Resource scheduling method and device, equipment and storage medium

Abstract

Provided are a resource scheduling method, apparatus and device, and a storage medium, applied to the technical field of artificial intelligence, the method comprising: receiving AI resource scheduling demand information carried in artificial intelligence AI task requests of N terminals sent by a first base station, N being an integer greater than 1; acquiring state information of the N terminals from the first base station; inputting the AI resource scheduling demand information and the state information of the N terminals into a pre-trained scheduling strategy decision model to obtain a resource scheduling strategy for the N terminals; and sending the resource scheduling policy to the first base station. The method is different from a current fair resource scheduling algorithm, and different quantities of resources are provided for each terminal according to own requirements and conditions of each terminal, so that the AI resource scheduling efficiency is integrally improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a resource scheduling method, apparatus, device, and storage medium. Background Art

[0002] In the current artificial intelligence (AI) services in the fourth-generation (4G) and fifth-generation (5G) mobile communication networks, various information required by the network side (such as base stations and core networks) and the terminal side is reported to the AI function nodes, and AI-related processing is performed outside the network side. However, in the sixth-generation (6G) mobile communication network, an in-network AI mechanism will be adopted, that is, a complete operating environment for the entire life cycle of the AI workflow, such as data collection, data preprocessing, model training, model inference, and model evaluation, will be provided inside the 6G architecture, and the computing power, data, algorithms, connections, network functions, protocols, and processes required for AI services will be deeply integrated and designed.

[0003] In the scenario of scheduling in-network AI service requests at a certain moment in the network, multiple user terminals send AI resource request information to the base station. In the prior art, the base station usually schedules the reserved AI resources equally to multiple user terminals based on the fair principle of first come, first served. However, due to the timeliness of network resources and the concurrency of many users in the communication network, the prior art lacks comprehensive consideration of these factors when scheduling the in-network AI resource request information of multiple users, resulting in low AI resource scheduling efficiency. Summary of the Invention

[0004] Embodiments of this application provide a resource scheduling method, apparatus, device, and storage medium to solve the problem of low resource scheduling efficiency in the existing resource scheduling method.

[0005] To solve the above technical problems, this application is implemented as follows:

[0006] In a first aspect, an embodiment of this application provides a resource scheduling method, which is executed by a Service Management Orchestrator (SMO). The method includes:

[0007] Receiving AI resource scheduling requirement information carried in the AI task requests of N terminals sent by a first base station, where N is an integer greater than 1;

[0008] Obtaining the status information of the N terminals from the first base station;

[0009] Input the AI resource scheduling requirement information and status information of the N terminals into a pre-trained scheduling policy decision model to obtain a resource scheduling policy for the N terminals. Among them, the scheduling policy decision model is obtained by training an initial scheduling policy decision model using a historical scheduling data set. The historical scheduling data set includes network system state data, resource scheduling action data, and actual resource scheduling efficiency data before a first moment. The resource scheduling action data is obtained from a pre-constructed digital twin, and the digital twin is constructed based on the network state information between the first base station and each terminal and historical AI resource scheduling information. The first moment is the moment when the AI task request is received;

[0010] Send the resource scheduling policy to the first base station.

[0011] Optionally, the AI resource scheduling requirement information includes at least one of the following: AI task type, AI model type, and AI task data volume size;

[0012] And / or, the status information of the terminal includes at least one of the following: the location information of the terminal, the transmission channel quality information of the terminal, the signal-to-noise ratio of the transmission channel of the terminal, the transmission channel power of the terminal, and the waiting time of the transmission data of the terminal in the first scheduling waiting queue of the first base station.

[0013] Optionally, the network system state data includes the AI task data volume and AI task limit conditions of each of the M terminals at each moment in a historical period, the total AI task data volume of the M terminals at each moment in the historical period, and the total AI task completion time of the M terminals at each moment in the historical period. M is an integer greater than 1;

[0014] The resource scheduling action data includes the resource scheduling decisions of each of the M terminals at each moment in the historical period:

[0015] The actual resource scheduling efficiency data includes the actual resource scheduling efficiency of each of the M terminals at each moment in the historical period.

[0016] Optionally, the actual resource scheduling efficiency of each of the M terminals at each moment in the historical period is determined according to the transmission efficiency of each terminal to the first base station, the AI task execution efficiency of each terminal, and the waiting time of the transmission data of each terminal in the second scheduling waiting queue of the first base station;

[0017] Among them, the transmission efficiency of each terminal to the first base station is determined according to the total task data volume of the M terminals, the distance from each terminal to the first base station, the Gaussian white noise power of the transmission channel of each terminal, and the transmission channel gain of each terminal;

[0018] The AI task execution efficiency of each terminal is determined according to the AI task type of each terminal, the AI model of each terminal, the size of the AI task data volume of each terminal, and the AI resources scheduled by each terminal.

[0019] Optionally, the scheduling policy decision model includes a policy Actor network and an evaluation Critic network; before inputting the AI resource scheduling requirement information and status information of the N terminals into the pre-trained scheduling policy decision model, the method further includes:

[0020] Based on the resource scheduling action data before the first moment and the initial Actor network, obtain the network system state data and actual resource scheduling efficiency data before the first moment;

[0021] Based on the network system state data and resource scheduling action data before the first moment, and the initial Critic network, determine the expected resource scheduling efficiency data;

[0022] According to the expected resource scheduling efficiency data and the actual resource scheduling efficiency data, determine the loss value of the Critic network;

[0023] Based on the loss value of the Critic network, adjust the weights of the initial Critic network to obtain a trained Critic network;

[0024] Based on the trained Critic network, adjust the network parameters of the initial Actor network.

[0025] Optionally, after sending the resource scheduling policy to the first base station, the method further includes:

[0026] Send the AI resource scheduling information at the first moment to the digital twin.

[0027] Optionally, the digital twin is located in the SMO, or the digital twin is a separate network element device.

[0028] Optionally, sending the resource scheduling policy to the first base station includes:

[0029] Delete the first resource from the AI resource list of the SMO, where the first resource is the resource indicated to be scheduled in the resource scheduling policy;

[0030] After deleting the first resource from the AI resource list, determine whether there is still AI resource scheduling requirement information waiting to be scheduled, or whether the AI resource list is empty;

[0031] In the case where there is no such AI resource scheduling requirement information waiting to be scheduled or the AI resource list is empty, send the resource scheduling policy to the first base station.

[0032] In a second aspect, an embodiment of the present application provides a resource scheduling method, which is executed by a first base station. The method includes:

[0033] Receive AI task requests sent by N terminals at a first moment, where the AI task requests carry AI resource scheduling requirement information;

[0034] Send the AI resource scheduling requirement information and status information of the N terminals to the SMO;

[0035] Receive the resource scheduling policies of the N terminals sent by the SMO, where the resource scheduling policies of the N terminals are obtained by the SMO inputting the AI resource scheduling requirement information and status information of the N terminals into a pre-trained scheduling policy decision model;

[0036] Based on the resource scheduling policies of the N terminals, schedule AI resources for each terminal among the N terminals;

[0037] Based on the AI resources scheduled for each terminal among the N terminals, execute the AI tasks of each terminal among the N terminals to obtain the AI task results of each terminal among the N terminals.

[0038] Optionally, the N terminals include a first terminal. After executing the AI tasks of each terminal among the N terminals based on the AI resources scheduled for each terminal among the N terminals to obtain the AI task results of each terminal among the N terminals, the method further includes at least one of the following:

[0039] In the case where the base station connected to the first terminal is the first base station, send the AI task result of the first terminal to the first terminal;

[0040] In the case where the base station connected to the first terminal is switched from the first base station to a second base station and the second base station is on the service function chain, directly send the AI task result of the first terminal to the second base station;

[0041] When the base station connected to the first terminal is switched from the first base station to the second base station and the second base station is not on the service function chain, broadcast the location information of the first terminal. After the second base station confirms that the first terminal is the terminal communicating with it based on the location information of the first terminal, send the AI task result of the first terminal to the second base station.

[0042] In a third aspect, an embodiment of the present application provides a resource scheduling method, which is executed by a terminal. The method includes:

[0043] Send an AI task request to a first base station at a first moment. The AI task request carries AI resource scheduling requirement information, and the AI resource scheduling requirement information includes at least one of the following: AI task type, AI model type, and size of AI task data volume;

[0044] When the base station connected to the terminal is the first base station, receive the AI task result executed by the first base station based on the AI task request.

[0045] When the base station connected to the terminal is switched from the first base station to a second base station, receive the AI task result executed by the first base station based on the AI task request sent by the second base station.

[0046] In a fourth aspect, an embodiment of the present application provides a resource scheduling method, which is executed by a digital twin. The method includes:

[0047] Send resource scheduling action data before the first moment to the SMO. The first moment is the moment when the SMO receives AI task requests of N terminals from the first base station, and N is an integer greater than 1;

[0048] Receive the AI resource scheduling information at the first moment sent by the SMO. The AI resource scheduling information at the first moment is obtained based on the resource scheduling strategy output by the SMO using a pre-trained scheduling strategy decision model. The scheduling strategy decision model is trained by using the resource scheduling action data before the first moment to an initial scheduling strategy decision model.

[0049] In a fifth aspect, an embodiment of the present application further provides a resource scheduling device, which is applied to the SMO. The resource scheduling device includes:

[0050] A first receiving module, configured to receive the AI resource scheduling requirement information carried in the AI task requests of N terminals sent by the first base station, where N is an integer greater than 1;

[0051] A first obtaining module, configured to obtain the status information of the N terminals from the first base station;

[0052] A second acquisition module, configured to input the AI resource scheduling requirement information and status information of the N terminals into a pre-trained scheduling policy decision model to obtain a resource scheduling policy for the N terminals, where the scheduling policy decision model is obtained by training an initial scheduling policy decision model with a historical scheduling data set, the historical scheduling data set includes network system status data, resource scheduling action data, and actual resource scheduling efficiency data before a first moment, the resource scheduling action data is obtained from a pre-constructed digital twin, and the digital twin is constructed based on the network status information and historical AI resource scheduling information between the first base station and each terminal; the first moment is the moment when the AI task request is received;

[0053] A first sending module, configured to send the resource scheduling policy to the first base station.

[0054] Optionally, the AI resource scheduling requirement information includes at least one of the following: AI task type, AI model type, and size of AI task data;

[0055] And / or, the status information of the terminal includes at least one of the following: location information of the terminal, transmission channel quality information of the terminal, signal-to-noise ratio of the transmission channel of the terminal, transmission channel power of the terminal, and waiting time of the transmission data of the terminal in the first scheduling waiting queue of the first base station.

[0056] Optionally, the network system status data includes the AI task data volume and AI task limit conditions of each of the M terminals at each moment in a historical period, the total AI task data volume of the M terminals at each moment in the historical period, and the total AI task completion time of the M terminals at each moment in the historical period, where M is an integer greater than 1;

[0057] The resource scheduling action data includes the resource scheduling decisions of each of the M terminals at each moment in the historical period:

[0058] The actual resource scheduling efficiency data includes the actual resource scheduling efficiency of each of the M terminals at each moment in the historical period.

[0059] Optionally, the actual resource scheduling efficiency of each of the M terminals at each moment in the historical period is determined according to the transmission efficiency of each terminal to the first base station, the AI task execution efficiency of each terminal, and the waiting time of the transmission data of each terminal in the second scheduling waiting queue of the first base station;

[0060] Among them, the transmission efficiency of each terminal to the first base station is determined according to the total task data volume of the M terminals, the distance from each terminal to the first base station, the Gaussian white noise power of the transmission channel of each terminal, and the transmission channel gain of each terminal;

[0061] The AI task execution efficiency of each terminal is determined according to the AI task type of each terminal, the AI model of each terminal, the size of the AI task data volume of each terminal, and the AI resources scheduled by each terminal.

[0062] Optionally, the device further includes:

[0063] A first processing module, configured to obtain the network system state data and the actual resource scheduling efficiency data before the first moment based on the resource scheduling action data and the initial Actor network before the first moment;

[0064] A first determination module, configured to determine the expected resource scheduling efficiency data based on the network system state data and the resource scheduling action data before the first moment, and the initial Critic network;

[0065] A second determination module, configured to determine the loss value of the Critic network according to the expected resource scheduling efficiency data and the actual resource scheduling efficiency data;

[0066] A first adjustment module, configured to adjust the weights of the initial Critic network based on the loss value of the Critic network to obtain a trained Critic network;

[0067] A second adjustment module, configured to adjust the network parameters of the initial Actor network based on the trained Critic network.

[0068] Optionally, the device further includes:

[0069] A second sending module, configured to send the AI resource scheduling information at the first moment to the digital twin.

[0070] Optionally, the digital twin is located in the SMO, or the digital twin is a separate network element device.

[0071] Optionally, the first sending module includes:

[0072] A first deletion unit, configured to delete a first resource from the AI resource list of the SMO, where the first resource is the resource indicated to be scheduled in the resource scheduling policy;

[0073] A first determination unit, configured to determine whether there is still AI resource scheduling requirement information waiting to be scheduled or whether the AI resource list is empty after deleting the first resource from the AI resource list;

[0074] A first sending unit, configured to send the resource scheduling policy to the first base station when there is no AI resource scheduling requirement information waiting to be scheduled or the AI resource list is empty.

[0075] In a sixth aspect, an embodiment of the present application further provides a resource scheduling device, which is applied to a first base station. The resource scheduling device includes:

[0076] A second receiving module, configured to receive AI task requests sent by N terminals at a first moment, where the AI task requests carry AI resource scheduling requirement information;

[0077] A third sending module, configured to send the AI resource scheduling requirement information and status information of the N terminals to the SMO;

[0078] A third receiving module, configured to receive the resource scheduling policies of the N terminals sent by the SMO, where the resource scheduling policies of the N terminals are obtained by the SMO inputting the AI resource scheduling requirement information and status information of the N terminals into a pre-trained scheduling policy decision model;

[0079] A first scheduling module, configured to schedule AI resources for each of the N terminals based on the resource scheduling policies of the N terminals;

[0080] A first execution module, configured to execute the AI tasks of each of the N terminals based on the AI resources scheduled for each of the N terminals, and obtain the AI task results of each of the N terminals.

[0081] Optionally, the N terminals include a first terminal, and the device further includes at least one of the following:

[0082] A fourth sending module, configured to send the AI task result of the first terminal to the first terminal when the base station connected to the first terminal is the first base station;

[0083] A fifth sending module, configured to directly send the AI task result of the first terminal to the second base station when the base station connected to the first terminal is switched from the first base station to the second base station and the second base station is on the service function chain;

[0084] A sixth sending module, configured to broadcast the location information of the first terminal when the base station connected to the first terminal is switched from the first base station to the second base station and the second base station is not on the service function chain, and after the second base station confirms that the first terminal is the terminal communicating with it based on the location information of the first terminal, send the AI task result of the first terminal to the second base station.

[0085] In a seventh aspect, an embodiment of the present application further provides a resource scheduling device, which is applied to a terminal. The resource scheduling device includes:

[0086] A seventh sending module, configured to send an AI task request to a first base station at a first moment, where the AI task request carries AI resource scheduling requirement information, and the AI resource scheduling request information includes at least one of the following: AI task type, AI model type, and size of AI task data volume;

[0087] A fourth receiving module, configured to receive the AI task result executed based on the AI task request sent by the first base station when the base station connected to the terminal is the first base station;

[0088] A fifth receiving module, configured to receive the AI task result executed by the first base station based on the AI task request sent by the second base station when the base station connected to the terminal is switched from the first base station to a second base station.

[0089] In an eighth aspect, an embodiment of the present application further provides a resource scheduling device, which is applied to a digital twin. The resource scheduling device includes:

[0090] An eighth sending module, configured to send resource scheduling action data before the first moment to an SMO, where the first moment is the moment when the SMO receives the AI task requests of N terminals from the first base station, and N is an integer greater than 1;

[0091] A sixth receiving module, configured to receive the AI resource scheduling information at the first moment sent by the SMO, where the AI resource scheduling information at the first moment is obtained based on the resource scheduling policy output by the SMO using a pre-trained scheduling policy decision model, and the scheduling policy decision model is trained using the resource scheduling action data before the first moment to an initial scheduling policy decision model.

[0092] In a ninth aspect, an embodiment of the present application further provides an SMO. The SMO includes a transceiver and a processor. The transceiver is configured to:

[0093] Receive the AI resource scheduling requirement information carried in the AI task requests of N terminals sent by the first base station, where N is an integer greater than 1;

[0094] Obtain the status information of the N terminals from the first base station;

[0095] The processor is configured to:

[0096] Input the AI resource scheduling requirement information and status information of the N terminals into a pre-trained scheduling policy decision model to obtain a resource scheduling policy for the N terminals, where the scheduling policy decision model is obtained by training an initial scheduling policy decision model using a historical scheduling data set, the historical scheduling data set includes network system status data, resource scheduling action data, and actual resource scheduling efficiency data before a first moment, the resource scheduling action data is obtained from a pre-constructed digital twin, the digital twin is constructed based on the network status information and historical AI resource scheduling information between the first base station and each terminal, and the first moment is the moment when the AI task request is received;

[0097] The transceiver is configured to:

[0098] Send the resource scheduling policy to the first base station.

[0099] Optionally, the AI resource scheduling requirement information includes at least one of the following: AI task type, AI model type, size of AI task data volume;

[0100] And / or, the status information of the terminal includes at least one of the following: location information of the terminal, transmission channel quality information of the terminal, signal-to-noise ratio of the transmission channel of the terminal, transmission channel power of the terminal, waiting time of the transmission data of the terminal in the first scheduling waiting queue of the first base station.

[0101] Optionally, the network system status data includes the AI task data volume and AI task limit conditions of each of the M terminals at each moment in a historical period, the total AI task data volume of the M terminals at each moment in the historical period, the total AI task completion time of the M terminals at each moment in the historical period, where M is an integer greater than 1;

[0102] The resource scheduling action data includes the resource scheduling decisions of each of the M terminals at each moment in the historical period:

[0103] The actual resource scheduling efficiency data includes the actual resource scheduling efficiency of each of the M terminals at each moment in the historical period.

[0104] Optionally, the actual resource scheduling efficiency of each of the M terminals at each moment in the historical period is determined according to the transmission efficiency of each terminal to the first base station, the AI task execution efficiency of each terminal, and the waiting time of the transmission data of each terminal in the second scheduling waiting queue of the first base station;

[0105] Among them, the transmission efficiency of each terminal to the first base station is determined according to the total task data volume of the M terminals, the distance from each terminal to the first base station, the Gaussian white noise power of the transmission channel of each terminal, and the transmission channel gain of each terminal;

[0106] The AI task execution efficiency of each terminal is determined according to the AI task type of each terminal, the AI model of each terminal, the size of the AI task data volume of each terminal, and the AI resources scheduled by each terminal.

[0107] Optionally, the scheduling policy decision model includes a policy Actor network and an evaluation Critic network; the processor is further configured to:

[0108] Based on the resource scheduling action data before the first moment and the initial Actor network, obtain the network system state data and the actual resource scheduling efficiency data before the first moment;

[0109] Based on the network system state data and resource scheduling action data before the first moment, and the initial Critic network, determine the expected resource scheduling efficiency data;

[0110] According to the expected resource scheduling efficiency data and the actual resource scheduling efficiency data, determine the loss value of the Critic network;

[0111] Based on the loss value of the Critic network, adjust the weights of the initial Critic network to obtain a trained Critic network;

[0112] Based on the trained Critic network, adjust the network parameters of the initial Actor network.

[0113] Optionally, the transceiver is further configured to:

[0114] Send the AI resource scheduling information at the first moment to the digital twin.

[0115] Optionally, the digital twin is located in the SMO, or the digital twin is a separate network element device.

[0116] Optionally, the transceiver is specifically configured to:

[0117] Delete a first resource from the AI resource list of the SMO, where the first resource is the resource indicated for scheduling in the resource scheduling policy;

[0118] After deleting the first resource from the AI resource list, determine whether there is still AI resource scheduling requirement information waiting for scheduling, or whether the AI resource list is empty;

[0119] When there is no AI resource scheduling requirement information waiting for scheduling or the AI resource list is empty, send the resource scheduling policy to the first base station.

[0120] In a tenth aspect, an embodiment of the present application further provides a first base station, where the first base station includes a transceiver and a processor, and the transceiver is used for:

[0121] Receive AI task requests sent by N terminals at a first moment, where the AI task requests carry AI resource scheduling requirement information;

[0122] Send the AI resource scheduling requirement information and status information of the N terminals to the SMO;

[0123] Receive the resource scheduling policy of the N terminals sent by the SMO, where the resource scheduling policy of the N terminals is obtained by the SMO inputting the AI resource scheduling requirement information and status information of the N terminals into a pre-trained scheduling policy decision model;

[0124] The processor is used for:

[0125] Schedule AI resources for each of the N terminals based on the resource scheduling policy of the N terminals;

[0126] Execute the AI tasks of each of the N terminals based on the AI resources scheduled for each of the N terminals to obtain the AI task results of each of the N terminals.

[0127] Optionally, the N terminals include a first terminal, and the transceiver is further used for:

[0128] When the base station connected to the first terminal is the first base station, send the AI task result of the first terminal to the first terminal;

[0129] When the base station connected to the first terminal is switched from the first base station to a second base station and the second base station is on the service function chain, directly send the AI task result of the first terminal to the second base station;

[0130] When the base station connected to the first terminal is switched from the first base station to the second base station and the second base station is not on the service function chain, broadcast the location information of the first terminal. After the second base station confirms that the first terminal is the terminal communicating with it based on the location information of the first terminal, send the AI task result of the first terminal to the second base station.

[0131] In a tenth aspect, an embodiment of the present application further provides a terminal, where the terminal includes a transceiver and a processor, and the transceiver is configured to:

[0132] Send an AI task request to a first base station at a first moment, where the AI task request carries AI resource scheduling requirement information, and the AI resource scheduling requirement information includes at least one of the following: AI task type, AI model type, size of AI task data volume;

[0133] When the base station connected to the terminal is the first base station, receive the AI task result executed based on the AI task request sent by the first base station;

[0134] When the base station connected to the terminal is switched from the first base station to a second base station, receive the AI task result executed by the first base station based on the AI task request sent by the second base station.

[0135] In an eleventh aspect, an embodiment of the present application further provides a digital twin, where the digital twin includes a transceiver and a processor, and the transceiver is configured to:

[0136] Send resource scheduling action data before the first moment to the SMO, where the first moment is the moment when the SMO receives AI task requests of N terminals from the first base station, and N is an integer greater than 1;

[0137] Receive the AI resource scheduling information at the first moment sent by the SMO, where the AI resource scheduling information at the first moment is obtained based on a resource scheduling policy output by a pre-trained scheduling policy decision model, and the scheduling policy decision model is obtained by training an initial scheduling policy decision model using the resource scheduling action data before the first moment.

[0138] In a twelfth aspect, an embodiment of the present application further provides an electronic device, including a processor, a memory, and a computer program stored on the memory and executable on the processor, where when the computer program is executed by the processor, the steps of the above resource scheduling method are implemented.

[0139] In a fourteenth aspect, an embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above resource scheduling method are implemented.

[0140] The resource scheduling method according to an embodiment of the present application includes receiving AI resource scheduling requirement information carried in AI task requests of N terminals sent by a first base station, where N is an integer greater than 1; obtaining status information of the N terminals from the first base station; inputting the AI resource scheduling requirement information and status information of the N terminals into a pre-trained scheduling policy decision model to obtain a resource scheduling policy for the N terminals, where the scheduling policy decision model is obtained by training an initial scheduling policy decision model using a historical scheduling data set, the historical scheduling data set includes network system state data, resource scheduling action data, and actual resource scheduling efficiency data before a first moment, the resource scheduling action data is obtained from a pre-constructed digital twin, and the digital twin is constructed based on network state information and historical AI resource scheduling information between the first base station and each terminal, and the first moment is the moment when the AI task request is received; sending the resource scheduling policy to the first base station. In this method, SMO deeply trains the scheduling policy decision model using the historical scheduling data set. When formulating a specific resource scheduling policy, in addition to combining the AI resource scheduling requirement information of each of the N terminals, it also combines the real-time status information of each of the N terminals, taking into account features such as network channel changes and the concurrency of multiple terminals. Different from the current fair resource scheduling algorithm, different amounts of resources are provided for each terminal according to the needs and situations of each terminal itself, thereby overall improving the efficiency of AI resource scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0141] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments of the present application. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0142] Figure 1 is a flowchart of the resource scheduling method executed by SMO provided by an embodiment of the present application;

[0143] Figure 2 is a flowchart of the resource scheduling method executed by the first base station provided by an embodiment of the present application;

[0144] Figure 3 is a flowchart of the resource scheduling method executed by the terminal provided by an embodiment of the present application;

[0145] Figure 4 It is a flowchart of a resource scheduling method executed by a digital twin provided by an embodiment of the present application;

[0146] Figure 5 It is a schematic diagram of the resource scheduling method provided by an embodiment of the present application;

[0147] Figure 6 It is a structural diagram of a resource scheduling device applied to SMO provided by an embodiment of the present application;

[0148] Figure 7 It is a structural diagram of a resource scheduling device applied to a first base station provided by an embodiment of the present application;

[0149] Figure 8 It is a structural diagram of a resource scheduling device applied to a terminal provided by an embodiment of the present application;

[0150] Figure 9 It is a structural diagram of a resource scheduling device applied to a digital twin provided by an embodiment of the present application;

[0151] Figure 10 It is a structural diagram of SMO provided by an embodiment of the present application;

[0152] Figure 11 It is a structural diagram of the first base station provided by an embodiment of the present application;

[0153] Figure 12 It is a structural diagram of the terminal provided by an embodiment of the present application;

[0154] Figure 13 It is a structural diagram of the digital twin provided by an embodiment of the present application. Detailed implementation manners

[0155] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts shall fall within the protection scope of the present application.

[0156] An embodiment of the present application provides a resource scheduling method, which is executed by service management and orchestration (SMO). Refer to Figure 1 , Figure 1 which is a flowchart of the resource scheduling method provided by an embodiment of the present application. As shown in Figure 1 , the method includes the following steps:

[0157] Step 101: Receive the AI resource scheduling requirement information carried in the AI task requests of N terminals sent by the first base station, where N is an integer greater than 1;

[0158] At a certain moment, first, N terminals send AI task requests to the first base station. Each AI task request of each terminal carries AI resource scheduling requirement information. The first base station schedules AI resources for the AI tasks of each terminal according to the AI resource scheduling requirement information of each terminal, and completes the corresponding AI tasks based on the scheduled AI resources. The AI task completion results are sent to the corresponding terminals. AI tasks can include AI training, AI inference, and AI verification, etc. In order to better schedule AI resources for N terminals, the first base station can send the AI resource scheduling requirement information of N terminals to the SMO, and the SMO formulates a suitable resource scheduling strategy by synthesizing the AI resource scheduling requirement information of N terminals. In addition, it is also necessary to determine whether there is AI resource scheduling requirement information that has not been scheduled before in the waiting queue of the first base station. If it exists, it is also obtained together.

[0159] Step 102: Obtain the status information of the N terminals from the first base station;

[0160] In this step, in order to better formulate a suitable resource scheduling strategy according to the characteristics of each terminal, in addition to obtaining AI resource requirement information, the SMO also needs to obtain the status information of each of the N terminals from the first base station. The status information of the terminal can specifically include the transmission channel quality information of the terminal, the location information of the terminal, and the signal-to-noise ratio of the transmission channel of the terminal, etc.

[0161] Step 103: Input the AI resource scheduling requirement information and status information of the N terminals into a pre-trained scheduling strategy decision model to obtain a resource scheduling strategy for the N terminals. Among them, the scheduling strategy decision model is obtained by training an initial scheduling strategy decision model using a historical scheduling data set. The historical scheduling data set includes network system state data, resource scheduling action data, and actual resource scheduling efficiency data before the first moment. The resource scheduling action data is obtained from a pre-constructed digital twin. The digital twin is constructed based on the network state information between the first base station and each terminal and historical AI resource scheduling information. The first moment is the moment when the AI task request is received;

[0162] In this step, to obtain the resource scheduling policies for the N terminals based on the AI resource scheduling requirement information and status information of each terminal among the N terminals, it is first necessary to train the scheduling policy decision model. The scheduling policy decision model is obtained by training the initial scheduling policy decision model using the historical scheduling data set. Exemplarily, if the SMO needs to formulate a resource scheduling policy for the AI task requests of the N terminals received by the first base station at nine o'clock, the SMO can use the resource scheduling data before nine o'clock to train the scheduling policy model. Specifically, the historical scheduling data set may include historical network system status data, historical resource scheduling action data, and historical actual resource scheduling efficiency data. It should be noted that in this step, it is also necessary to pre-construct a digital twin based on the network status information and historical AI resource scheduling information between the first base station and each terminal, and then obtain the resource scheduling action data from the constructed digital twin.

[0163] Step 104: Send the resource scheduling policy to the first base station.

[0164] In this step, after the SMO obtains the resource scheduling policies for the N terminals through the trained policy scheduling decision model, it sends the resource scheduling policy to the first base station. Then, the first base station allocates AI resources to each of the N terminals specifically according to the resource scheduling policy, and completes the corresponding AI tasks based on the allocated AI resources, and then sends the completion results of the AI tasks to the corresponding terminals.

[0165] In this implementation manner, the SMO deeply trains the scheduling policy decision model using the historical scheduling data set. When formulating a specific resource scheduling policy, in addition to combining the AI resource scheduling requirement information of each terminal among the N terminals, it also combines the real-time status information of each terminal among the N terminals, takes into account features such as network channel changes and the concurrency of many terminals, and is different from the current fair resource scheduling algorithm. It provides different amounts of resources for each terminal according to the needs and situations of each terminal, thereby overall improving the efficiency of AI resource scheduling.

[0166] Optionally, the AI resource scheduling requirement information includes at least one of the following: AI task type, AI model type, size of AI task data volume;

[0167] And / or, the status information of the terminal includes at least one of the following: location information of the terminal, transmission channel quality information of the terminal, signal-to-noise ratio of the transmission channel of the terminal, transmission channel power of the terminal, waiting time of the transmission data of the terminal in the first scheduling waiting queue of the first base station.

[0168] In one implementation, the AI resource scheduling requirement information may include the AI task type, the AI model type, and the size of the AI task data volume. Among them, the AI tasks may include AI training, AI inference, data processing, and other AI-related computing tasks, and the AI model is the model used by the first base station when executing AI tasks using AI resources.

[0169] The status information of the terminal may include the location information of the terminal, the transmission channel quality information of the terminal, the signal-to-noise ratio of the transmission channel of the terminal, the transmission channel power of the terminal, and the waiting time of the transmission data of the terminal in the first scheduling waiting queue of the first base station. Moreover, these status information of the terminal are all changing in real time.

[0170] In this implementation, based on the AI resource scheduling requirement information and status information of the terminal, a resource scheduling strategy that meets the needs and characteristics of the terminal itself is formulated, thereby improving the efficiency of AI resource scheduling as a whole.

[0171] Optionally, the network system status data includes the AI task data volume and AI task limit conditions of each of the M terminals at each moment in the historical period, the total AI task data volume of the M terminals at each moment in the historical period, the total AI task completion time of the M terminals at each moment in the historical period, where M is an integer greater than 1;

[0172] The resource scheduling action data includes the resource scheduling decisions of each of the M terminals at each moment in the historical period:

[0173] The actual resource scheduling efficiency data includes the actual resource scheduling efficiency of each of the M terminals at each moment in the historical period.

[0174] In one implementation, since the terminals that initiate AI task requests each time are different, in order to distinguish the N terminals that send AI task requests at the first moment mentioned above, it is determined that the M terminals that send AI task requests at the moment before the first moment. The network system status data can be represented as S t ={U i , B i , T i}, where U i is composed of the AI task data volume and AI task limit conditions of each of the M terminals at each moment in the historical period, and the AI task limit conditions may include the limit time for completing the AI task, etc. B i is the total AI task data volume of the M terminals at each moment in the historical period, and T iis the total AI task completion time of M terminals at each moment in the historical period. Then, the network system state data at each moment in the historical period is arranged in chronological order to obtain the historical network system state data set.

[0175] Exemplarily, if the first moment is nine o'clock, the historical network system state data set can be composed of the network system state data at eight o'clock (specifically including the AI task data volume and AI task limit conditions of each of the M terminals at eight o'clock, the total AI task data volume of the M terminals at eight o'clock, and the total AI task completion time of the M terminals at eight o'clock), the network system state data at eight minutes and one second (specifically including the AI task data volume and AI task limit conditions of each of the M terminals at eight minutes and one second, the total AI task data volume of the M terminals at eight minutes and one second, and the total AI task completion time of the M terminals at eight minutes and one second) …… the network system state data at eight minutes and fifty-nine seconds.

[0176] The resource scheduling action data includes the resource scheduling decisions of each of the M terminals at each moment in the historical period, which can be expressed as a t , referring to the foregoing example, a t can also be arranged in chronological order to form the historical resource scheduling action data set A{a t}.

[0177] The actual resource scheduling efficiency data includes the actual resource scheduling efficiency of each of the M terminals at each moment in the historical period, and the actual resource scheduling efficiency can be obtained through a predefined resource scheduling efficiency model.

[0178] In this embodiment, by using the historical network system state data, historical resource scheduling action data, and historical actual resource scheduling efficiency data as references to train the scheduling policy decision model, the accuracy of the subsequent scheduling policy decision model for predicting resource scheduling policies can be improved.

[0179] Optionally, the actual resource scheduling efficiency of each of the M terminals at each moment in the historical period is determined according to the transmission efficiency of each terminal to the first base station, the AI task execution efficiency of each terminal, and the waiting time of the transmission data of each terminal in the second scheduling waiting queue of the first base station;

[0180] wherein, the transmission efficiency of each terminal to the first base station is determined according to the total task data volume of the M terminals, the distance of each terminal to the first base station, the Gaussian white noise power of the transmission channel of each terminal, and the transmission channel gain of each terminal;

[0181] The AI task execution efficiency of each terminal is determined according to the AI task type of each terminal, the AI model of each terminal, the size of the AI task data volume of each terminal, and the AI resources scheduled by each terminal.

[0182] In one implementation, the actual resource scheduling efficiency of each terminal among the M terminals at each moment in the historical period can be obtained through a predefined resource scheduling efficiency model where T i all represents the resource scheduling efficiency, and T i tra represents the transmission efficiency from the terminal to the first base station, and α i represents the weight of T i tra Specifically, T i tra can be calculated through the following formula:

[0183] T i tra = (total task data volume of M terminals * distance from the terminal to the first base station / Gaussian white noise power of the transmission channel of the terminal) * transmission channel gain of each terminal.

[0184] represents the AI task execution efficiency of the terminal, and β i represents the weight of Specifically,

[0185]

[0186] represents the waiting time of the transmission data of the terminal in the second scheduling waiting queue of the first base station, and γ i represents the weight of Specifically, it can be obtained by subtracting the time when the transmission data of the terminal is scheduled from the time when it starts waiting in the second scheduling waiting queue of the first base station.

[0187] In this implementation, the resource scheduling efficiency is calculated through a predefined resource scheduling efficiency model, which improves the accuracy of the calculated resource scheduling efficiency.

[0188] Optionally, the scheduling policy decision model includes a policy Actor network and an evaluation Critic network; before inputting the AI resource scheduling requirement information and status information of the N terminals into the pre-trained scheduling policy decision model, the method further includes:

[0189] Based on the resource scheduling action data before the first moment and the initial Actor network, obtain the network system state data and the actual resource scheduling efficiency data before the first moment;

[0190] Based on the network system state data and resource scheduling action data before the first moment, and the initial Critic network, determine the expected resource scheduling efficiency data;

[0191] According to the expected resource scheduling efficiency data and the actual resource scheduling efficiency data, determine the loss value of the Critic network;

[0192] Based on the loss value of the Critic network, adjust the weights of the initial Critic network to obtain a trained Critic network;

[0193] Based on the trained Critic network, adjust the network parameters of the initial Actor network.

[0194] In one implementation, the scheduling policy decision model includes an Actor network and a Critic network. The SMO first initializes the Actor network, and then uses the Actor network to communicate with the digital twin. At time t (a time before the first moment), it receives the resource scheduling action data a t = πθ(S) + N (where θ represents the probability of executing a in state S, and N can be understood as a normalized constant). After that, the network system state data S t at time t changes to the network state data S (t+1) at time (t + 1), and calculates the reward r t (i.e., the above-mentioned actual resource scheduling efficiency data). In this way, the quadruple composed of S t , a t , S (t+1) , and r t can be obtained using the Actor network. Repeat this process to obtain multiple quadruples in the historical period. Then use the multiple quadruples to train the Critic network. Specifically, the current parameters w of the Critic network can be updated by the gradient backpropagation of the loss function , where Q(S t , a t ; w) represents the expected resource scheduling efficiency data, y t represents the actual resource scheduling data r t , and K represents the number of action executions in each round of updating the loss function.

[0195] Then, the trained Critic network is used to adjust the network parameters of the initial Actor network, so that the Actor network can generate higher Q-values and finally obtain a behavioral policy with more rewards. Specifically, the constructed loss function

[0196] is used to optimize the current parameters w of the Actor network.

[0197] In this implementation, the initial scheduling policy decision model is trained with the historical scheduling data set before the first moment, so that the trained scheduling policy decision model can formulate the optimal resource scheduling policy based on the AI resource scheduling requirement information and status information of N terminals at the first moment, thereby improving the efficiency of resource scheduling.

[0198] Optionally, after sending the resource scheduling policy to the first base station, the method further includes:

[0199] Sending the AI resource scheduling information at the first moment to the digital twin.

[0200] In one implementation, after SMO sends the resource scheduling policy at the first moment to the first base station, it is also necessary to store the AI resource scheduling information at the first moment in the digital twin. The AI resource scheduling information at the first moment may specifically include network status data, resource scheduling action data, actual resource scheduling efficiency data, etc. at the first moment, for subsequent training networks to select training data.

[0201] Optionally, the digital twin is located in the SMO, or the digital twin is a separate network element device.

[0202] In one implementation, as a digital mirror of physical network facilities, the digital twin has the same network elements, topology, and data as the physical network, and can achieve full-process fine-grained "replication" of the network and devices. The digital twin network will also record and manage the behaviors of the digital twins of the network, support traceability and playback of them, complete pre-verification without affecting network operation, and reduce the cost of trial and error. The digital twin can be located in the SMO or be a separate network element device, and its main functions include base stations, decision module management on the network side, decision module data analysis, and decision module operating environment, etc.

[0203] Optionally, the sending the resource scheduling policy to the first base station includes:

[0204] Deleting the first resource from the AI resource list of the SMO, where the first resource is the resource indicated to be scheduled in the resource scheduling policy;

[0205] After deleting the first resource from the AI resource list, determine whether there is still AI resource scheduling requirement information waiting to be scheduled, or whether the AI resource list is empty;

[0206] In the case where there is no AI resource scheduling requirement information waiting to be scheduled or the AI resource list is empty, send the resource scheduling policy to the first base station.

[0207] In one implementation, after SMO determines the resource scheduling policy for N terminals that send AI task requests at the first moment, it is necessary to delete the AI resources that have been scheduled from the AI resource list. Further, after deleting the AI resources that have been scheduled from the AI resource list, determine whether there is still AI resource scheduling requirement information waiting to be scheduled, or whether the AI resource list is empty. In the case where there is no AI resource scheduling requirement information waiting to be scheduled or the AI resource list is empty, send the resource scheduling policy to the first base station to end the scheduling process. In the case where there is still AI resource scheduling requirement information waiting to be scheduled or the AI resource list is not empty, continue to complete the resource scheduling task to ensure that all resource scheduling tasks can be completed in a timely manner.

[0208] The embodiments of the present application provide a resource scheduling method, which is executed by the first base station. Refer to Figure 2 , Figure 2 is the flowchart of the resource scheduling method provided by the embodiments of the present application, as shown in Figure 2 shown, including the following steps:

[0209] Step 201, receive AI task requests sent by N terminals at the first moment, where the AI task requests carry AI resource scheduling requirement information;

[0210] Step 201, send the AI resource scheduling requirement information and status information of the N terminals to the SMO;

[0211] Step 202, receive the resource scheduling policy of the N terminals sent by the SMO, where the resource scheduling policy of the N terminals is obtained by the SMO inputting the AI resource scheduling requirement information and status information of the N terminals into a pre-trained scheduling policy decision model;

[0212] Step 203, schedule AI resources for each terminal among the N terminals based on the resource scheduling policy of the N terminals;

[0213] Step 204, execute the AI tasks of each terminal among the N terminals based on the AI resources scheduled for each terminal among the N terminals to obtain the AI task results of each terminal among the N terminals.

[0214] It should be noted that, as the implementation manner of the first base station corresponding to the embodiment shown in Figure 1 , the specific implementation manner can be referred to the relevant description of the embodiment shown in Figure 1 . To avoid repeated description, this embodiment will not be elaborated herein, and the same beneficial effects can also be achieved.

[0215] Optionally, the N terminals include a first terminal. After performing the AI tasks of each of the N terminals based on the AI resources scheduled for each terminal among the N terminals and obtaining the AI task results of each of the N terminals, the method further includes at least one of the following:

[0216] When the base station connected to the first terminal is the first base station, sending the AI task result of the first terminal to the first terminal;

[0217] When the base station connected to the first terminal is switched from the first base station to a second base station and the second base station is on the service function chain, directly sending the AI task result of the first terminal to the second base station;

[0218] When the base station connected to the first terminal is switched from the first base station to the second base station and the second base station is not on the service function chain, broadcasting the location information of the first terminal, and after the second base station confirms that the first terminal is the terminal communicating with it based on the location information of the first terminal, sending the AI task result of the first terminal to the second base station.

[0219] In one implementation manner, after the first base station completes the AI task by using the scheduled AI resources, it is necessary to feedback the AI task completion result to the terminal. If the terminal does not switch the connected base station, the first base station directly sends the AI task completion result to the corresponding terminal. If the terminal moves and switches the connected base station, it can be specifically divided into the following two cases:

[0220] When the base station connected to the terminal is switched from the first base station to the second base station and the second base station is on the service function chain, directly sending the AI task result of the first terminal to the second base station;

[0221] When the base station connected to the terminal is switched from the first base station to the second base station and the second base station is not on the service function chain, broadcasting the location information of the first terminal, and after the second base station confirms that the first terminal is the terminal communicating with it based on the location information of the first terminal, the first base station then sends the AI task result of the first terminal to the second base station.

[0222] Then, the second base station sends the AI task completion result to the corresponding terminal to ensure that the terminal can receive the AI task completion result in a timely manner.

[0223] An embodiment of the present application provides a resource scheduling method, which is executed by a terminal. Refer to Figure 3 , Figure 3 which is a flowchart of the resource scheduling method provided by the embodiment of the present application. As Figure 3 shown, it includes the following steps:

[0224] Step 301: Send an AI task request to the first base station at the first moment. The AI task request carries AI resource scheduling requirement information, and the AI resource scheduling requirement information includes at least one of the following: AI task type, AI model type, and AI task data volume size;

[0225] Step 302: Receive the AI task result executed based on the AI task request sent by the first base station when the base station connected to the terminal is the first base station;

[0226] Step 303: Receive the AI task result executed by the first base station based on the AI task request sent by the second base station when the base station connected to the terminal is switched from the first base station to the second base station.

[0227] It should be noted that, as an implementation manner of the terminal corresponding to the embodiment Figure 1 shown, the specific implementation manner can refer to the relevant description of the embodiment Figure 1 shown. To avoid repeated description, this embodiment will not be elaborated herein, and the same beneficial effects can also be achieved.

[0228] An embodiment of the present application provides a resource scheduling method, which is executed by a digital twin. Refer to Figure 4 , Figure 4 which is a flowchart of the resource scheduling method provided by the embodiment of the present application. As Figure 4 shown, it includes the following steps:

[0229] Step 401: Send an AI task request to the first base station at the first moment. The AI task request carries AI resource scheduling requirement information, and the AI resource scheduling requirement information includes at least one of the following: AI task type, AI model type, and AI task data volume size;

[0230] Step 402: Receive the AI task result executed based on the AI task request sent by the first base station when the base station connected to the terminal is the first base station;

[0231] Step 403: Receive the AI task result executed by the first base station based on the AI task request sent by the second base station when the base station connected to the terminal is switched from the first base station to the second base station.

[0232] It should be noted that, as an implementation manner of the digital twin corresponding to the embodiment shown in Figure 1 , the specific implementation manner can refer to the relevant description of the embodiment shown in Figure 1 . To avoid repeated description, this embodiment will not be elaborated here, and the same beneficial effects can also be achieved.

[0233] Refer to Figure 5 , Figure 5 is a schematic diagram of the resource scheduling method provided by the embodiment of the present application. The implementation process of the specific resource scheduling method can be as follows: First, N terminals send AI task requests to the first base station at the first moment. The AI task requests carry AI resource scheduling requirement information. Then, the first base station forwards the AI resource scheduling requirement information to the SMO. After receiving the AI task requirement information, in order to better formulate a resource scheduling strategy, the SMO also needs to obtain the status information of the N terminals from the first base station. After obtaining the AI resource scheduling requirement information and status information of the N terminals, the SMO obtains the resource scheduling action data before the first moment, the network system status data before the first moment, and the actual resource scheduling efficiency data before the first moment from the digital twin (the digital twin is constructed based on the network status information and historical AI resource scheduling information between the first base station and each terminal) to train the initial scheduling strategy decision model. Then, the SMO inputs the obtained AI resource scheduling requirement information and status information of the N terminals into the trained scheduling strategy decision model to obtain the resource scheduling strategy for the N terminals. The SMO sends the obtained resource scheduling strategy to the first base station. The first base station allocates AI resources to the N terminals according to the resource scheduling strategy received from the SMO, and executes the corresponding AI tasks based on the allocated AI resources for the N terminals, and finally feeds back the completed results of the executed AI tasks to the corresponding terminals.

[0234] The embodiment of the application uses the digital twin to optimize the scheduling strategy decision model, completes pre-verification without affecting network operation, and reduces the trial-and-error cost. The resource scheduling strategy obtained through the scheduling strategy decision model meets the own needs and actual situations of the terminals, and allocates AI resources to each terminal specifically, thereby improving the efficiency of resource scheduling.

[0235] Refer to Figure 6 , Figure 6 is a structural diagram of a resource scheduling device applied to the SMO provided by an embodiment of the present application. The resource scheduling device 600 includes:

[0236] A first receiving module 601, configured to receive the AI resource scheduling requirement information carried in the AI task requests of N terminals sent by the first base station, where N is an integer greater than 1;

[0237] The first acquisition module 602 is configured to acquire the status information of the N terminals from the first base station;

[0238] The second acquisition module 603 is configured to input the AI resource scheduling requirement information and status information of the N terminals into a pre-trained scheduling policy decision model to obtain a resource scheduling policy for the N terminals. The scheduling policy decision model is obtained by training an initial scheduling policy decision model using a historical scheduling data set. The historical scheduling data set includes network system status data, resource scheduling action data, and actual resource scheduling efficiency data before a first moment. The resource scheduling action data is obtained from a pre-constructed digital twin. The digital twin is constructed based on the network status information and historical AI resource scheduling information between the first base station and each terminal. The first moment is the moment when the AI task request is received;

[0239] The first sending module 604 is configured to send the resource scheduling policy to the first base station.

[0240] Optionally, the AI resource scheduling requirement information includes at least one of the following: AI task type, AI model type, and size of AI task data volume;

[0241] And / or, the status information of the terminal includes at least one of the following: location information of the terminal, transmission channel quality information of the terminal, signal-to-noise ratio of the transmission channel of the terminal, transmission channel power of the terminal, and waiting time of the transmission data of the terminal in the first scheduling waiting queue of the first base station.

[0242] Optionally, the network system status data includes the AI task data volume and AI task limit conditions of each of the M terminals at each moment in a historical period, the total AI task data volume of the M terminals at each moment in the historical period, and the total AI task completion time of the M terminals at each moment in the historical period, where M is an integer greater than 1;

[0243] The resource scheduling action data includes the resource scheduling decisions of each of the M terminals at each moment in the historical period:

[0244] The actual resource scheduling efficiency data includes the actual resource scheduling efficiency of each of the M terminals at each moment in the historical period.

[0245] Optionally, the actual resource scheduling efficiency of each of the M terminals at each moment in the historical period is determined according to the transmission efficiency of each terminal to the first base station, the AI task execution efficiency of each terminal, and the waiting time of the transmission data of each terminal in the second scheduling waiting queue of the first base station;

[0246] Among them, the transmission efficiency of each terminal to the first base station is determined according to the total task data volume of the M terminals, the distance from each terminal to the first base station, the Gaussian white noise power of the transmission channel of each terminal, and the transmission channel gain of each terminal;

[0247] The AI task execution efficiency of each terminal is determined according to the AI task type of each terminal, the AI model of each terminal, the size of the AI task data volume of each terminal, and the AI resources scheduled by each terminal.

[0248] Optionally, the device further includes:

[0249] A first processing module, configured to obtain the network system state data and the actual resource scheduling efficiency data before the first moment based on the resource scheduling action data and the initial Actor network before the first moment;

[0250] A first determination module, configured to determine the expected resource scheduling efficiency data based on the network system state data and the resource scheduling action data before the first moment, and the initial Critic network;

[0251] A second determination module, configured to determine the loss value of the Critic network according to the expected resource scheduling efficiency data and the actual resource scheduling efficiency data;

[0252] A first adjustment module, configured to adjust the weights of the initial Critic network based on the loss value of the Critic network to obtain a trained Critic network;

[0253] A second adjustment module, configured to adjust the network parameters of the initial Actor network based on the trained Critic network.

[0254] Optionally, the device further includes:

[0255] A second sending module, configured to send the AI resource scheduling information at the first moment to the digital twin.

[0256] Optionally, the digital twin is located in the SMO, or the digital twin is a separate network element device.

[0257] Optionally, the first sending module includes:

[0258] A first deletion unit, configured to delete a first resource from the AI resource list of the SMO, where the first resource is the resource indicated to be scheduled in the resource scheduling policy;

[0259] A first judgment unit, configured to determine whether there is still AI resource scheduling requirement information waiting to be scheduled or whether the AI resource list is empty after deleting the first resource from the AI resource list;

[0260] A first sending unit, configured to send the resource scheduling policy to the first base station when there is no AI resource scheduling requirement information waiting to be scheduled or the AI resource list is empty.

[0261] It should be noted that, as the corresponding embodiment of the device side in Figure 1 , the specific implementation manner can refer to the relevant description of the embodiment shown in Figure 1 . To avoid repeated description, this embodiment will not be elaborated here, and the same beneficial effects can still be achieved.

[0262] Refer to Figure 7 , Figure 7 FIG. is the structural diagram of a resource scheduling device applied to a first base station provided in an embodiment of the present application. The resource scheduling device 700 includes:

[0263] A second receiving module 701, configured to receive AI task requests sent by N terminals at a first moment, where the AI task requests carry AI resource scheduling requirement information;

[0264] A third sending module 702, configured to send the AI resource scheduling requirement information and status information of the N terminals to the SMO;

[0265] A third receiving module 703, configured to receive the resource scheduling policy of the N terminals sent by the SMO, where the resource scheduling policy of the N terminals is obtained by the SMO inputting the AI resource scheduling requirement information and status information of the N terminals into a pre-trained scheduling policy decision model;

[0266] A first scheduling module 704, configured to schedule AI resources for each of the N terminals based on the resource scheduling policy of the N terminals;

[0267] A first execution module 705, configured to execute the AI tasks of each of the N terminals based on the AI resources scheduled for each of the N terminals, and obtain the AI task results of each of the N terminals.

[0268] Optionally, the N terminals include a first terminal, and the device further includes at least one of the following:

[0269] A fourth sending module, configured to send the AI task result of the first terminal to the first terminal when the base station connected to the first terminal is the first base station;

[0270] A fifth sending module, configured to directly send the AI task result of the first terminal to the second base station when the base station connected to the first terminal is switched from the first base station to the second base station and the second base station is on the service function chain;

[0271] A sixth sending module, configured to broadcast the location information of the first terminal when the base station connected to the first terminal is switched from the first base station to the second base station and the second base station is not on the service function chain, and send the AI task result of the first terminal to the second base station after the second base station confirms that the first terminal is the terminal communicating with it based on the location information of the first terminal.

[0272] It should be noted that, as the specific implementation manner of the corresponding device side in Figure 2 the specific implementation manner can refer to the relevant description of the corresponding Figure 2 corresponding Figure 1 embodiment shown in, in order to avoid repeated description, this embodiment will not be elaborated, and the same beneficial effects can also be achieved.

[0273] Refer to Figure 8 , Figure 8 which is the structural diagram of a resource scheduling device applied to a terminal provided in an embodiment of the present application. The resource scheduling device 800 includes:

[0274] A seventh sending module 801, configured to send an AI task request to a first base station at a first moment, where the AI task request carries AI resource scheduling requirement information, and the AI resource scheduling request information includes at least one of the following: AI task type, AI model type, and AI task data volume size;

[0275] A fourth receiving module 802, configured to receive the AI task result executed based on the AI task request sent by the first base station when the base station connected to the terminal is the first base station;

[0276] A fifth receiving module 803, configured to receive the AI task result executed by the first base station based on the AI task request sent by the second base station when the base station connected to the terminal is switched from the first base station to the second base station.

[0277] It should be noted that, as the specific implementation manner of the corresponding device side in Figure 3 the specific implementation manner can refer to the relevant description of the corresponding Figure 3 corresponding Figure 1 embodiment shown in, in order to avoid repeated description, this embodiment will not be elaborated, and the same beneficial effects can also be achieved.

[0278] See Figure 9 , Figure 9 FIG. Figure 9 is a structural diagram of a resource scheduling device applied to a digital twin provided in an embodiment of the present application. The resource scheduling device 900 includes:

[0279] An eighth sending module 901, configured to send resource scheduling action data before a first moment to the SMO, where the first moment is the moment when the SMO receives AI task requests of N terminals from a first base station, and N is an integer greater than 1;

[0280] A sixth receiving module 902, configured to receive the AI resource scheduling information at the first moment sent by the SMO, where the AI resource scheduling information at the first moment is obtained based on a resource scheduling policy output by a pre-trained scheduling policy decision model, and the scheduling policy decision model is obtained by training an initial scheduling policy decision model using the resource scheduling action data before the first moment.

[0281] It should be noted that, as a specific implementation manner of the corresponding device side in Figure 4 , the specific implementation manner can refer to the relevant description of the embodiment shown in Figure 4 corresponding Figure 1 . To avoid repeated description, this embodiment will not be elaborated herein, and the same beneficial effects can also be achieved.

[0282] An embodiment of the present application further provides an SMO. Since the principle of the SMO to solve problems is similar to that of the resource scheduling method in the embodiment of the present application, the implementation of the SMO can refer to the implementation of the method, and the repeated parts will not be elaborated. As Figure 10 shown, the SMO of the embodiment of the present application includes: A transceiver 1010 is configured to:

[0283] Receive AI resource scheduling requirement information carried in AI task requests of N terminals sent by a first base station, where N is an integer greater than 1;

[0284] Obtain the status information of the N terminals from the first base station;

[0285] A processor 1000 is configured to read a program in a memory 1020 and execute the following steps:

[0286] Input the AI resource scheduling requirement information and status information of the N terminals into a pre-trained scheduling policy decision model to obtain a resource scheduling policy for the N terminals. Among them, the scheduling policy decision model is obtained by training an initial scheduling policy decision model using a historical scheduling data set. The historical scheduling data set includes network system state data, resource scheduling action data, and actual resource scheduling efficiency data before a first moment. The resource scheduling action data is obtained from a pre-constructed digital twin, and the digital twin is constructed based on the network state information between the first base station and each terminal and historical AI resource scheduling information. The first moment is the moment when the AI task request is received;

[0287] The transceiver 1010 is used for:

[0288] Send the resource scheduling policy to the first base station.

[0289] Among them, in Figure 10 The bus architecture may include any number of interconnected buses and bridges, specifically, various circuits of one or more processors represented by the processor 1000 and the memory represented by the memory 1020 are linked together. The bus architecture can also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art. Therefore, they will not be further described herein. The bus interface provides an interface. The transceiver 1010 may be multiple elements, that is, including a transmitter and a transceiver, and provides a unit for communicating with various other devices on the transmission medium. The processor 1000 is responsible for managing the bus architecture and general processing, and the memory 1020 can store the data used by the processor 1000 when performing operations.

[0290] Optionally, the AI resource scheduling requirement information includes at least one of the following: AI task type, AI model type, AI task data volume size;

[0291] And / or, the status information of the terminal includes at least one of the following: the location information of the terminal, the transmission channel quality information of the terminal, the signal-to-noise ratio of the transmission channel of the terminal, the transmission channel power of the terminal, the waiting time of the transmission data of the terminal in the first scheduling waiting queue of the first base station.

[0292] Optionally, the network system state data includes the AI task data volume and AI task limit conditions of each of the M terminals at each moment in a historical period, the total AI task data volume of the M terminals at each moment in the historical period, the total AI task completion time of the M terminals at each moment in the historical period, where M is an integer greater than 1;

[0293] The resource scheduling action data includes the resource scheduling decisions of each of the M terminals at each moment in the historical period:

[0294] The actual resource scheduling efficiency data includes the actual resource scheduling efficiency of each of the M terminals at each moment in the historical period.

[0295] Optionally, the actual resource scheduling efficiency of each of the M terminals at each moment in the historical period is determined according to the transmission efficiency of each terminal to the first base station, the AI task execution efficiency of each terminal, and the waiting time of the transmission data of each terminal in the second scheduling waiting queue of the first base station;

[0296] Wherein, the transmission efficiency of each terminal to the first base station is determined according to the total task data volume of the M terminals, the distance of each terminal to the first base station, the Gaussian white noise power of the transmission channel of each terminal, and the transmission channel gain of each terminal;

[0297] The AI task execution efficiency of each terminal is determined according to the AI task type of each terminal, the AI model of each terminal, the size of the AI task data volume of each terminal, and the AI resources scheduled by each terminal.

[0298] Optionally, the scheduling policy decision model includes a policy Actor network and an evaluation Critic network; the processor 1000 is further configured to read the program in the memory 1020 and execute the following steps:

[0299] Based on the resource scheduling action data before the first moment and the initial Actor network, obtain the network system state data and the actual resource scheduling efficiency data before the first moment;

[0300] Based on the network system state data and the resource scheduling action data before the first moment, and the initial Critic network, determine the expected resource scheduling efficiency data;

[0301] According to the expected resource scheduling efficiency data and the actual resource scheduling efficiency data, determine the loss value of the Critic network;

[0302] Based on the loss value of the Critic network, adjust the weights of the initial Critic network to obtain a trained Critic network;

[0303] Based on the trained Critic network, adjust the network parameters of the initial Actor network.

[0304] Optionally, the transceiver 1010 is further configured to:

[0305] Send the AI resource scheduling information at the first moment to the digital twin.

[0306] Optionally, the digital twin is located in the SMO, or the digital twin is a separate network element device.

[0307] Optionally, the transceiver 1010 is specifically configured to:

[0308] Delete a first resource from the AI resource list of the SMO, where the first resource is the resource indicated for scheduling in the resource scheduling policy;

[0309] After deleting the first resource from the AI resource list, determine whether there is still AI resource scheduling requirement information waiting for scheduling, or whether the AI resource list is empty;

[0310] In the case where there is no AI resource scheduling requirement information waiting for scheduling or the AI resource list is empty, send the resource scheduling policy to the first base station.

[0311] The SMO provided by the embodiments of the present application can execute the above method embodiments, and its implementation principle and technical effects are similar, which will not be elaborated here in this embodiment.

[0312] The embodiments of the present application also provide a first base station. Since the principle of the first base station to solve problems is similar to the resource scheduling method in the embodiments of the present application, the implementation of the first base station can refer to the implementation of the method, and the repeated parts will not be elaborated. As Figure 11 shown, the first base station of the embodiments of the present application includes: a transceiver 1110, configured to:

[0313] Receive AI task requests sent by N terminals at a first moment, where the AI task requests carry AI resource scheduling requirement information;

[0314] Send the AI resource scheduling requirement information and status information of the N terminals to the SMO;

[0315] Receive the resource scheduling policy of the N terminals sent by the SMO, where the resource scheduling policy of the N terminals is obtained by the SMO inputting the AI resource scheduling requirement information and status information of the N terminals into a pre-trained scheduling policy decision model;

[0316] A processor 1100, configured to read a program in a memory 1120 and execute the following processes:

[0317] Based on the resource scheduling policy of the N terminals, schedule AI resources for each of the N terminals;

[0318] Execute the AI tasks of each of the N terminals based on the AI resources scheduled for each of the N terminals, and obtain the AI task results of each of the N terminals.

[0319] Among them, in Figure 11 the bus architecture may include any number of interconnected buses and bridges, specifically various circuits represented by one or more processors represented by processor 1100 and a memory represented by memory 1120 are linked together. The bus architecture can also link together various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art, and therefore, they will not be further described herein. The bus interface provides an interface. The transceiver 1110 can be multiple components, that is, including a transmitter and a transceiver, and provides a unit for communicating with various other devices on the transmission medium. The processor 1100 is responsible for managing the bus architecture and general processing, and the memory 1120 can store the data used by the processor 1100 when performing operations.

[0320] Optionally, the N terminals include a first terminal, and the transceiver 1110 is further configured to:

[0321] When the base station connected to the first terminal is the first base station, send the AI task result of the first terminal to the first terminal;

[0322] When the base station connected to the first terminal is switched from the first base station to a second base station and the second base station is on the service function chain, directly send the AI task result of the first terminal to the second base station;

[0323] When the base station connected to the first terminal is switched from the first base station to the second base station and the second base station is not on the service function chain, broadcast the location information of the first terminal, and after the second base station confirms that the first terminal is the terminal communicating with it based on the location information of the first terminal, send the AI task result of the first terminal to the second base station.

[0324] The first base station provided in the embodiments of the present application can execute the above method embodiments, and its implementation principles and technical effects are similar, and will not be elaborated herein.

[0325] The embodiments of the present application also provide a terminal. Since the principle of the terminal to solve problems is similar to the resource scheduling method in the embodiments of the present application, the implementation of the terminal can refer to the implementation of the method, and the repeated parts will not be elaborated. As Figure 12 shown, the computing node of the embodiments of the present application includes: The transceiver 1210 is configured to:

[0326] Send an AI task request to the first base station at the first moment, where the AI task request carries AI resource scheduling requirement information, and the AI resource scheduling requirement information includes at least one of the following: AI task type, AI model type, and size of AI task data volume;

[0327] When the base station connected to the terminal is the first base station, receive the AI task result executed by the first base station based on the AI task request;

[0328] When the base station connected to the terminal is switched from the first base station to the second base station, receive the AI task result executed by the first base station based on the AI task request and sent by the second base station.

[0329] Among them, in Figure 12 The bus architecture may include any number of interconnected buses and bridges, specifically, various circuits of one or more processors represented by the processor 1200 and the memory represented by the memory 1220 are linked together. The bus architecture can also link together various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art, so they will not be further described herein. The bus interface provides an interface. The transceiver 1210 may be multiple elements, that is, including a transmitter and a transceiver, and provides a unit for communicating with various other devices on the transmission medium. The processor 1200 is responsible for managing the bus architecture and general processing, and the memory 1220 may store data used by the processor 1200 when executing operations.

[0330] The terminal provided in the embodiment of the present application can execute the above method embodiment, and its implementation principle and technical effect are similar, which will not be elaborated herein.

[0331] The embodiment of the present application also provides a digital twin. Since the principle of solving problems by the digital twin is similar to the resource scheduling method in the embodiment of the present application, the implementation of the digital twin can refer to the implementation of the method, and the repeated parts will not be elaborated. As Figure 13 shown, the digital twin of the embodiment of the present application includes: The transceiver 1310 is used for:

[0332] Send resource scheduling action data before the first moment to the SMO, where the first moment is the moment when the SMO receives the AI task requests of N terminals from the first base station, and N is an integer greater than 1;

[0333] Receive the AI resource scheduling information at the first moment sent by the SMO, where the AI resource scheduling information at the first moment is obtained based on the resource scheduling policy output by the pre-trained scheduling policy decision model, and the scheduling policy decision model is obtained by training the initial scheduling policy decision model with the resource scheduling action data before the first moment.

[0334] Among them, in Figure 13 , the bus architecture may include any number of interconnected buses and bridges, specifically various circuits of one or more processors represented by the processor 1300 and the memory represented by the memory 1320 are linked together. The bus architecture may also link together various other circuits such as peripheral devices, voltage regulators, and power management circuits, etc., which are well known in the art, so they will not be further described herein. The bus interface provides an interface. The transceiver 1310 may be a plurality of components, that is, including a transmitter and a transceiver, and provides a unit for communicating with various other devices on the transmission medium. The processor 1300 is responsible for managing the bus architecture and general processing, and the memory 1320 may store data used by the processor 1300 when executing operations.

[0335] The digital twin provided by the embodiment of the present application can execute the above method embodiment, and its implementation principle and technical effect are similar, which will not be elaborated here in this embodiment.

[0336] The embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it realizes each process of the above resource scheduling method embodiment and can achieve the same technical effect. To avoid repetition, it will not be elaborated here. Among them, the computer-readable storage medium is, for example, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disc, etc.

[0337] It should be noted that in this article, the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. Without more limitations, an element defined by the statement "including a..." does not exclude the existence of another identical element in the process, method, article or device including that element.

[0338] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-described example methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions for causing a terminal (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in various embodiments of the present application.

[0339] The embodiments of the present application have been described above in conjunction with the accompanying drawings. However, the present application is not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative and not restrictive. Under the inspiration of the present application, those of ordinary skill in the art can also make many forms without departing from the purpose of the present application and the scope protected by the claims, and all of them belong to the protection scope of the present application.

Claims

1. A resource scheduling method, characterized in that, Executed by Service Management Orchestration (SMO), the method includes: Receiving AI resource scheduling requirement information carried in AI task requests of N terminals sent by a first base station, where N is an integer greater than 1; Obtaining the status information of the N terminals from the first base station; Inputting the AI resource scheduling requirement information and status information of the N terminals into a pre-trained scheduling policy decision model to obtain a resource scheduling policy for the N terminals. Among them, the scheduling policy decision model is obtained by training an initial scheduling policy decision model using a historical scheduling data set. The historical scheduling data set includes network system state data, resource scheduling action data, and actual resource scheduling efficiency data before a first moment. The resource scheduling action data is obtained from a pre-constructed digital twin. The digital twin is constructed based on network state information and historical AI resource scheduling information between the first base station and each terminal. The first moment is the moment when the AI task request is received; Sending the resource scheduling policy to the first base station.

2. The resource scheduling method according to claim 1, wherein The AI resource scheduling requirement information includes at least one of the following: AI task type, AI model type, size of AI task data volume; And / or, the status information of the terminal includes at least one of the following: location information of the terminal, transmission channel quality information of the terminal, signal-to-noise ratio of the transmission channel of the terminal, transmission channel power of the terminal, waiting time of the transmission data of the terminal in the first scheduling waiting queue of the first base station.

3. The resource scheduling method according to claim 1, wherein The network system state data includes the AI task data volume and AI task limit conditions of each of the M terminals at each moment in a historical period, the total AI task data volume of the M terminals at each moment in the historical period, and the total AI task completion time of the M terminals at each moment in the historical period, where M is an integer greater than 1; The resource scheduling action data includes the resource scheduling decisions of each of the M terminals at each moment in the historical period: The actual resource scheduling efficiency data includes the actual resource scheduling efficiency of each of the M terminals at each moment in the historical period.

4. The resource scheduling method according to claim 3, wherein The actual resource scheduling efficiency of each of the M terminals at each moment in the historical period is determined according to the transmission efficiency of each terminal to the first base station, the AI task execution efficiency of each terminal, and the waiting time of the transmission data of each terminal in the second scheduling waiting queue of the first base station; Among them, the transmission efficiency of each terminal to the first base station is determined according to the total task data volume of the M terminals, the distance from each terminal to the first base station, the Gaussian white noise power of the transmission channel of each terminal, and the transmission channel gain of each terminal; The AI task execution efficiency of each terminal is determined according to the AI task type of each terminal, the AI model of each terminal, the size of the AI task data volume of each terminal, and the AI resources scheduled by each terminal.

5. The resource scheduling method according to claim 1, wherein The scheduling policy decision model includes a policy Actor network and an evaluation Critic network; before inputting the AI resource scheduling requirement information and status information of the N terminals into the pre-trained scheduling policy decision model, the method further includes: Based on the resource scheduling action data before the first moment and the initial Actor network, obtaining the network system state data and actual resource scheduling efficiency data before the first moment; Based on the network system state data and resource scheduling action data before the first moment, and the initial Critic network, determining the expected resource scheduling efficiency data; According to the expected resource scheduling efficiency data and the actual resource scheduling efficiency data, determining the loss value of the Critic network; Adjusting the weights of the initial Critic network based on the loss value of the Critic network to obtain a trained Critic network; Adjusting the network parameters of the initial Actor network based on the trained Critic network.

6. The resource scheduling method according to claim 1, wherein After sending the resource scheduling policy to the first base station, the method further includes: Sending the AI resource scheduling information at the first moment to the digital twin.

7. The resource scheduling method according to claim 1, wherein The digital twin is located in the SMO, or the digital twin is a separate network element device.

8. The resource scheduling method according to claim 1, wherein Sending the resource scheduling policy to the first base station includes: Deleting a first resource from the AI resource list of the SMO, where the first resource is the resource indicated to be scheduled in the resource scheduling policy; After deleting the first resource from the AI resource list, determining whether there is still AI resource scheduling requirement information waiting to be scheduled, or whether the AI resource list is empty; In the case where there is no AI resource scheduling requirement information waiting to be scheduled or the AI resource list is empty, sending the resource scheduling policy to the first base station.

9. A resource scheduling method, characterized in that, Executed by the first base station, the method includes: Receiving AI task requests sent by N terminals at the first moment, where the AI task requests carry AI resource scheduling requirement information; Sending the AI resource scheduling requirement information and status information of the N terminals to the SMO; Receiving the resource scheduling policy of the N terminals sent by the SMO, where the resource scheduling policy of the N terminals is obtained by the SMO based on inputting the AI resource scheduling requirement information and status information of the N terminals into the pre-trained scheduling policy decision model; Based on the resource scheduling policy of the N terminals, scheduling AI resources for each of the N terminals; Based on the AI resources scheduled for each of the N terminals, executing the AI tasks of each of the N terminals to obtain the AI task results of each of the N terminals.

10. The resource scheduling method according to claim 9, wherein The N terminals include a first terminal. After executing the AI tasks of each of the N terminals based on the AI resources scheduled for each of the N terminals to obtain the AI task results of each of the N terminals, the method further includes at least one of the following: When the base station connected to the first terminal is the first base station, send the AI task result of the first terminal to the first terminal; When the base station connected to the first terminal is switched from the first base station to the second base station and the second base station is on the service function chain, directly send the AI task result of the first terminal to the second base station; When the base station connected to the first terminal is switched from the first base station to the second base station and the second base station is not on the service function chain, broadcast the location information of the first terminal, and after the second base station confirms that the first terminal is the terminal communicating with it based on the location information of the first terminal, send the AI task result of the first terminal to the second base station.

11. A resource scheduling method, characterized in that, Executed by the terminal, the method includes: Send an AI task request to the first base station at the first moment, where the AI task request carries AI resource scheduling requirement information, and the AI resource scheduling requirement information includes at least one of the following: AI task type, AI model type, size of AI task data volume; When the base station connected to the terminal is the first base station, receive the AI task result executed based on the AI task request sent by the first base station; When the base station connected to the terminal is switched from the first base station to the second base station, receive the AI task result executed by the first base station based on the AI task request sent by the second base station.

12. A resource scheduling method, characterized in that, Executed by the digital twin, the method includes: Send the resource scheduling action data before the first moment to the SMO, where the first moment is the moment when the SMO receives the AI task requests of N terminals from the first base station, and N is an integer greater than 1; Receive the AI resource scheduling information at the first moment sent by the SMO, where the AI resource scheduling information at the first moment is obtained based on the resource scheduling policy output by the SMO using the pre-trained scheduling policy decision model, and the scheduling policy decision model is obtained by training the initial scheduling policy decision model using the resource scheduling action data before the first moment.

13. A resource scheduling device, characterized in that, Applied to the SMO, the resource scheduling device includes: A first receiving module, configured to receive the AI resource scheduling requirement information carried in the AI task requests of N terminals sent by the first base station, where N is an integer greater than 1; A first obtaining module, configured to obtain the status information of the N terminals from the first base station; A second acquisition module, configured to input the AI resource scheduling requirement information and status information of the N terminals into a pre-trained scheduling policy decision model to obtain a resource scheduling policy for the N terminals. The scheduling policy decision model is obtained by training an initial scheduling policy decision model using a historical scheduling data set. The historical scheduling data set includes network system state data, resource scheduling action data, and actual resource scheduling efficiency data before a first moment. The resource scheduling action data is obtained from a pre-constructed digital twin. The digital twin is constructed based on the network state information between the first base station and each terminal and historical AI resource scheduling information. The first moment is the moment when the AI task request is received. A first sending module, configured to send the resource scheduling policy to the first base station.

14. A resource scheduling device, characterized in that, Applied to the first base station, the resource scheduling device includes: A second receiving module, configured to receive an AI task request sent by N terminals at a first moment. The AI task request carries AI resource scheduling requirement information. A third sending module, configured to send the AI resource scheduling requirement information and status information of the N terminals to the SMO. A third receiving module, configured to receive the resource scheduling policy of the N terminals sent by the SMO. The resource scheduling policy of the N terminals is obtained by the SMO inputting the AI resource scheduling requirement information and status information of the N terminals into a pre-trained scheduling policy decision model. A first scheduling module, configured to schedule AI resources for each of the N terminals based on the resource scheduling policy of the N terminals. A first execution module, configured to execute the AI tasks of each of the N terminals based on the AI resources scheduled for each of the N terminals to obtain the AI task results of each of the N terminals.

15. A resource scheduling device, characterized in that, Applied to terminal execution, the resource scheduling device includes: A seventh sending module, configured to send an AI task request to the first base station at a first moment. The AI task request carries AI resource scheduling requirement information. The AI resource scheduling request information includes at least one of the following: AI task type, AI model type, and AI task data volume size. A fourth receiving module, configured to receive the AI task result executed based on the AI task request sent by the first base station when the base station connected to the terminal is the first base station. A fifth receiving module, configured to receive the AI task result executed by the first base station based on the AI task request sent by the second base station when the base station connected to the terminal is switched from the first base station to the second base station.

16. A resource scheduling device, characterized in that, Applied to digital twin execution, the resource scheduling device includes: An eighth sending module, configured to send resource scheduling action data before a first moment to the SMO. The first moment is the moment when the SMO receives an AI task request of N terminals from the first base station, and N is an integer greater than 1. A sixth receiving module, configured to receive the AI resource scheduling information at the first moment sent by the SMO, where the AI resource scheduling information at the first moment is obtained based on a resource scheduling policy output by a pre-trained scheduling policy decision model, and the scheduling policy decision model is obtained by training an initial scheduling policy decision model with resource scheduling action data before the first moment.

17. A SMO, characterized in that, The SMO includes a transceiver and a processor, and the transceiver is configured to: Receive AI resource scheduling requirement information carried in AI task requests of N terminals sent by a first base station, where N is an integer greater than 1; Obtain the status information of the N terminals from the first base station; The processor is configured to: Input the AI resource scheduling requirement information and status information of the N terminals into a pre-trained scheduling policy decision model to obtain a resource scheduling policy for the N terminals. Among them, the scheduling policy decision model is obtained by training an initial scheduling policy decision model with a historical scheduling data set. The historical scheduling data set includes network system state data, resource scheduling action data, and actual resource scheduling efficiency data before the first moment. The resource scheduling action data is obtained from a pre-constructed digital twin, and the digital twin is constructed based on the network state information between the first base station and each terminal and historical AI resource scheduling information. The first moment is the moment when the AI task request is received; The transceiver is configured to: Send the resource scheduling policy to the first base station.

18. A first base station, characterized in that, The first base station includes a transceiver and a processor, and the transceiver is configured to: Receive AI task requests sent by N terminals at the first moment, where the AI task requests carry AI resource scheduling requirement information; Send the AI resource scheduling requirement information and status information of the N terminals to the SMO; Receive the resource scheduling policy of the N terminals sent by the SMO, where the resource scheduling policy of the N terminals is obtained by the SMO inputting the AI resource scheduling requirement information and status information of the N terminals into a pre-trained scheduling policy decision model; The processor is configured to: Schedule AI resources for each of the N terminals based on the resource scheduling policy of the N terminals; Execute the AI tasks of each of the N terminals based on the AI resources scheduled for each of the N terminals to obtain the AI task results of each of the N terminals.

19. A terminal, characterized in that, The terminal includes a transceiver, and the transceiver is configured to: Send an AI task request to the first base station at the first moment, where the AI task request carries AI resource scheduling requirement information, and the AI resource scheduling requirement information includes at least one of the following: AI task type, AI model type, and size of AI task data volume; Receive the AI task result executed based on the AI task request sent by the first base station when the base station to which the terminal is connected is the first base station; When the base station connected to the terminal is switched from the first base station to the second base station, receive the AI task result executed by the first base station based on the AI task request sent by the second base station.

20. A digital twin, characterized in that, The digital twin includes a transceiver, and the transceiver is configured to: Send resource scheduling action data before the first moment to the SMO, where the first moment is the moment when the SMO receives AI task requests of N terminals from the first base station, and N is an integer greater than 1; Receive the AI resource scheduling information at the first moment sent by the SMO, where the AI resource scheduling information at the first moment is obtained based on the resource scheduling policy output by the pre-trained scheduling policy decision model, and the scheduling policy decision model is obtained by training the initial scheduling policy decision model using the resource scheduling action data before the first moment.

21. An electronic device, characterized in that, Includes a processor, a memory, and a computer program stored on the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the resource scheduling method according to any one of claims 1 to 8, or when the computer program is executed by the processor, it implements the steps of the base station control method according to any one of claims 9 to 10, or when the computer program is executed by the processor, it implements the steps of the resource scheduling method according to claim 11, or when the computer program is executed by the processor, it implements the steps of the resource scheduling method according to any one of claims 12.

22. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium. When the computer program is executed by the processor, it implements the steps of the resource scheduling method according to any one of claims 1 to 8, or when the computer program is executed by the processor, it implements the steps of the base station control method according to any one of claims 9 to 10, or when the computer program is executed by the processor, it implements the steps of the resource scheduling method according to claim 11, or when the computer program is executed by the processor, it implements the steps of the resource scheduling method according to any one of claims 12.

Images