A clustering method and system based on long short-term memory and attention

Abstract

The present invention proposes a clustering method and system based on long short-term memory and attention, which relates to the technical field of wireless communication networks. The method includes: a user device measures the channel state information of a radio unit it serves and several radio units adjacent to the radio unit it serves, and feeds back the channel state information to each radio unit; calculates the current load factor of each radio unit, and uses a cloud computing platform to pre-process the load factor and the channel state information of several radio units to obtain the channel state information of M radio units with the strongest received signals; uses the channel state information of the M radio units as the input of a spectrum efficiency prediction model based on long short-term memory and attention, and outputs the spectrum efficiency of the user device; clusters the radio units communicating with the user device based on the spectrum efficiency of the user device. The present invention effectively reduces the computational complexity, improves the clustering efficiency, and can be expanded to large-scale networks.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless communication networks, and in particular to a clustering method and system based on long short-term memory and attention. Background Art

[0002] One of the most critical challenges facing the next generation of wireless communication networks is the need to support a wide range of new applications, such as augmented reality, 3D holographic telepresence, and industrial robotics. Against this backdrop, B5G wireless networks can provide enhanced flexibility through dynamic allocation of system resources while achieving a perceived zero-latency vision to ensure a smooth transition to 6G wireless networks. In order to meet the above requirements, increasingly dense network deployment is a very promising technology. Since cell-free networks can provide the high spectrum efficiency required for 6G, combining dense deployment with cell-free networks can provide future 6G networks with extremely high flexibility while meeting various challenges in next-generation wireless networks.

[0003] The corresponding standards of open radio access network (O-RAN) are transforming access networks into open, intelligent, virtualized and fully interoperable architectures, which brings the required flexibility and intelligent control opportunities to cell-free networks. At present, most of the research work on cell-free wireless networks is based on centralized central processing units. The basic assumption of this architecture is that there are single or multiple central processing units in the network, which are connected to access point (AP) units distributed in the network through forward and backhaul links. In this case, the design of access nodes is relatively simple because they only need to receive signals and transmit them to the central processing unit through forward and backhaul links. However, for larger networks with more access nodes and users, the centralized approach means that the access nodes and share all user-related data signals, and the required fronthaul capacity and computing resources grow nonlinearly with the number of users. This means that the centralized approach of large-scale antenna systems without cells is very challenging for practical networks, that is, it is impossible to expand to large-scale networks. The prior art discloses a user-centric AP clustering scheme based on user trajectory prediction, taking into account user channels, interference and loads, and obtains the spectrum efficiency of user equipment UE in a centralized manner. However, this method has two prerequisites: on the one hand, all radio units provide services for user equipment UE; on the other hand, the channel state information and transmission power estimated by all radio units need to be transmitted to a centralized processing module. The above prerequisites lead to high computational complexity and low clustering efficiency when the scheme is implemented, and it cannot be extended to large-scale networks. Summary of the invention

[0004] In order to solve the problems of high computational complexity, low clustering efficiency and inability to be expanded to large-scale networks in the above-mentioned prior art, the present invention proposes a clustering method and system based on long short-term memory and attention, which effectively reduces the computational complexity, improves the clustering efficiency, and can be expanded to large-scale networks.

[0005] In order to achieve the above technical effects, the technical solution of the present invention is as follows:

[0006] A clustering method based on long short-term memory and attention comprises the following steps:

[0007] S1. Use the user equipment UE to measure the channel state information of its serving radio unit O-RU and several radio units O-RU adjacent to the current serving radio unit O-RU, and feed back the measured channel state information to each radio unit O-RU;

[0008] S2. Calculate the current load factor of each radio unit O-RU using the radio unit O-RU, and transmit the load factor and the channel state information of each radio unit O-RU to the cloud computing platform;

[0009] S3. Preprocess the load factor and the channel state information of several radio units O-RU using the cloud computing platform to obtain the channel state information of M radio units O-RU with the strongest received signals, where M is an integer;

[0010] S4. Using the channel state information of the M radio units O-RU as the input of a pre-trained spectrum efficiency prediction model based on long short-term memory and attention, and outputting the spectrum efficiency of the user equipment UE by the spectrum efficiency prediction model;

[0011] S5. Clustering the radio units O-RUs for communication with the user equipment UE based on the spectrum efficiency of the user equipment UE.

[0012] Preferably, the preprocessing of the load factor and the channel state information of the plurality of radio units O-RUs by using the cloud computing platform comprises:

[0013] S31. Determine whether the load factor is greater than a preset load factor threshold, and if so, filter out the radio unit O-RU whose load factor is not greater than the load factor threshold;

[0014] S32. Based on the screened radio units O-RU, the channel state information of M radio units O-RU with the strongest received signals is preferably selected from the channel state information of the plurality of radio units O-RU.

[0015] Preferably, the spectral effect prediction model includes an embedding layer, an attention network layer, a long short-term memory network layer and an activation function layer, the output end of the embedding layer is connected to the input end of the attention network layer, the output end of the attention network layer is connected to the input end of the long short-term memory network layer, and the output end of the long short-term memory network layer is connected to the activation function layer.

[0016] Preferably, the using the channel state information of the M radio units O-RU as the input of a pre-trained spectrum efficiency prediction model based on long short-term memory and attention, and the spectrum efficiency prediction model outputting the spectrum efficiency of the user equipment UE, comprises:

[0017] S41. The embedding layer receives the channel state information of the M radio units O-RU, converts the channel state information of the M radio units O-RU into vector features and inputs them into the attention network layer;

[0018] S42. The attention network layer receives the vector features and outputs a first vector feature sequence to the long short-term memory network layer;

[0019] S43. The long short-term memory network layer receives the first vector feature sequence, outputs a second vector feature sequence, the second vector feature sequence is regularized and input into the activation function layer, and the activation function layer outputs the spectrum efficiency of the M radio units O-RU.

[0020] Preferably, the long short-term memory network layer includes a first unidirectional long short-term memory network, a regularization layer and a second unidirectional long short-term memory network, the output end of the attention network layer is connected to the input end of the first unidirectional long short-term memory network, the output end of the first unidirectional long short-term memory network is connected to the input end of the regularization layer, the output end of the regularization layer is connected to the input end of the second unidirectional long short-term memory network, and the output end of the second unidirectional long short-term memory network is connected to the activation function layer.

[0021] Preferably, clustering the user equipment UE based on the spectrum efficiency of the user equipment UE includes:

[0022] S51. Determine whether the gain of the spectrum efficiency of the user equipment UE predicted by the M radio units O-RU relative to the spectrum efficiency of the user equipment UE predicted by the M-1 radio units O-RU is greater than a preset gain threshold, if so, execute S52; if not, stop clustering the user equipment UE;

[0023] S52. Add the radio unit O-RU with the strongest received signal among the remaining M radio units O-RU as the radio unit O-RU serving the user equipment UE, and record the radio unit O-RU serving the user equipment UE as the service cluster corresponding to the user equipment UE.

[0024] The present invention also proposes a clustering system based on long short-term memory and attention, comprising: a plurality of user equipments UE, a plurality of radio units O-RU and a cloud computing platform;

[0025] A user equipment UE is configured to measure channel state information of a serving radio unit O-RU and a plurality of radio units O-RU adjacent to the currently serving radio unit O-RU, and feed back the measured channel state information to the respective radio units O-RU;

[0026] A radio unit O-RU, configured to calculate a current load factor of each of the radio units O-RU, and transmit the load factor and channel state information of a plurality of radio units O-RU to a cloud computing platform;

[0027] The cloud computing platform is used to pre-process the load factor and the channel state information of several radio units O-RU to obtain the channel state information of M radio units O-RU with the strongest received signals, wherein M is an integer. The channel state information of the M radio units O-RU is used as the input of a pre-trained spectrum efficiency prediction model based on long short-term memory and attention. The spectrum efficiency prediction model outputs the spectrum efficiency of the user equipment UE. Based on the spectrum efficiency of the user equipment UE, the radio units O-RU communicating with the user equipment UE are clustered.

[0028] Preferably, when the user equipment UE measures channel state information of several radio units O-RU adjacent to the user equipment UE, the user equipment UE selects one radio unit O-RU from the several radio units O-RU as a serving radio unit O-RU, and the serving radio unit O-RU is used to determine a measurement cluster of the user equipment UE, and the measurement cluster includes several radio units O-RU adjacent to the serving radio unit O-RU.

[0029] Preferably, the cloud computing platform includes a centralized unit O-CU for upper-layer network operations and a distributed unit O-DU for lower-layer network operations, and the centralized unit O-CU and the distributed unit O-DU are bidirectionally connected.

[0030] The present invention also proposes a computer device, comprising: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus;

[0031] The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform operations of the clustering method based on long short-term memory and attention.

[0032] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0033] The present invention proposes a clustering method and system based on long short-term memory and attention. First, a user equipment UE measures the channel state information of a radio unit O-RU it serves and several radio units O-RU adjacent to the currently serving radio unit O-RU. Then, a cloud computing platform is used to pre-process the current load factors of the radio units O-RU and the channel state information of several radio units O-RU. During the processing, the channel state information, interference and load of the user equipment UE are considered to obtain the channel state information of M radio units O-RU with the strongest received signals. Then, the channel state information of the M radio units O-RU is used as the input of a spectrum efficiency prediction model based on long short-term memory and attention. The spectrum efficiency is output by the spectrum efficiency prediction model. The radio units O-RU communicating with the user equipment UE are clustered based on the spectrum efficiency, which effectively reduces the computational complexity, improves the clustering efficiency, and can be expanded to large-scale networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A flow chart of a clustering method based on long short-term memory and attention proposed in an embodiment of the present invention is shown;

[0035] Figure 2 Another flow chart of a clustering method based on long short-term memory and attention proposed in an embodiment of the present invention is shown;

[0036] Figure 3 A structural diagram showing a spectral effect prediction model proposed in an embodiment of the present invention;

[0037] Figure 4 A block diagram of a clustering system based on long short-term memory and attention proposed in an embodiment of the present invention is shown;

[0038] Figure 5 A structural diagram of a computer device proposed in an embodiment of the present invention is shown.

[0039] 501. Processor; 502. Memory; 503. Communication interface; 504. Communication bus; 505. Executable instructions. DETAILED DESCRIPTION

[0040] The accompanying drawings are only used for illustrative purposes and are not to be construed as limiting the present invention;

[0041] It is understandable to those skilled in the art that some well-known contents may be omitted in the drawings;

[0042] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0043] Example 1

[0044] like Figure 1 As shown, this embodiment proposes a clustering method based on long short-term memory and attention, including the following steps:

[0045] S1. Use the user equipment UE to measure the channel state information of its serving radio unit O-RU and several radio units O-RU adjacent to the current serving radio unit O-RU, and feed back the measured channel state information to each radio unit O-RU;

[0046] S2. Calculate the current load factor of each radio unit O-RU using the radio unit O-RU, and transmit the load factor and the channel state information of each radio unit O-RU to the cloud computing platform;

[0047] S3. Preprocess the load factor and the channel state information of several radio units O-RU using the cloud computing platform to obtain the channel state information of M radio units O-RU with the strongest received signals, where M is an integer;

[0048] In S3, see Figure 1 and Figure 2 , the preprocessing of the load factor and the channel state information of the plurality of radio units O-RU by using the cloud computing platform includes:

[0049] S31. Determine whether the load factor is greater than a preset load factor threshold, and if so, filter out the radio unit O-RU whose load factor is not greater than the load factor threshold;

[0050] S32. Based on the screened radio units O-RU, the channel state information of M radio units O-RU with the strongest received signals is preferably selected from the channel state information of the plurality of radio units O-RU.

[0051] S4. Using the channel state information of the M radio units O-RU as the input of a pre-trained spectrum efficiency prediction model based on long short-term memory and attention, and outputting the spectrum efficiency of the user equipment UE by the spectrum efficiency prediction model;

[0052] In S4, see Figure 3The spectral effect prediction model includes an embedding layer, an attention network layer, a long short-term memory network layer and an activation function layer, wherein the output end of the embedding layer is connected to the input end of the attention network layer, the output end of the attention network layer is connected to the input end of the long short-term memory network layer, and the output end of the long short-term memory network layer is connected to the activation function layer; wherein the attention network layer adopts a Transformer network, the long short-term memory network layer adopts an LSTM network, and the activation function layer adopts a softmax activation function;

[0053] The method of using the channel state information of the M radio units O-RU as an input of a pre-trained spectrum efficiency prediction model based on long short-term memory and attention, and outputting the spectrum efficiency of the user equipment UE by the spectrum efficiency prediction model, comprises:

[0054] S41. The embedding layer receives the channel state information of the M radio units O-RU, converts the channel state information of the M radio units O-RU into vector features and inputs them into the attention network layer; wherein the channel state information of each radio unit O-RU includes the information of the most recent t moments;

[0055] S42. The attention network layer receives the vector features and outputs a first vector feature sequence to the long short-term memory network layer;

[0056] S43. The long short-term memory network layer receives the first vector feature sequence, outputs a second vector feature sequence, the second vector feature sequence is regularized and input into the activation function layer, and the activation function layer outputs the spectrum efficiency of the M radio units O-RU.

[0057] The long short-term memory network layer includes a first unidirectional long short-term memory network, a regularization layer dropout and a second unidirectional long short-term memory network, the output end of the attention network layer is connected to the input end of the first unidirectional long short-term memory network, the output end of the first unidirectional long short-term memory network is connected to the input end of the regularization layer, the output end of the regularization layer is connected to the input end of the second unidirectional long short-term memory network, and the output end of the second unidirectional long short-term memory network is connected to the activation function layer; the long short-term memory network layer adopts an LSTM network, which is the basis for sequential data understanding. The first unidirectional long short-term memory network and the second unidirectional long short-term memory network both include a number of LSTM storage units, and these single LSTM storage units are good at capturing complex time dynamics, ensuring that key historical data is retained while being sensitive to new information;

[0058] The attention network layer uses the Transformer network to enhance the long short-term memory network layer. The Transformer network is the key to contextual understanding. The attention network layer is a collection of multiple self-attention heads. Each head pays special attention to different aspects of the data sequence to ensure overall understanding. Figure 2 The spectral effect prediction model shown performs well in capturing sequential patterns and broader contextual information, and this ability is mainly attributed to the synergistic fusion of LSTM memory units and the self-attention mechanism of the Transformer network; the LSTM memory unit captures long-term dependencies, while the self-attention mechanism of the Transformer network understands the context at both fine-grained and macro-levels.

[0059] S5. Clustering the radio units O-RUs for communication with the user equipment UE based on the spectrum efficiency of the user equipment UE.

[0060] The user equipment UE is clustered based on the spectrum efficiency of the user equipment UE, see Figure 2 ,include:

[0061] S51. Determine whether the gain of the spectrum efficiency of the user equipment UE predicted by the M radio units O-RU relative to the spectrum efficiency of the user equipment UE predicted by the M-1 radio units O-RU is greater than a preset gain threshold, where the gain threshold is 20%. If so, execute S52; if not, stop clustering the user equipment UE;

[0062] S52. Add the radio unit O-RU with the strongest received signal among the remaining M radio units O-RU as the radio unit O-RU serving the user equipment UE, and record the radio unit O-RU serving the user equipment UE as the service cluster corresponding to the user equipment UE.

[0063] It should be noted that, for clustering the user equipment UE, a user-centric AP clustering solution is proposed here based on the trajectory prediction of the user equipment UE, while considering the user channel, interference and load;

[0064] If the minimum mean square error (MMSE) channel estimation is used for the channel state information of the user equipment UE, the signal-to-noise ratio of the k-th user equipment is expressed as:

[0065]

[0066] Among them, p k is the transmission power of the kth user equipment UE, h kl is the channel state information between the kth user equipment UE and the lth radio unit O-RU estimated based on the minimum mean square error, w klis the beamforming vector of the kth user equipment UE, D kl is the selection matrix between the kth user equipment UE and the lth radio unit O-RU, E is the expectation, σ is the error term, if the radio unit O-RU provides service for the user equipment UE, then D kl =I N Otherwise, D kl =0; To maximize the signal-to-noise ratio in formula (1), w k The values ​​are as follows:

[0067]

[0068] Among them, p kl is the transmission power used by the lth radio unit O-RU to transmit to the kth user equipment UE, C il is the covariance matrix of the channel state information transmitted between the lth radio unit O-RU and the kth user equipment UE. Based on formula (1) and formula (2), the spectrum efficiency ω of the kth user equipment UE is obtained. k for:

[0069] ω k =E[ln(1+SINR k )](3)

[0070] Although the user spectrum efficiency obtained by the centralized method above is the best, it also has two prerequisites:

[0071] First, it is assumed that all O-RUs provide services to users;

[0072] Second, the channel state information and transmit power estimated by all O-RUs need to be transmitted to a centralized processing module, where the centralized processing module is an O-CU serving the user equipment UE;

[0073] The above prerequisites require the following reasons that lead to high computational complexity of the algorithm:

[0074] 1) To invert the matrix as in formula (2), we need to calculate the inverse of an N×N matrix. The complexity of matrix inversion is In actual dense deployment, N is usually 32 or 64, or larger, and N is obtained. 3 =32,768, the computational complexity is difficult to achieve.

[0075] 2) The amount of information transmitted is large, and the bandwidth requirement for the front-haul of O-RAN is large, which is not feasible in actual O-RAN deployment.

[0076] Therefore, although the centralized processing algorithm can obtain the optimal spectrum efficiency, it has high computational complexity and high information transmission, and is not feasible in large-scale network applications. In order to meet the feasibility requirements of actual large-scale networks without sacrificing performance, this embodiment proposes a clustering method based on long short-term memory and attention. First, the user equipment UE measures the channel state information of the radio unit O-RU it serves and several radio units O-RU adjacent to the currently serving radio unit O-RU. Then, the cloud computing platform is used to pre-process the current load factors of the radio units O-RU and the channel state information of several radio units O-RU. In the processing process, the channel state information, interference and load of the user equipment UE are considered to obtain the channel state information of the M radio units O-RU with the strongest received signals. Then, the channel state information of the M radio units O-RU is used as the input of the spectrum efficiency prediction model based on long short-term memory and attention, and the spectrum efficiency is output by the spectrum efficiency prediction model. The radio units O-RU communicating with the user equipment UE are clustered based on the spectrum efficiency, which effectively reduces the computational complexity, improves the clustering efficiency, and can be expanded to large-scale networks. In this embodiment, a distributed architecture based on O-RAN is proposed, where each (higher-level) access node locally estimates the channels of its associated users and uses this information to locally process data signals using a federated learning scheme.

[0077] Example 2

[0078] Considering a cell-free network, a clustering scheme centered on user equipment UE, the user equipment UE is served by nearby distributed antenna units, which provide the best channel, for example, the lowest path loss. If only the path loss of the user equipment UE is considered, the purpose is to maximize the spectrum efficiency of the user equipment UE. In an actual system, the mobile channel of the user equipment UE changes rapidly, and at the same time, the loads of different access nodes are different. Even if the access node adjacent to the user equipment UE has a good channel, it cannot provide service for it due to its heavy load, resulting in access failure. Therefore, this embodiment proposes a clustering system based on long short-term memory and attention, which considers a cell-free large-scale MIMO system built on the O-RAN architecture. The system is in line with the next-generation virtualized C-RAN (centralized radio access network) ecosystem, see Figure 3 , the clustering system includes: K user equipments UE, L radio units O-RU and a cloud computing platform; wherein the K user equipments UE are arbitrarily distributed in the coverage area; each radio unit O-RU is equipped with N antennas, K, L and N are integers;

[0079] A user equipment UE is configured to measure channel state information of a serving radio unit O-RU and a plurality of radio units O-RU adjacent to the currently serving radio unit O-RU, and feed back the measured channel state information to the respective radio units O-RU;

[0080] A radio unit O-RU, configured to calculate a current load factor of each of the radio units O-RU, and transmit the load factor and channel state information of a number of radio units O-RU to a cloud computing platform; wherein all the radio units O-RU are connected to the cloud computing platform O-Cloud via a fronthaul, and have virtualization and processing resource sharing functions;

[0081] The cloud computing platform is used to pre-process the load factor and the channel state information of several radio units O-RU to obtain the channel state information of M radio units O-RU with the strongest received signals, wherein M is an integer. The channel state information of the M radio units O-RU is used as the input of a pre-trained spectrum efficiency prediction model based on long short-term memory and attention. The spectrum efficiency prediction model outputs the spectrum efficiency of the user equipment UE. Based on the spectrum efficiency of the user equipment UE, the radio units O-RU communicating with the user equipment UE are clustered.

[0082] The cloud computing platform includes a centralized unit O-CU for upper-layer network operations and a distributed unit O-DU for lower-layer network operations, and the centralized unit O-CU and the distributed unit O-DU are bidirectionally connected.

[0083] According to the O-RAN specification, the centralized unit O-CU is responsible for upper layer operations, which include functions of PDCP, RRC and other layers. The PDCP layer is a protocol located between the RRC layer and the RLC layer, and is mainly responsible for the security, reliability and efficiency of packet data transmission. In the O-RAN architecture, the centralized unit O-CU is responsible for the functions of the PDCP layer, which include compression, encryption, integrity protection, reordering, segmentation and header compression, etc. These functions ensure the efficiency, security and reliability of data transmission; the functions of the RRC layer include paging process, RRC connection establishment process, security activation process, RRC connection configuration process, RRC connection reestablishment process and release process, etc. These processes ensure the effective management and control of wireless resources. The centralized unit O-CU is responsible for processing upper layer operations in the O-RAN architecture, ensuring the efficiency, security and reliability of packet data transmission, as well as the effective management and control of wireless resources. Through the collaborative work with the distributed unit O-DU, the wireless access network function of the O-RAN architecture is jointly realized;

[0084] The distributed unit O-DU is responsible for the lower-layer network operations, which include the functions of MAC, RLC and PHY layers. The MAC layer is part of the data link layer and is responsible for establishing a logical connection between the physical layer and the network layer, and processing data encapsulation, decapsulation, addressing and error control functions. The RLC layer is the radio link control layer, which is located above the MAC layer and is mainly responsible for data segmentation and reassembly, retransmission and flow control functions. The PHY layer is the physical layer, which is the bottom layer of the communication system and is responsible for processing all functions related to the transmission medium, such as signal modulation, demodulation, encoding, decoding, synchronization, etc. In the O-RAN architecture, the distributed unit O-DU is responsible for processing the underlying network operations such as MAC, RLC and PHY, and works with the radio unit O-RU to realize the functions of the wireless access network. It supports multiple technologies and virtualization technologies, and has the characteristics of high performance and flexibility.

[0085] The cloud computing platform further includes a near-real-time RAN intelligent controller logical node (Near-RT RIC) and a non-real-time RAN intelligent controller logical node (Non-RT RIC). The near-real-time RAN intelligent controller logical node is connected to the non-real-time RAN intelligent controller logical node through an interface. The near-real-time RAN intelligent controller logical node is used to support near-real-time radio resource control and optimization. The non-real-time RAN intelligent controller logical node is used to support non-real-time radio resource control and optimization. The non-real-time RAN intelligent controller logical node is located in a service management and orchestration unit (SMO). The service management and orchestration unit is responsible for non-real-time intelligent orchestration. The two logical nodes support fully virtualized end-to-end resource optimization.

[0086] RAN has multiple deployment options, in some of which the O-DU and O-CU are co-located, while in some deployment options they can be logically and geographically separated;

[0087] When the user equipment UE measures channel state information of several radio units O-RU adjacent to the user equipment UE,

[0088] That is, when the user equipment UE is connected to the wireless network for the first time,

[0089] The user equipment UE selects a radio unit O-RU from several radio units O-RU as a serving radio unit O-RU, and the distributed unit O-DU controlling the serving radio unit O-RU logically becomes its corresponding main distributed unit O-DU; this selection is that the user equipment UE calculates the channel state information based on the downlink pilot; the serving radio unit O-RU is used to determine the measurement cluster of the user equipment UE. The concept of the serving radio unit O-RU is very important because the serving radio unit O-RU determines the measurement cluster, that is, the set of O-RUs that measure the channel state information of the user equipment UE. For the kth user equipment UE, its measurement cluster is represented as The measurement cluster A plurality of radio units O-RUs are included which are adjacent to the serving radio unit O-RU.

[0090] Then, the service cluster serving the kth user equipment UE is represented by That is, the set of radio units O-RU to which data is transmitted, where for Subset; this service cluster Joint precoding / decoding is performed for the kth user equipment UE; these radio units O-RU are in It can belong to one or more distributed units O-DU. If it belongs to multiple distributed units O-DU, the distributed units O-DU need to cooperate, that is, the distributed units O-DU need to transmit the corresponding joint precoding information; when the user equipment UE moves, the radio unit O-RU cluster that measures and serves it will change continuously. The change of the service cluster is called the switching of the service cluster. The service cluster is a collection of service clusters;

[0091] In this embodiment, first, the user equipment UE measures the channel state information of the radio unit O-RU it serves and several radio units O-RU adjacent to the currently served radio unit O-RU, and then uses the cloud computing platform to pre-process the current load factors of the radio units O-RU and the channel state information of several radio units O-RU. In the processing process, the channel state information, interference and load of the user equipment UE are considered to obtain the channel state information of M radio units O-RU with the strongest received signals, and then the channel state information of the M radio units O-RU is used as the input of the spectrum efficiency prediction model based on long short-term memory and attention, the spectrum efficiency is output by the spectrum efficiency prediction model, and the radio units O-RU communicating with the user equipment UE are clustered based on the spectrum efficiency, which effectively reduces the calculation complexity, improves the clustering efficiency, and can be expanded to large-scale networks.

[0092] Example 3

[0093] See also Figure 4 This embodiment also proposes a computer device, see Figure 5 , comprising: a processor 501, a memory 502, a communication interface 503 and a communication bus 504, wherein the processor 501, the memory 502 and the communication interface 503 communicate with each other through the communication bus 504;

[0094] Wherein: the processor 501, the memory 502 and the communication interface 503 communicate with each other through the communication bus 504. The communication interface 503 is used for network communication with other devices such as a client or other servers. The processor 501 is used to execute executable instructions 505, which can specifically execute the operations of the abnormal transaction data monitoring method. Specifically, the executable instructions 505 may include program codes. The processor 501 may be a central processing unit CPU, or an application specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement an embodiment of the present invention. The computer device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0095] The memory 502 is used to store executable instructions 505. The memory 502 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0096] The executable instruction 505 can be specifically called by the processor 501 to enable the computer device to perform the following operations:

[0097] S1. Use the user equipment UE to measure the channel state information of its serving radio unit O-RU and several radio units O-RU adjacent to the current serving radio unit O-RU, and feed back the measured channel state information to each radio unit O-RU;

[0098] S2. Calculate the current load factor of each radio unit O-RU using the radio unit O-RU, and transmit the load factor and the channel state information of each radio unit O-RU to the cloud computing platform;

[0099] S3. Preprocess the load factor and the channel state information of several radio units O-RU using the cloud computing platform to obtain the channel state information of M radio units O-RU with the strongest received signals, where M is an integer;

[0100] S4. Using the channel state information of the M radio units O-RU as the input of a pre-trained spectrum efficiency prediction model based on long short-term memory and attention, and outputting the spectrum efficiency of the user equipment UE by the spectrum efficiency prediction model;

[0101] S5. Clustering the radio units O-RUs for communication with the user equipment UE based on the spectrum efficiency of the user equipment UE.

[0102] In this embodiment, first, the user equipment UE measures the channel state information of the radio unit O-RU it serves and several radio units O-RU adjacent to the currently served radio unit O-RU, and then uses the cloud computing platform to pre-process the current load factors of the radio units O-RU and the channel state information of several radio units O-RU. In the processing process, the channel state information, interference and load of the user equipment UE are considered to obtain the channel state information of M radio units O-RU with the strongest received signals, and then the channel state information of the M radio units O-RU is used as the input of the spectrum efficiency prediction model based on long short-term memory and attention, the spectrum efficiency is output by the spectrum efficiency prediction model, and the radio units O-RU communicating with the user equipment UE are clustered based on the spectrum efficiency, which effectively reduces the calculation complexity, improves the clustering efficiency, and can be expanded to large-scale networks.

[0103] Obviously, the above embodiments of the present invention are only examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A clustering method based on long short-term memory and attention, characterized in that: The following steps are involved: S1. Use the user equipment UE to measure the channel state information of its serving radio unit O-RU and several radio units O-RU adjacent to the current serving radio unit O-RU, and feed back the measured channel state information to each radio unit O-RU; S2. Calculate the current load factor of each radio unit O-RU using the radio unit O-RU, and transmit the load factor and the channel state information of each radio unit O-RU to the cloud computing platform; S3. Preprocess the load factor and the channel state information of several radio units O-RU using the cloud computing platform to obtain the channel state information of M radio units O-RU with the strongest received signals, where M is an integer; S4. Using the channel state information of the M radio units O-RU as the input of a pre-trained spectrum efficiency prediction model based on long short-term memory and attention, and outputting the spectrum efficiency of the user equipment UE by the spectrum efficiency prediction model; S5. Clustering the radio units O-RUs for communication with the user equipment UE based on the spectrum efficiency of the user equipment UE.

2. The clustering method based on long short-term memory and attention according to claim 1, characterized in that: The preprocessing of the load factor and the channel state information of the plurality of radio units O-RU by using the cloud computing platform includes: S31. Determine whether the load factor is greater than a preset load factor threshold, and if so, filter out the radio unit O-RU whose load factor is not greater than the load factor threshold; S32. Based on the screened radio units O-RU, the channel state information of M radio units O-RU with the strongest received signals is preferably selected from the channel state information of the plurality of radio units O-RU.

3. The clustering method based on long short-term memory and attention according to claim 1, characterized in that: The spectral effect prediction model includes an embedding layer, an attention network layer, a long short-term memory network layer and an activation function layer. The output end of the embedding layer is connected to the input end of the attention network layer, the output end of the attention network layer is connected to the input end of the long short-term memory network layer, and the output end of the long short-term memory network layer is connected to the activation function layer.

4. The clustering method based on long short-term memory and attention according to claim 3, characterized in that: The method of using the channel state information of the M radio units O-RU as an input of a pre-trained spectrum efficiency prediction model based on long short-term memory and attention, and outputting the spectrum efficiency of the user equipment UE by the spectrum efficiency prediction model, comprises: S41. The embedding layer receives the channel state information of the M radio units O-RU, converts the channel state information of the M radio units O-RU into vector features and inputs them into the attention network layer; S42. The attention network layer receives the vector features and outputs a first vector feature sequence to the long short-term memory network layer; S43. The long short-term memory network layer receives the first vector feature sequence, outputs a second vector feature sequence, the second vector feature sequence is regularized and input into the activation function layer, and the activation function layer outputs the M radio units O-RU to predict the spectrum efficiency of the user equipment UE.

5. The clustering method based on long short-term memory and attention according to claim 4, characterized in that: The long short-term memory network layer includes a first unidirectional long short-term memory network, a regularization layer and a second unidirectional long short-term memory network, the output end of the attention network layer is connected to the input end of the first unidirectional long short-term memory network, the output end of the first unidirectional long short-term memory network is connected to the input end of the regularization layer, the output end of the regularization layer is connected to the input end of the second unidirectional long short-term memory network, and the output end of the second unidirectional long short-term memory network is connected to the activation function layer.

6. The clustering method based on long short-term memory and attention according to claim 4, characterized in that: The clustering of the user equipment UE based on the spectrum efficiency of the user equipment UE includes: S51. Determine whether the gain of the spectrum efficiency of the user equipment UE predicted by the M radio units O-RU relative to the spectrum efficiency of the user equipment UE predicted by the M-1 radio units O-RU is greater than a preset gain threshold, if so, execute S52; if not, stop clustering the user equipment UE; S52. Add the radio unit O-RU with the strongest received signal among the remaining M radio units O-RU as the radio unit O-RU serving the user equipment UE, and record the radio unit O-RU serving the user equipment UE as the service cluster corresponding to the user equipment UE.

7. A clustering system based on long short-term memory and attention, characterized in that: include: Several user equipments UE, several radio units O-RU and a cloud computing platform; A user equipment UE is configured to measure channel state information of a serving radio unit O-RU and a plurality of radio units O-RU adjacent to the currently serving radio unit O-RU, and feed back the measured channel state information to the respective radio units O-RU; A radio unit O-RU, configured to calculate a current load factor of each of the radio units O-RU, and transmit the load factor and channel state information of a plurality of radio units O-RU to a cloud computing platform; The cloud computing platform is used to pre-process the load factor and the channel state information of several radio units O-RU to obtain the channel state information of M radio units O-RU with the strongest received signals, wherein M is an integer. The channel state information of the M radio units O-RU is used as the input of a pre-trained spectrum efficiency prediction model based on long short-term memory and attention. The spectrum efficiency prediction model outputs the spectrum efficiency of the user equipment UE. Based on the spectrum efficiency of the user equipment UE, the radio units O-RU communicating with the user equipment UE are clustered.

8. The clustering system based on long short-term memory and attention according to claim 7, characterized in that: When the user equipment UE measures channel state information of several radio units O-RU adjacent to the user equipment UE, the user equipment UE selects a radio unit O-RU from the several radio units O-RU as a serving radio unit O-RU, and the serving radio unit O-RU is used to determine a measurement cluster of the user equipment UE, and the measurement cluster includes several radio units O-RU adjacent to the serving radio unit O-RU.

9. The clustering system based on long short-term memory and attention according to claim 7, characterized in that: The cloud computing platform includes a centralized unit O-CU for upper-layer network operations and a distributed unit O-DU for lower-layer network operations, and the centralized unit O-CU and the distributed unit O-DU are bidirectionally connected.

10. A computer device, characterized in that: include: A processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform the operation of the clustering method based on long short-term memory and attention as described in any one of claims 1-6.

Images