides a power control method for high energy efficiency D2D cooperative multi-hop relay communication based on deep neural network, which maximizes the power of D2D s and D2D r. i The energy efficiency of the system composed of (i=1, ..., N-1) is the goal, which can effectively solve the D2D s and D2D r of the underlay mode D2D cooperative multi-hop relay communication. i Optimal transmit power control problem.

[0010] The D2D cooperative multi-hop relay communication power control method based on deep neural network provided in the present disclosure mainly includes the following steps:

[0011] S1, construct the objective function and constraint model of D2D cooperative multi-hop relay communication power control, where the control goal is to maximize the energy efficiency of the system;

[0012] S2, uses a deep neural network with normalized channel gain and transmit power as input to find the optimal solution.

[0013] Furthermore, the step S1 specifically includes:

[0014] Assume that the system includes a source D2D s and a relay D2D r i (i=1, ..., N-1), D2D s shares the authorized spectrum with the reused cell user CUE1 (cellular user equipment), D2D r i Shares licensed spectrum with CUEi+1; D2D r0 and D2D r N They represent the source D2D s and the sink D2D d respectively, N is the total number of system hops, and BS (base station) is the base station;

[0015] Then the objective function and constraint model of power control are:

[0016]

[0017]

[0018] in:

[0019] η is the system energy efficiency,

[0020] D2D r i-1 , the transmit power of CUE i;

[0021] Q d2d The performance of CUE i and D2D cooperative multi-hop relay communication system is the QoS (quality of service) requirement of achievable spectrum efficiency;

[0022] They are CUE i and BS, D2D r i-1 Channel gain with BS;

[0023] γ N D2D N , that is, the equivalent SINR at D2D d.

[0024] Furthermore, in step S1, γ N The specific calculation methods include:

[0025] The equivalent SINR from D2D r0 to D2D r2 via D2D r1, that is, at D2D r2, is:

[0026]

[0027] in, They are the SINR at D2D r1 and D2D r2 respectively; is the transmit power of D2D r0, D2D r1, CUE 1, and CUE 2; are the channel gains between D2D r0 and D2D r1, D2D r1 and D2D r2, CUE 1 and D2D r1, and CUE 2 and D2D r2 respectively;

[0028] Based on the above formula, the equivalent SINR at D2D r3 is:

[0029]

[0030] in, is the SINR at D2D r3; is the transmit power of D2D r2 and CUE 3; They are the channel gains between D2D r2 and D2D r3, and between CUE 3 and D2D r3 respectively;

[0031] Combining the above two equations, we get D2D r i Equivalent SINR at (i=1,...,N):

[0032]

[0033] in, For D2D i SINR at For D2D i-1 , the transmit power of CUE i; D2D r i-1 With D2D i , CUE i and D2D r i The channel gain between .

[0034] Furthermore, in step S2, the specific method for obtaining the optimal solution includes:

[0035] S21, channel gain Perform normalization;

[0036] S22, for the transmission power pr i-1 Perform normalization;

[0037] S23, normalize the channel gain and transmit power As input to deep neural networks;

[0038] S24, after training, find the minimum loss function That is the optimal power control scheme.

[0039] Furthermore, the normalization method of step S21 includes:

[0040] make

[0041] but |A| represents the number of elements in set A;

[0042] The normalized channel gain is

[0043]

[0044] in:

[0045] It means the expectation of finding a.

[0046] Furthermore, the normalization method of step S22 includes:

[0047]

[0048] Furthermore, in step S23, the deep neural network used includes:

[0049] P fully connected hidden layers in series, each layer contains Q neurons; the activation function of these P hidden layers is the ReLU function;

[0050] A softmax layer is used at the output layer of the network;

[0051] The result obtained after processing by the network is multiplied by P max That is, the transmission power {p ri-1 |i=1,...,N}.

[0052] Furthermore, in step S24, the loss function used is:

[0053]

[0054] Among them, λ i (i=1,...,N) and γ are positive control parameters; [a] + =max(a,0).

[0055] Compared with the prior art, the present invention has the following beneficial effects: (1) based on deep neural networks, by utilizing the property that it can approximate any function, a general (for any N hops) solution is provided for D2D cooperative multi-hop relay communication systems, and a nearly optimal power control solution can be quickly obtained, avoiding the traditional highly complex iterative solution method for non-convex optimization problems;

[0056] (2) By setting the energy efficiency objective function, the ratio of data transmission to energy consumption in the underlay mode D2D cooperative multi-hop relay communication system can be significantly improved, which is suitable for energy-constrained systems;

[0057] (3) By setting the control parameter λ i (i=1, ..., N) and γ, considering the QoS requirements of CUE i and D2D cooperative multi-hop relay communication system, a good balance is achieved between algorithm effectiveness, complexity and reliability;

[0058] (4) It has good versatility. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The above and other objects, features and advantages of the present disclosure will become more apparent through a more detailed description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, wherein like reference numerals generally represent like components throughout the exemplary embodiments of the present disclosure.

[0060] Figure 1 for two-hop D2D relay communication in cellular networks;

[0061] Figure 2 Schematic diagram of D2D cooperative multi-hop relay communication in N-hop underlay mode in cellular network;

[0062] Figure 3 2 is a radio frame structure in an exemplary embodiment. DETAILED DESCRIPTION

[0063] The preferred embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the preferred embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0064] The present disclosure provides a power control method for energy-efficient D2D cooperative multi-hop relay communication based on deep neural network, the purpose of which is to maximize the power consumption of D2D s and D2D r. i Energy efficiency of the system composed of (i=1,...,N-1).

[0065] In an exemplary embodiment, the system structure is as shown in the attached Figure 2 As shown in the figure, D2D s shares the authorized spectrum (represented by a certain number of RBs) with CUE1, and D2D r i Share the licensed spectrum with CUEi+1. For the convenience of description, D2D r0 and D2D r N Represents D2D s and D2D d.

[0066] According to the present disclosure, the power control method for the system mainly includes the following steps:

[0067] 1. Construction of power control model

[0068] According to the above structure of this system, we can get:

[0069] The equivalent SINR from D2D r0 to D2D r2 via D2D r1, that is, at D2D r2, is:

[0070]

[0071] in, They are the SINRs at D2D r1 and D2D r2 respectively. is the transmit power of D2D r0, D2D r1, CUE 1, and CUE 2. They are the channel gains between D2D r0 and D2D r1, D2D r1 and D2D r2, CUE 1 and D2D r1, and CUE 2 and D2D r2 respectively.

[0072] Based on (8), the equivalent SINR at D2D r3 is:

[0073]

[0074] in is the SINR at D2D r3. is the transmit power of D2D r2 and CUE 3. They are the channel gains between D2D r2 and D2D r3, and between CUE 3 and D2D r3 respectively.

[0075] Combining (8) and (9), we can get D2D r i The equivalent SINR at:

[0076]

[0077] in For D2D i SINR at the location. For D2D i-1 , the transmit power of CUE i. D2D r i-1 With D2D i , CUE i and D2D r i The channel gain between .

[0078] Therefore, the power control problem of maximizing D2D cooperative multi-hop relay communication can be expressed as follows:

[0079]

[0080]

[0081] in Qd2d are the QoS requirements of CUE i and D2D cooperative multi-hop relay communication system respectively (measured by achievable spectrum efficiency, SE). They are CUE i and BS, D2D r i-1 The channel gain between the BS and the GNSS. N It is obtained by iterative calculation of (8) and (10).

[0082] 2. Solution based on deep neural network

[0083] Obviously, problem (11) is a non-convex optimization problem. The solution of this problem has the aforementioned defects, there is no unified solution, and it is difficult to find the analytical optimal solution in a low computing time. The power control strategy based on deep neural network can provide a near-optimal solution for (11) with a low computing time.

[0084] For this system, power control needs to know the channel gain between the transmitter and the receiver. Therefore, pilot transmission and channel state information (CSI) feedback are required in the wireless frame. When the BS (or other network center control node) obtains the channel gain through CSI feedback, it performs power control and starts data transmission. The wireless frame structure diagram of the system is shown in the figure below. Figure 3 shown.

[0085] The specific solution process includes:

[0086] (1) Channel gain Normalize.

[0087] make

[0088] but |A| represents the number of elements in set A.

[0089] The normalized channel gain is

[0090]

[0091] in

[0092]

[0093]

[0094] It means the expectation of finding a.

[0095] (2) Transmitting power Normalize it and get:

[0096]

[0097] (3) Normalize the channel gain and transmit power As input to the following deep neural network.

[0098] The deep neural network consists of P fully connected hidden layers in series, each containing Q neurons. The activation function of the P hidden layers is the ReLU function. A softmax layer is used as the output layer of the network. The result obtained after processing by the network is multiplied by P max That is, the transmission power

[0099] (3) After (unsupervised learning) training, the loss function can be minimized That is the optimal power control solution.

[0100] The loss function is

[0101]

[0102] where λ i (i=1,...,N) and γ are positive control parameters. [a] + =max(a,0).

[0103] Therefore, this embodiment uses a deep neural network and its ability to approximate any function to provide a general solution for any N-hop D2D cooperative multi-hop relay communication system, which can quickly obtain an approximately optimal power control solution and avoid the traditional high-complexity iterative solution method for non-convex optimization problems.

[0104] The above technical scheme is only an exemplary embodiment of the present invention. For those skilled in the art, it is easy to make various types of improvements or modifications based on the application methods and principles disclosed in the present invention, and it is not limited to the method described in the above specific embodiment of the present invention. Therefore, the method described above is only preferred and does not have a restrictive meaning.

Claims

1. A power control method for D2D cooperative multi-hop relay communication in Underlay mode based on deep neural network, comprising the following steps: S1, construct the objective function and constraint model of D2D cooperative multi-hop relay communication power control, where the control goal is to maximize the energy efficiency of the system; S2, uses a deep neural network to obtain the optimal solution with normalized channel gain and transmit power as input.

2. The method according to claim 1, characterized in that The step S1 specifically includes: Assume that the system includes source D2Ds and relay D2Dr i (i=1,…,N-1), D2Ds shares the authorized spectrum with the reused cell user CUE1, D2Dr i Shares licensed spectrum with CUEi+1; D2Dr0 and D2Dr N They represent the source D2Ds and the sink D2Dd respectively, N is the total number of system hops, and BS is the base station; Then the objective function and constraint model of power control are: in: η is the system energy efficiency, D2Dr i-1 , the transmit power of CUEi; Q d2d The performance of CUEi and D2D cooperative multi-hop relay communication system is the QoS (quality of service) requirement of achievable spectrum efficiency; They are CUEi and BS, D2Dr i-1 Channel gain with BS; γ N D2Dr N , that is, the equivalent SINR at D2Dd.

3. The method according to claim 2, characterized in that In step S1, γ N The specific calculation methods include: The equivalent SINR from D2Dr0 to D2Dr2 via D2Dr1, that is, at D2Dr2, is: in, are the SINRs at D2Dr1 and D2Dr2 respectively; is the transmit power of D2Dr0, D2Dr1, CUE 1, and CUE 2; They are the channel gains between D2Dr0 and D2Dr1, D2Dr1 and D2Dr2, CUE 1 and D2Dr1, and CUE 2 and D2Dr2 respectively; Based on the above formula, the equivalent SINR at D2Dr3 is: in, is the SINR at D2Dr3; is the transmission power of D2Dr2 and CUE 3; They are the channel gains between D2Dr2 and D2Dr3, and between CUE 3 and D2Dr3 respectively; Combining the above two equations, we get D2Dr i Equivalent SINR at (i=1,…,N): in, For D2Dr i SINR at For D2Dr i-1 , the transmit power of CUEi; D2Dr i-1 With D2Dr i , CUEi and D2Dr i The channel gain between .

4. The method according to claim 2 or 3, characterized in that: In step S2, the specific method for obtaining the optimal solution includes: S21, channel gain Perform normalization; S22, for transmit power Perform normalization; S23, normalize the channel gain and transmit power As input to deep neural networks; S24, after training, find the minimum loss function That is the optimal power control scheme.

5. The method according to claim 4, characterized in that The normalization method of step S21 includes: make but |A| represents the number of elements in set A; The normalized channel gain is in: It means the expectation of finding a.

6. The method according to claim 4, characterized in that The normalization method of step S22 includes:

7. The method according to claim 4, characterized in that In step S23, the deep neural network used includes: P fully connected hidden layers in series, each layer contains Q neurons; the activation function of these P hidden layers is the ReLU function; A softmax layer is used at the output layer of the network; The result obtained after processing by the network is multiplied by P max That is, the transmission power 8. The method according to claim 4, characterized in that In step S24, the loss function used is: Among them, λ i (i=1,…,N) and γ are positive control parameters; [a] + =max(a,0).

Images