A task migration method based on joint uplink and downlink user clustering

Abstract

A task migration method based on joint uplink and downlink user clustering is proposed. When the downlink migration task result is transmitted back to the terminal, a large-scale antenna array millimeter wave non-orthogonal multiple access technology is used, and wireless energy transmission is performed to the terminal at the same time. In the downlink clustering method, the initialization cluster center selection and cluster center update method are designed according to the channel characteristics, and then based on the downlink user clustering result, the users are selected between different clusters for uplink user virtual MIMO pairing by taking advantage of the strong correlation of the channels of users in the same cluster, the large difference of the channels of users between clusters, and the low interference. Through experimental simulation, it is verified that when migrating edge computing tasks, the proposed joint uplink and downlink user clustering task migration method has lower latency and energy consumption than other traditional methods.

Description

Technical Field

[0001] The present invention belongs to the field of edge computing, and in particular relates to a task migration algorithm based on joint uplink and downlink user clustering. Background Art

[0002] In the past few years, with the exponential growth of mobile devices such as smartphones, handheld game consoles, and in-car multimedia computers, more and more new mobile applications, such as augmented reality, image processing, natural language processing, face recognition, and interactive games, have emerged and gradually occupied a considerable proportion of usage. These types of mobile applications are usually computationally intensive, latency sensitive, and have high energy consumption. However, due to physical size constraints, mobile devices are usually equipped with low-capacity batteries and limited computing power, which has become a bottleneck for the development of mobile applications.

[0003] One solution is to let mobile devices transfer intensive computing tasks to a remote cloud center with high computing power and large storage capacity. However, existing mobile cloud computing faces various challenges, including high latency and low scalability caused by the long propagation distance from mobile devices to remote cloud centers, as well as high burden on front-end links caused by centralized deployment of cloud centers. The cost of cloud computing is declining slowly, and the complex network environment makes it difficult to achieve a breakthrough in network latency. The linear growth of its centralized computing power can no longer match the explosive growth of massive edge data. In short, uploading large-scale data to cloud computing centers for processing is time-consuming and labor-intensive, resulting in untimely data processing and loss of the meaning and value of utilization. In addition, the real-time and flexibility requirements of data processing bring unsolvable challenges to the traditional cloud computing model. In this application context, edge computing came into being.

[0004] As the future mobile network B5G / 6G moves towards intelligence, the emergence of different computing-intensive and high-energy consumption application services has led to the widespread use of edge computing, in which mobile terminals migrate tasks to edge servers for processing. However, due to physical size limitations, mobile devices are usually limited in battery capacity and computing power. During the migration of edge computing tasks, a large number of intensive computing tasks will accelerate the energy consumption of the terminal and shorten the battery life of the mobile device. Therefore, how to migrate tasks energy-efficiently and reduce the energy consumption of mobile terminals has always been an open problem in edge computing. In the current research on energy-efficient edge computing task migration, existing research focuses on the optimization of task migration in a single scenario of centimeter waves or millimeter waves. The future network will be a scenario of mixed application of high-speed millimeter waves and long-distance centimeter waves, which is one of the very important candidate technologies for 6G. From the above, it can be seen that the current problems to be solved are as follows:

[0005] (1) The traditional edge computing system has low energy efficiency when migrating tasks in a single cm-wave or mm-wave scenario;

[0006] (2) In the uplink centimeter-wave and downlink millimeter-wave edge computing systems, the uplink centimeter-wave MIMO user pairing and downlink millimeter-wave user clustering separation lead to high implementation complexity and low energy efficiency. Summary of the invention

[0007] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a task migration method based on joint uplink and downlink user clustering, which involves flexible and scalable user clustering to achieve management of mobile devices in the face of massive mobile devices, and aims at the task migration needs of users and base stations that require multiple uplink and downlink interactions in wireless edge computing oriented to 6G. It includes a downlink millimeter wave and uplink centimeter wave task migration method and an uplink centimeter wave virtual MIMO user pairing algorithm based on downlink millimeter wave user clustering. The initialization cluster center selection and cluster center update method are designed according to the channel characteristics of the millimeter wave large-scale antenna array NOMA system. Based on the downlink user clustering results, a low-complexity uplink virtual MIMO user pairing algorithm based on downlink user clustering is proposed by utilizing the characteristics of strong channel correlation and low inter-cluster interference of users in the same cluster, and finally the task migration of joint uplink and downlink user clustering is realized. Through simulation, the experiment proves that when edge computing tasks are migrated, the user clustering results of the proposed improved K-means algorithm are used in downlink transmission, and the uplink user virtual MIMO pairing algorithm has lower energy-delay product and lower energy consumption than other traditional methods. It can effectively optimize the energy consumption and delay of mobile terminal task migration in edge computing scenarios.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is:

[0009] A task migration method based on joint uplink and downlink user clustering includes the following steps:

[0010] 1. A task migration algorithm based on joint uplink and downlink user clustering, characterized in that the task migration algorithm method based on joint uplink and downlink user clustering comprises the following steps:

[0011] 1) Multiple receiving antennas of different terminals with different channel gains are grouped into NOMA clusters, and all user receiving antennas in each cluster are scheduled based on NOMA;

[0012] 2) When selecting the initial cluster center, follow the principle of maximizing the channel gain of the cluster center and minimizing the channel similarity of the cluster center users. First, sort the users according to the channel gain, and select the user with the largest channel gain as the cluster center of the first cluster. Then set a threshold as the maximum value of the channel similarity between cluster centers. The threshold should be as small as possible at the beginning to ensure that there is low channel interference between the users of the initial cluster centers. Then, include the users whose channel similarity with the first cluster center is less than the threshold into the candidate cluster center set, and select the user with the largest channel gain as the new cluster center;

[0013] 3) When updating the K-means cluster center, the clustering criterion is the user channel direction angle. The cluster center of each user cluster is updated to the user with the lowest correlation with other user clusters. The correlation between the user and other user clusters is defined as the sum of the normalized channel correlations between the user and other cluster users, that is:

[0014]

[0015] Among them, M (k) represents the cluster containing user k. The similarity of user channel vectors can be measured using the normalized direction, i.e.:

[0016]

[0017] User clustering algorithm for millimeter wave downlink massive antenna array NOMA energy-carrying system based on improved K-means:

[0018] Step 1: Calculate the channel gains of K users

[0019] Step 2: Arrange the user channels in reverse order according to their channel gains. The channel gain set is A = {a1, a2, ..., a k}, select the user with the largest channel gain as the cluster center of a cluster, the cluster center set is H, and the remaining users are placed in the set Initialize the candidate cluster center set l = 1;

[0020] Step 3: When Ω is not empty, users whose channel correlation with the selected cluster center is less than the threshold δ are selected as candidate cluster centers. The candidate cluster center set is In Ω, select the user with the largest channel gain as the center of a cluster and put it into H, and the rest of the users are put into the set l = l + 1. When Ω is empty, increase the threshold by δ = δ + (1-δ) / 10, and update the candidate cluster center set. When l<L, repeat step 3;

[0021] Step 4: When l=L, select L initial cluster center user sets H;

[0022] Step 5: compare the angle difference between each user and the L initial cluster center users, and include user i into the cluster with the smallest angle difference;

[0023] Step 6: The cluster center of the lth cluster is updated as:

[0024]

[0025] Step 7: Repeat steps 5 to 6 to update the cluster members until the result no longer changes.

[0026] 2. An uplink centimeter wave virtual MIMO user pairing algorithm based on downlink clustering, characterized in that the uplink centimeter wave virtual MIMO user pairing algorithm based on downlink clustering comprises the following steps:

[0027] 1) The configuration of the uplink multi-user virtual MIMO system consists of K users, each with one antenna, sending to the base station through a flat fading channel;

[0028] 2) Pairing scheduling builds a 2×2 virtual MIMO wireless channel between K users and the base station. Time-frequency resource sharing is achieved through spatial multiplexing of the uplink, and information is sent to the base station;

[0029] 3) The user clusters obtained by the improved K-means algorithm for downlink have clustering characteristics in terms of geographical location. Users in different clusters are far apart, the channel correlation is low, and the orthogonality is relatively high. Therefore, virtual MIMO pairing between users in different clusters can achieve a higher uplink system capacity. Based on this, an uplink virtual MIMO user pairing algorithm based on downlink user clustering is proposed.

[0030] Uplink virtual MIMO user pairing algorithm based on downlink user clustering:

[0031] Step 1: Arrange the L cluster users obtained by the downlink improved K-means algorithm in reverse order according to their channel gains. The channel gain set of the lth cluster user is l∈1,2,...,L, the user with the largest channel gain is the cluster head of the cluster, and the cluster head group is H;

[0032] Step 2: In the cluster head cluster H, the base station randomly selects two users i and j for pairing;

[0033] Step 3: According to the channel gain set A i , A j The users in the two clusters are paired in pairs according to the user order, until the users in one of the clusters are paired;

[0034] Step 4: Repeat steps 2 and 3 until no new cluster head pairing is generated;

[0035] Step 5: If there are unpaired users, data transmission is performed according to the SIMO mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic diagram of the downlink multi-user massive antenna array NOMA system beam of the present invention;

[0037] Figure 2 This is the uplink virtual MIMO user pairing scenario of the present invention;

[0038] Figure 3 The relationship between the energy consumption and the number of users of the present invention;

[0039] Figure 4 The relationship between the energy consumption and the task migration size of the present invention; DETAILED DESCRIPTION

[0040] The present invention is further described below in conjunction with the accompanying drawings, but the present invention is not limited to the following embodiments.

[0041] like Figure 1 As shown, a task migration method based on joint uplink and downlink user clustering includes the following steps:

[0042] Step 1: construct a dual-mode base station and K mobile devices to form a hybrid transmission system of centimeter waves and millimeter waves, using centimeter wave transmission in uplink transmission and millimeter wave transmission in downlink transmission;

[0043] Step 2: Uplink task migration virtual MIMO transmission initialization: For tasks that cannot be performed locally, the single-antenna user terminal uses virtual MIMO uplink to migrate local tasks to the + base station edge server. For the initial uplink MIMO pairing, a random pairing method is used for transmission;

[0044] Step 3: The base station edge server calculates the task and transmits the result back to the user terminal through the downlink transmission of the massive antenna millimeter wave NOMA. The downlink sets hybrid precoding to reduce the number of required RF chains, allowing one beamforming vector to serve multiple users at the same time. When updating the K-means cluster center, the clustering criterion is the user channel direction angle. The cluster center of each user cluster is updated to the user with the lowest correlation with other user clusters. The correlation between the user and other user clusters is defined as the sum of the normalized channel correlations of the user and other cluster users, that is:

[0045]

[0046] Among them, M (k) represents the cluster containing user k. The similarity of user channel vectors can be measured using the normalized direction, i.e.:

[0047]

[0048] Improved K-means is used to cluster user devices, NOMA technology is used for transmission within the cluster, and wireless energy transmission is used to charge terminal users; users of downlink large-scale antenna millimeter waves are clustered and NOMA technology is used for transmission within the cluster, including:

[0049] 1) The number of mobile users is K. Calculate the channel gain a of K users k =||h k ||2;

[0050] 2) Downlink channel vector h k ,k∈1,...,K,

[0051] 3) Arrange the user channels in reverse order according to their channel gains. The channel gain set is A = {a1, a2, ..., a k}, select the user with the largest channel gain as the cluster center of a cluster, the cluster center set is H, and the remaining users are placed in the set

[0052] 4) Initialize the candidate cluster center set l = 1;

[0053] 5) Initial threshold δ, when Ω is not empty, users whose channel correlation with the selected cluster center is less than the threshold δ are selected as candidate cluster centers. The candidate cluster center set is In Ω, select the user with the largest channel gain as the center of a cluster and put it into H, and the rest of the users are put into the set l = l + 1;

[0054] 6) When Ω is empty, increase the threshold by δ = δ + (1-δ) / 10, and update the candidate cluster center set.

[0055] 7) The number of clustered beams is L. When l<L, repeat steps 5) to 6);

[0056] 8) When l=L, select L initial cluster core user sets H;

[0057] 9) Compare the angle difference between each user and the L initial cluster center users, and include user i into the cluster with the smallest angle difference.

[0058] 10) The cluster center of the lth cluster is updated as:

[0059]

[0060] 11) Repeat steps 9) to 10) to update the members in the cluster until the result does not change;

[0061] Step 4: When the user terminal receives the task result returned by the downlink and needs to migrate the task again, it selects users in different clusters from the downlink clustering result to perform virtual MIMO pairing and migrate the uplink task;

[0062] The uplink centimeter-wave virtual MIMO user pairing algorithm will be based on the downlink millimeter-wave user clustering, including:

[0063] 1) Arrange the L cluster users obtained by the downlink improved K-means algorithm in reverse order of their channel gains. The channel gain set of the lth cluster user is l∈1,2,...,L, the user with the largest channel gain is the cluster head of the cluster, and the cluster head group is H;

[0064] 2) In the cluster head cluster H, the base station randomly selects two users i and j for pairing;

[0065] 3) According to the channel gain set A i , A j The users in the two clusters are paired in pairs according to the user order, until the users in one of the clusters are paired;

[0066] 4) Repeat steps 2)-3) until no new pairing is generated;

[0067] Step 5: Repeat steps 2 and 3 until the user terminal confirms that the task is completed.

[0068] The base station is equipped with millimeter wave and centimeter wave antenna arrays.

[0069] The mobile device is equipped with a millimeter wave antenna, a centimeter wave antenna and a power split receiver for SWIPT technology.

[0070] Example

[0071] Task migration algorithm based on joint uplink and downlink user clustering:

[0072] The mobile user equipment is located around L parent points with R l (l∈L) is a radius independent distribution. Assume is the user position in the lth cluster Relative to the probability density function of the parent point,

[0073]

[0074] In the downlink massive antenna array NOMA system, such as Figure 1As shown, multiple receiving antennas of different terminals with different channel gains are grouped into NOMA clusters, and all user receiving antennas in each cluster are scheduled based on NOMA.

[0075] In the downlink millimeter wave large antenna array NOMA system, the base station is equipped with N mm The number of transmitting antennas is used for beamforming. The total number of mobile devices in the cell is K, and each device is equipped with a millimeter wave receiving antenna. Assume that the mobile devices are divided into L clusters. Since there are only N RF different simulated precoding beam vectors are available, so L = N RF , M l represents the set of users served by the lth beam. Each beam can serve multiple user devices. The channel modeling of the kth user in the downlink millimeter wave system can be expressed as:

[0076]

[0077] in, is the complex channel gain, d k represents the distance from the mobile device to the base station, η mm,LOS represents the path loss exponent of the millimeter wave channel line-of-sight, Represents the array direction vector:

[0078]

[0079] in is the departure angle, λ is the wavelength, and d=λ / 2 is the spacing between the antennas.

[0080] For the mth (m∈M l The received signal of each user is modeled as:

[0081]

[0082] Among them, s l,m Represents the transmission signal, p l,m represents the transmission power, v l,m represents the noise of the mth user in the lth beam. A is the analog precoding matrix, and d is the digital precoding matrix.

[0083] There is inter-cluster interference and intra-cluster interference in NOMA user clusters. In order to minimize network interference and maximize system capacity, a “robust” millimeter-wave massive antenna array NOMA user clustering algorithm will be designed by utilizing the channel gain differences and correlations between NOMA users. Let h i and h j denote the downlink millimeter wave channel vectors of user i and user j respectively, then the channel difference and correlation between any two users can be defined as:

[0084]

[0085] Since the considered mmWave massive antenna array NOMA system uses a single beam to communicate with a cluster of users, the users in the cluster should have strong channel correlation. In mmWave systems, the cosine similarity of two user channel vectors is proven to be an effective metric for determining the similarity of two user channels, as shown below:

[0086]

[0087] The derivation steps of this formula follow the definition of Feyer kernel. When the input parameter increases, the Feyer kernel converges to zero quickly, which means that the similarity of two user channel vectors can be measured by the normalized direction, such as and When the difference in the normalized direction is zero, that is, in the same normalized direction, the kernel value increases, and when the difference increases, the kernel value decreases.

[0088] Due to the combined characteristics of user clustering and the non-convexity of power allocation in the millimeter wave NOMA system, machine learning is used to quickly locate the global optimal solution. On the other hand, an appropriate real scenario is designed for the user spatial distribution model, in which users are located close to each other, and the user distribution is modeled as a mixture model of Gaussian distribution using the Poisson cluster process to form real clusters. People are distributed in clusters and there are centroids. The K-means algorithm is more applicable when dealing with user distribution scenario clustering. The traditional millimeter wave NOMA user clustering K-means algorithm requires the number of clusters to run the algorithm, and the clustering results depend on the random selection of the initial cluster center and the cluster center update method. Since the number of users K is greater than the number of RF chains N RF , and at the same time only N RF different simulated precoding vectors are available, so the number of clusters is determined by the number of RF chains N RF Determination. In the improved K-means algorithm of the present invention, when selecting the initial cluster center, the principle of maximizing the channel gain of the cluster center and minimizing the channel similarity of the cluster center users is followed. First, the users are sorted according to the channel gain, and the user with the largest channel gain is selected as the cluster center of the first cluster. Then a threshold is set as the maximum value of the channel similarity between the cluster centers. The threshold should be as small as possible at the beginning to ensure that there is lower channel interference between the initialized cluster center users. Then, the users whose channel similarity with the first cluster center is less than the threshold are included in the candidate cluster center set, and the user with the largest channel gain is selected as the new cluster center. Repeat the above steps. When the latest candidate cluster center set is empty, appropriately increase the threshold and lower the cluster center channel difference standard until a predetermined number of cluster centers are selected. Next, the selected cluster center users are used as the initial cluster centers for K-means user clustering.

[0089] When updating the K-means cluster center, the clustering criterion is the user channel direction angle. The cluster center of each user cluster is updated to the user with the lowest correlation with other user clusters. The correlation between a user and other user clusters is defined as the sum of the normalized channel correlations between the user and other cluster users, that is:

[0090]

[0091] When the clustering is completed, the user cluster set M is obtained. l ,l∈1,2,...L. Next, in each user cluster, the user with the largest channel gain is selected as the cluster head of the user cluster designed with the millimeter-wave large antenna array beam. The precoding is simulated according to the channel design of the cluster head user to obtain the antenna array gain of all beams. At the same time, in each cluster, the users are sorted in descending order according to the channel gain. Therefore, in the cluster composed of |M l In a MIMO-NOMA cluster consisting of | users, the strongest user is the first user (also defined as the cluster head), and the weakest user is the |Mth user. l | users. On the other hand, the difference in channel gain between the cluster head and other users in the MIMO-NOMA cluster is large enough, and the equivalent channel gain is close to the channel gain of the cluster head. Therefore, under this channel condition, the cluster head of each MIMO-NOMA cluster can almost completely eliminate inter-cluster interference. All cluster members except the cluster head can achieve good inter-cluster interference elimination effect. Afterwards, the digital precoding is designed according to the clustering of the remaining users.

[0092] Uplink centimeter-wave virtual MIMO user pairing algorithm based on downlink clustering:

[0093] Virtual MIMO technology can obtain additional multi-user diversity gain through reasonable user grouping strategies, achieving a significant increase in uplink throughput, such as Figure 2 As shown in Figure 1, the user pairing method is a key issue that directly affects its performance. In the traditional uplink and downlink user clustering and separation task migration system, the virtual MIMO pairing algorithm fails to achieve a good balance between computational complexity and system performance, resulting in high system complexity and low energy efficiency.

[0094] The virtual MlMO pairing method takes into account the orthogonality of the channel matrix, that is, it gives priority to pairing users with more orthogonal channel matrices. In the improved K-means algorithm, users are clustered into clusters with high correlation in channel relevance and geographical location, and users in different clusters have relatively low inter-cluster interference and high channel orthogonality.

[0095] Therefore, a low-complexity uplink virtual MIMO user pairing strategy is proposed.

[0096] The configuration of the uplink multi-user virtual MIMO system consists of K users, each with one antenna, sending to the base station through a flat fading channel. Pairing scheduling constructs a 2×2 virtual MIMO wireless channel between the K users and the base station. Time-frequency resource sharing is achieved through spatial multiplexing of the uplink to send information to the base station. The virtual MIMO system considered in this section can be understood as a point-to-point MIMO system with two transmit antennas and two receive antennas. The channel matrix can be described as:

[0097]

[0098] Among them, H 1w and H 2w are the complex channel matrices from the first user and the second user to the base station. ij represents the channel response factor from the jth transmit antenna to the ith receive antenna. The signal received by the base station can be described as:

[0099]

[0100] in

[0101]

[0102] Where x is the transmitted signal vector of the two pairs of users, and y is the received signal vector of the base station. t YesN t ×N r In the virtual MIMO pairing algorithm proposed in this section, the two paired users are located in different user clusters, and the distance between the user clusters is far enough, so this section ignores R t . N t and N r Respectively represent the number of transmitting antennas and the number of receiving antennas, N t =N r =2. n is a normalized complex additive Gaussian white noise vector with a mean of zero, where I is the identity matrix.

[0103] The user clusters obtained by the improved K-means algorithm for downlink have clustering characteristics in terms of geographical location. Users in different clusters are far apart, with low channel correlation and relatively high orthogonality. Therefore, virtual MIMO pairing between users in different clusters can achieve a higher uplink system capacity.

[0104] Performance evaluation:

[0105] In this simulation, we simulated by using MATLAB_2018. When modeling the user location, we considered the Poisson clustering process to capture the spatial correlation between base stations and mobile devices. Mobile user devices are clustered around L parent points with Rl (l∈L) is a radius independent distribution, where the parent point radius of the mobile user based on the Poisson cluster is 5m. When clustering downlink, the initial threshold of channel similarity is set to 0.3. Setting the channel similarity as small as possible can ensure the difference of the selected initial cluster center users. The centimeter wave bandwidth is set to 20MHz and the carrier frequency is 2GHz. The millimeter wave bandwidth is set to 200MHz and the carrier frequency is 28GHz. The experimental parameters are detailed in Table 1.

[0106] Table 1 Simulation parameters

[0107]

[0108]

[0109] The proposed algorithm is compared with the following two methods:

[0110] 1) The downlink uses the traditional K-means user clustering algorithm, while the uplink users are not paired and centimeter-wave multi-user OFDM technology is used for transmission

[0111] 2) Uplink centimeter-wave users are randomly paired, and downlink millimeter-wave users are randomly clustered.

[0112] Figure 3 The relationship between energy consumption and the number of users is shown, where the transmission task B k The size is 5×10 6 bits. It can be found that compared with the other two schemes, the task migration algorithm based on joint uplink and downlink user clustering proposed in this chapter can effectively reduce the energy consumption of mobile terminals during edge computing task migration. The relationship between the energy-delay product and the number of users is shown. Compared with the other two schemes, the proposed user clustering algorithm can effectively improve the uplink and downlink system throughput, reduce the task migration delay, and achieve a lower energy-delay product than the other two schemes.

[0113] Figure 4 The relationship between the energy consumption and energy-delay product of the proposed joint uplink and downlink user clustering algorithm and the transmission task size is shown when the number of users is 8. The simulation results show that compared with the typical uplink and downlink user clustering and separation task migration algorithm, this chapter improves the K-means user clustering for downlink users and performs uplink user virtual MIMO pairing based on the downlink user clustering results, which can achieve lower energy consumption and energy-delay product.

Claims

1. A task migration method based on joint uplink and downlink user clustering, characterized in that: The following steps are involved: Step 1: construct a dual-mode base station and K mobile devices to form a hybrid transmission system of centimeter waves and millimeter waves, using centimeter wave transmission in uplink transmission and millimeter wave transmission in downlink transmission; Step 2: Uplink task migration virtual MIMO transmission initialization: For tasks that cannot be performed locally, the single-antenna user terminal uses virtual MIMO uplink to migrate local tasks to the + base station edge server. For the initial uplink MIMO pairing, a random pairing method is used for transmission; Step 3: The base station edge server calculates the task and transmits the result back to the user terminal through the downlink transmission of the massive antenna millimeter wave NOMA. The downlink sets hybrid precoding to reduce the number of required RF chains, allowing one beamforming vector to serve multiple users at the same time. When updating the K-means cluster center, the clustering criterion is the user channel direction angle. The cluster center of each user cluster is updated to the user with the lowest correlation with other user clusters. The correlation between the user and other user clusters is defined as the sum of the normalized channel correlations of the user and other cluster users, that is: Among them, M (k) represents the cluster containing user k. The similarity of user channel vectors can be measured using the normalized direction, i.e.: Improved K-means is used to cluster user devices, NOMA technology is used for transmission within the cluster, and wireless energy transmission is used to charge terminal users; users of downlink large-scale antenna millimeter waves are clustered and NOMA technology is used for transmission within the cluster, including: 1) The number of mobile users is K. Calculate the channel gain a of K users k =||h k ||2; 2) Downlink channel vector h k ,k∈1,...,K, 3) Arrange the user channels in reverse order according to their channel gains. The channel gain set is A = {a1, a2, ..., a k }, select the user with the largest channel gain as the cluster center of a cluster, the cluster center set is H, and the remaining users are placed in the set 4) Initialize the candidate cluster center set l = 1; 5) Initial threshold δ, when Ω is not empty, users whose channel correlation with the selected cluster center is less than the threshold δ are selected as candidate cluster centers. The candidate cluster center set is In Ω, select the user with the largest channel gain as the center of a cluster and put it into H, and the rest of the users are put into the set l = l + 1; 6) When Ω is empty, increase the threshold by δ = δ + (1-δ) / 10, and update the candidate cluster center set. 7) The number of clustered beams is L. When l<L, repeat steps 5) to 6); 8) When l=L, select L initial cluster core user sets H; 9) Compare the angle difference between each user and the L initial cluster center users, and include user i into the cluster with the smallest angle difference; 10) The cluster center of the lth cluster is updated as: 11) Repeat steps 9) to 10) to update the members in the cluster until the result does not change; Step 4: When the user terminal receives the task result returned by the downlink and needs to migrate the task again, it selects users in different clusters from the downlink clustering result to perform virtual MIMO pairing and migrate the uplink task; The uplink centimeter-wave virtual MIMO user pairing algorithm will be based on the downlink millimeter-wave user clustering, including: 1) Arrange the L cluster users obtained by the downlink improved K-means algorithm in reverse order of their channel gains. The channel gain set of the lth cluster user is The user with the largest channel gain is the cluster head of the cluster, and the cluster head group is H; 2) In the cluster head cluster H, the base station randomly selects two users i and j for pairing; 3) According to the channel gain set A i , A j The users in the two clusters are paired in pairs according to the user order, until the users in one of the clusters are paired; 4) Repeat steps 2)-3) until no new pairing is generated; Step 5: Repeat steps 2 and 3 until the user terminal confirms that the task is completed.

2. The task migration method based on joint uplink and downlink user clustering according to claim 1, characterized in that: The base station is equipped with millimeter wave and centimeter wave antenna arrays.

3. The task migration method based on joint uplink and downlink user clustering according to claim 1, characterized in that: The mobile device is equipped with a millimeter wave antenna, a centimeter wave antenna and a power split receiver for SWIPT technology.

Images