A 6G access network autonomous control method based on deep learning

Abstract

The present invention belongs to the field of artificial intelligence technology, and specifically relates to a method for autonomous control of 6G access networks based on deep learning. The present invention first periodically collects traffic data between base stations and user devices as input to a traffic prediction model based on a residual network and an attention mechanism, and the prediction model outputs short-term prediction results of base station traffic. According to the prediction results, the future traffic change trend of the base station is compared and judged with the base station carrying capacity at the current moment to confirm whether the base station should be expanded, reduced, or maintained unchanged. Finally, if the base station needs to be expanded or reduced, the corresponding instructions are issued to the base station management system to complete the dynamic management of the number of base stations, realize autonomous control of the 6G access network based on deep learning, and avoid resource waste while improving the base station bandwidth utilization.

Description

Technical Field

[0001] The present invention belongs to the field of 6G access network artificial intelligence technology, and specifically provides a 6G access network autonomous control method based on deep learning. Background Art

[0002] With the explosive growth of mobile applications, user data usage habits have gradually shifted from text messaging to online video streaming, and customer services have become more diverse. To further meet the emerging demands of communication networks and leverage the convenience and high speed of mobile internet to drive the development of various industries, operators and communications equipment manufacturers have been actively engaged in 6G network research and development. Compared to traditional networks, 6G networks offer numerous architectural improvements, including more flexible networking methods and widespread adoption of technologies such as Network Function Virtualization (NFV), which are based on computer software thinking. However, due to the limited coverage area of ​​6G base stations, achieving seamless coverage across regions, airspaces, and oceans requires relocation and large-scale deployment of base stations. If the traffic demand of user devices in a certain area falls short of the corresponding base station's carrying capacity, this will lead to waste of energy and hardware resources. Therefore, finding a method that can adaptively adjust and allocate resources in the 6G access network based on user device traffic demand, improve access network efficiency, save resources during low-demand periods, and enhance network carrying capacity during peak demand periods is an urgent issue.

[0003] The application of autonomous management and control methods for 6G access networks can balance user traffic demand with base station carrying capacity. This approach rationally manages the number of operating base stations based on user traffic usage habits. When user traffic usage is high and existing base stations can no longer meet bandwidth demands, the number of base stations can be increased to meet user quality of service requirements. When user traffic usage is low and existing base stations far exceed bandwidth demands, the number of base stations can be reduced, achieving rational allocation of redundant resources and controlling base station energy consumption. Base station number management is based on traffic thresholds. If the expansion threshold is set at 80%, base station capacity will be expanded when the total user device traffic reaches 80% of the base station bandwidth. Threshold-based base station management also has significant drawbacks: If the threshold is set too high, the time required for base station expansion will affect user experience. If the threshold is set too low, some base station bandwidth may remain unutilized, leading to resource waste if future traffic levels do not show an upward trend but only fluctuate slightly.

[0004] In order to address the shortcomings of autonomous control of 6G access networks based on traffic thresholds, it is possible to consider real-time monitoring of traffic between base stations and user devices, while introducing a deep learning model to predict the future short-term traffic conditions of base stations. The predicted results are compared with the traffic carrying capacity of the base station cluster at the current moment to determine whether to expand or shrink the base station, thereby realizing autonomous control of the 6G access network based on deep learning. This invention is based on the current and historical traffic data of the base station cluster. By establishing a deep learning traffic prediction model based on a multi-layer residual network and an attention mechanism, it predicts future short-term traffic changes and dynamically adjusts the number of base stations. It is a new idea for autonomous control of 6G access networks. Summary of the Invention

[0005] The purpose of the present invention is to predict the future short-term change trend of base station traffic through a traffic prediction model based on deep learning technology, and use the prediction results to judge the expansion and contraction of base stations, and dynamically adjust the number of base stations based on the judgment results. While avoiding resource waste and improving bandwidth utilization, it ensures the access network carrying capacity, meets user needs, and realizes the integration of deep learning technology and 6G access network autonomous management and control methods.

[0006] The technical solution of the 6G access network autonomous control method based on deep learning of the present invention is:

[0007] S1. In a base station cluster, periodically monitor the traffic transmission between each base station and the user equipment, collect and normalize the base station traffic data to obtain a training set.

[0008] S2. Construct a base station traffic prediction model based on deep learning to predict the future short-term traffic situation of each base station. The base station traffic prediction model consists of a multi-layer residual network and an attention mechanism; the multi-layer residual network is composed of a cascade of N layers of residual networks, and the single-layer residual network is composed of the first dilated causal convolution layer, the first weight normalization layer, the first ReLU activation function layer, the first Dropout layer, the second dilated causal convolution layer, the second weight normalization layer, the second ReLU activation function layer, and the second Dropout layer from input to output. At the same time, the input of the single-layer residual network also passes through a one-dimensional full convolution layer and is summed with the output of the second Dropout layer to obtain the output of the single-layer residual network; the attention mechanism is a scaled dot product attention mechanism. After the output of the multi-layer residual network is mapped into the feature space by the Dense layer, the mapping result is used as the query Q of the scaled dot product attention mechanism;

[0009] S3. Use the training set to train the base station traffic prediction model. Specifically, the training set is input into the base station traffic prediction model to obtain the forward propagation output result. At the same time, the training set is weighted and biased to obtain the key K and the corresponding value V of the scaled dot product attention mechanism. The mean square error between training sets is used as the error function of the model training process:

[0010]

[0011] Among them, loss is the mean square error, M is the length of the training set sequence, x i For the training set traffic data, Output the results for the model forward propagation; use the Adam optimizer to optimize the network parameters, gradually reduce the loss function value, update the network weights and biases of each layer, and finally obtain a trained base station traffic prediction model;

[0012] S4. Using the trained base station traffic prediction model, the future short-term base station traffic prediction value V is obtained based on the base station traffic history value of the previous L hours. pred ;

[0013] S5. Calculate the base station cluster expansion threshold V max and the shrinkage threshold V min , the base station short-term traffic prediction value V pred With V max 、V min Comparison is made to enable judgment on base station expansion and contraction behavior in the autonomous management and control of 6G access networks;

[0014] S6. Based on the result of the base station expansion and contraction judgment, the number of base stations is dynamically managed. At the same time, the configuration information and user equipment information stored in each base station in the base station cluster are synchronized, and the base station to which the user equipment is connected is switched.

[0015] Furthermore, the specific method of S1 is:

[0016] Collect historical traffic data between the base station cluster and user equipment, with a data interval of i hours; and sum the historical traffic data of multiple groups of user equipment under the same base station to form a traffic data set X;

[0017] The flow data set X is normalized using the range normalization method so that the data in the flow data set are at the same order of magnitude, thereby obtaining the training set.

[0018] After obtaining the training set, a deep learning traffic prediction model is constructed based on the residual network and attention mechanism. The model structure is as follows: Figure 1 As shown in the figure, it includes two parts: a multi-layer residual network and an attention mechanism. The process specifically includes the following steps:

[0019] Define the model input flow sequence as {x1,x2,…,x L}, L is the length of the model input flow sequence, the total number of residual network layers in the multi-layer residual network is N, and each layer of the residual network uses the ReLU activation function to calculate the output features of this layer and connect them in sequence; the transformation of the flow data in the single-layer residual network is: y = H(x,ω h )+ω s x, where y is the output of a single-layer residual network, x is the input of a single-layer residual network, H is the transformation that the input flow data undergoes in the single-layer residual network, ω h 、ω s is the weight parameter; in order to ensure that the width of the input flow data of the single-layer residual network is consistent with the width of the output result, the single-layer residual network also includes a one-dimensional full convolution operation;

[0020] The specific structure of the single-layer residual network is as follows Figure 2 As shown, it includes two layers of dilated causal convolution layers, two layers of weight normalization layers, two layers of ReLU activation function layers and two layers of Dropout layers; the flow data first undergoes dilated causal convolution operation, and the operation result is input into the weight normalization layer to normalize the weights in the network. After weight normalization, the result is input into the ReLU activation function layer for nonlinear mapping. The output of the ReLU activation function layer is recycled and the above steps are performed once, and the result is summed with the one-dimensional full convolution operation result to finally obtain the single-layer residual network output. The role of the one-dimensional full convolution operation is to ensure that the width of the single-layer residual network input flow data is consistent with the output result;

[0021] The operation in the causal expansion convolution layer uses the method of skipping part of the input flow sequence to increase the model receptive field. The convolution calculation process is: Among them, x′ is the convolution output, x is the convolution input, {f0,f1,…,f k-1} is the convolution kernel, d is the expansion factor, and b is the bias; when id*j<0, let x i-d*j =0, so that the traffic data forecast value is only related to the historical value, ensuring the causal nature of the traffic forecast result;

[0022] The weight normalization operation is: Among them, w is the weight vector, v is the vector parameter, and g is the scalar parameter. The convergence speed of stochastic gradient descent is accelerated by reparameterizing the weight vector.

[0023] The output of the multi-layer residual network is mapped into the feature space through the Dense layer, and the mapping result is used as the query Q of the attention mechanism. The attention mechanism is specifically the scaled dot product attention mechanism: Among them, O is the final output result of the deep learning traffic prediction model, the key K and value V are obtained by weighting and biasing the traffic dataset X, and softmax is the activation function;

[0024] The forward propagation output of the traffic prediction model and the mean square error between the training sets are used as the error function of the model training process: Among them, loss is the mean square error, M is the length of the training set sequence, x i For the training set traffic data, Output results for the model forward propagation;

[0025] The Adam optimizer is used to optimize network parameters, gradually reduce the loss function value, update the network weights and biases of each layer, and finally obtain a trained deep learning traffic prediction model;

[0026] Using the trained deep learning traffic prediction model, input the traffic data in the test set, output the model traffic prediction results, and predict the future short-term base station traffic prediction value V based on the historical value of the base station traffic in the previous L hours. pred .

[0027] Furthermore, the specific method of S5 is:

[0028] Calculate the base station cluster expansion threshold: V max =n bs ·V bs , where n bs is the total number of base stations in the base station cluster, V bs is the theoretical downlink bandwidth of a single 6G base station; calculate the base station cluster capacity reduction threshold: V min =(n bs -1)·V bs ;

[0029] When the flow prediction value V pred Greater than the base station cluster expansion threshold V max When , the result of determination is base station expansion;

[0030] When the flow prediction value V pred Less than the base station cluster capacity reduction threshold V min When , the result of determination is that the base station is reduced in capacity;

[0031] When the flow prediction value V pred Between the base station cluster expansion threshold V max and the shrinking threshold V min When , the result of determination is that the number of base stations does not change.

[0032] Furthermore, the specific method of S6 is:

[0033] When the result of the determination is that the base station capacity is to be expanded or reduced, an expansion or reduction instruction is issued to the base station management system to guide the change in the number of base stations;

[0034] Based on the current number of base stations, gradually increase or decrease the number of base stations, and recalculate the base station cluster expansion threshold and base station cluster reduction threshold according to the total number of base stations after increase or decrease, so that V max ≤V pred ≤V min After that, the modulation of the number of base stations is ended, and then the configuration information and user equipment status information saved by each base station in the base station cluster are synchronized, the base station to which the user equipment is connected is reallocated, and the base station switching and re-access of the user equipment are performed.

[0035] The beneficial effects of the present invention are:

[0036] (1) Design a deep learning traffic prediction model based on a multi-layer residual network and an attention mechanism to predict future traffic levels at base stations. Compared with the baseline prediction model, the RMSE index is improved by an average of 14%;

[0037] (2) Based on the prediction results of the deep learning traffic prediction model, the prediction results are compared with the expansion and contraction thresholds of the base station cluster to realize the expansion and contraction judgment of the base station;

[0038] (3) Through the deep learning traffic prediction model and the dynamic adjustment of the number of base stations, the balance between user traffic demand and base station carrying capacity is ensured, the base station bandwidth utilization is improved, and resource waste is avoided. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 In the embodiment, the deep learning traffic prediction model structure is based on a multi-layer residual network and an attention mechanism.

[0040] Figure 2 In the embodiment, the single-layer residual network structure in the deep learning traffic prediction model.

[0041] Figure 3 In the embodiment, the autonomous control process of the 6G access network is performed based on traffic prediction.

[0042] Figure 4 2 is a comparison of prediction results of different traffic prediction models in the embodiment.

[0043] Figure 5 The figure shows a comparison of bandwidth utilization of different 6G access network control methods in the embodiment. DETAILED DESCRIPTION

[0044] The present invention will be described in further detail below in conjunction with the embodiments and accompanying drawings.

[0045] Example

[0046] This example uses OSM as the base station operation and management system, employing Free5GC and UERANSIM to simulate a 6G network environment. Both software programs are deployed to OSM as virtualized network functions (VNFs). Free5GC simulates various core network functional modules, while UERANSIM simulates base stations and user equipment in a 6G access network. This example simulates traffic transmission services between base stations and user equipment using the TCP packet transmission and reception mode of the iPerf3 client and server. The service traffic volume is set based on the traffic model data.

[0047] The underlying hardware environment in this example consists of three Linux virtual machines, one of which is installed with Kubernetes and OSM. The other two are used to install the OpenStack cloud infrastructure management platform. OpenStack is linked to OSM via API calls. Once linked, the real-time status of OpenStack can be viewed in the OSM user interface.

[0048] The deep learning traffic prediction model based on the multi-layer residual network and attention mechanism is built in Python, and the model code is developed under the TensorFlow machine learning framework. The parameter settings of the model are shown in Table 1.

[0049] Table 1 Base station traffic prediction model parameter settings

[0050]

[0051] The process of autonomous control method of 6G access network based on deep learning traffic prediction model is as follows Figure 3 The specific steps are as follows:

[0052] S1. Using a base station traffic simulation model, generate 120,000 data points at one-hour intervals. Divide the data into 10 groups and use the mean to ensure the accuracy of the results of the embodiment. Each group of data is normalized using the range normalization method to form a traffic data set X, which is divided into a training set and a test set.

[0053] S2. Based on the constructed traffic dataset X, a deep learning traffic prediction model based on a multi-layer residual network and an attention mechanism is trained. The model has an input sequence length of 120, and uses traffic data at time t, t+1, ..., t+119 as input and traffic data at time t+120 as output. The trained model is used to predict future base station traffic conditions.

[0054] S3. Launch the trained deep learning traffic prediction model, set the model access address and the prediction API interface;

[0055] S4. Create OSM VNF templates corresponding to the Free5GC and UERANSIM programs, and integrate traffic monitoring-related codes into the VNF templates so that the 6G access network autonomous control method can be put into operation as a whole.

[0056] S5. Create a network service, which mainly defines the network connection relationship between the core network and the access network;

[0057] S6. After the network service is online, check the running status of the corresponding VNF instance in OpenStack to verify that the network service is normal. If the network service is in normal operation, you can observe whether the core network and access network functional modules are running normally and whether the uesimtun network card for emulating the user device has been generated. If the above checks are normal, use the Ping command to check whether the user device can connect to the Internet normally.

[0058] S7. Use the iPerf3 tool to write commands related to traffic transmission between the base station and the user equipment through shell scripts. The traffic size setting in the command is provided by the base station traffic simulation model data;

[0059] S8. Observe the changes in the number of base stations in OSM and calculate the bandwidth utilization index: It means the system's resource utilization. The higher the result, the more fully the resources are utilized.

[0060] In order to evaluate the prediction performance of the traffic prediction model based on the multi-layer residual network and the attention mechanism, the embodiment also compares it with the benchmark prediction model. The benchmark prediction model includes six commonly used models: ARIMA model, LSTM model, GRU model, DeepAR model, TCN model, and TCN-LSTM model. It verifies whether the prediction results of the traffic prediction model used in the present invention are better than those of the benchmark model. The test results are as follows: Figure 4 At the same time, the MAE, MAPE, and RMSE indicators of the prediction results of each model are statistically analyzed, and the results are shown in Table 2.

[0061] Table 2 Performance of prediction indicators of each model

[0062]

[0063]

[0064] As can be seen from Table 2, the ARIMA model performs worst in terms of MAE and RMSE, and the MAPE index is also higher than that of other models, indicating that the ARIMA model has poor prediction effect; the traffic prediction model based on the multi-layer residual network and attention mechanism outperforms other models in terms of MAE and RMSE, and is not much different from the best-performing DeepAR model in terms of MAPE index. This shows that this model has higher prediction accuracy than the other six benchmark models, proving that this model has certain advantages over the benchmark models in terms of prediction accuracy.

[0065] The method of the present invention is also compared with an access network control method based on threshold settings to verify the performance of the present invention. To facilitate the description of the embodiment process, the 6G access network autonomous control method based on deep learning is referred to as the experimental group, and the access network control method based on threshold settings is referred to as the control group. The control group sets the base station expansion and contraction strategy to expand the base station when the base station traffic load reaches 80% of the upper limit; and scale the base station down when it falls below 50% of the base station VNF traffic load upper limit.

[0066] The changes in bandwidth utilization of the experimental group and the control group over 24 hours a day are as follows Figure 5 As shown. Figure 5 It can be seen that the bandwidth utilization of the experimental group is slightly worse than that of the control group at 10:00 and 20:00, but is significantly better than that of the control group at 9:00 and 14:00. This proves that the bandwidth utilization of the 6G access network autonomous control method based on deep learning is better than the access network control method based on threshold setting.

[0067] In summary, the present invention can effectively predict the traffic change trend of 6G base stations, rationally allocate the number of base stations, effectively utilize bandwidth, realize the intelligent autonomous control of 6G access networks, improve the service quality of mobile communication networks under the 6G NFV architecture, and achieve a balance between operator services and energy consumption.

Claims

1. A 6G access network autonomous control method based on deep learning, characterized in that: The following steps are involved: S1. In a base station cluster, periodically monitor the traffic transmission between each base station and the user equipment, collect and normalize the base station traffic data to obtain a training set. S2. Construct a base station traffic prediction model based on deep learning to predict the future short-term traffic situation of each base station. The base station traffic prediction model consists of a multi-layer residual network and an attention mechanism; the multi-layer residual network is composed of a cascade of N layers of residual networks, and the single-layer residual network is composed of the first dilated causal convolution layer, the first weight normalization layer, the first ReLU activation function layer, the first Dropout layer, the second dilated causal convolution layer, the second weight normalization layer, the second ReLU activation function layer, and the second Dropout layer from input to output. At the same time, the input of the single-layer residual network also passes through a one-dimensional full convolution layer and is summed with the output of the second Dropout layer to obtain the output of the single-layer residual network; the attention mechanism is a scaled dot product attention mechanism. After the output of the multi-layer residual network is mapped into the feature space by the Dense layer, the mapping result is used as the query Q of the scaled dot product attention mechanism; S3. Use the training set to train the base station traffic prediction model. Specifically, the training set is input into the base station traffic prediction model to obtain the forward propagation output result. At the same time, the training set is weighted and biased to obtain the key K and the corresponding value V of the scaled dot product attention mechanism. The mean square error between training sets is used as the error function of the model training process: Among them, loss is the mean square error, M is the length of the training set sequence, x i For the training set traffic data, Output the results for the model forward propagation; use the Adam optimizer to optimize the network parameters, gradually reduce the loss function value, update the network weights and biases of each layer, and finally obtain a trained base station traffic prediction model; S4. Using the trained base station traffic prediction model, the future short-term base station traffic prediction value V is obtained based on the base station traffic history value of the previous L hours. pred ; S5. Calculate the base station cluster expansion threshold V max and the shrinkage threshold V min , the base station short-term traffic prediction value V pred With V max 、V min Comparison is made to enable judgment on base station expansion and contraction behavior in the autonomous management and control of 6G access networks; S6. Based on the result of the base station expansion and contraction judgment, the number of base stations is dynamically managed. At the same time, the configuration information and user equipment information stored in each base station in the base station cluster are synchronized, and the base station to which the user equipment is connected is switched.

2. A 6G access network autonomous control method based on deep learning according to claim 1, characterized in that: The specific method of S1 is: Collect historical traffic data between the base station cluster and user equipment, with a data interval of i hours; and sum the historical traffic data of multiple groups of user equipment under the same base station to form a traffic data set X; The flow data set X is normalized using the range normalization method so that the data in the flow data set are at the same order of magnitude, thereby obtaining the training set.

3. The method for autonomous control of 6G access network based on deep learning according to claim 1, characterized in that: In S2, the single-layer residual network processes the input data as follows: y=H(x,ω) h )+ω s x Among them, y is the output of a single-layer residual network, x is the input of a single-layer residual network, H is the transformation that the input data undergoes in the single-layer residual network, ω h 、ω s is the weight parameter; The convolution calculation process of the causal dilated convolution layer is: Among them, x′ is the convolution output, {f0,f1,…,f k-1 } is the convolution kernel, d is the expansion factor, and b is the bias; when id*j<0, let x i-d*j =0, so that the traffic data forecast value is only related to the historical value; The calculation process of the weight normalization layer is: Among them, w is the weight vector, v is the vector parameter, and g is the scalar parameter.

4. The method for autonomous control of 6G access network based on deep learning according to claim 1, characterized in that: The specific method of S5 is: Calculate the base station cluster expansion threshold: V max =n bs ·V bs , where n bs is the total number of base stations in the base station cluster, V bs is the theoretical downlink bandwidth of a single 6G base station; calculate the base station cluster capacity reduction threshold: V min =(n bs -1)·V bs ; When the flow prediction value V pred Greater than the base station cluster expansion threshold V max When , the result of determination is base station expansion; When the flow prediction value V pred Less than the base station cluster capacity reduction threshold V min When , the result of determination is that the base station is reduced in capacity; When the flow prediction value V pred Between the base station cluster expansion threshold V max and the shrinking threshold V min When , the result of determination is that the number of base stations does not change.

5. The method for autonomous control of 6G access network based on deep learning according to claim 4, characterized in that: The specific method of S6 is: When the result of the determination is that the base station capacity is to be expanded or reduced, an expansion or reduction instruction is issued to the base station management system to guide the change in the number of base stations; Based on the current number of base stations, gradually increase or decrease the number of base stations, and recalculate the base station cluster expansion threshold and base station cluster reduction threshold according to the total number of base stations after increase or decrease, so that V max ≤V pred ≤V min After that, the modulation of the number of base stations is ended, and then the configuration information and user equipment status information saved by each base station in the base station cluster are synchronized, the base station to which the user equipment is connected is reallocated, and the base station switching and re-access of the user equipment are performed.

Images