Resource configuration information determination method and device, electronic equipment and storage medium

By using an LSTM model to predict frame data size in 5G NR communication and dynamically adjusting resource configuration, the waste caused by inaccurate resource configuration is solved, and resource utilization efficiency is improved, especially in XR data transmission.

CN118921697APending Publication Date: 2024-11-08SHANGHAI CYGNUS SEMICON CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411043782.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-31
Publication Date
2024-11-08

Smart Images

  • Figure CN118921697A_ABST
    Figure CN118921697A_ABST
Patent Text Reader

Abstract

The invention provides a resource configuration information determination method and device, electronic equipment and a storage medium, and relates to the technical field of communication. According to the method, the predicted frame size of the to-be-received frame data is accurately predicted through the neural network model, and then the resource configuration information is determined according to the predicted frame size, so that resource configuration can be dynamically performed according to the frame size of the to-be-received frame data, accurate configuration of resources can be realized, and the problem of resource waste is avoided.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of communication technology, and in particular to a method, device, electronic device and storage medium for determining resource configuration information. Background Art

[0002] 5G NR (New Radio) technology is a new global communication standard based on orthogonal frequency division multiplexing. With the introduction of the new 5G standard, a series of emerging technologies are gradually coming into the public eye. With the continuous development of 5G technology, signal quality continues to improve, and new signals with large data volumes and large changes pose challenges to communication networks.

[0003] In the related art, communication resources are often pre-configured to adapt to the high changes in signals. However, such a processing method may waste resources because it needs to fully cover the amount of data that may appear in the signal. Summary of the invention

[0004] The purpose of the embodiments of the present application is to provide a method, device, electronic device and storage medium for determining resource configuration information, so as to improve the problem of resource waste caused by inaccurate resource allocation in the prior art.

[0005] In a first aspect, an embodiment of the present application provides a method for determining resource configuration information, the method comprising:

[0006] Inputting frame information of the received multiple frames of data into the neural network model to obtain a predicted frame size of the frame data to be received predicted by the neural network model, wherein the last frame of data in the multiple frames of data is separated from the first frame of data to be received by one time step;

[0007] Resource configuration information is determined according to the predicted frame size, and the resource configuration information is used to instruct the terminal to perform corresponding resource configuration in the next scheduling period.

[0008] In the above implementation process, a neural network model is used to accurately predict the predicted frame size of the frame data to be received, and then the resource configuration information is determined according to the predicted frame size. In this way, resources can be dynamically configured according to the frame size of the frame data to be received, thereby achieving accurate configuration of resources and avoiding the problem of resource waste.

[0009] Optionally, the neural network model is an LSTM model, and the frame information of the received multiple frames of data is input into the neural network model to obtain the predicted frame size of the frame data to be received predicted by the neural network model, including:

[0010] The frame information of the received multiple frames of data is input into the LSTM model to obtain the predicted frame size of the frame data to be received predicted by the LSTM model.

[0011] In the above implementation process, the LSTM model performs well in processing long sequence data. It is good at capturing long-term dependencies, and multi-frame data is a long sequence data. Therefore, more accurate prediction results can be obtained by using the LSTM model for prediction.

[0012] Optionally, the LSTM model includes multiple modules, each module includes an input gate, a forget gate, an output gate and a memory unit, and the frame information of the received multiple frames of data is input into the LSTM model to obtain the predicted frame size of the frame data to be received predicted by the LSTM model, including:

[0013] The frame information of the frame data obtained at each moment is input into a corresponding module for calculation to obtain the predicted frame size of the frame data output at the current moment; wherein, the frame information input at moment t is calculated according to the following steps to obtain the predicted frame size output at moment t:

[0014] Calculate the forgetting coefficient at time t using the forgetting gate based on the frame information input at time t and the predicted frame size output at time t-1;

[0015] Utilizing the input gate to calculate the input coefficient at time t and the candidate cell state at time t based on the frame information input at time t and the predicted frame size output at time t-1;

[0016] The memory unit is used to perform weighted summation on the cell state at time t-1 and the candidate cell state at time t to obtain the cell state at time t; wherein the weighted coefficient corresponding to the cell state at time t-1 is the forgetting coefficient at time t, and the weighted coefficient corresponding to the candidate cell state at time t is the input coefficient at time t;

[0017] The output gate is used to calculate the output coefficient at time t based on the frame information input at time t and the predicted frame size output at time t-1, and the predicted frame size output at time t is calculated based on the output coefficient at time t and the cell state at time t.

[0018] In the above implementation process, in the LSTM model, only useful input data will be further processed and the corresponding results will be output, so it can filter out more useful information from the input data, making the final output result more accurate, thereby achieving accurate prediction of the frame size.

[0019] Optionally, the training process of the LSTM model includes the following:

[0020] Acquire training data, where the training data includes frame information of multiple frames of training data, where the multiple frames of training data are sent by different servers, and a time step is spaced between two adjacent frames of training data;

[0021] Normalizing the training data to obtain processed training data;

[0022] The processed training data is input into the LSTM model for training, and after the training conditions are met, a trained LSTM model is obtained.

[0023] In the above implementation process, the LSTM model is trained by collecting frame information of multiple frames of data sent by different servers, which can improve the training effect of the LSTM model.

[0024] Optionally, the acquiring training data includes:

[0025] Receive multiple frames of training data sent by different servers;

[0026] storing frame information of the multiple frames of training data in a buffer of the base station, wherein the multiple frames of training data are classified and stored according to the corresponding relationship between data rate and frame rate;

[0027] When the buffer is full of data, the data of different categories stored in the buffer are extracted as training data.

[0028] In the above implementation process, multiple frames of training data are classified and stored according to the correspondence between data rate and frame rate, so that the types of training data can be increased. When the data volume is sufficient, data of different categories can be extracted for model training to meet the data volume requirements of model training and improve the training effect of the model.

[0029] Optionally, after obtaining the trained LSTM model, the method further includes:

[0030] The LSTM model is evaluated according to the actual frame size of the received frame data and the predicted frame size output by the LSTM model, and it is determined whether to update the LSTM model based on the evaluation result.

[0031] In the above implementation process, the LSTM model is evaluated according to the actual frame size and the predicted frame size, so that the training effect of the LSTM can be evaluated. If the effect is not good, the LSTM model is updated, thereby improving the prediction accuracy of the LSTM model.

[0032] Optionally, the multiple frames of data refer to the latest M frames of data stored in the buffer of the base station, and M is a set value. Using the latest M frames of data to predict the frame size of subsequent frame data can avoid the influence of too old data on the prediction result, so as to improve the prediction accuracy.

[0033] Optionally, determining resource configuration information according to the predicted frame size includes:

[0034] Determining a maximum predicted frame size among predicted frame sizes of frame data to be received;

[0035] Determining the predicted number of resource blocks required for transmission according to the ratio between the maximum predicted frame size and the size of data transmittable by a single resource block;

[0036] The resource configuration information is determined according to the predicted quantity.

[0037] In the above implementation process, the maximum predicted frame size can be used to determine the reasonable number of resource blocks, thereby achieving accurate resource configuration and avoiding resource waste.

[0038] Optionally, the multi-frame data includes extended reality XR data transmitted between the server, the base station, and the terminal. Since the size of XR data is not fixed, the prediction method of this solution can predict the frame size of the frame data to be received, so that the size of the XR data can be predicted, and the determined resource configuration information can dynamically match the amount of resources required by the XR data packet size, effectively reducing the waste of resources.

[0039] In a second aspect, an embodiment of the present application provides a device for determining resource configuration information, the device comprising:

[0040] A prediction module, used for inputting frame information of the received multiple frames of data into a neural network model to obtain a predicted frame size of the frame data to be received predicted by the neural network model, wherein the last frame of data in the multiple frames of data is separated from the first frame of data to be received by one time step;

[0041] The information determination module is used to determine resource configuration information according to the predicted frame size, and the resource configuration information is used to instruct the terminal to perform corresponding resource configuration in the next scheduling period.

[0042] In a third aspect, an embodiment of the present application provides an electronic device, comprising a processor and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the steps in the method provided in the first aspect are performed.

[0043] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps in the method provided in the first aspect are performed.

[0044] In a fifth aspect, an embodiment of the present application provides a computer program product, comprising computer program instructions, which, when read and executed by a processor, execute the steps in the method provided in the first aspect above.

[0045] Other features and advantages of the present application will be described in the following description, and partly become apparent from the description, or be understood by practicing the embodiments of the present application. The purpose and other advantages of the present application can be realized and obtained by the structures specifically pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0047] Figure 1 A flowchart of a method for determining resource configuration information provided in an embodiment of the present application;

[0048] Figure 2 A schematic diagram of the structure of an LSTM model provided in an embodiment of the present application;

[0049] Figure 3 A schematic diagram of a data input and output process provided in an embodiment of the present application;

[0050] Figure 4 A flow chart of a training process for training an LSTM model provided in an embodiment of the present application;

[0051] Figure 5 A schematic diagram of the fitting of predicted data and real data using a BP neural network model provided in an embodiment of the present application;

[0052] Figure 6 A schematic diagram of the fitting of predicted data and real data using a CNN model provided in an embodiment of the present application;

[0053] Figure 7 A schematic diagram of the fitting of predicted data and real data using an LSTM model provided in an embodiment of the present application;

[0054] Figure 8 A structural block diagram of a device for determining resource configuration information provided in an embodiment of the present application;

[0055] Fig. 9A schematic diagram of the structure of an electronic device for executing a method for determining resource configuration information provided in an embodiment of the present application. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present application.

[0057] It should be noted that the terms "system" and "network" in the embodiments of the present invention can be used interchangeably. "Multiple" means two or more. In view of this, in the embodiments of the present invention, "multiple" can also be understood as "at least two". "And / or" describes the association relationship of associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / ", unless otherwise specified, generally indicates that the previously associated objects are in an "or" relationship.

[0058] An embodiment of the present application provides a method for determining resource configuration information. The method predicts the predicted frame size of frame data to be received through a neural network model, and then determines resource configuration information based on the predicted frame size. In this way, resource configuration can be dynamically performed based on the frame size of the frame data to be received, thereby solving the problem of resource waste caused by inaccurate resource allocation.

[0059] Please refer to Figure 1 , Figure 1 A flowchart of a method for determining resource configuration information provided in an embodiment of the present application, the method comprising the following steps:

[0060] Step S110: inputting the frame information of the received multiple frames of data into the neural network model to obtain the predicted frame size of the frame data to be received predicted by the neural network model.

[0061] Step S120: Determine resource configuration information according to the predicted frame size.

[0062] The execution subject of the method for determining resource configuration information in this solution may be a base station.

[0063] In some embodiments, a semi-persistent scheduling (Semi-Persistent Scheduling, SPS) downlink transmission mode and a configured authorized uplink transmission mode may be used between the base station and the terminal. Semi-persistent scheduling refers to pre-deploying a portion of resources to the terminal on the base station side, and collecting the terminal's demand information and current network conditions in real time when executing scheduling. The specific process of the base station using semi-static scheduling is as follows: the base station uses radio resource control (Radio Resource Control, RRC) to configure resources, and the resource configuration information is transmitted to the terminal side through the downlink control channel, and then activated through downlink control information (Downlink Control Information, DCI). When the network condition or user request changes, real-time scheduling is performed based on the resources pre-allocated in the next SPS cycle. Therefore, the use of semi-static scheduling can maximize the utilization of system resources.

[0064] However, during the semi-static scheduling period, the resource configuration remains unchanged, which may cause the configured resources to fail to meet the terminal's service transmission requirements. Therefore, during the semi-static scheduling period, the base station can also use dynamic scheduling for scheduling. The dynamic scheduling method means that the base station dynamically adjusts the resource allocation and scheduling strategy. In other words, the scheduling method can be a combination of semi-static scheduling and dynamic scheduling.

[0065] In other implementations, the base station and the terminal may also use a dynamic scheduling method, so that the base station can adjust resource allocation in a timely manner according to the predicted frame size obtained by prediction.

[0066] The implementation process of the three scheduling methods is introduced below.

[0067] (1) Semi-static scheduling method.

[0068] In this scheduling mode, the base station adjusts resource allocation only when the scheduling cycle arrives. In this case, the base station can initially configure resources for the terminal according to demand and activate the initial resource configuration. In the first scheduling cycle, the terminal transmits frame data to the base station according to the initial resource configuration. In this scheduling cycle, the base station receives multiple frames of data. Before the next scheduling cycle arrives, the base station can predict the predicted frame size of the frame data to be received based on the multiple frames of data, and determine the resource configuration information based on the predicted frame size. Then, when the next scheduling cycle arrives, the resource configuration information can be sent to the terminal and the resource configuration of the terminal can be activated. In this way, the terminal can transmit data according to the newly received resource configuration information in the next scheduling cycle.

[0069] In some embodiments, the base station can input the frame information of all received frame data (or the frame information of the latest M frame data) into the neural network model in real time to predict the frame size of the next frame data, so that when the next scheduling period arrives, the base station can predict the predicted frame size of the next frame data to be received, and then determine the resource configuration information based on the predicted frame size, and then perform resource configuration in the scheduling period. Alternatively, in this implementation method, the base station predicts the frame sizes of multiple frame data within a scheduling period, at which time the frame size of the frame data sent by the terminal can be known, and a maximum frame size can be determined therefrom to determine the resource configuration information.

[0070] Alternatively, the base station's prediction timing may be before the scheduling cycle, such as at time t1. The base station predicts the frame size of the frame data to be received at the next time t2 based on the frame information of the frame data received before time t1 (or the frame information of the latest M frame data), and determines the resource configuration information based on the frame size. At this time, when the next time, such as t2 (i.e., the scheduling cycle), arrives, resource configuration can be performed directly.

[0071] In some implementations, since the resource configuration information is determined by the base station based on the predicted frame size of the frame data to be received, the frame data to be received here may refer to a frame of data, and then the resource configuration information may be determined based on the predicted frame size of the frame data. However, during the scheduling period, the frame size of the frame data sent by the terminal may change, so the resource configuration information may not match the frame size of other frame data, so one way to solve this problem is to leave a certain margin when determining the resource configuration information, for example, the resource configuration information includes the number of resource blocks, if the number of resource blocks determined based on the predicted frame size is N, then the number of resource blocks finally determined may be N+i, where i is a positive integer.

[0072] Another implementation method is that the frame data to be received refers to multi-frame data, that is, the frame size of the multi-frame data is predicted. For example, the output data volume of the neural network model can be pre-set. In order to match the service transmission requirements of the terminal in the next scheduling period, the output data volume here can be related to the number of frames transmitted by the terminal in the period. For example, the output data volume is greater than or equal to the number of frames, so that the base station can predict the frame size of more frame data, and then determine the resource configuration information based on the maximum frame size in the multi-frame data. The resource configuration information can be determined, so as to ensure that the data transmitted by the terminal in the next scheduling period matches the configured resources.

[0073] (2) A scheduling method that combines semi-static scheduling and dynamic scheduling.

[0074] Similar to the above-mentioned semi-static scheduling, at the beginning, the base station allocates initial resource configuration information to the terminal, and then in the first scheduling period, the base station predicts the frame size of the frame data to be received in real time based on the frame size of the received multi-frame data. If the resource configuration information determined based on the predicted frame size of the frame data to be received does not match the currently configured resource configuration information, dynamic scheduling is performed (at this time, the base station can predict the frame size of the frame data to be received in real time based on the frame information of the received frame data, and then determine the resource configuration information in real time, and perform dynamic scheduling when the resource configuration information changes). Generally, dynamic scheduling is performed when the currently configured resource configuration information is insufficient. At this time, the base station sends the currently determined resource configuration information to the terminal and activates it. When the next scheduling period arrives, the base station can send the latest resource configuration information (which can be the latest resource configuration information allocated to the terminal before the scheduling period, or the latest resource configuration information determined during the scheduling period) to the terminal and activate it.

[0075] This scheduling method can ensure that if the transmission data of the terminal changes during the scheduling period of semi-static scheduling, resource allocation can be readjusted in time.

[0076] (3) Dynamic scheduling method.

[0077] In this scheduling method, the base station can predict the frame size of the frame data to be received in real time or regularly based on the frame size of the received multi-frame data, and then adjust the resources according to the latest determined resource configuration information. If the latest determined resource configuration information is changed compared to the previously allocated resource configuration information, the resource allocation is readjusted according to the latest determined resource configuration information. This scheduling method allows the terminal to adjust resources in a timely manner when the transmission data changes to adapt to changes in transmission requirements.

[0078] Regardless of the scheduling method, after the base station sends the initial resource configuration information to the terminal and activates it, the terminal can transmit data to the base station according to the initial resource configuration information. After the base station receives a certain number of frame data, it can input the frame information of the received multiple frames of data into the neural network model for prediction, and then determine the resource configuration information based on the predicted frame size of the frame data to be received. The resource configuration information is used to instruct the terminal to perform corresponding resource configuration in the next scheduling period.

[0079] Among them, the last frame data in the multi-frame data is separated from the first frame data to be received by a time step. The time step can be understood as the time interval between two frames of data. The frame data sent by the terminal is arranged in chronological order. For example, the multi-frame data includes M frames of data, and there is a time step between adjacent frames of data. Therefore, when predicting, it can be predicted that the frame size of the M+1th frame data can be predicted, or the frame size of the M+1th frame data - the M+1+ith frame data can be predicted, wherein the value of i can be related to the output data volume of the neural network model.

[0080] It can be understood that in the semi-static scheduling mode, the base station can make a prediction before the next scheduling cycle. At this time, the base station can predict the predicted frame size of the next frame of data to be received, and then determine the resource configuration information based on the predicted frame size, or the base station can predict the predicted frame size of multiple frames of data to be received, and then select the maximum predicted frame size from them to determine the resource configuration information. In the dynamic scheduling mode, the base station can make a prediction in real time. At this time, the base station can predict the predicted frame size of the next frame of data to be received, and then determine the resource configuration information based on the predicted frame size. If the resource configuration information changes compared to the current resource configuration information of the terminal, dynamic scheduling is performed and the resource configuration is reconfigured.

[0081] In some embodiments, frame information may refer to the specific content of frame data, such as frame size, frame time, frame type and other information. Of course, frame information may also only include frame size, that is, the frame size of subsequent frame data can be predicted based on the frame size of multiple frames of data.

[0082] In some embodiments, resource configuration information includes but is not limited to a time series period, the number of resource blocks, time-frequency resources, etc., where the time series period can be determined by the receiving time of multiple frames of data, and the number of resource blocks can be determined by the ratio of the frame size to the size of data that can be transmitted by a single resource block.

[0083] In the above implementation process, a neural network model is used to accurately predict the predicted frame size of the frame data to be received, and then the resource configuration information is determined according to the predicted frame size. In this way, resources can be dynamically configured according to the frame size of the frame data to be received, thereby achieving accurate configuration of resources and avoiding the problem of resource waste.

[0084] Based on the above embodiment, the base station can store the received data in a buffer. The storage space of the buffer is limited. If the buffer is full, the newly received data is stored in the buffer by overwriting the historical data, which is equivalent to storing the latest data in the buffer. Therefore, when making predictions, the latest M frames of data can be obtained from the buffer as the multi-frame data that has been received, and the value of M can be flexibly set according to needs.

[0085] In this manner, the base station may predict the frame size of subsequent frame data after receiving each M frame data (for example, if M is 10, the frame size of the 11th frame or the 11th frame to the 20th frame data may be predicted by extracting the 11th frame data to the 20th frame data, and then the frame size of the 21st frame or the 21st frame to the 30th frame data may be predicted by extracting the M frame data to predict the frame size of subsequent frame data after receiving each new frame data (for example, if M is 10, the frame size of the 11th frame data may be predicted by extracting the 10th frame data to the 10th frame data, and then the frame size of the 12th frame data may be predicted by extracting the 2nd frame data to the 11th frame data), and when the time to adjust resource allocation is reached, the resource configuration information is determined according to the frame size of the latest predicted frame data.

[0086] In the above implementation process, the latest multi-frame data is used to predict the frame size of subsequent frame data, which can avoid the influence of too old data on the prediction result, so as to improve the prediction accuracy.

[0087] Based on the above embodiment, the neural network model can be a long short-term memory network LSTM model, and the frame information of the received multiple frames of data can be input into the LSTM model to obtain the predicted frame size of the frame data to be received predicted by the LSTM model.

[0088] The LSTM model is a special unit called a memory cell, similar to an accumulator and a gated neuron. It will have a weight and connect to itself in the next time step, copying the true value of its own state and the accumulated external signal, but this self-connection is controlled by a multiplication gate that learns from another unit and decides when to clear the memory content. In addition, the LSTM model is a time recurrent neural network that is suitable for processing and predicting important events with relatively long intervals and delays in time series.

[0089] In the LSTM model, only useful input data will be processed further and the corresponding results will be output, so it can filter out more useful information from the input data, making the final output result more accurate.

[0090] Therefore, the LSTM model performs well in processing long sequence data. It is good at capturing long-term dependencies, and multi-frame data is a long sequence data. Therefore, more accurate prediction results can be obtained by using the LSTM model for prediction.

[0091] Based on the above embodiment, the core of the LSTM model is the memory cell, which is responsible for storing and transmitting information. The model controls the cell state of the memory cell through three gate structures. The three gates are the forget gate, input gate and output gate. These three gates and the memory cell can be understood as the network layer of the LSTM model.

[0092] Among them, the forget gate is used to decide which information to discard from the memory unit. It reads the hidden state h of the previous moment. t-1 and the current input x t , and then a sigmoid function outputs a value between 0 and 1, which indicates how much information in the memory unit at the previous moment needs to be retained.

[0093] The input gate is used to determine what new information needs to be added to the memory unit. It also reads the hidden state h at the previous moment. t-1 and the current input x t , and then it is divided into two parts for processing. One part uses a sigmoid function to determine which information needs to be updated, and the other part uses a tanh function to generate an update value to represent the new information content. Finally, the two parts are combined to update the cell state of the memory unit.

[0094] The output gate is used to determine the output value at the current moment. First, a sigmoid function is used to determine which parts of the memory cell will be output. Then, the value of the memory cell processed by the tanh function is multiplied by the output of the sigmoid function to obtain the final output value.

[0095] Please refer to Figure 2 , Figure 2 The structural diagram of the LSTM model is shown in Figure 2 In the figure, each black line transmits an entire vector from the output of one node to the input of other nodes. The circles represent pointwise operations, such as the sum or product of vectors, and the rectangular boxes are neural network layers. The combined lines represent the connection of vectors, and the separated lines represent that the content is copied and then distributed to different locations.

[0096] The top horizontal arrow represents the cell state of the memory unit, which runs through each network layer. The cell state is similar to a conveyor belt, running directly on the entire chain, with only a small amount of linear interaction. It is easy for information to flow unchanged on it. Of course, the conveyor belt itself cannot control which information is remembered or not. The forget gate plays a controlling role.

[0097] The forget gate is used to remove or add information to the cell state. The gate is a method to selectively pass information. They contain a sigmoid neural network layer and a pointwise multiplication operation. The sigmoid layer outputs a value between 0 and 1, describing how much of each part can pass through, 0 means "no amount is allowed to pass", and 1 means "any amount is allowed to pass".

[0098] There are three control gates in the LSTM model to protect and control the cell state, namely the input gate, forget gate, and output gate. Through these three control gates, the LSTM model makes the weight of the self-loop changeable. At the same time, when the model parameters are fixed, the integral scale at different times can be changed dynamically, avoiding the gradient disappearance or gradient expansion during deep learning training.

[0099] The following describes the frame size prediction scheme in this application in combination with the implementation principle of the LSTM model.

[0100] The LSTM model includes multiple modules, each of which includes an input gate, a forget gate, an output gate, and a memory unit. When making predictions, the frame information of the frame data obtained at each moment can be input into a corresponding module for calculation to obtain the predicted frame size of the frame data output at the current moment, where the frame information input at moment t is calculated according to the following steps to obtain the predicted frame size output at moment t:

[0101] The forgetting function at time t is calculated using the forget gate based on the frame information input at time t and the predicted frame size output at time t-1;

[0102] The input gate is used to calculate the input coefficient at time t and the candidate cell state at time t based on the frame information input at time t and the predicted frame size output at time t-1;

[0103] The cell state at time t-1 and the candidate cell state at time t are weighted and summed using the memory unit to obtain the cell state at time t; the weight coefficient corresponding to the cell state at time t-1 is the forgetting coefficient at time t, and the weight coefficient corresponding to the candidate cell state at time t is the input coefficient at time t;

[0104] The output gate is used to calculate the output coefficient at time t based on the frame information input at time t and the predicted frame size output at time t-1, and the predicted frame size output at time t is calculated based on the output coefficient at time t and the cell state at time t.

[0105] Specifically, the forget gate controls whether to forget. In the LSTM model, it controls whether to forget the hidden cell state of the previous layer with a certain probability. Its input is the hidden state h of the previous sequence. t-1 (i.e., the predicted frame size output at time t-1) and the frame information x of the frame data input at time t t , through an activation function, usually a sigmoid function, to get the output f of the forget gate t Since the output f of the sigmoid function t Between [0,1], so the output f here is t Represents the probability of the state of the hidden cell in the previous layer, which is also the forgetting function, which can be expressed mathematically as:

[0106] f t =sigmoid(W f h t-1 +U f x t +b f );

[0107] Among them, W f Represents the parameters of the current neuron, U f is the weight of the frame data input at the current moment, b f is the bias. The expression of the sigmoid function is:

[0108] The input gate consists of two parts. The first part uses the sigmoid activation function, and the output is the input coefficient i at time t. t The second part uses the tanh activation function, and the output is the candidate cell state a at time t t The results of the two will be multiplied together to update the cell state, which can be expressed mathematically as:

[0109] i t =sigmoid(W i h t-1 +U i x t +b i )

[0110] a t =tanh(W a h t-1 +U a x t +b a )

[0111] Among them, W i ,W a is the parameter of the corresponding neuron, U i ,U a is the weight of the frame data input at the current moment, b i ,b a For bias.

[0112] Among them, the expression of tanh function is:

[0113]

[0114] The results of the forget gate and the input gate will act on the cell state C t , the memory unit is the cell state C at time t-1 t-1 and the candidate cell state a at time t tPerform weighted summation to obtain the cell state C at time t t , that is, update the cell state, that is, from the cell state C t-1 Update to C t , cell state C t The mathematical expression is:

[0115]

[0116] in, is the Hadamard product.

[0117] The predicted frame size h output at time t t The update part (which can be used as the prediction result output by the model at time t) consists of two parts. The first part is to obtain the output coefficient o at time t. t , which is determined by the predicted frame size h output at the previous moment t-1 and the frame information x of the frame data input at time t t (It can be considered as the frame information of the last frame of data in multiple frames. When the model is processing, each moment corresponds to the processing of one frame of data), and the activation function sigmoid. The second part is obtained by the cell state C at time t t And tanh activation function, that is:

[0118] o t =sigmoid(W o h t-1 +U o x t +b o )

[0119]

[0120] Among them, W o represents the parameters of the corresponding neuron, U o is the weight of the frame data input at the current moment, b o For bias.

[0121] Therefore, through the internal calculation processing of the above LSTM model, the predicted frame size of the next frame of data can be obtained.

[0122] In the above implementation process, in the LSTM model, only useful input data will be further processed and the corresponding results will be output, so it can filter out more useful information from the input data, making the final output result more accurate, thereby achieving accurate prediction of the frame size.

[0123] On the basis of the above embodiment, the above LSTM model is obtained after pre-training, so before using the LSTM model to predict the frame size, the LSTM model needs to be trained, and the training process includes: obtaining training data, the training data includes frame information of multiple frames of training data, the multiple frames of training data are sent by different servers, and there is a time step between two adjacent frames of training data. The training data is normalized to obtain the processed training data, and then the processed training data is input into the LSTM model for training. After the training conditions are met, the trained LSTM model is obtained.

[0124] The training process is similar to the above prediction process, except that the training data obtained during the training process is the frame information of a large amount of frame data. During training, each frame data can be marked with the frame size of the next frame data, and then these data are input into the LSTM model for training. Of course, the frame information input into the LSTM model includes not only the frame size, but also the relevant information of the frame data, such as frame time, frame type, etc., so that the LSTM model can learn the frame size change law of the frame data through more information, so as to train the parameters in the LSTM model to achieve the best, thereby obtaining the best output result. The training process can refer to the above prediction process. After the forward propagation in the training process is completed, the parameters of the model can be updated by the gradient descent method. After a period of training, the loss function converges (that is, the training conditions are met), and a trained LSTM model can be obtained.

[0125] In some embodiments, in order to evaluate the training effect of the LSTM model, the LSTM model may be evaluated according to the actual frame size of the received frame data and the predicted frame size output by the LSTM model, and it is determined whether to update the LSTM model based on the evaluation result.

[0126] For example, a threshold ε can be set and an indicator rerr can be used to evaluate the prediction results. The calculation method of rerr is as follows:

[0127]

[0128] Among them, S t Indicates the actual frame size of the frame data received at time t, S t ' represents the predicted frame size of the frame data output by the LSTM model at time t. If the value of the indicator rerr is greater than the threshold ε, the LSTM model is updated, that is, the training parameters of the LSTM model are adjusted, that is, new frame data is collected again to continue training the LSTM model until rerr is less than or equal to the threshold ε (this condition can also refer to the training condition). If the value of the indicator rerr is less than or equal to the threshold ε, the LSTM model is not updated, and the training ends at this time.

[0129] Here, the LSTM model is evaluated based on the actual frame size and the predicted frame size, so as to evaluate the training effect of the LSTM. If the effect is not good, the LSTM model is updated to improve the prediction accuracy of the LSTM model.

[0130] In the above implementation process, the LSTM model is trained by collecting frame information of multiple frames of data sent by different servers, which can improve the training effect of the LSTM model.

[0131] On the basis of the above embodiment, in order to avoid the interference of training data on model training, before inputting the training data into the model for training, the training data can also be normalized using the Z-score normalization operation. After the normalization operation, the training data eliminates the dimension, and the convergence speed of the model is faster during training. The normalized training data can be input into the above Figure 2 In the LSTM network layer shown in the figure, the output of the network layer can be connected to a fully connected layer and an output layer, and the final prediction result is output through the output layer. The specific implementation process is as follows Figure 3 As shown. Among them, the normalization operation can be expressed as follows:

[0132]

[0133] Among them, S t+i represents the original training data, S t ' +i represents the normalized data, μ represents the mean of the frame size of the original training data (such as the frame size), and δ represents the standard deviation of the frame size of the original training data (such as the frame size). After the normalization operation, the data satisfies the normal distribution, with a mean of 0 and a standard deviation of 1.

[0134] Based on the above embodiment, the training data can be extracted from the buffer of the base station. The base station can receive multiple frames of data sent by different servers, and then store the frame information of the multiple frames of data in the buffer of the base station, wherein the multiple frames of data are classified and stored according to the correspondence between the data rate and the frame rate, and then when the buffer is full of data, the data of different categories stored in the buffer are extracted as training data.

[0135] The multi-frame data in this application may refer to XR data. The base station may store the frame information of XR data sent by different servers in the same buffer. For different services, the frame size and frame rate of XR data are different. According to the data rate-frame rate, it can be divided into several types: 10Mbps-30fps, 10Mbps-60fps, 20Mbps-30fps, 20Mbps-60fps, 30Mbps-30fps, 30Mbps-60fps, 40Mbps-30fps, 40Mbps-60fps, 50Mbps-30fps, 50Mbps-60fps, 60Mbps-30fps, and 60Mbps-60fps.

[0136] The base station stores different types of XR data in the cache. When training the model, these types of data are shuffled and input into the model for training, which can greatly improve the versatility of the model.

[0137] During training, the training data can be divided into training data and test data according to a set ratio, such as 8:2. For the training data, multiple frames of data can be taken from the collected time t0 in the order of time received as a batch input. Then, one time step is moved, and the frame size of the 16th frame data is taken as the frame size to be predicted by the model. t0 Start by taking 15 data points as a batch and then t0+1 Start taking from the back, and follow this order until the end of the sequence S t0+c , input the model for training, the overall training process can be as follows Figure 4 shown.

[0138] In the above implementation process, multiple frames of data are classified and stored according to the correspondence between data rate and frame rate, so that the types of training data can be increased. When the data volume is sufficient, data of different categories can be extracted for model training to meet the data volume requirements of model training and improve the training effect of the model.

[0139] On the basis of the above embodiment, the neural network model used for frame size prediction may be a backpropagation neural network model (BP network) and a convolutional neural network model (CNN) in addition to the above LSTM model.

[0140] The following introduces the training process of the BP neural network model and the convolutional neural network model.

[0141] (1) Training process of BP neural network model:

[0142] Step 1: The base station collects a large amount of data (such as XR data) related information, such as: the i-th frame data b i , frame size wait.

[0143] Step 2: Set the frame size of each frame data Send it to the input layer of the BP neural network model and initialize the parameters.

[0144] Step 3: Forward propagation process, set the weight between node i and node j to w ij , the input value of each node is x j , the threshold of node j is b j , then the output value S of each node j for:

[0145]

[0146] x j =f(S j );

[0147] Among them, f(i) is the activation function.

[0148] Step 4: Back propagation process, assuming that the result of the output layer is d j , the error function is:

[0149]

[0150] Among them, d j is the frame size of the predicted next frame data, y j is the true value size, and the purpose of training is to minimize the error E(w,b).

[0151] Step 5: Update the weights and biases of neurons in the BP neural network model through the gradient descent method. The training ends when the error is less than the set threshold.

[0152] (2) Training process of convolutional neural network model:

[0153] Step 1: Frame information (such as frame size) S of multiple frames of data collected by the base station t Processing is performed, and the data starting from the i-th moment is placed in the buffer of the base station, and the buffer size is set to C. The data in the buffer is used to predict the frame size of the data at the next moment.

[0154] Step 2: Data in the buffer Normalization operation is performed according to a certain mapping method.

[0155] Step 3: Initialize the parameters of the convolutional neural network model, set the maximum number of training times, initial learning rate, learning rate reduction factor and other parameters.

[0156] Step 4: Input the normalized data into the input layer of the neural network model.

[0157] Step 5: The convolutional neural network model can perform feature extraction, which maps the time-frequency feature distribution from high dimensions to low-order feature space. The convolution layer consists of three levels. The first level performs convolution operation on the input to extract features. The convolution operation can be expressed as:

[0158]

[0159] Among them, (i, j) represents the position of the feature point, W represents the input frame data, X represents the convolution kernel, X(m, n) represents the weight of the convolution kernel at position (m, n), and W(i, j) represents the feature value at position (i, j).

[0160] The nonlinear activation function of the second level acts on the output of the first level as follows:

[0161] RELU(a)=max(a,0), where a represents the input value;

[0162] The third level further adjusts the output of the convolutional layer through the pooling function.

[0163] Step 5: Predicted frame size Denormalization is performed according to the original normalization method, and the error is calculated by comparing with the actual i+c+1 time data, and then the model parameters are updated based on the error.

[0164] When making a prediction, the base station may input the frame size of the received multiple frames of data into a BP neural network model or a convolutional neural network model to predict and obtain the predicted frame size of the next frame of data.

[0165] In practical applications, the LSTM model, the BP neural network model and the convolutional neural network model can also be used in combination. For example, multiple frames of data can be input into the three models respectively. At this time, three prediction results can be obtained. The final prediction result can be obtained based on the three prediction results. For example, the three prediction results can be weighted summed or averaged to obtain the final prediction result. The final prediction result is the predicted frame size of the final frame data to be received. In this way, the prediction results of the three models can be combined to determine a more accurate predicted frame size, thereby determining more accurate resource configuration information.

[0166] The inventors of this application conducted relevant experiments on the prediction effects of the three models, taking the mean prediction as the benchmark. Figure 5 , Figure 6 , Figure 7 The BP neural network model ( Figure 5 ), convolutional neural network model ( Figure 6 ), LSTM model( Figure 7 ) under the prediction data and the real data fitting situation, where the red line represents the real value and the blue line represents the predicted value. For example, the horizontal axis is the predicted sample and the vertical axis is the predicted result, such as the predicted frame size. An indicator err is used to represent the error between the predicted data and the real data. The mathematical expression is as follows:

[0167]

[0168] Among them, s(n) represents the actual size of the data frame at time n, and e(n) represents the difference between the actual size and the predicted size of the data frame at time n. Table 1 below shows the performance comparison results of different algorithms under the err indicator.

[0169] Table 1 Comparison of error results of various algorithms

[0170] data Mean prediction Weiner Prediction BP network prediction CNN network prediction LSTM network prediction 10Mbps 60fps 0.200290 0.1504 0.125760 0.093050 0.0860 10Mbps 30fps 0.080414 0.0689 0.057495 0.043223 0.0298 20Mbps 60fps 0.057892 0.0475 0.021031 0.018718 0.0168 20Mbps 30fps 0.055364 0.0345 0.028977 0.020476 0.0185 30Mbps 60fps 0.042341 0.0408 0.027742 0.023315 0.0229 30Mbps 30fps 0.035302 0.0291 0.028346 0.019680 0.0195 40Mbps 60fps 0.064786 0.0383 0.036188 0.023977 0.0241 40Mbps 30fps 0.015837 0.0129 0.011517 0.010644 0.0102 50Mbps 60fps 0.028141 0.0161 0.017142 0.016461 0.0158 50Mbps 30fps 0.024722 0.0204 0.018156 0.015294 0.0136

[0171] Since the prediction occurs during the process of base station resource allocation, the delay generated by the prediction should also be considered within the PDB (Packet Delay Budget) of the data packet. If the prediction delay is too large, it will greatly affect the PDB of each data packet. The data in Table 2 below is the time required to predict a frame of data using the LSTM model:

[0172] Table 2. Time required to predict one frame using LSTM network

[0173] data 10Mbps 60fps 20Mbps 60fps 30Mbps 60fps 40Mbps 60fps Predicted latency 0.5210μs 0.5195μs 0.5234μs 0.5219μs

[0174] When the frame rate of the transmitted data is 60fps, the time to transmit one frame is about 16.67ms, and the PDB of the XR data packet is millisecond-level data, so the predicted delay basically does not affect the PDB of the data packet.

[0175] In the single-stream 30Mbps-60fps scenario, when the base station allocates too many resources, the redundant resources are released. When the resources are allocated too little, the base station uses dynamic scheduling. Comparing the redundant resources, the number of dynamic scheduling times, and the amount of resources that need to be scheduled when scheduling 10,000 frames under the three neural network models, the following table data is obtained from the statistics of gNB scheduling 10,000 frames of 30Mbps-60fps data packets:

[0176] Table 3. Scheduler performance analysis under base station scheduling of 10,000 frames

[0177]

[0178] It can be seen from the above table that when too many resources are allocated, the amount of redundant resources of the prediction scheme based on the LSTM network model is the smallest among the three algorithms; when too few resources are allocated, the amount of resources required to be scheduled by the prediction scheme based on the LSTM network algorithm is also the smallest among the three algorithms, and the number of dynamic scheduling required is also the least.

[0179] On the basis of the above-mentioned embodiment, in the method of determining resource configuration information according to the predicted frame size, the maximum predicted frame size among the predicted frame sizes of the frame data to be received can be determined first, and then the predicted number of resource blocks required for transmission can be determined according to the ratio between the maximum predicted frame size and the data size that can be transmitted by a single resource block, and then the resource configuration information can be determined according to the predicted number.

[0180] If the frame size of a frame of data is predicted, then the maximum predicted frame size is the frame size of the frame of data; if the frame size of multiple frames of data is predicted, then the maximum predicted frame size is determined from the multiple frame sizes. The data size that can be transmitted by a single resource block can be determined by the configuration of the resource blocks in the base station. The ratio between the maximum predicted frame size and the data size that can be transmitted by a single resource block can determine the predicted number of resource blocks required, so that when performing resource configuration, the predicted number of resource blocks can be configured for the terminal.

[0181] In addition, the resource configuration information may also include a time series period, and the average period of the received multiple frames of data may be calculated by the arrival time of each frame of data. Thus, when performing resource configuration, the average period of data transmission may be configured to the terminal.

[0182] In the above implementation process, the maximum predicted frame size can be used to determine the reasonable number of resource blocks, thereby achieving accurate resource configuration and avoiding resource waste.

[0183] On the basis of the above embodiments, the data mentioned in this scheme is XR data, which has the characteristics of large data scale, jitter in transmission, and high latency requirements. The size of the XR data packet is determined by the given data rate and frame rate, which is modeled as a random variable following a truncated Gaussian distribution. In a single-stream scenario, the maximum size of a frame of data is 3 times the minimum size. In a multi-stream scenario, the maximum difference in the size of the data frame is 9 times, so the size of the XR data is changing dynamically. The prediction method of this scheme can predict the size of the next frame of data, so that the size of the XR data can be predicted, so that the resource configuration information determined can dynamically match the amount of resources required for the size of the XR data packet, effectively reducing the waste of resources.

[0184] Please refer to Figure 8 , Figure 8The structure block diagram of a device 200 for determining resource configuration information provided in an embodiment of the present application, the device 200 may be a module, program segment or code on an electronic device. It should be understood that the device 200 is similar to the above-mentioned Figure 1 The method embodiment corresponds to and can be executed Figure 1 The various steps involved in the method embodiment and the specific functions of the device 200 can be found in the above description. To avoid repetition, the detailed description is appropriately omitted here.

[0185] Optionally, the device 200 includes:

[0186] The prediction module 210 is used to input the frame information of the received multiple frames of data into the neural network model to obtain the predicted frame size of the frame data to be received predicted by the neural network model, wherein the last frame of data in the multiple frames of data is separated from the first frame of data to be received by one time step;

[0187] The information determination module 220 is used to determine resource configuration information according to the predicted frame size, and the resource configuration information is used to instruct the terminal to perform corresponding resource configuration in the next scheduling period.

[0188] Optionally, the neural network model is an LSTM model, and the prediction module 210 is used to input frame information of the received multiple frames of data into the LSTM model to obtain the predicted frame size of the frame data to be received predicted by the LSTM model.

[0189] Optionally, the LSTM model includes multiple modules, each module includes an input gate, a forget gate, an output gate and a memory unit, and the prediction module 210 is used to input the frame information of the frame data obtained at each moment into a corresponding module for calculation to obtain the predicted frame size of the frame data output at the current moment; wherein, the frame information input at time t is calculated according to the following steps to obtain the predicted frame size output at time t:

[0190] Calculate the forgetting coefficient at time t using the forgetting gate based on the frame information input at time t and the predicted frame size output at time t-1;

[0191] Utilizing the input gate to calculate the input coefficient at time t and the candidate cell state at time t based on the frame information input at time t and the predicted frame size output at time t-1;

[0192] The memory unit is used to perform weighted summation on the cell state at time t-1 and the candidate cell state at time t to obtain the cell state at time t; wherein the weighted coefficient corresponding to the cell state at time t-1 is the forgetting coefficient at time t, and the weighted coefficient corresponding to the candidate cell state at time t is the input coefficient at time t;

[0193] The output gate is used to calculate the output coefficient at time t based on the frame information input at time t and the predicted frame size output at time t-1, and the predicted frame size output at time t is calculated based on the output coefficient at time t and the cell state at time t.

[0194] Optionally, the device 200 further includes:

[0195] The training module is used to obtain training data, wherein the training data includes frame information of multiple frames of training data, wherein the multiple frames of training data are sent by different servers, and there is a time step between two adjacent frames of training data; normalize the training data to obtain processed training data; input the processed training data into the LSTM model for training, and obtain a trained LSTM model after the training conditions are met.

[0196] Optionally, the training module is used to receive multiple frames of training data sent by different servers; store frame information of the multiple frames of training data in a buffer of the base station, wherein the multiple frames of training data are classified and stored according to the correspondence between data rate and frame rate; when the buffer is full of data, extract differently classified data stored in the buffer as training data.

[0197] Optionally, the training module is used to evaluate the LSTM model according to an actual frame size of the received frame data and a predicted frame size output by the LSTM model, and determine whether to update the LSTM model based on the evaluation result.

[0198] Optionally, the multiple frames of data refer to the latest M frames of data stored in a buffer of the base station, where M is a set value.

[0199] Optionally, the information determination module 220 is used to determine the maximum predicted frame size among the predicted frame sizes of the frame data to be received; determine the predicted number of resource blocks required for transmission based on the ratio between the maximum predicted frame size and the data size that can be transmitted by a single resource block; and determine the resource configuration information based on the predicted number.

[0200] Optionally, the multiple frames of data include extended reality XR data transmitted between a server, a base station, and a terminal.

[0201] It should be noted that those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0202] Please refer to Fig. 9 , Fig. 9A structural diagram of an electronic device for executing a method for determining resource configuration information provided in an embodiment of the present application, the electronic device may include: at least one processor 310, such as a CPU, at least one communication interface 320, at least one memory 330 and at least one communication bus 340. Among them, the communication bus 340 is used to realize the connection and communication between these components. Among them, the communication interface 320 of the device in the embodiment of the present application is used to communicate signaling or data with other node devices. The memory 330 can be a high-speed RAM memory or a non-volatile memory (non-volatile memory), such as at least one disk storage. The memory 330 can optionally also be at least one storage device located away from the aforementioned processor. Computer-readable instructions are stored in the memory 330. When the computer-readable instructions are executed by the processor 310, the electronic device executes the above-mentioned Figure 1 The method process is shown.

[0203] Understandably, Fig. 9 The structure shown is for illustration only, and the electronic device may also include Fig. 9 More or fewer components as shown, or with Fig. 9 Different configurations are shown. Fig. 9 Each component shown in the figure can be implemented by hardware, software or a combination thereof.

[0204] The present application embodiment provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the following is performed: Figure 1 The method process in the method embodiment shown is executed by the electronic device.

[0205] This embodiment discloses a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer can perform the methods provided by the above method embodiments, for example, including:

[0206] Inputting frame information of the received multiple frames of data into the neural network model to obtain a predicted frame size of the frame data to be received predicted by the neural network model, wherein the last frame of data in the multiple frames of data is separated from the first frame of data to be received by one time step;

[0207] Resource configuration information is determined according to the predicted frame size, and the resource configuration information is used to instruct the terminal to perform corresponding resource configuration in the next scheduling period.

[0208] In summary, the embodiments of the present application provide a method, device, electronic device and storage medium for determining resource configuration information. The method uses a neural network model to accurately predict the predicted frame size of the frame data to be received, and then determines the resource configuration information based on the predicted frame size. In this way, resources can be dynamically configured according to the frame size of the frame data to be received, thereby achieving accurate configuration of resources and avoiding the problem of resource waste.

[0209] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0210] In addition, the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0211] Furthermore, the functional modules in the various embodiments of the present application may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.

[0212] In this document, relational terms such as first and second, etc. are used merely to distinguish one entity or operation from another entity or operation, but do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0213] The above description is only an embodiment of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for determining resource configuration information, characterized in that: The method comprises: Inputting frame information of the received multiple frames of data into the neural network model to obtain a predicted frame size of the frame data to be received predicted by the neural network model, wherein the last frame of data in the multiple frames of data is separated from the first frame of data to be received by one time step; Resource configuration information is determined according to the predicted frame size, and the resource configuration information is used to instruct the terminal to perform corresponding resource configuration in the next scheduling period.

2. The method according to claim 1, characterized in that The neural network model is an LSTM model, and the frame information of the received multiple frames of data is input into the neural network model to obtain the predicted frame size of the frame data to be received predicted by the neural network model, including: The frame information of the received multiple frames of data is input into the LSTM model to obtain the predicted frame size of the frame data to be received predicted by the LSTM model.

3. The method according to claim 2, characterized in that The LSTM model includes multiple modules, each of which includes an input gate, a forget gate, an output gate and a memory unit. The frame information of the received multiple frames of data is input into the LSTM model to obtain the predicted frame size of the frame data to be received predicted by the LSTM model, including: The frame information of the frame data obtained at each moment is input into a corresponding module for calculation to obtain the predicted frame size of the frame data output at the current moment; wherein, the frame information input at moment t is calculated according to the following steps to obtain the predicted frame size output at moment t: Calculate the forgetting coefficient at time t using the forgetting gate based on the frame information input at time t and the predicted frame size output at time t-1; Utilizing the input gate to calculate the input coefficient at time t and the candidate cell state at time t based on the frame information input at time t and the predicted frame size output at time t-1; The memory unit is used to perform weighted summation on the cell state at time t-1 and the candidate cell state at time t to obtain the cell state at time t; wherein the weighted coefficient corresponding to the cell state at time t-1 is the forgetting coefficient at time t, and the weighted coefficient corresponding to the candidate cell state at time t is the input coefficient at time t; The output gate is used to calculate the output coefficient at time t based on the frame information input at time t and the predicted frame size output at time t-1, and the predicted frame size output at time t is calculated based on the output coefficient at time t and the cell state at time t.

4. The method according to claim 2, characterized in that: The training process of the LSTM model includes the following: Acquire training data, where the training data includes frame information of multiple frames of training data, where the multiple frames of training data are sent by different servers, and a time step is spaced between two adjacent frames of training data; Normalizing the training data to obtain processed training data; The processed training data is input into the LSTM model for training, and after the training conditions are met, a trained LSTM model is obtained.

5. The method according to claim 4, characterized in that The obtaining of training data comprises: Receiving multiple frames of training data sent by different servers; storing frame information of the multiple frames of training data in a buffer of a base station, wherein the multiple frames of training data are classified and stored according to a corresponding relationship between data rate and frame rate; when the buffer is full of data, extracting data of different classifications stored in the buffer as training data; And / or, after obtaining the trained LSTM model, the method further includes: The LSTM model is evaluated according to the actual frame size of the received frame data and the predicted frame size output by the LSTM model, and it is determined whether to update the LSTM model based on the evaluation result.

6. The method according to any one of claims 1 to 5, characterized in that: The determining resource configuration information according to the predicted frame size includes: Determine the maximum predicted frame size among the predicted frame sizes of the frame data to be received; determine the predicted number of resource blocks required for transmission according to the ratio between the maximum predicted frame size and the size of data that can be transmitted by a single resource block; determine the resource configuration information according to the predicted number; And / or, the multiple frames of data include extended reality XR data transmitted between the server, the base station, and the terminal.

7. A device for determining resource configuration information, characterized in that: The device comprises: A prediction module, used for inputting frame information of the received multiple frames of data into a neural network model to obtain a predicted frame size of the frame data to be received predicted by the neural network model, wherein the last frame of data in the multiple frames of data is separated from the first frame of data to be received by one time step; The information determination module is used to determine resource configuration information according to the predicted frame size, and the resource configuration information is used to instruct the terminal to perform corresponding resource configuration in the next scheduling period.

8. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the method according to any one of claims 1 to 6 is executed.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is performed.

10. A computer program product, characterized in that The method comprises computer program instructions, and when the computer program instructions are read and executed by a processor, the method according to any one of claims 1 to 6 is executed.