Multi-point channel spectrum assisted base station switching method and system

Abstract

The invention discloses a multipoint channel spectrum assisted base station switching method and system. In consideration of a scene of dense deployment of large-scale antenna array base stations, in an offline training stage, a plurality of adjacent base stations jointly acquire channel information of user samples in an edge region of the base stations, and dissimilarity is extracted and fused, so that a multi-point channel map of the region is constructed; training an SNR prediction model based on the atlas by using a machine learning technology; in the online prediction stage, a new user is mapped to a multi-point channel map through a sampling external expansion method, the SNR prediction model is used for predicting the channel SNR between the new user and a neighbor base station, and efficient base station switching is achieved according to a prediction result. According to the method, the multi-point channel map is utilized, the relative position relation of the users in the edge area is accurately described on the premise that real position information is not needed, and therefore a more accurate SNR prediction model is trained. According to the invention, the overhead of the traditional switching method can be reduced, and the accurate base station switching rate can be provided at the same time.

Description

Technical Field

[0001] The present invention relates to the field of multiple-input multiple-output (MIMO) and millimeter wave (mmWave) wireless communications using large-scale antenna arrays, and in particular to a base station switching method and system assisted by a multi-point channel spectrum. Background Art

[0002] Massive MIMO and millimeter wave are key factors in achieving high data rates and widespread connectivity in cellular communications. Although the millimeter wave band can support data transmission rates of several Gb / s, it will experience severe path loss during signal propagation, so dense deployment of base stations is required to ensure seamless coverage. Due to user mobility and densely deployed base stations, users may frequently switch between different serving base stations to maintain link quality, which requires the relevant base stations to perform a handover procedure (HO) with the core network. The traditional handover procedure is performed based on the user's received signal strength measurement, which indicates the signal-to-noise ratio (SNR) of the transmission channel, i.e., link quality. This process involves additional signaling overhead and may reduce the throughput of the communication network. In order to effectively manage base station switching, it is required to obtain the channel SNR in an intelligent way instead of the traditional method of frequently sending measurement signals. The user's real location can be used to assist HO management in millimeter wave networks, but obtaining the user's location brings practical challenges and privacy risks.

[0003] On the other hand, channel graph (CC), as an unsupervised learning method, can determine the relative position of users based entirely on channel state information (CSI). The reason behind this is that the MIMO antenna array provides high-dimensional CSI, which contains sufficient geometric information of the propagation environment. This channel graph represents the relative position information of users, thereby relaxing the requirements for the real position, and it has been proven to effectively support many wireless resource management applications. However, users located at the edge of the cell or blocked by obstacles may be inaccurately mapped to the channel graph due to poor link quality and channel acquisition capabilities, thereby degrading the performance of subsequent applications. Therefore, it is necessary to study a new technology based on a channel graph that can accurately characterize the relative position relationship of edge user groups to achieve efficient base station switching management. Summary of the invention

[0004] Purpose of the invention: In view of the shortcomings of the prior art, the purpose of the present invention is to provide a base station switching method and system assisted by a multi-point channel map. For the scenario of densely deployed base stations in the mmWave network, a multi-point channel map is first constructed, and then the user is mapped to the map through the out-of-sampling expansion method, and the channel SNR between the user and the adjacent base station is predicted according to the user's position in the map. Finally, the base station that provides the highest channel SNR is selected for active and efficient switching, so as to reduce switching overhead and improve switching accuracy.

[0005] Technical solution: To achieve the above-mentioned invention object, the present invention adopts the following technical solution:

[0006] A base station switching method assisted by a multi-point channel spectrum comprises the following steps:

[0007] In the offline training phase, multiple distributed base stations jointly obtain channel information of user samples in edge areas. Each base station calculates the dissimilarity between users based on a dissimilarity measure that is insensitive to small-scale channel fading, constructs a local dissimilarity matrix, and sends the local dissimilarity matrix and the measured channel average signal-to-noise ratio (SNR) to the central unit. The central unit fuses the local dissimilarity matrices according to the channel average SNRs of different base stations to form a global dissimilarity matrix, and draws a multi-point channel map based on the global dissimilarity matrix. The coordinates of each user sample on the multi-point channel map are marked with the channel SNR between it and the neighboring base stations.

[0008] In the offline training phase, machine learning technology is used to train the SNR prediction model based on the multi-point channel spectrum. The model input is the user's spectrum coordinates, and the output is the channel SNR between the user and each neighboring base station.

[0009] In the online prediction stage, the graph coordinates of new users are obtained through the sampling out-of-band method, and the SNR prediction model is used to predict the channel SNR between the new users and the neighboring base stations based on the position of the new users in the graph. The next serving base station is selected according to the SNR prediction results to achieve active and efficient base station switching.

[0010] Furthermore, multiple distributed base stations deploy MIMO-OFDM systems, working in time division duplex TDD or frequency division duplex FDD mode.

[0011] Furthermore, the dissimilarity is defined as d ★ , whose expression is The minimization operation is used to overcome the drastic phase change of the channel due to small-scale fading, h l 、h m are the channel vectors of the lth and mth edge area users respectively, and ||·||2 is the binary norm operation of the vector.

[0012] Furthermore, each distributed base station constructs a local dissimilarity matrix D b , [D b ] l,m =d ★ (h l b ,h m b ), where l,m∈{1,…,N UE} is the index of users in the edge area, NUE is the number of users, h l b 、h m b are the channel vectors of the lth and mth edge area users obtained by base station b respectively; the average channel SNR of the kth edge area user calculated by base station b in To obtain the expected operation, σ 2 is the noise energy at the base station receiving end, The channel vector of the kth edge area user obtained by base station b.

[0013] Furthermore, the central unit fuses the local dissimilarity according to the average SNR of the channels of different base stations to form the global dissimilarity D, which is expressed as in ξ is the weight index, B is the number of base stations, The channel average SNRs of the lth and mth edge area users calculated for base station b respectively.

[0014] Furthermore, the central unit performs a dimensionality reduction operation on the global dissimilarity matrix to obtain the virtual position coordinate z of each user sample on the multi-point map and draws it into a multi-point channel map

[0015] Furthermore, the out-of-sampling expansion method calculates the dissimilarity between the new user and the existing user samples based on the channel information of the new user in the current serving base station, and calculates the coordinates of the new user in the graph by averaging the corresponding coordinates of the closest preset number of user samples in the graph.

[0016] Furthermore, a deep neural network (DNN) is used to train a graph-based channel SNR prediction model. The trained model is stored in the central unit. The central unit selects the base station that provides the highest channel SNR as the next serving base station based on the prediction results, and informs the relevant base stations to perform the switching process.

[0017] The present invention also provides a base station switching system assisted by a multi-point channel spectrum, comprising:

[0018] An offline training module is used in the offline training phase. Multiple distributed base stations calculate the dissimilarity between users based on the channel information of user samples in the edge areas obtained, based on a dissimilarity measure that is insensitive to small-scale channel fading, and construct a local dissimilarity matrix. The local dissimilarity matrix and the measured channel average signal-to-noise ratio (SNR) are sent to the central unit. The central unit fuses the local dissimilarity matrices according to the channel average SNRs of different base stations to form a global dissimilarity matrix, and draws a multi-point channel map according to the global dissimilarity matrix. The coordinates of each user sample on the multi-point channel map are marked with the channel SNR between it and the neighboring base stations. The SNR prediction model based on the multi-point channel map is trained using machine learning technology. The model input is the user's map coordinates, and the output is the channel SNR between the user and each neighboring base station.

[0019] The online prediction module is used in the online prediction stage to obtain the spectrum coordinates of the new user through the sampling expansion method, and use the SNR prediction model to predict the channel SNR between the new user and the neighboring base station based on the position of the new user in the spectrum;

[0020] And a base station switching module is used for the central unit to select the next serving base station according to the SNR prediction result.

[0021] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of the offline training phase and / or the steps of the online prediction phase in the multi-point channel spectrum assisted base station switching method.

[0022] Beneficial effects: The present invention uses a multi-point channel map to predict the SNR of the channel between the user and the neighboring base station. The central unit directly selects the switching base station according to the prediction result. Compared with the traditional switching process of frequently sending measurement signals to obtain channel quality, the present invention reduces the switching signaling overhead of the multi-base station network and improves the throughput of the communication system. In view of the fact that the user group that needs to be switched usually appears in the edge area of ​​the base station, the multi-point channel map provides a more accurate relative position relationship of the user group, so it is expected to obtain a better SNR prediction accuracy, that is, the correct switching rate of the base station. The present invention further adopts the method of calculating low-complexity dissimilarity and sampling expansion, which reduces the difficulty of physical layer implementation and improves the flexibility of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a schematic diagram of the overall method flow of an embodiment of the present invention;

[0024] Figure 2 A sample distribution diagram of users who participated in generating the graph in an embodiment of the present invention;

[0025] Figure 3Schematic diagram of an SNR prediction model in an embodiment of the present invention;

[0026] Figure 4 Schematic diagram of a multi-point channel spectrum assisted base station switching process in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] like Figure 1 As shown, the multi-point channel spectrum assisted base station switching method disclosed in the embodiment of the present invention can be divided into an offline training stage and an online prediction stage. In the offline training stage, a distributed MIMO base station group collaboratively constructs a multi-point channel spectrum of multiple cell edge collection areas, and uses machine learning technology to train a spectrum-based channel SNR prediction model. Specifically, multiple distributed base stations jointly obtain channel information of user samples in edge areas, and each base station calculates the dissimilarity between each user based on a dissimilarity metric that is insensitive to small-scale channel fading, constructs a local dissimilarity matrix, and sends the local dissimilarity matrix and the measured channel average SNR to the central unit; the central unit fuses each local dissimilarity matrix according to the channel average SNR of different base stations to form a global dissimilarity matrix, and draws a multi-point channel spectrum according to the global dissimilarity matrix; the coordinates of each user sample on the multi-point channel spectrum are marked with the channel SNR between it and the neighboring base station; the SNR prediction model trained based on the multi-point channel spectrum is used to capture the relationship between the user's virtual position in the spectrum and the channel quality to the neighboring base station. The model input is the user's spectrum coordinates, and the output is the channel SNR between the user and each neighboring base station.

[0029] In the online prediction stage, the spectrum coordinates of the new user are obtained by the sampling out-of-sample expansion method for the new user, and the SNR prediction model is used to predict the channel SNR between the new user and the neighboring base station based on the position of the new user in the spectrum, and the next serving base station is selected according to the SNR prediction result to achieve active and efficient base station switching. In the embodiment of the present invention, the offline training stage needs to store a large amount of user channel data and perform a series of operations to complete the construction of the multi-point channel spectrum and the training of the SNR prediction model. In the online prediction stage, based on the results achieved in the offline training process, the current serving base station quickly calculates the channel spectrum coordinates of the new user (that is, the user other than the offline training user sample) and sends it to the central unit. The central unit predicts the SNR of the channel between it and other neighboring base stations, thereby selecting the base station that provides the highest channel SNR as the base station to be switched in the next transmission cycle. After completing the offline training stage, due to the quasi-static characteristics of the actual environment, the channel spectrum and SNR prediction model assist in performing the functions of the online prediction stage in the next multiple transmission cycles. Therefore, the offline training process can be updated over a long period of time to ensure the timeliness of the method of the embodiment of the present invention.

[0030] The present invention uses multiple base stations to work together and integrates multiple different observation angles, which helps to build a more accurate multi-point channel map. The user's position in the channel map is strongly correlated with the channel quality between it and the base station. This correlation is captured through classical machine learning technology. A large number of labeled data sets can be used to train a map-based SNR prediction model to directly obtain the channel's SNR. The present invention uses a multi-point channel map to accurately depict the relative position relationship of users in the edge area without the need for real position information, thereby training a more accurate SNR prediction model. The present invention can provide a more accurate correct base station switching rate while reducing the overhead of traditional switching methods.

[0031] The method of this embodiment is described in more detail below with reference to a specific scenario.

[0032] Part 1: Building a distributed MIMO-OFDM base station system model and dividing base station switching target user groups

[0033] Specifically, consider a system of multiple large-scale antenna array base stations deployed in a distributed manner in a millimeter wave network. Each base station is equipped with more than dozens of antennas, and its array can adopt a linear array, a circular array, a plate array or other array structures. Each antenna unit can adopt an omnidirectional antenna or a sector antenna. When each antenna unit adopts an omnidirectional antenna, a 120-degree sector antenna and a 60-degree sector antenna, the spacing between each antenna can be configured as 1 / 2 wavelength, 1 / 3 wavelength and 1 wavelength. Each antenna unit can adopt a single-polarization or multi-polarization antenna, and different base stations can use different antenna array configurations; the base station uses broadband OFDM technology and works in time division duplex (TDD) or frequency division duplex (FDD) mode. The central unit is located in the data center or cloud computing center, responsible for processing data from distributed base stations, and is also responsible for the management and coordination of the overall network.

[0034] In this embodiment, the user end uses a single antenna. Consider deploying B base stations, each equipped with A antennas, and a total of S subcarriers evenly distributed at the center frequency f C The total bandwidth occupied is W. The user sends an uplink pilot signal, and each base station obtains the corresponding uplink channel based on the received signal. The uplink channel vector between user k and base station b is recorded as Assuming that the signal propagation follows the plane wave model, the channel information includes the multipath arrival angle and propagation delay. Represents the channel on the ath antenna and the sth subcarrier.

[0035] Figure 2 This is a schematic diagram of the distribution of user samples participating in generating the map. The selected user samples are located in the edge areas of each base station service, which can also be interpreted as the boundary areas of multiple adjacent cells.

[0036] In order to accurately capture the relative positions of users in the multi-cell boundary area and construct a complete channel map, it is necessary to divide the base station switching target user groups. Suppose the set of serviceable users covered by base station s is This set includes not only the user groups of this cell, but also the user groups of some neighboring base stations. The user groups located in the boundary area of ​​multiple cells can be expressed as where {t1,...,t p} is the neighboring cell of base station s. In addition, by calculating the cumulative probability distribution function of the SNR of all user channel samples, the threshold γ is set th The user samples below the threshold are also classified as the target user group. This method adds some users who are not in the edge area of ​​the base station service but have poor communication channel quality with the base station (due to factors such as obstacles) to the training data set to ensure the integrity of the data.

[0037] Part II: Offline training phase and online prediction phase in efficient base station switching method

[0038] Based on the multi-point architecture of multiple base stations with large-scale antenna arrays working together and the division of base station switching target groups, all channel samples of user groups in the multi-cell boundary area and their neighboring base stations are obtained in the offline training stage, that is, Where N UE is the number of user samples, and B is the number of base stations.

[0039] Each distributed base station performs the construction of the local dissimilarity matrix D b The operation is as follows: b ] l,m =d ★ (h l b ,h m b ). ★ Characterizes a dissimilarity measure that is insensitive to small-scale channel fading, and its expression is The minimization operation is used to overcome the dramatic phase changes of the channel due to small-scale fading, so that the dissimilarity can effectively reflect the actual physical distance between users. Further, the objective function is rewritten as And use Get ＊ The equivalent analytical expression is Where |·| is the modulus operation, and ||·||2 is the vector bi-norm operation. Calculate the channel average SNR The expectation operation This can be accomplished by averaging multiple channels on coherent time resources or frequency resources, σ 2 is the noise energy at the base station receiving end. After completing the above local operations, each distributed base station unit converts its local dissimilarity matrix D b and the measured channel average Send to the central unit.

[0040] The central unit fuses the local dissimilarity according to the channel observation quality of different base stations to form a global dissimilarity matrix D, which is expressed as in ξ is the weight index. The weight index shows a certain robustness in the process of dissimilarity fusion, and is generally set to 2 or 3. For the special case where a user sample n cannot connect to base station b due to severe channel fading or blocking, set SNRγ n b =0 and T Inf is a large normal number.

[0041] The Isomap algorithm is used to reduce the dimension of the global dissimilarity matrix. Specifically, the matrix D is regarded as a weighted graph containing the distances of each point, and its k-nearest neighbor graph G is found. k (k nearest neighbors). k Dijkstra algorithm is used to search for the shortest path, and the shortest path distance between any two nodes in the graph is added to the final dissimilarity matrix D f In, based on D f To implement dimensionality scaling, we need to solve the following problem:

[0042]

[0043] Among them, z k That is, the virtual position coordinates of user sample k on the multi-point map. The multi-point channel map is expressed as After the multi-point channel map is constructed at the central unit, the central unit sends the map data to each distributed base station and stores it in the local unit.

[0044] The above process of constructing a multi-point map can be summarized as follows:

[0045] Step S1, the distributed base station calculates the local dissimilarity matrix and SNR, and sends the calculation results to the central unit.

[0046] Step S2: The central unit performs data fusion to form a global dissimilarity matrix.

[0047] Step S3: The central unit uses the Isomap algorithm on the global dissimilarity matrix to generate a multi-point channel map.

[0048] The coordinates of each sample on the map are marked with the SNR of the channel between it and the neighboring base station, forming a labeled data set to train the SNR prediction model. The loss function is defined as the error between the predicted SNR and the actual SNR.

[0049] Figure 3 This is a schematic diagram of the SNR prediction model. The model adopts a deep neural network structure, with a fully connected layer plus a RELU activation function, with a total of two hidden layers. The input dimension is consistent with the dimension of the channel spectrum coordinates, and the output dimension is consistent with the number of neighboring base stations.

[0050] Figure 4 It is a schematic diagram of the multi-point channel map assisted base station switching process, in which the multi-point channel map and SNR prediction model implemented in the offline training stage are saved in the central unit and applied in the online prediction stage.

[0051] After completing offline training, when the next switching cycle arrives, the current serving base station b first obtains the CSI of the new user q, which is recorded as According to the CSI, the sampling expansion method is implemented to map the user to the global spectrum. Specifically, p user samples (whose CSI is has the smallest dissimilarity), forming a set The coordinates of the new user in the graph are the average of the coordinates of the samples in the set, and the calculation formula is: The current serving base station will It is sent to the central unit and input into the SNR training model. The central unit then obtains the SNR prediction value of the channel between user q and other neighboring base stations. Based on the result, the central unit determines the base station that currently provides the highest channel SNR, selects it as the base station to be switched, and informs the relevant base stations to execute the switching process.

[0052] An embodiment of the present invention discloses a base station switching system assisted by a multi-point channel spectrum, including:

[0053] An offline training module is used in the offline training phase. Multiple distributed base stations calculate the dissimilarity between users based on the channel information of user samples in the edge areas obtained, based on a dissimilarity measure that is insensitive to small-scale channel fading, and construct a local dissimilarity matrix. The local dissimilarity matrix and the measured channel average signal-to-noise ratio (SNR) are sent to the central unit. The central unit fuses the local dissimilarity matrices according to the channel average SNRs of different base stations to form a global dissimilarity matrix, and draws a multi-point channel map according to the global dissimilarity matrix. The coordinates of each user sample on the multi-point channel map are marked with the channel SNR between it and the neighboring base stations. The SNR prediction model based on the multi-point channel map is trained using machine learning technology. The model input is the user's map coordinates, and the output is the channel SNR between the user and each neighboring base station.

[0054] The online prediction module is used in the online prediction stage to obtain the spectrum coordinates of the new user through the sampling expansion method, and use the SNR prediction model to predict the channel SNR between the new user and the neighboring base station based on the position of the new user in the spectrum;

[0055] And a base station switching module is used for the central unit to select the next serving base station according to the SNR prediction result.

[0056] A computer program product disclosed in an embodiment of the present invention includes a computer program, and when the computer program is executed by a processor, the steps of the offline training phase and / or the steps of the online prediction phase in the base station switching method assisted by a multi-point channel spectrum are implemented. Anything not described in detail in the present invention is a well-known technology to those skilled in the art.

[0057] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A base station switching method assisted by a multi-point channel spectrum, characterized in that: The steps include: In the offline training phase, multiple distributed base stations jointly obtain channel information of user samples in edge areas. Each base station calculates the dissimilarity between users based on a dissimilarity metric that is insensitive to small-scale channel fading, constructs a local dissimilarity matrix, and sends the local dissimilarity matrix and the measured channel average signal-to-noise ratio (SNR) to the central unit. The central unit fuses the local dissimilarity matrices according to the average channel SNRs of different base stations to form a global dissimilarity matrix, and draws a multi-point channel map according to the global dissimilarity matrix; the coordinates of each user sample on the multi-point channel map are marked with the channel SNR between it and the neighboring base stations; In the offline training phase, machine learning technology is used to train the SNR prediction model based on the multi-point channel spectrum. The model input is the user's spectrum coordinates, and the output is the channel SNR between the user and each neighboring base station. In the online prediction stage, the graph coordinates of new users are obtained through the sampling expansion method, and the SNR prediction model is used to predict the channel SNR between the new user and the neighboring base stations based on the position of the new user in the graph, and the next serving base station is selected according to the SNR prediction results.

2. The base station switching method assisted by a multi-point channel map according to claim 1, characterized in that: Multiple distributed base stations deploy MIMO-OFDM systems, working in time division duplex TDD or frequency division duplex FDD mode.

3. The base station switching method assisted by a multi-point channel map according to claim 1, characterized in that: The dissimilarity is defined as d ★ , whose expression is The minimization operation is used to overcome the drastic phase change of the channel due to small-scale fading, h l 、h m are the channel vectors of the lth and mth edge area users respectively, and ||·||2 is the binary norm operation of the vector.

4. The base station switching method assisted by a multi-point channel map according to claim 1, characterized in that: Each distributed base station builds a local dissimilarity matrix D b , [D b ] l,m =d ★ (h l b ,h m b ), where l,m∈{1,...,N UE } is the index of users in the edge area, N UE is the number of users, d ＊ is the dissimilarity, h l b 、h m b The channel vectors of the lth and mth edge area users obtained by base station b are respectively; the channel average of the kth edge area user calculated by base station b is in To obtain the expected operation, σ 2 is the noise energy at the base station receiving end, is the channel vector of the kth edge area user obtained by base station b, and ||·||2 is the binary norm operation of the vector.

5. The base station switching method assisted by a multi-point channel map according to claim 1, characterized in that: The central unit fuses the local dissimilarity according to the average SNR of the channels of different base stations to form the global dissimilarity D, which is expressed as: in ξ is the weight index, B is the number of base stations, D b The local dissimilarity matrix constructed for base station b, l,m∈{1,…,N UE } is the index of users in the edge area, N UE is the number of users, The channel average SNRs of the lth and mth edge area users calculated for base station b respectively.

6. The base station switching method assisted by a multi-point channel map according to claim 1, characterized in that: The central unit performs dimensionality reduction on the global dissimilarity matrix to obtain the virtual position coordinate z of each user sample on the multi-point map and draws it into a multi-point channel map. N UE is the number of users.

7. The base station switching method assisted by a multi-point channel map according to claim 1, characterized in that: The out-of-sampling expansion method is to calculate the dissimilarity between the new user and the existing user samples based on the channel information of the new user in the current serving base station, and calculate the coordinates of the new user in the spectrum by averaging the corresponding coordinates of the closest preset number of user samples in the spectrum.

8. The base station switching method assisted by a multi-point channel map according to claim 1, characterized in that: A deep neural network is used to train a graph-based channel SNR prediction model. The trained SNR prediction model is stored in the central unit. The central unit selects the base station that provides the highest channel SNR as the next serving base station based on the prediction results, and informs the relevant base stations to perform the switching process.

9. A base station switching system assisted by a multi-point channel map, characterized in that: include: Offline training module, used in the offline training stage, multiple distributed base stations calculate the dissimilarity between users based on the channel information of user samples in the edge areas obtained, based on a dissimilarity metric that is insensitive to small-scale channel fading, and construct a local dissimilarity matrix, and send the local dissimilarity matrix and the measured channel average signal-to-noise ratio (SNR) to the central unit; The central unit fuses the local dissimilarity matrices according to the average channel SNRs of different base stations to form a global dissimilarity matrix, and draws a multi-point channel map according to the global dissimilarity matrix; the coordinates of each user sample on the multi-point channel map are marked with the channel SNR between it and the neighboring base stations; and the machine learning technology is used to train the SNR prediction model based on the multi-point channel map, the model input is the user's map coordinates, and the output is the channel SNR between the user and each neighboring base station; The online prediction module is used in the online prediction stage to obtain the spectrum coordinates of the new user through the sampling expansion method, and use the SNR prediction model to predict the channel SNR between the new user and the neighboring base station based on the position of the new user in the spectrum; And a base station switching module is used for the central unit to select the next serving base station according to the SNR prediction result.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the offline training phase and / or the steps of the online prediction phase in the base station switching method assisted by a multi-point channel map according to any one of claims 1-8 are implemented.

Images