KR20240028019A - Resource Allocation for Transmit Power Minimization in IoT Cellular Networks - Google Patents

Abstract

According to one aspect of the present invention, in a resource allocation method for minimizing transmission power by controlling transmission power in a non-orthogonal multiple access (NOMA) IoT cellular network system,
The system is,
base station and
An IoT device connected to the base station through uplink communication;
It includes a control device that controls transmission power to IoT devices connected to the base station through communication,
The control device is
(a) controlling all IoT devices connected to the base station to minimize the sum of the transmission power of each IoT device while keeping the transmission rate of each IoT device above the threshold transmission rate; and
(B) For all IoT devices connected to the base station, controlling the transmission rate of each IoT device to be higher than the threshold transmission rate and minimizing the maximum transmission power among the IoT devices.

Description

Resource allocation method for transmit power minimization in IoT cellular networks {Resource Allocation for Transmit Power Minimization in IoT Cellular Networks}

The present invention relates to a resource allocation method for minimizing transmission power in an IoT cellular network.

The Internet of Things (IoT) is receiving great attention in the 5th generation (5G) and future 6th generation (6G) communication systems. Specifically, IoT-based communication technology can be applied to various industrial fields that require large-scale connectivity, such as smart home, smart grid, and smart transportation, and can dramatically improve existing services. Additionally, IoT technology is attracting attention in areas such as medical care and disaster management that can improve quality of life and save lives. As demand for IoT-based services in various fields increases significantly, the number of connected IoT devices per person is expected to reach approximately 20 by 2030.

Therefore, it is important to develop a communication framework that can fairly and efficiently support multiple IoT devices with limited resources.

Unlike OMA (Orthogonal Multiple Access), which requires multiple users to use different frequency resources, Non-orthogonal Multiple Access (NOMA) allows multiple users to simultaneously occupy the same frequency resource. Resources can be used more efficiently. It is a multiple access method that allows allocation of non-orthogonal resources by sharing frequency, time, and code resources for each user terminal, while allowing more terminals to be accommodated with designated resources. This refers to a technology that improves spectral efficiency by simultaneously transmitting data from two or more terminals sharing the same resources (time slot, subcarrier, spreading code, etc.). The purpose is to increase resource efficiency by simultaneously transmitting data for two or more terminals on the same time, frequency, and space resources. NOMA is a technology used to increase the capacity of 5G cells. It breaks the orthogonality property of the existing orthogonal multiple access (OFDMA) method in terms of frequency resource allocation and simultaneously overlaps data from two or more terminals on the same frequency resource. Resource efficiency can be improved through allocation. For example, when a base station transmits overlapping data signals to a terminal in the cell center area and a terminal in the cell border area with a large difference in channel quality, small power is allocated to the cell center area terminal and high power is allocated to the cell border area terminal. do. Terminals within a cell first decode terminals in the central area, which are cells with high signal strength, and then decode cell edge terminals, according to the Sequential Interference Cancellation (SIC) method. A terminal in the center of a cell can decode its own signal, considering it as interference because the signal from the terminal at the cell border is relatively weak. Since the terminal at the center of the decoded cell is removed from the received signal before decoding the terminal at the cell border, the signal from the terminal at the cell border can be successfully decoded.

Because of these advantages, NOMA is considered one of the promising technologies to support large-scale connectivity in IoT networks.

In prior research (see Non-Patent Document 1), many studies have been conducted on resource allocation considering the quality of service (QoS) requirements of each user in a NOMA-supported IoT network.

Additionally, in a prior study (see Non-Patent Document 2), the authors proposed a low-complexity joint radio and computational resource allocation method that maximizes energy efficiency in the NOMA mobile edge computing system with intra- and inter-cell interference.

Additionally, in prior research (see Non-Patent Document 3), joint optimization of user scheduling and power allocation was proposed to minimize total power consumption while meeting the long-term data rate requirements of each user in a downlink NOMA IoT network.

Additionally, in prior research (see Non-Patent Document 4), the author studied the power allocation problem using SQP (sequential quadratic programming) method to maximize spectral efficiency in the NOMA IoT network.

Additionally, in prior research (see Non-Patent Document 5), a new algorithm was proposed to jointly optimize power allocation, energy harvesting, and time scheduling to improve the performance of unmanned aerial vehicle (UAV) communication in the NOMA IoT network.

The target data rate for each user required for IoT-based services is typically much lower than applications that require high data rates, such as real-time high-definition (HD) video streaming and virtual reality. Additionally, Iota devices typically have no external power source, so their operating life is limited by their internal battery capacity. Therefore, in some cases, it is important to minimize energy consumption to extend the operating life of battery-powered IoT devices.

In prior research (see Non-Patent Document 6), joint optimization of user clustering and power allocation was studied to minimize the total transmission power of a multi-cell, multi-carrier NOMA network.

Additionally, previous research (see Non-Patent Document 7) considered the overall power minimization problem for a multi-cell uplink NOMA network under the assumption of imperfect sequential interference cancellation (SIC). In this prior study, an appropriate low-complexity iterative optimization algorithm was proposed using convex relaxation to solve the considered resource allocation problem.

In a previous study, a low-complexity transmission power allocation technique was proposed to allocate power according to the channel gain of each user in a downlink NOMA network.

Recently, deep learning (DL) has been shown to show excellent performance in solving various types of complex resource allocation problems considered in communication systems. In particular, it was suggested that DL-based algorithms can be successfully applied to various communication scenarios where real-time processing is important because they can infer near-optimal solutions with simple linear operations using a pre-trained deep neural network (DNN) model (B (see patent document 8).

Additionally, in prior research (see Non-Patent Document 9), the authors proposed a DL-based method that maximizes the achievable data rate by solving the joint optimization problem of subband allocation and power control in an uplink IoT cellular network where spectrum leakage exists between subbands. An approach has been proposed. In the preceding study above, the performance of the proposed method according to the type of DNN structure was also examined by comparing the performance of the case of applying the Fully Connected Neural Network (FNN) structure and the case of applying the Convolutional Neural Network (CNN).

Additionally, in prior research (see Non-Patent Document 10), a recurrent neural network (RNN)-based resource allocation technique was used to solve joint optimization of subchannel allocation and power control to maximize energy efficiency of NOMA-based heterogeneous IoT networks in imperfect environments. It has been suggested.

Additionally, in prior research (see Non-Patent Document 11), joint optimization of user association, sub-channel allocation, and power control was studied based on semi-supervised learning to maximize the energy efficiency of NOMA-based mmWave heterogeneous networks.

Genetic Algorithm (GA) can be used as a method to find a solution to an optimization problem. Genetic algorithm is an optimization technique first introduced by John Holland in his 1975 book "Adaptation on Natural and Artificial Systems." This genetic algorithm is a method of evolutionary computation that solves optimization problems by imitating actual biological evolution.

Genetic algorithms list possible solutions to a problem and then gradually change genes to create highly accurate and appropriate solutions. In genetic algorithms, the solution to a problem is called a chromosome, and a chromosome is composed of genes. Therefore, evolution can be seen as obtaining a good solution by modifying the genes in the chromosome. In other words, it can be said to be a search algorithm that imitates evolution to find a better answer.

It was the most widely used algorithm before the advent of NN (Neural Network), and even after the advent of artificial neural networks, it still plays an important role, such as being used to set initial values in deep learning. Research on finding solutions to network optimization problems using genetic algorithms has also been presented in previous studies.

In prior studies (Non-Patent Documents 13 and 14), a method of finding a solution by applying a genetic algorithm was proposed to reduce the complexity of calculations for optimization in a non-orthogonal multiple access (NOMA) system.

To support multiple IoT devices with limited frequency resources, research on Non-orthogonal multiple access (NOMA), which allows different users to use the same frequency resources, has been actively conducted.

Because IoT devices typically do not have an external power source, their operating life is limited by internal battery capacity. In addition, in most cases in IoT communication, it is analyzed that it is more important to minimize power consumption under the prerequisite of achieving a minimum transmission rate, and consequently increase the operating life of IoT devices, than to achieve a high transmission rate. Based on this, a specific implementation method for an energy-efficient information transmission technique that minimizes power consumption while satisfying the minimum transmission rate required by the system for all IoT devices is required.

Republic of Korea Patent Publication No. 10-2196546 (drx and scheduling optimization method to reduce terminal power consumption in IoT environment)

[1] L. Dai, B. Wang, Z. Ding, Z. Wang, S. Chen, and L. Hanzo, “A survey of non-orthogonal multiple access for 5G,IEEE Commun. Surveys Tuts., vol. 20, no. 3, pp. 2294-2323, 3rd Quart., 2018. [2] B. Liu, C. Liu, and M. Peng, “Resource allocation for energy-efficient MEC in NOMA-enabled massive IoT networks,” IEEE J. Sel. Areas Commun., vol. 39, no. 4, pp. 1015-1027, Apr. 2020. [3] D. Zhai, R. Zhang, L. Cai, B. Li, and Y. Jiang, “Energy-efficient user scheduling and power allocation for NOMA-based wireless networks with massive IoT devices,” IEEE Internet Things J. , vol. 5, no. 3, pp. 1857-1868, Jun. 2018. [4] W. U. Khan, J. Liu, F. Jameel, V. Sharma, R. Jantti, and Z. Han, ¨“Spectral efficiency optimization for next generation NOMA-enabled IoT networks,” IEEE Trans. Veh. Technol., vol. 69, no. 12, pp. 15 284-15 297, Dec. 2020. [5] H. Li, J. Li, M. Liu, Z. Ding, and F. Gong, “Energy harvesting and resource allocation for cache-enabled UAV based IoT NOMA networks,” IEEE Trans. Veh. Technol, vol. 70, no. 9, pp. 9625-9630, Sep. 2021. [6] D. Ni, L. Hao, Q. T. Tran, and X. Qian, “Transmit power minimization for downlink multi-cell multi-carrier NOMA networks,” IEEE Comm. Lett., vol. 22, no. 12, pp. 2459-2462, Dec. 2018. [7] M. Zeng, W. Hao, O. A. Dobre, Z. Ding, and H. V. Poor, “Power minimization for multi-cell uplink NOMA with imperfect SIC,” IEEE Comm. Lett., vol. 22, no. 12, pp. 2459-2034, Dec. 2020. [8] H. Huang, S. Guo, G. Gui, Z. Yang, J. Zhang, H. Sari, and F. Adachi, “Deep learning for physical-layer 5G wireless techniques: Opportunities, challenges and solutions ,”IEEE Wireless Commun., vol. 27, no. 1, pp. 214-222, Feb. 2020. [9] H. W. Kim, H. J. Park, and S. H. Chae, “Sub-band assignment and power control for IoT cellular networks via deep learning,” IEEE Access, 2022. [10] M. Liu, T. Song, and G. Gui, “Deep cognitive perspective: Resource allocation for NOMA-based heterogeneous IoT with imperfect SIC,” IEEE Internet Things J., vol. 6, no. 2, pp. 2885-2894, Apr. 2019. [11] H. Zhang, H. Zhang, K. Long, and G. K. Karagiannidis, “Deep learning based radio resource management in NOMA networks: User association, subchannel and power allocation,” IEEE Trans. Netw. Sci. Eng., vol. 7, no. 4, pp. 2406-2415, Jun. 2020. [12] A. Benjebbovu, A. Li, Y. Saito, Y. Kishiyama, A. Harada, and T. Nakamura, “System-level performance of downlink NOMA for future LTE enhancements,” in Proc. IEEE Globecom Workshops (GC Wkshps), Dec. 2013, pp. 66-70. [13] J. Luo, J. Tang, D. K. So, G. Chen, K. Cumanan, and J. A. Chambers, “A deep learning-based approach to power minimization in multi-carrier NOMA with SWIPT,” IEEE Access, vol. 7, pp. 17, 450-17 460, 2019. [14] Hanliang You, Zhiwen Pan, Nan Liu, Xiaohu You, “Clustering Scheme for Downlink Hybrid NOMA Systems Based on Genetic Algorithm.” IEEE Access, vol. 8, pp. 129461-129468, 2020.

The purpose of the present invention is to provide a resource allocation method for IoT devices to minimize transmission power while satisfying the minimum transmission rate of an IoT cellular network in an IoT system.

The present invention is not limited to the purposes mentioned above, and other purposes not mentioned can be clearly understood from the description below.

According to one aspect of the present invention, in a resource allocation method for minimizing transmission power by controlling transmission power in a non-orthogonal multiple access (NOMA) IoT cellular network system,

The system is,

base station and

An IoT device connected to the base station through uplink communication;

It includes a control device that controls transmission power for IoT devices connected to the base station through communication,

The control device is

(a) controlling all IoT devices connected to the base station to minimize the sum of the transmission power of each IoT device while keeping the transmission rate of each IoT device above the threshold transmission rate; and

(b) For all IoT devices connected to the base station, controlling the transmission rate of each IoT device to be higher than the threshold transmission rate and minimizing the maximum transmission power among the IoT devices.

Additionally, step (a) is

(a-1) calculating a sub-band allocation that maximizes the minimum transmission rate among the IoT devices in order to ensure that all IoT devices satisfy the threshold transmission rate condition, - where the transmission power of all IoT devices is the maximum transmission power Characterized by setting;

(a-2) based on the calculated sub-band allocation, calculating the transmission power of each IoT device that minimizes the sum of the transmission powers of the IoT devices,

Step (b) above is

(b-1) Calculating a sub-band allocation that maximizes the minimum transmission rate among the IoT devices in order to ensure that all IoT devices satisfy the critical transmission rate condition - where the transmission power of all IoT devices is set to the maximum transmission power. Characterized by setting;

(b-2) based on the calculated sub-band allocation, calculating the transmission power of each IoT device that minimizes the maximum transmission power among the IoT devices.

Additionally, step (a) is

It is characterized by including finding and controlling the solution of the following equation 5.

[Equation 5]

here is the kth element represents the power control vector for all IoT devices with , and R _th is the critical data rate that all IoT devices must achieve, is the (k,n)th element. It is a matrix for sub-band allocation with .

In addition, step (b) is characterized by including finding and controlling the solution of the following equation (6).

[Equation 6]

In addition, step (a-1) includes obtaining a SAP solution of Equation 7, and step (a-2) includes obtaining a PCP1 solution of Equation 8,

[Equation 7]

[Equation 8]

here is the kth element represents the power control vector for all IoT devices with , and R _th is the critical data rate that all IoT devices must achieve, is the (k,n)th element. It is a matrix for sub-band allocation with is a sub-band allocation matrix obtained from SAP, which maximizes the minimum transmission rate, is the power control vector when all IoT devices use the maximum transmission power P _max .

Step (b-1) includes obtaining a SAP solution of Equation 7, and step (b-2) includes obtaining a PCP2 solution of Equation 9.

[Equation 9]

here is the kth element represents the power control vector for all IoT devices with , and R _th is the critical data rate that all IoT devices must achieve, is the (k,n)th element. It is a matrix for sub-band allocation with is the sub-band allocation matrix obtained from SAP, and is the sub-band allocation matrix that maximizes the minimum transmission rate, is the power control vector when all IoT devices use the maximum transmission power P _max .

Additionally, in the step (a-1), each possible sub-band allocation vector (a1-1) is regarded as a chromosome, and each element of the chromosome, called a gene, follows a discontinuous uniform distribution in the range [1, N]. Randomly generating an initial set of chromosomes to establish a set of sub-band allocation vectors as the initial population; (a1-2) calculating a fitness score for the sub-band allocation vector set in response to the fitness evaluation step of the genetic algorithm; - the fitness score applying the genetic algorithm (GA) is calculated using a fixed transmission power control vector (P _max ) is defined as the minimum transmission rate among all IoT devices, and each IoT device corresponding to each chromosome Transmission rate ( ) is Equation 4 ( , here is an IoT device ( ) is characterized by calculating and obtaining the transmission rate of all IoT devices by applying (representing the signal-to-interference plus noise ratio).

(a1-3) sorting the sub-band allocation vector set in descending order according to the fitness score to classify the upper group as the parent group and the lower group as the successor group;

(a1-4) Crossover step to cross over some chromosomes of the parent group;

(a1-5) Mutation step to increase the diversity of the sub-band allocation vector generated in the previous step;

(a1-6) forming a new progeny group and updating the sub-band allocation vector set, which is the population;

(a1-7) calculating the best fitness score among the updated population;

(a1-8) determining whether the highest fitness score increases over 10 consecutive generations, and repeating steps (a1-3) to (a1-8) until it does not increase; and

(a1-9) By applying the genetic algorithm including the step of setting the chromosome giving the best fitness score in the final population as the final solution for the sub-band allocation vector and calculating the optimal sub-band allocation matrix, the above [mathematics] The solution to [Equation 7] includes finding.

In addition, PCP1 in step (a-2) and PCP2 in step (b-2) obtain a solution through unsupervised learning (USL) applying a deep neural network (DNN) model of the deep learning processing unit embedded in the control device. Characterized in that, in the method of finding a solution through unsupervised learning applying the deep neural network (DNN) model, the normalized channel gain of the transmission power of the IoT device in forward propagation is input to the DNN model, and the given Based on the bias vector and weight matrix, the power control vector is predicted as the output of the DNN, and in reverse propagation, the loss function L1 or L2 is calculated based on the predicted power control vector and the given sub-band allocation matrix (C*). , in order to minimize the loss function, a process of updating the weight matrix and bias vector using gradient descent is performed, and the forward and backward propagation operations include being performed by repeatedly learning the loss function until it converges.

In addition, PCP1 in step (a-2) and PCP2 in step (b-2) obtain a solution through unsupervised learning (USL) applying a deep neural network (DNN) model of the deep learning processor built into the control device. Characterized in that, the deep neural network (DNN) model includes one input layer with K nodes, one output layer with K nodes, two hidden layers with 200 nodes, and the deep neural network ( The method of finding a solution through unsupervised learning (USL) applying the DNN model is to obtain the solution of the SAP and then standardize the transmission power of each IoT device for the given optimal sub-band allocation matrix (C*). Input data as a vector consisting of channel gains ( ), the output vector from the ℓth layer where ℓ ∈ {1, 2} is a normalized power control vector ( ), and the unsupervised learning is performed on the optimal sub-band allocation matrix (C*) and the transmission power (p*) determined by forward propagation. ) is the minimum signal-to-interference-plus-noise ratio (SINR) that achieves In the loss function L1 calculated by Equation 15 for PCP1 and L2 calculated by Equation 16 above, α and β are the entire transmission in the range of 0≤α and β≤1. It involves minimizing power and learning to minimize the maximum transmission power.

[Equation 15]

[Equation 16]

here Is It is the kth element of .

In addition, the DNN model is M multiple DNN models, each with different loss functions, and the control device selects the one with the best performance for each channel case among the power control vectors predicted by the M models, and selects the PCP1 and PCP2 It includes finding a solution to .

In addition, when the power control vector predicted through unsupervised learning (USL) is applied to the given sub-band allocation matrix (C*), if the minimum rate constraint condition is not met, the transmission determined by forward propagation It is characterized by setting the power (p*) to the maximum transmission power (p _max ) and finding the solution of PCP1 and PCP2.

The resource allocation method for minimizing transmission power while satisfying the minimum transmission rate of an IoT cellular network according to an embodiment of the present invention achieves excellent performance with very low computational complexity.

The resource allocation method according to an embodiment of the present invention can reduce computational complexity by separating a non-convex integer programming problem that is difficult to solve directly into a two-step sequential optimization problem.

Additionally, by applying the joint DNN model in the proposed approach, performance can be significantly improved without increasing computational complexity.

Figure 1 shows an example of a NOMA-based uplink IoT cellular network system according to an embodiment of the present invention.
Figure 2 shows an example of a control device according to an embodiment of the present invention.
Figure 3 shows an example of a genetic algorithm (GA).
Figure 4 shows an example of crossover and mutation when N = 3 and K = 5 in a system according to an embodiment of the present invention.
Figure 5 shows an example of the structure of a fully-connected DNN model for solving PCP1 and PCP2 according to an embodiment of the present invention.
Figure 6 shows the outage probability for the threshold data rate (R _th ) when N = 4 and K = 10 in the DNN model according to an embodiment of the present invention.
Figure 7 shows the average total transmission power for the threshold data rate (R _th ) when N = 4 and K = 10 in the DNN model according to an embodiment of the present invention.
Figure 8 shows the outage probability for the threshold data rate (R _th ) when the maximum transmission power among the transmission powers of all IoT devices is minimized in the DNN model according to an embodiment of the present invention.
Figure 9 shows the outage probability for the threshold data rate (R _th ) when the maximum transmission power among the transmission powers of all IoT devices is minimized in the DNN model according to an embodiment of the present invention.
Figure 10 shows the outage probability and average total transmission power for M when N = 4, K = 10, and R _th = 2.0bps/Hz in the joint DNN model according to an embodiment of the present invention.
Figure 11 shows the outage probability and average maximum transmission power for M when N = 4, K = 10, and R _th = 2.0bps/Hz in the joint DNN model according to an embodiment of the present invention.
Figure 12 shows the outage probability for minimizing the total transmission power for R _th .
Figure 13 shows the average total transmit power for R _th .
Figure 14 shows the outage probability for minimizing the maximum transmission power for R _th .
Figure 15 shows the average maximum transmission power for R _th .
Figure 16 shows a comparison of outage probabilities for minimizing the total transmission power for R _th for each method.
Figure 17 shows the average total transmission power comparison for R _th for each scheme.
Figure 18 shows a comparison of outage probabilities for minimizing the maximum transmission power for R _th for each method.
Figure 19 shows the average maximum transmission power comparison for R _th for each scheme.

The terms used in this application are only used to describe specific embodiments and are not intended to limit the invention. Singular expressions include plural expressions unless the context clearly dictates otherwise.

In this application, when a part “includes” a certain component, this means that it may further include other components rather than excluding other components unless specifically stated to the contrary.

In addition, terms such as "...unit", "...unit", and "device" used in the specification refer to a unit that processes at least one function or operation, which may be implemented as hardware, software, or a combination of hardware and software. there is.

Additionally, when describing components of embodiments of the present invention, terms such as first and second may be used. These terms are only used to distinguish the component from other components, and the nature, order, or order of the component is not limited by the term. When a component is described as being 'connected', 'coupled' or 'connected' to another component, that component may be directly connected, coupled or connected to that other component, but that component and that other component It should be understood that another component may be 'connected', 'combined', or 'connected' between elements.

Hereinafter, a resource allocation method and system for minimizing transmission power of an IoT cellular network according to the implementation of the present invention will be described in detail.

In this specification, bold lowercase letters (e.g. a, b, . . . ) are used to indicate vectors, and bold uppercase letters A, B, . . . Used to represent matrices. Calligraphy characters (e.g. X, Y , . . . ) are used to represent finite sets. The circularly symmetric complex Gaussian distribution with mean 0 and variance σ ² is It is defined as represents the non-negative part of x, i.e. [x] ⁺ = max(x, 0). Also, [1:N] = {1, 2, . . . , N }. All 1-row vectors of length K are It is displayed as

In one embodiment of the present invention, an uplink non-orthogonal multiple access (hereinafter referred to as 'NOMA') IoT cellular network meets the quality of service (QoS) requirements of the user of each IoT device while meeting the IoT device. Provides a method to implement joint optimization of sub-band allocation and transmission power control to minimize transmission power. More specifically, in one embodiment of the present invention, there are two optimization problems that minimize the total transmission power of all users and the maximum transmission power among all users, respectively, while satisfying the minimum transmission rate required for all IoT devices connected to the system. Provides a method to solve the problem.

The technical motivation of the first optimization problem here is to make IoT networks more energy efficient by minimizing the overall power consumption of the network. Additionally, the technical motivation of the second optimization problem is to ensure that the battery life of all IoT devices is above a certain level.

Figure 1 shows an example of a NOMA-based uplink IoT cellular network system according to an embodiment of the present invention.

Referring to FIG. 1, the NOMA-based uplink IoT cellular network system according to an embodiment of the present invention uses a base station (BS) and IoT devices (10-1, 10-2, ... 10-K) with the same bandwidth. It includes an uplink IoT cellular network using a total of N sub-bands and a control device 900 that controls it.

According to one embodiment of the present invention, IoT devices are all There are total number of sub-bands. In one embodiment of the present invention, it is assumed that IoT devices are transmitters and the base station is a receiver, performing uplink communication.

According to an embodiment of the present invention, IoT devices (10-1, 10-2, ...10-K) are assigned to one sub-band for communication with the base station (at this time, multiple IoT devices are connected to the same sub-band) Accessible (i.e. NOMA), IoT devices assigned to each sub-band transmit their data to the base station.

The control device 900 according to an embodiment of the present invention (a) sets the transmission rate of each IoT device to a threshold transmission rate or higher for all IoT devices connected through communication with the base station (BS), and transmits the IoT devices. controlling to minimize the sum of power; and (b) controlling the transmission rate of each IoT device to be above the threshold transmission rate for all IoT devices connected to the base station, and setting the maximum transmission rate among each IoT device. It includes performing a control function to minimize the value of transmission power.

In the present invention, the critical data rate refers to the minimum data rate required by the system.

The control device according to an embodiment of the present invention may be implemented in a computer system, for example, as a computer-readable recording medium.

Figure 2 shows an example of a control device according to an embodiment of the present invention.

As shown in FIG. 2, the control device 900 includes at least one of one or more processors 910, a memory 920, a storage unit 930, a user interface input unit 940, and a user interface output unit 950. elements, which can communicate with each other via bus 960. Additionally, the computer system 900 may also include a network interface 970 for accessing a network to which a base station (BS) and IoT devices 10-1, 10-2,...10-K are connected. The processor 910 may be a CPU or a semiconductor device that executes processing instructions stored in the memory 920 and/or the storage unit 930. The memory 920 and storage unit 930 may include various types of volatile/non-volatile storage media. For example, memory may include ROM 924 and RAM 925.

Accordingly, embodiments of the present invention may be implemented using a computer-implemented method or a non-volatile computer recording medium storing computer-executable instructions. When executed by a processor, the instructions may perform a method according to at least one embodiment of the present invention.

For example, when the instructions are executed by the processor 910, the sub-band allocation and transmission power calculation processing unit 980 calculates the optimal sub-band allocation according to a programmed algorithm, and the deep learning processing unit ( 990), the optimal transmission power can be calculated. The deep learning processing unit 990 contains a deep learning program and learns according to a programmed algorithm to find the optimal solution for controlling the transmission power of each IoT device.

According to an embodiment of the present invention, considering NOMA communication, it is considered that interference between users may occur between IoT devices in the same sub-band. The base station (BS) can perform successive interference cancellation (SIC) to remove interference between IoT devices for data decoding.

In one embodiment of the present invention, it is assumed that the base station (BS) and all IoT devices each have one antenna. Additionally, each IoT device can transmit data by accessing only one of the N sub-bands, and NOMA, which allows multiple IoT devices to be assigned to the same sub-band, may be allowed.

In one embodiment of the present invention and represents the set of all IoT devices and the set of all sub-bands, respectively.

also, is a user go It is defined as indicating whether it has been assigned to . in other words, means that IoT device k is assigned to sub-band n, otherwise am.

also, IoT devices using sub-band n ∈N at time t, denoted by The transmission signal of can be expressed by the following equation 1.

here is the data symbol of user device k at time t and p _k is the power constraint P _max , i.e. , This is the transmission power of the IoT device (k).

Since each IoT device can only access one sub-band, am. Referring to Figure 1, sub-band at time t, denoted by The received signal of the base station (BS) can be expressed as Equation 2 below.

here represents the channel coefficient from user device k to the base station (BS) in sub-band n. is the additive noise in sub-band n.

In one embodiment of the present invention It is assumed that is independent of the channel coefficient and the signal transmitted from the user. All related to Channel State Information (CSI) and By receiving a pilot signal from an IoT device before communication begins, Assume that the base station knows.

Additionally, in the case of IoT communication, since the packet size is generally small, each channel coefficient is assumed to remain unchanged during communication. In this case, by omitting the time index t, cast It can be expressed simply as

Recalling that multiple IoT devices can be assigned to the same sub-band of the considered network, it can be said that such NOMA networks can use spectrum more efficiently than OMA networks, where multiple user devices cannot use the same sub-band simultaneously. It is known. However, in the NOMA network, interference between user devices within a sub-band that does not exist in the Orthogonal Multiple Access (OMA) network may occur, so it is important to properly manage interference between IoT devices in order to increase the achievable transmission rate in the NOMA network. do.

To this end, the base station (BS) performs sequential interference cancellation (hereinafter referred to as 'SIC') by sequentially decoding data streams of the same sub-band one by one. In SIC operation, the decoded stream may be removed from the received signal before decoding the remaining stream, effectively increasing the signal to interference plus noise ratio (SINR) of the remaining stream.

In one embodiment of the present invention, specifically given About and It is said that

Additionally, in one embodiment of the present invention, it is assigned to sub-band(n), and all (here About) satisfying Consider a set of indices of IoT devices sorted in descending order according to channel strength, denoted by .

Then, for a given sub-band n, the base station (BS) device first streams the remaining streams according to the set process. By treating the set as noise, Recover . Afterwards, the base station (BS) device follows the set process. at remove the contribution from the remaining stream By processing the set of Attempt to recover .

This recursive process continues until all streams in sub-band n are decrypted. As a result, the stream The signal to interference-plus-noise ratios (hereinafter referred to as 'SINR') for is expressed in Equation 3 below.

and is an IoT device It represents the signal to interference plus noise ratio. IoT devices The transmission rate is given by Equation 4.

[Step 1 optimization method]

In one embodiment of the present invention, the goal is to provide a control method that minimizes the transmission power of IoT devices while satisfying the minimum transmission rate constraints for all users by jointly optimizing sub-band allocation and transmission power control of IoT devices. . Specifically, we provide a method to solve two optimization problems: minimizing the total transmission power of all users and minimizing the maximum transmission power among all users.

According to an embodiment of the present invention, the optimization problem for minimizing the total transmission power under minimum transmission rate constraints can be optimized by solving the following Equation 5.

here represents the power control vector for all users with p _k as the kth element, R _th is the threshold data rate that all users must achieve, is the (k,n)th element. It is a matrix for sub-band allocation with . For fixed C and p is given as Equation 4.

In [Equation 5], (5a) represents the goal of minimizing the sum of the transmission power of IoT devices. (5b) shows the transmission power range conditions of IoT devices, as 5b is due to transmission power constraints. (5c) represents the critical transmission rate condition. (5d) indicates which sub-band the IoT device is assigned to. (5e) indicates the condition that IoT devices can be assigned to one sub-band. In (5d) and (5e), each IoT device can be assigned to only one sub-band.

P1 is about a method of minimizing the sum of the transmission power of IoT devices while ensuring that the transmission rate of each IoT device is above the threshold transmission rate.

By solving the above P1 optimization problem, the total power consumption in the IoT communication network can be minimized and energy-efficient communication can be provided.

Next, the optimization problem to minimize the maximum transmission power among the transmission powers of all IoT devices in the system under minimum transmission rate constraints can be expressed as [Equation 6].

In the following equation 6, P2 relates to a method of minimizing the value of the maximum transmission power among IoT devices while ensuring that the transmission rate of each IoT device is above the threshold transmission rate for all IoT devices.

The control device 900 obtains a solution to the P2 optimization problem using Equation 6 above and performs control, thereby providing a method of ensuring the battery life of all IoT devices at a certain level or higher.

Here (6b)-(6e) are the same as (5b)-(5e) of P1.

P2 is almost identical to P1, but only the objective function of the transmission power allocation part is different.

In Equation 6, (6a) represents the goal of minimizing the maximum transmission power among IoT devices. (6b) represents the transmission power range conditions of IoT devices. (6c) represents the critical transmission rate condition. (6d) indicates which sub-band the IoT device is assigned to. (6e) indicates the condition that IoT devices can be assigned to one sub-band.

Power minimization problem considered P1 and P2 are binary variables Because it concerns a non-convex integer programming problem due to the ratio expression of and [Equation 4], it is difficult to find an optimal solution for this problem using existing optimization techniques. In one embodiment of the present invention, a new two-stage optimization method consisting of a GA algorithm and unsupervised learning (USL) was designed to handle these difficulties. As a result of many simulations, it was analyzed that the proposed two-step optimization method can provide a nearly optimal solution for the considered optimization problem with low computational complexity.

[2-step optimization method]

In order to reduce the computational complexity of solving P1 and P2, instead of directly finding the optimal sub-band allocation matrix and transmission power control vector, in one embodiment of the present invention, sub-band allocation and transmission power control are performed sequentially in two steps. An optimization method is proposed.

More specifically, first, a method of solving the sub-band assignment problem (hereinafter referred to as 'SAP') is presented while fixing the transmission power of IoT devices as shown in Equation 7 below.

That is, in one embodiment of the present invention, in order to reduce the computational complexity of solving P1 and P2, a two-step optimization technique is proposed that sequentially performs sub-band allocation and transmission power control for each IoT device.

First, the P1 problem in Equation 5 is transformed into the following Equations 7 and 8.

According to an embodiment of the present invention, step (a) can be obtained through (a-1) corresponding to Equation 7 and (a-2) corresponding to Equation 8.

here is the power control vector when all IoT devices use the maximum transmission power P _max .

[Equation 7] is (a-1) to find a sub-band allocation that maximizes the minimum transmission rate among IoT devices in order to ensure that all IoT devices satisfy the critical transmission rate condition (here, the transmission power of all IoT devices is is set to the maximum transmission power.) Solving this sub-band allocation problem (SAP) to maximize the minimum transmission rate among all IoT devices can help meet the minimum transmission rate constraints of P1 and P2.

After solving SAP of Equation 7 above and determining C as C*, the appropriate power control vectors for P1 and P2 are obtained by solving the following power control problem (hereinafter referred to as 'PCP'), i.e., PCP1, respectively. can. Specifically, for a given C*, PCP1 can be expressed as Equation 8.

PCP1 is the process of finding the transmission power of each IoT device that minimizes the sum of the transmission powers of IoT devices, based on the sub-band allocation obtained from SAP in [Equation 7], the previous step (a-2).

[Equation 8]

here is a sub-band allocation matrix obtained from SAP, and is a sub-band allocation matrix that maximizes the minimum transmission rate.

Additionally, the P2 problem in Equation 6 can be solved by modifying the following Equations 7 and 9 to obtain a solution to the PCP2 problem.

PCP2, like PCP1, first obtains SAP from Equation 7, and based on the sub-band allocation obtained from SAP of Equation 7 for a given C*, the transmission of each IoT device minimizes the maximum transmission power among IoT devices. It is a process of finding power.

According to another embodiment of the present invention, the proposed approach applies genetic algorithm to sub-band allocation problem (SAP) and unsupervised learning (USL) to power control problem (PCP). We propose a two-step algorithm, which is another embodiment.

For example, (1) the process of finding the optimal sub-band allocation is performed by applying the genetic algorithm (GA), an adaptive heuristic search algorithm, to SAP. At this time, the transmission power of each IoT device is fixed to the maximum power.

(2) Then, based on the sub-band allocation determined in (1), PCP1 and PCP2 are learned using deep learning based on unsupervised learning to find the optimal solution. Additionally, the loss function required for learning is Discover and learn new things.

Below, the proposed two-step algorithm is explained in detail.

A. [Genetic algorithm-based sub-band allocation]

An exhaustive search that considers all possible sub-band allocations can always find the optimal solution to the sub-band allocation problem (SAP). However, the total number of all possible sub-band assignments is Given that , it is practically impossible to implement such an exhaustive search even if N and/or K are slightly large.

To overcome this problem, a subsection, which is an embodiment of the present invention, proposes applying a genetic algorithm-based search method to solve the sub-band allocation problem (SAP) with low computational complexity.

Genetic algorithms are adaptive heuristic search algorithms inspired by the theory of natural evolution. The genetic algorithm uses the fitness score obtained directly from the objective function to find the optimal solution without differential operations, so it can be appropriately applied to optimization problems in the discrete search space and optimization problems in the continuous search space.

In genetic algorithms, the solution to a problem is called a chromosome, and a chromosome is composed of genes. Therefore, evolution can be seen as obtaining a good solution by modifying the genes in the chromosome.

Additionally, genetic algorithms have the ability to perform unique operations such as crossover and mutation in parallel, ensuring diversity of solutions and avoiding being trapped in local optima. Thanks to these advantages, it is known from previous research that genetic algorithm-based search methods can achieve excellent performance in various integer combination optimization problems.

In addition, a method of applying a genetic algorithm to reduce the complexity of calculations for optimization in a non-orthogonal multiple access (NOMA) system has been proposed in previous research (see Non-Patent Document 14).

As an example of a genetic algorithm according to an embodiment of the present invention, first, to establish an initial population, an initial chromosome set (i.e., solution set) is randomly generated and fitness scores for all chromosomes in the population are calculated.

Then, the chromosomes in the population are divided into parent and progeny groups according to their fitness scores, with chromosomes with higher scores belonging to the parent group and chromosomes with lower scores belonging to the progeny group. Next, the chromosomes of the parent group are randomly crossed over and mutated to create a new progeny group.

Then, the population is updated by replacing the previous progeny group with a new progeny group. This process continues until the highest fitness score of the population converges. As a result, the optimal population can be obtained under the genetic algorithm, and the chromosome with the highest fitness score becomes the final solution. When solving an optimization problem using a genetic algorithm (GA), the fitness score can have a significant impact on performance, so it is important to set the fitness score appropriately.

In one embodiment of the present invention, motivated by the fact that all IoT devices in the system must meet a minimum transmission rate constraint, when fixing the transmission power control vector to p _max , among the transmission rates that all IoT devices can achieve, sub -The suitability score to solve the band allocation problem (SAP) was set to the minimum transmission rate that all IoT devices can achieve.

For convenience of notation in this specification, the kth element It is defined that represents a sub-band allocation vector representing the sub-band index assigned to the kth IoT device. For example, when N= 2 and K= 4, u = [1 2 2 1] means that IoT devices 1 and 4 are assigned to sub-band 1 and the remaining IoT devices are assigned to sub-band 2.

Figure 3 shows an example of a genetic algorithm (GA).

The method of finding a solution by applying the proposed genetic algorithm (GA) to the sub-band allocation problem (SAP) is performed as follows.

1) First, the initial population setting step 510 is performed. In step 510, each possible sub-band allocation vector is considered a chromosome of the genetic algorithm (GA). Then, the initial population is established by randomly generating an initial set of chromosomes (a set of sub-band allocation vectors) in which each element of the chromosome, called a gene, follows a discontinuous uniform distribution in the range [1, N].

2) Next, a step (step 520) of calculating fitness scores for all chromosomes in the population (sub-band allocation vector set) is performed in response to the fitness evaluation step of the genetic algorithm (GA).

According to an embodiment of the present invention, the fitness score applying the genetic algorithm (GA) is defined as the minimum transmission rate among the rates that all IoT devices can achieve using a fixed transmission power control vector (P _max ), and is defined as the minimum transmission rate for each chromosome. Each corresponding IoT device ( ) transmission rate ( ) is Equation 4 ( ) can be applied to calculate the transmission rate of all IoT devices.

3) Next, a step 530 is performed to classify the upper group into the parent group and the lower group into the progeny group by sorting the chromosomes of the population in descending order of fitness scores (the number of chromosomes in the parent and progeny groups is shown in the table below). (respectively specified in I).

4) Next, a step 540 is performed in which some chromosomes of the parent group are selected as chromosomes to cross over according to a given crossover probability.

These chromosomes are randomly selected to form pairs, and a random intersection point is set for each pair.

Figure 4 shows an example of crossover and mutation when N = 3 and K = 5 in a system according to an embodiment of the present invention.

Referring to FIG. 4, a crossover step 540 is performed after the crossover point in which genes of two chromosomes are exchanged to create a new offspring as shown in FIG. 4(a). This process is called the crossover step.

5) Next, the chromosome generated in the previous step undergoes a mutation step (step 550) in which some genes are randomly changed with a given mutation probability, as shown in Figure 4(b). Step 550 is called a mutation step, and in one embodiment of the present invention, mutation step 550 is performed to increase the diversity of sub-band allocation vectors within the population (sub-band allocation vector set).

6) Next, steps 540 and 550 above are repeated to form a new generation of new generations. A population update step 560 is then performed to update the current population by replacing previous progeny groups with new groups.

7) Next, calculate the best fitness score within the current generation population. In one embodiment of the present invention, a step 570 of calculating the highest fitness score in the current updated population is performed in response to the step of calculating a fitness score in the current generation population.

8) Next, determine whether the highest fitness score increases over 10 consecutive generations, and repeat steps 530 to 570 until it does not increase (580).

9) Next, a step 590 is performed in which the chromosome with the best fitness score in the final population is set as the final solution for the sub-band allocation vector and denoted by u*. In step 590, the sub-band allocation matrix C corresponding to u*, denoted by C*, can be easily determined. For example, when N= 3, K= 5, u* = [2 3 1 2 3], C* is specified as follows.

In the flowchart of FIG. 4, S is the number of consecutive generations in which the best fitness score has not increased, G is the number of generations counted up to the current generation, and max(fG) means the best fitness score in the G generation population. After going through judgment in step 590 of Figure 3, the entire algorithm is terminated when S = 10.

B. Deep neural network (DNN)-based power control

According to an embodiment of the present invention, the solution of SAP is obtained through the proposed genetic algorithm (GA), and then the process of finding the solutions of PCP1 and PCP2 for a given C* is performed.

To this end, one embodiment of the present invention is characterized by implementation through unsupervised learning (USL) implemented as a programmed DNN model to find the solution to PCP1 and PCP2. In one embodiment of the present invention, a fully-connected DNN model is applied as the DNN model.

Figure 5 shows an example of the structure of a fully-connected DNN model for solving PCP1 and PCP2 according to an embodiment of the present invention.

Figure 5 go It represents the structure of the fully-connected DNN model, which is the kth element of .

The fully-connected DNN proposed in an embodiment of the present invention consists of one input layer with K nodes, one output layer with K nodes, and two hidden layers with 200 nodes (hidden layer and The appropriate number of nodes can be determined numerically).

First, all and About , here is C*'s is the first element, and the original channel gain is standardized as shown in Equation 10 below.

This normalization can help the learning process converge faster and prevent convergence to a local optimal solution. Then, the input data of the fully-connected DNN model inserted into the input layer is given as a vector consisting of the normalized channel gain of the transmission power of each IoT device, and is expressed as Equation 101 below.

This equation 101 represents the output vector, weight matrix, and bias vector of the ℓth layer of the DNN, respectively, where ℓ∈{1, 2, 3} and Nℓ represent the number of nodes of the ℓth layer given by the following equation 11: .

In the DNN applied according to an embodiment of the present invention, the first layer and the second layer are hidden layers, and the 0th layer and the third layer are the input layer and output layer, respectively.

Each hidden layer can be seen as preventing the real problem by preventing the gradient loss problem using the ReLU function. Here, the ReLU function is defined as ReLU(x) = [x] ⁺ . Specifically, the output vector in the ℓth layer where ℓ∈{1, 2} can be expressed as Equation 12.

Here, the ReLU function is a vector It is applied for each element.

Additionally, referring to Figure 5, the transmission power constraints of each IoT device can be more easily met by using the sigmoid function in the output layer. Here, the sigmoid function is defined as equation 102 below.

Specifically, given the input vector About The output vector of the output layer, expressed as , is given by Equation 13 below.

Here, the sigmoid function is a vector It is applied element by element, and for DNN, L = 3. elements have values between 0 and 1 is the normalized power control vector predicted by the DNN model.

Therefore, the transmission power control vector is finally determined as shown in Equation 14.

Next, we describe the loss function settings proposed for unsupervised learning (USL) to solve PCP1 and PCP2. Unlike supervised learning (SL), which requires labeled data to train a DNN model, unsupervised learning (USL) trains a DNN model using unlabeled data without supervision. On the other hand, when training a DNN model with unsupervised learning (USL), the loss function can have a significant impact on performance and must be set carefully.

In one embodiment of the present invention, a prior document ([33] W. Lee, “Resource allocation for multi-channel underlay cognitive radio network based on deep neural network,” IEEE Commun. Lett, vol. 22, no. 9, pp. 1942-1945, Sep. 2018. [34] W. Lee and K. Lee, “Resource allocation scheme for guarantee of QoS in D2D communications using deep neural network”,IEEE Commun. Lett, vol. 25, no. 3, pp 887-891, Dec. 2021) Referring to the applied example, we consider an intuitive method of directly reflecting the optimization problem in the loss function through appropriate transformation.

In one embodiment of the present invention, each IoT device The minimum signal-to-interference-plus-noise ratios (hereinafter referred to as 'SINR') required to meet the minimum transmission rate constraint (R _th ) given by were considered.

Additionally, for the fixed C* determined in the previous step and p* determined by forward propagation, It can be expressed as It represents the SINR of IoT device k that achieves a transmission rate of .

Then, set the loss function for PCP1, denoted as L1, as shown in Equation 15.

here Is It is the kth element of and α and β are controllable parameters that satisfy 0 ≤ α, β ≤ 1.

Equation 15 represents the loss function for PCP1.

Here, the first term in Equation 15 is directly derived from the objective function of PCP1, and serves to train the DNN in the direction of minimizing the total transmission power as the sum of transmission powers.

Meanwhile, the second term in Equation 15 is a transmission rate condition and is related to the normalized SINR margin for the minimum required SINR of each IoT device k. In one embodiment of the present invention, all IoT devices are learned to meet the minimum SINR constraints by activating only IoT devices that do not meet the constraints.

Additionally, in one embodiment of the present invention, α and β can be controlled in the proposed loss function to determine which of the two effects mentioned above to focus on. Specifically, when α is relatively large, the proposed unsupervised learning (USL) trains the DNN with more emphasis on minimizing the sum transmission power, whereas when β is relatively large, the transmission rate during the learning process Learning is more focused on satisfying constraints.

Similarly, the loss function of PCP2, denoted by L2, is expressed as Equation 16.

Here, the first term in [Equation 16] is the maximum transmission power, which is derived from the objective function of PCP2 that minimizes the maximum transmission power among the transmission powers of all IoT devices, while the second term is the transmission rate condition, in the case of the loss function L1 Same as

Equation 16 represents the loss function for PCP2.

Equations 15 and 16 for the loss functions L1 and L2, respectively (α, β pairs can be numerically optimized to achieve better performance. For example, if α and β are well adjusted, the transmission power can be minimized A good balance can be achieved between satisfying the critical data rate conditions.

That is, in the unsupervised learning method using the DNN model, in the loss function L1 calculated by Equation 15 for PCP1 and L2 calculated by Equation 16 for PCP2, α and β are 0≤α, β≤ It involves learning to find conditions that minimize the total transmission power and the maximum transmission power in the range of 1.

In the unsupervised learning method using the DNN model, a series of normalized channel gains in forward propagation are inserted into the DNN as input, and the power control vector can be predicted as the output of the DNN based on the given bias vector and weight matrix.

On the other hand, in reverse propagation, the loss function L1 or L2 is calculated based on the predicted power control vector and the given sub-band allocation matrix (C*), and the weight matrix and bias vector are updated using gradient descent to minimize the loss function. The process is carried out. Forward and backward propagation operations are performed by repeatedly learning the loss function until it converges.

Additionally, according to another embodiment of the present invention, the performance of the proposed unsupervised learning (USL) can be further improved by using multiple DNN models instead of a single DNN model.

Specifically, in one embodiment of the present invention applying a multiple DNN model, the prior literature ([27] F. Liang, C. Shen, W. Yu, and F. Wu, “Towards optimal power control via ensembling deep neural networks, “IEEE Trans. Commun., vol. 68, no. 3, pp. 1760-1776, Mar. 2019.) Referring to the applied example, model i ∈ [1: M] is (α _i , Consider a DNN model trained using β _i ) pairs. where M is the number of DNN models considered and It has conditions. Then, among the power control vectors predicted by the M models, we can select the one that achieves the best performance for each channel case. In this specification, the combination of these DNN models is defined as a 'joint model'. Using these joint models can achieve strictly better performance than using a single model.

And when the power control vector predicted through the proposed unsupervised learning (USL) is applied to the given C*, if the minimum rate constraint condition is not met, p* can be simply set as p*=p _max .

As explained in the following discussion, allowing p* = p _max does not significantly increase the transmission power on average unless R _th is too large. Additionally, even if p* = p _max is set, the minimum transmission rate constraint may still not be met for some IoT devices depending on channel conditions. This phenomenon is defined as ‘outage’.

[One. Numerical simulation]

The performance of the algorithm proposed according to an embodiment of the present invention was numerically evaluated. Referring to FIG. 1 for simulation according to an embodiment of the present invention, the base station (BS) is located in the center of the cell and each IoT device (10-1, 10-2, 10-3, ... 10-K) has a radius. It is assumed that they are uniformly and randomly distributed in cells of 1 km. Additionally, it is assumed that the bandwidth of each sub-band is 15kHz, σ ² = -168dBm/Hz, and P _max = 23dBm.

Additionally, the path loss for IoT device (k) is, , (see HW Kim, HJ Park, and SH Chae, “assignment and power control for IoT cellular networks via deep learning,” IEEE Access, 2022).

here and is the distance between the base station BS and the IoT device (k). The channel coefficient is It is modeled as silver Follow.

In the unsupervised learning (USL) process performed in the power control stage, the early stopping method was adopted to increase learning efficiency. That is, in one embodiment of the present invention, if the validation loss for validation data does not decrease during 30 consecutive learning epochs, the learning process is terminated.

Additionally, in order to accelerate the convergence of the learning process, the initial learning rate is set to decrease by 0.2 times whenever the validation loss on the validation data does not decrease during 10 consecutive learning epochs.

Additionally, in one embodiment of the present invention, the dropout method is applied to the hidden layer of the proposed DNN model to avoid the overfitting problem by randomly ignoring the output of the hidden node.

Simulation parameters according to one embodiment of the present invention are summarized in Table 1 below.

[1-A Optimization of α and β in DNN according to an embodiment of the present invention]

The performance of the resource allocation method algorithm for minimizing transmission power according to an embodiment of the present invention is greatly affected by the α and β values of the loss function used in DNN. Therefore, the present inventors first investigated the performance trends of the proposed algorithm for α and β.

Figure 6 shows the outage probability for the threshold data rate (R _th ) when N = 4 and K = 10 in the DNN model according to an embodiment of the present invention.

Figure 7 shows the average total transmission power for the threshold data rate (R _th ) when N = 4 and K = 10 in the DNN model according to an embodiment of the present invention.

Figures 6 and 7 show the results of minimizing the total transmission power and maximum transmission power among the transmission powers of all IoT devices for different pairs of (α, β), respectively.

In this simulation, only cases in which no outage occurs for each method are considered when plotting the average total transmission power or average maximum transmission power.

Additionally, in order to reduce the probability of outage, the case where p*= p _max is allowed in the power control process is indicated as '+P _max '. Additionally, for comparison, the 'NOMA search/genetic algorithm (GA)' method, which performs sub-band allocation through full search and power control through genetic algorithm (GA), was considered.

As shown in Figures 6 and 7, the outage probability and the average total transmission power consumption of all methods increase with the minimum transmission rate constraint (R _th ), as expected. More interestingly, when α is large as α = 0.9 and β is small as β = 0.13, the proposed method can achieve relatively low total transmission power, but in the learning process, minimizing the total transmission power is the minimum transmission rate constraint that can be satisfied. Focus more on Conversely, when β is large as β = 0.9 and α is small as α = 0.13, the outage probability is relatively reduced, but the total transmission power consumption is relatively increased because the minimum transmission rate constraint is considered more important.

Additionally, allowing it increases the overall transmission power but can greatly reduce the outage probability, especially when α is large.

Figure 8 shows the outage probability for the threshold data rate (R _th ) when the maximum transmission power among the transmission powers of all IoT devices is minimized in the DNN model according to an embodiment of the present invention.

Figure 9 shows the outage probability for the threshold data rate (R _th ) when the maximum transmission power among the transmission powers of all IoT devices is minimized in the DNN model according to an embodiment of the present invention.

Referring to Figures 8 and 9, similar performance trends can be observed even when the maximum transmission power among the transmission powers of all IoT devices is minimized.

Based on these observations, it can be seen that by controlling α and β, a good balance can be obtained between minimizing transmission power and satisfying the quality of service (QoS).

Therefore, as we attempt to numerically optimize the (α, β pair) that achieves good performance in these two aspects for each loss function L1 and L2, the optimal pair for L1 and L2 is given by

L1 and L2 are given by (α, β= (0.41, 0.62) and (α, β= (0.76, 0.27)) when N = 2 and K = 4, and ( are given as α, β= (0.34, 0.69) and (α, β= (0.41, 0.62).

In addition, later, a method of finding a solution can be presented by considering both cases in which p* = p _max is allowed or not allowed in the power control process.

[1-B Joint DNN Model]

As described above, another embodiment of the present invention proposes a joint DNN model using multiple DNN models with different pairs (α, β) to further improve the performance of the proposed unsupervised learning (USL). In particular, , considering the tradeoff observed in the previous embodiment, the (α _i , β _i ) pair, which is a controllable parameter for each model i ∈ [1: M ], is determined as shown in Equations 17 and 18. .

For each channel implementation, the joint model can select the best performing vector among the power control vectors predicted by the M model. Therefore, using multiple unsupervised learning (USL) models learned with different pairs of (α, β) can improve performance by increasing the diversity of predicted power control vectors.

Figure 10 shows the outage probability and average total transmission power for M when N = 4, K = 10, and R _th = 2.0bps/Hz in the joint DNN model according to an embodiment of the present invention.

Figure 11 shows the outage probability and average maximum transmission power for M when N = 4, K = 10, and R _th = 2.0bps/Hz in the joint DNN model according to an embodiment of the present invention.

Referring to Figures 10 and 11, as M increases, the average transmission power consumption of the joint DNN model decreases and is almost saturated when M = 8.

Additionally, Table 2 shows the simulation time (sec) per 100 samples for M ((N, K) = (4, 10)). Referring to Table 2, the required simulation time for M, the joint DNN model hardly increases with M, because the pre-trained DNN model infers the final result with simple linear operations, as summarized in Table 2. Additionally, similar results can be obtained even when the values of (N, K) are different. Considering this, in one embodiment of the present invention, M is set to 8 for the joint DNN model.

[1-C. Performance evaluation]

To compare performance evaluations according to an embodiment of the present invention, benchmark systems according to the following methods were considered.

- The joint global search method finds the optimal set of sub-band allocation matrices and discretized power control vectors for optimization problems P1 and P2, respectively. The quantization step size for the discretized power vector is defined to be sufficiently small.

- The search/GA method first considers all possible sub-band allocations to find the optimal sub-band allocation matrix for SAP. Then, given the sub-band allocation matrix, a genetic algorithm (GA) is used to find the optimal power control vector for PCP1 or PCP2. Here, the fitness score function of the genetic algorithm (GA) is set using the loss function described above (for example, the genetic algorithm (GA) performs an operation to find a chromosome that maximizes the fitness score, so the fitness score function applied is - It is set to L1 or -L2.)

- The GA/GA method first searches for the optimal sub-band allocation matrix for SAP using the proposed genetic algorithm (GA) as described previously. Then, we find the optimal power control vector for PCP for a given sub-band allocation matrix in the same way as search/GA.

- The GA/FPC (fractional power control) method finds the optimal sub-band allocation matrix in the same way as the proposed algorithm. Meanwhile, in the power control step, FPC is performed to compensate for path loss by allocating power in inverse proportion to the channel gain of each user for a given sub-band allocation.

- The GA/no power control (GA/NPC) method finds the best sub-band allocation matrix in the same way as the proposed genetic algorithm (GA) and then simply sets the power control vector p to p* = p _max .

Additionally, the performance of the algorithm proposed according to an embodiment of the present invention is evaluated in an orthogonal multiple access (OMA) scenario. We considered a time division multiple access (TDMA)-OMA system where the achievable transmission rate for IoT device k is given by the following equation for a given sub-band allocation (J. Cui, Z. Ding, and P. Fan, “ (see “novel power allocation scheme under outage constraints in NOMA systems,” IEEE Signal Process. Lett., vol. 23, no. 9, pp. 1226-1230, 2016)

here and lim.

Figures 12 to 15 show a performance comparison of minimizing the total transmission power and maximum transmission power when N = 2 and K = 4 for each method, respectively.

Figure 12 shows the outage probability for minimizing the total transmission power for R _th .

Figure 13 shows the average total transmit power for R _th .

Figure 14 shows the outage probability for minimizing the maximum transmission power for R _th .

Figure 15 shows the average maximum transmission power for R _th .

According to an embodiment of the present invention, it can be observed that transmission power consumption and outage probability increase according to R _th in all methods.

Table 3 shows This shows the simulation time (sec) for each method per 100 samples using . In Table 3, GA/FPC [39] refers to prior research 12 (see non-patent document 12).

Referring to Table 3, although the proposed joint DNN model approach has much lower complexity as shown in Table 3, the results achieve performance close to the joint full search performance in terms of minimizing outage probability and strictly outperform other benchmark methods. It shows that it can be done.

and transmit power. Additionally, the performance of the proposed algorithm was found to be excellent even when using a single DNN model. Additionally, it can be observed that even if the transmission power increases slightly, the outage probability can be lowered by allowing p*=p _max in the power control process.

I did. Additionally, referring to Figures 6 and 7, the method proposed in Figures 12 and 14 is analyzed to have a much lower outage probability than the GA/FPC method. It is analyzed that the method according to an early embodiment of the present invention can have better performance because, unlike the GA/FPC method, it predicts the power control vector by considering R _th during the learning process. In addition, the GA/FPC method was excluded when plotting transmission power performance because the possibility of outage was too high.

Additionally, it can be seen that NOMA can achieve a much lower outage probability than OMA, especially when R _th is large. This means that the transmit power required to achieve the minimum rate constraint is It was analyzed that this is because it is exponentially proportional to and R _th .

Finally, the poor performance of the GA/NPC scheme clearly shows that power control is essential for energy-efficient communication.

Figures 16 to 19 show performance comparisons for minimizing the total transmission power and maximum transmission power, respectively, when N = 4 and K = 10 for each method.

Figure 16 shows a comparison of outage probabilities for minimizing the total transmission power for R _th for each method.

Figure 17 shows the average total transmission power comparison for R _th for each scheme.

Figure 18 shows a comparison of outage probabilities for minimizing the maximum transmission power for R _th for each method.

Figure 19 shows the average maximum transmission power comparison for R _th for each scheme.

Referring to Figures 16 to 19, it can be seen that the outage probability of OMA is much higher than that of NOMA as the ratio of K to N increases compared to the case of N = 2 and K = 4. Additionally, it shows that the performance improvement of NOMA increases compared to OMA as the number of IoT devices exceeds the number of sub-bands. From this, it can be seen that the allocation method proposed according to an embodiment of the present invention improves performance in NOMA. In addition, the overall performance trend is similar to the case of N = 2 and K = 4.

Additionally, we compared the computational complexity of the methods considered in Table 3 in terms of the required simulation time (the time required to train the DNN model was not considered here). The entire simulation was performed using python 3.8 with Tensorflow 2.7.0, Intel i7-10700F CPU, 32GB RAM, and NVIDIA GeForce. Table 3 shows that the proposed method has much lower computational complexity than the full search-based approach and GA/GA, and has similar computational complexity to the existing method. Additionally, it can be seen that the simulation time of the proposed technique hardly changes depending on whether p* = p _max is allowed or a joint model is used in the power control process. Therefore, the method proposed according to an embodiment of the present invention achieves excellent performance with very low computational complexity, and using a combined model in the proposed approach can significantly improve performance with little increase in computational complexity.

As described above, according to an embodiment of the present invention, a method for minimizing transmission power of a NOMA-based uplink IoT cellular network under minimum transmission rate constraints was presented. The approach specifically proposed in one embodiment of the present invention utilizes a genetic algorithm (GA) to perform sub-band allocation and determines power control for each IoT device by unsupervised learning (USL) implemented as a DNN model. It includes sequentially performing sub-band allocation and power control.

Additionally, we presented a loss function that can achieve a good balance between minimizing transmission power and satisfying minimum transmission rate constraints while training a DNN.

Another embodiment of the present invention further suggests that performance can be further improved by using several joint DNN models with different loss functions. The joint DNN model maintains computational complexity compared to using a single DNN model. Numerical simulations of the proposed approach show that the proposed two-stage algorithm performs better in both aspects: minimizing transmission power and reducing computational complexity. You can get it.

In the future, IoT devices according to an embodiment of the present invention can expand and apply the resource allocation method for minimizing transmission power while satisfying the minimum transmission rate of the IoT cellular network in the IoT system to the case where the base station (BS) has multiple antennas. There will be. Additionally, extending the proposed approach to multi-cell networks would also be a valuable future research direction.

10-1, 10-2, … 10-K: IoT devices
900: Control device
BS: base station

Claims (11)

In the resource allocation method for minimizing transmission power by controlling transmission power in a non-orthogonal multiple access (NOMA) IoT cellular network system,
The system is,
base station and
An IoT device connected to the base station through uplink communication;
It includes a control device that controls transmission power to IoT devices connected to the base station through communication,
The control device is
(a) controlling all IoT devices connected to the base station to minimize the sum of the transmission power of each IoT device while keeping the transmission rate of each IoT device above the threshold transmission rate; and
(b) for all IoT devices connected to the base station, controlling the transmission rate of each IoT device to be higher than the threshold transmission rate and minimizing the maximum transmission power among the IoT devices. Allocation method. According to paragraph 1,
The step (a) is
A resource allocation method comprising obtaining and controlling the solution of Equation 5 below.
[Equation 5]

here is the kth element represents the power control vector for all IoT devices with , and R _th is the critical data rate that all IoT devices must achieve, is the (k,n)th element. It is a matrix for sub-band allocation with . According to paragraph 2,
Step (b) above is
A resource allocation method comprising obtaining and controlling the solution of Equation 6 below.
[Equation 6]
According to paragraph 1,
The step (a) is
(a-1) calculating a sub-band allocation that maximizes the minimum transmission rate among the IoT devices under conditions for ensuring that all IoT devices satisfy a threshold transmission rate condition; and
(a-2) A resource allocation method comprising calculating the transmission power of each IoT device to minimize the sum of the transmission powers of the IoT devices, based on the calculated sub-band allocation. According to paragraph 1,
Step (b) above is
(b-1) calculating a sub-band allocation that maximizes the minimum transmission rate among the IoT devices in order to ensure that all IoT devices satisfy the threshold transmission rate condition; and
(b-2) A resource allocation method comprising calculating the transmission power of each IoT device that minimizes the maximum transmission power among the IoT devices, based on the calculated sub-band allocation. According to paragraph 4,
Step (a-1) includes obtaining the SAP solution of Equation 7, and step (a-2) includes obtaining the PCP1 solution of Equation 8,
[Equation 7]

[Equation 8]

here is the kth element represents the power control vector for all IoT devices with , and R _th is the critical data rate that all IoT devices must achieve, is the (k,n)th element. It is a matrix for sub-band allocation with is a sub-band allocation matrix obtained from SAP, which maximizes the minimum transmission rate, is the power control vector when all IoT devices use the maximum transmission power P _max .
The step (b-1) includes obtaining a SAP solution of Equation 7, and the step (b-2) includes obtaining a PCP2 solution of Equation 9.
[Equation 9]

here is the kth element represents the power control vector for all IoT devices with , and R _th is the critical data rate that all IoT devices must achieve, is the (k,n)th element. It is a matrix for sub-band allocation with is the sub-band allocation matrix obtained from SAP, and is the sub-band allocation matrix that maximizes the minimum transmission rate, is the power control vector when all IoT devices use the maximum transmission power P _max . According to clause 6,
The step (a-1) is
(a1-1) Consider each possible sub-band allocation vector as a chromosome, and randomly generate an initial set of chromosomes in which each element of the chromosome, called a gene, follows a discrete uniform distribution in the range [1, N] as the initial population. Establishing a sub-band allocation vector set;
(a1-2) calculating a fitness score for the sub-band allocation vector set in response to the fitness evaluation step of the genetic algorithm; - the fitness score applying the genetic algorithm (GA) is calculated using a fixed transmission power control vector (P _max ) is defined as the minimum transmission rate among all IoT devices, and each IoT device corresponding to each chromosome Transmission rate ( ) is Equation 4 ( , here is an IoT device ( ) is characterized by calculating and obtaining the transmission rate of all IoT devices by applying (representing the signal-to-interference plus noise ratio).
(a1-3) sorting the sub-band allocation vector set in descending order according to the fitness score to classify the upper group as the parent group and the lower group as the successor group;
(a1-4) Crossover step to cross over some chromosomes of the parent group;
(a1-5) Mutation step to increase the diversity of the sub-band allocation vector generated in the previous step;
(a1-6) forming a new progeny group and updating the sub-band allocation vector set, which is the population;
(a1-7) calculating the best fitness score among the updated population;
(a1-8) determining whether the highest fitness score increases over 10 consecutive generations, and repeating steps (a1-3) to (a1-8) until it does not increase; and
(a1-9) By applying the genetic algorithm including the step of setting the chromosome giving the best fitness score in the final population as the final solution for the sub-band allocation vector and calculating the optimal sub-band allocation matrix, the above [mathematics] A resource allocation method that includes finding the solution to [Equation 7]. According to clause 6,
PCP1 in step (a-2) and PCP2 in step (b-2) are characterized by finding a solution through unsupervised learning (USL) applying a deep neural network (DNN) model of the deep learning processing unit with the control device built-in. and
The method of finding a solution through unsupervised learning applying the deep neural network (DNN) model is:
In forward propagation, the normalized channel gain of the transmission power of the IoT device is input to the DNN model, and the power control vector is predicted as the output of the DNN based on the given bias vector and weight matrix,
In reverse propagation, the loss function L1 or L2 is calculated based on the predicted power control vector and the given sub-band allocation matrix (C*), and the weight matrix and bias vector are updated using gradient descent to minimize the loss function. This is done,
A resource allocation method including performing the forward and backward propagation operations by repeatedly learning the loss function until it converges. In paragraph 6
PCP1 in step (a-2) and PCP2 in step (b-2) are characterized by finding a solution through unsupervised learning (USL) applying a deep neural network (DNN) model of the deep learning processor built into the control device. and
The deep neural network (DNN) model includes one input layer with K nodes, one output layer with K nodes, two hidden layers with 200 nodes,
The method of finding a solution through unsupervised learning (USL) applying the deep neural network (DNN) model is:
After solving the SAP, the transmission power of each IoT device is converted into a vector composed of normalized channel gains for the given optimal sub-band allocation matrix (C*) as input data ( ), the output vector from the ℓth layer where ℓ ∈ {1, 2} is a normalized power control vector ( ) is output,
The unsupervised learning is performed on the optimal sub-band allocation matrix (C*) and the transmission power (p*) determined by forward propagation. ) is the minimum signal-to-interference-plus-noise ratio (SINR) that achieves When you say,
In the loss function L1 calculated by Equation 15 for the PCP1 and L2 calculated by the loss function Equation 16 above for the PCP2, α and β minimize the total transmission power in the range of 0≤α, β≤1, and A resource allocation method that includes learning to minimize the maximum transmission power.
[Equation 15]

[Equation 16]

here Is It is the kth element of . According to clause 8,
The DNN collection is M multiple DNN models, each with a different loss function,
The resource allocation method includes the control device selecting the one with the best performance for each channel case among the power control vectors predicted by M models to obtain solutions for the PCP1 and PCP2. According to clause 8,
When the power control vector predicted through unsupervised learning (USL) is applied to the given sub-band allocation matrix (C*), if the minimum rate constraint condition is not met, the transmission power determined by forward propagation ( A resource allocation method including determining the solution of PCP1 and PCP2 by setting p*) to the maximum transmission power (p _max ).