Artificial intelligence or machine learning (ai / ML)-based wireless communication method and wireless communication device

Abstract

Provided in the present disclosure is an artificial intelligence or machine learning (AI / ML)-based wireless communication method, which is executed at a user equipment. The method comprises at least one of the following: reporting first capability information supporting AI / ML; receiving a first reporting configuration of channel state information (CSI); receiving a CSI request or a model activation message; and determining AI / ML-based CSI processing time, and reporting the CSI on the basis of the AI / ML-based CSI processing time.

Description

A wireless communication method and wireless communication device based on artificial intelligence or machine learning (AI / ML) Technical Field

[0001] This disclosure relates to the field of wireless communication, and more particularly to a wireless communication method and wireless communication device based on artificial intelligence or machine learning (AI / ML). Background Technology

[0002] AI / ML-based CSI prediction can prevent channel aging. Its flowchart is shown in Figure 1. By taking historical CSI information as model input, it can output predicted CSI at multiple future time points.

[0003] AI / ML-based CSI compression aims to increase CSI accuracy while reducing the air interface overhead of CSI reporting. AI / ML-based CSI space-frequency domain compression is based on an encoder-decoder model. The encoder, deployed at the UE, uses space-frequency domain channel information (such as precoding matrices and eigenvectors) obtained from channel measurements as model input and outputs compressed channel information. The decoder, deployed at the base station, uses the compressed channel information as model input and outputs complete channel information. A quantizer can be deployed after the encoder to further reduce the transmission overhead of CSI feedback, forming a complete CSI generation module. Correspondingly, a dequantizer needs to be deployed before the decoder to output the unquantized compressed channel information, forming a complete CSI reconstruction module. The introduction of AI / ML will significantly impact CSI design; therefore, it is necessary to propose a wireless communication method and device based on artificial intelligence or machine learning (AI / ML) to improve existing technologies. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a wireless communication method based on artificial intelligence or machine learning (AI / ML) to address the above-mentioned deficiencies in the prior art, thereby solving the problems existing in the prior art.

[0005] According to one aspect of this disclosure, a wireless communication method based on artificial intelligence or machine learning (AI / ML) is provided, executed in a user equipment, the method comprising:

[0006] Report information on the first capabilities that support AI / ML.

[0007] According to one aspect of this disclosure, a wireless communication method based on artificial intelligence or machine learning (AI / ML) is provided, executed in a user equipment, the method comprising:

[0008] Receive a CSI request or activation model message; wherein the message includes at least one of the following: Downlink Control Information (DCI), Media Access Control Element (MACCE), and / or Radio Resource Control (RRC).

[0009] According to one aspect of this disclosure, a wireless communication method based on artificial intelligence or machine learning (AI / ML) is provided, executed in a user equipment, wherein the method includes:

[0010] Determine the CPU required for Channel State Information (CSI) reporting. Based on the total number of CPUs, the CPUs already occupied, and the CPUs required for CSI reporting, determine the CSIs that need to be reported.

[0011] Report the required CSI information.

[0012] According to one aspect of this disclosure, a wireless communication method based on artificial intelligence or machine learning (AI / ML) is provided, executed in a user equipment, wherein the method includes:

[0013] Report information on supporting AI / ML-related secondary capabilities.

[0014] According to one aspect of this disclosure, a wireless communication method based on artificial intelligence or machine learning (AI / ML) is provided, executed in a user equipment, wherein the method includes:

[0015] A first processing unit is determined, wherein the first processing unit is used to calculate the computing resources available to the user equipment;

[0016] Based on the first processing unit, the user equipment determines the tasks to be processed in parallel.

[0017] According to one aspect of this disclosure, a wireless communication method based on artificial intelligence or machine learning (AI / ML) is provided, executed in an access network device, the method comprising:

[0018] Receive first capability information supporting AI / ML;

[0019] A first reporting configuration for transmitting Channel State Information (CSI), wherein the first reporting configuration is determined based on the first capability information.

[0020] According to one aspect of this disclosure, a wireless communication method based on artificial intelligence or machine learning (AI / ML) is provided, executed in an access network device, the method comprising:

[0021] Send a message requesting a CSI or activating a model; wherein the message includes at least one of the following: Downlink Control Information (DCI), Media Access Control Element (MACCE), and / or Radio Resource Control (RRC).

[0022] According to one aspect of this disclosure, a wireless communication method based on artificial intelligence or machine learning (AI / ML) is provided, wherein the method is performed on an access network device, the method comprising:

[0023] Receive information on secondary capabilities related to AI / ML.

[0024] According to one aspect of this disclosure, a wireless communication method based on artificial intelligence or machine learning (AI / ML) is provided, executed in an access network device, wherein the method includes:

[0025] Receive mapping information, third capability information, and / or fourth capability information;

[0026] According to one aspect of this disclosure, a wireless communication device is provided, including a processor and a memory for storing a computer program, the processor for calling and running the computer program stored in the memory to perform steps in the data processing method as described in any of the preceding claims.

[0027] According to one aspect of this disclosure, a readable storage medium is provided for storing a computer program that is invoked and executed by a processor to perform any of the methods described above. Attached Figure Description

[0028] To more clearly illustrate the embodiments of this disclosure or related technologies, the following figures will be briefly described in the embodiments. Obviously, the figures are merely some embodiments of this disclosure, and those skilled in the art can obtain other figures based on these figures without creative effort.

[0029] Figure 1 illustrates a flowchart of CSI prediction in the prior art provided in this disclosure.

[0030] Figure 2 illustrates a schematic diagram of the wireless communication system architecture provided in this disclosure.

[0031] Figure 3 illustrates a flowchart of the unified CSI processing provided in this disclosure.

[0032] Figure 4 illustrates a flowchart of the CSI parallel processing provided in this disclosure.

[0033] Figure 5 illustrates a flowchart of the wireless communication method provided in this disclosure.

[0034] Figure 6 illustrates an exemplary block diagram of a wireless communication system provided in this disclosure. Detailed Implementation

[0035] The embodiments of this disclosure have been described in detail with reference to the accompanying drawings, outlining technical aspects, structural features, objectives, and effects, as described below. Specifically, the terminology used in the embodiments of this disclosure is for the purpose of describing particular embodiments only and is not intended to limit the disclosure.

[0036] In this disclosure, “A or B” may mean “A only”, “B only”, or “both A and B”.

[0037] In other words, in this disclosure, “A or B” can be interpreted as “A and / or B”. For example, in this disclosure, “A, B or C” can mean “A only”, “B only”, “C only” or “any combination of A, B, and C”.

[0038] The forward slash ( / ) or comma used in this disclosure can mean "and / or". For example, "A / B" can mean "A and / or B". Therefore, "A / B" can mean "A only", "B only", or "both A and B". For example, "A, B, C" can mean "A, B, or C".

[0039] In this disclosure, "at least one of A and B" may mean "only A", "only B" or "both A and B". Furthermore, in this disclosure, the expression "at least one of A or B" or "at least one of A and / or B" may be interpreted as "at least one of A and B".

[0040] Additionally, in this disclosure, "at least one of A, B, and C" may mean "A only", "B only", "C only" or "any combination of A, B, and C". Furthermore, "at least one of A, B, or C" or "at least one of A, B, and / or C" may mean "at least one of A, B, and C".

[0041] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0042] Those skilled in the art will recognize and understand that the details of the described examples are merely illustrative of some embodiments, and that the teachings set forth herein are applicable to various alternative settings.

[0043] The technical solutions disclosed herein can be applied to various wireless communication systems, such as: Long Term Evolution (LTE) systems, LTE Frequency Division Duplex (FDD) systems, LTE Time Division Duplex (TDD) systems, 5G communication systems, or future wireless communication systems, etc.

[0044] For example, the wireless communication system 100 of this disclosure is shown in FIG2. The wireless communication system 100 may include a base station 110, which may be a device communicating with user equipment (UE) 120. The base station 110 can provide communication coverage for a specific geographical area and can communicate with user equipment located within that coverage area. Optionally, the base station 110 may be an evolved Node B (eNB or eNodeB) in an LTE system, or it may be a mobile switching center, relay station, access point, vehicle-mounted equipment, wearable device, hub, switch, bridge, router, network-side equipment in a 5G network, or a base station in a future communication system, etc.

[0045] The wireless communication system 100 also includes at least one user equipment 120 located within the coverage area of ​​the base station 110. "User equipment" as used herein includes, but is not limited to, devices configured to receive / transmit communication signals via wired connections, such as via Public Switched Telephone Networks (PSTN), Digital Subscriber Line (DSL), digital cable, direct cable connection; and / or another data connection / network; and / or via a wireless interface, such as for cellular networks, Wireless Local Area Networks (WLAN), digital television networks such as DVB-H networks, satellite networks, AM-FM broadcast transmitters; and / or other user equipment. User equipment configured to communicate via a wireless interface may be referred to as a "wireless communication terminal," "wireless terminal," or "mobile terminal." Examples of mobile terminals include, but are not limited to, satellite or cellular phones; personal communications system (PCS) terminals that can combine cellular radiotelephone with data processing, fax, and data communication capabilities; PDAs that may include radiotelephones, pagers, Internet / intranet access, web browsers, notebooks, calendars, and / or Global Positioning System (GPS) receivers; and conventional laptop and / or handheld receivers or other electronic devices that include radiotelephone transceivers. User equipment can refer to access terminals, user units, user stations, mobile stations, mobile stations, remote stations, remote user equipment, mobile devices, wireless communication equipment, or user agents. Access terminals can be cellular phones, cordless phones, Session Initiation Protocol (SIP) phones, Wireless Local Loop (WLL) stations, personal digital assistants (PDAs), handheld devices with wireless communication capabilities, computing devices or other processing devices connected to a wireless modem, in-vehicle equipment, wearable devices, user equipment in 5G networks, or user equipment in future PLMN evolutions, etc.

[0046] Optionally, user equipment 120 can perform device-to-device (D2D) communication with each other.

[0047] Alternatively, 5G communication systems or 5G networks may also be referred to as New Radio (NR) systems or NR networks.

[0048] The wireless communication system 100 also includes a core network 130. The core network 130 may be an IP mobile communication network operated by a mobile communication operator. For example, the core network 130 may be the core network used by the mobile communication operator that operates and manages the wireless communication system 100, or it may be the core network used by a virtual mobile communication operator such as an MVNO (Mobile Virtual Network Operator).

[0049] The core network 130 can connect to the base station 110, serving as a relay device for transmitting user data. User equipment 120 transmits and receives user data via the core network 130. It should be noted that user data communication is not limited to IP communication; it can also be non-IP communication.

[0050] Figure 2 exemplarily illustrates a base station 110, two user equipment 120 and a core network 130. Optionally, the wireless communication system 100 may include multiple base stations and each base station may include other numbers of user equipment within its coverage area, which is not limited in this disclosure.

[0051] Optionally, the wireless communication system 100 may also include other network entities such as a network controller, a mobility management entity, and network elements, and this disclosure does not limit this. For example, the core network 130 may include other network entities such as a network controller, a mobility management entity, and network elements, and this disclosure does not limit this.

[0052] It should be understood that devices with wireless communication capabilities in the network / system of this disclosure may be referred to as wireless communication devices. Taking the wireless communication system 100 shown in Figure 2 as an example, the wireless communication device may include a base station 110, a user equipment 120, and a core network 130 with communication capabilities. The base station 110 and the user equipment 120 may be the specific devices described above, which will not be repeated here. The wireless communication device may also include other devices in the wireless communication system 100 (core network 130). For example, the core network 130 may include other network entities such as network controllers and mobility management entities, which are not limited in this disclosure.

[0053] The prior art has at least one of the following problems:

[0054] Question 1: If AI / ML models are applied to AI / ML-based CSI enhancement use cases such as beam management, CSI prediction, and CSI compression, the resources (e.g., computing units, memory, storage) required to run AI / ML models differ from those required by traditional non-AI / ML-based CSI computation methods. Therefore, current methods for quantifying CPU usage in non-AI / ML-based CSI computation are not applicable to AI / ML-based CSI computation. Furthermore, AI / ML models can run on different hardware modules. For example, models can run alongside traditional non-AI algorithms on a central processing unit (CPU), or they can run independently on a graphics processing unit (GPU) or neural network processing unit (NPU), or on modules designed for different functions. These different running methods will affect the total available CPU and CPU usage.

[0055] Question 2: With significant advancements in hardware capabilities in the future, user devices will have multiple built-in models that can be activated simultaneously to obtain multiple inference results (e.g., results obtained from the same measurement set, but different inference values ​​obtained through models corresponding to different use cases). Parallel model processing will reduce processing latency. How will the relevant computing resources be determined, along with the corresponding model selection mechanism and UE behavior when computing resources are insufficient?

[0056] Question 3: Compared to traditional CSI measurements, AI / ML-based CSI calculations require consideration of model migration and activation to ensure model usability, which increases CSI calculation time. Therefore, the CSI calculation time corresponding to non-AI / ML-based CSI measurements is not applicable to AI / ML-based CSI calculations. Furthermore, since different functions, features, and model architectures require different computational complexities, the corresponding CSI calculation times will also differ. Therefore, fixed quantification methods for non-AI / ML-based CSI calculation times are not applicable to AI / ML-based CSI calculations. Additionally, in model performance monitoring scenarios, the device workflow can be divided into an AI part (model data collection, model inference, and post-processing of model output) and a non-AI part (verification of model output results).

[0057] Therefore, compared to model inference, the CSI Report for model performance monitoring needs to consider both the processing time of Ground Truth and how to design the processing time of model inference. Additionally, for UE-sided models, the computation time of performance metrics also needs to be considered.

[0058] To address the aforementioned problems in the existing technology, the following technical solution is adopted.

[0059] The information sending method provided in this application will be described in detail below with reference to the accompanying drawings and through some embodiments and application scenarios.

[0060] Figure 3 illustrates one of the flowcharts for the AI / ML-based wireless communication method provided in this disclosure. Figure 3 shows a unified flowchart for CSI processing under the AI / ML air interface, applicable to different air interface use cases and different LCM stages. The parts marked with dashed boxes are optional, and their selection depends on the different LCM stages and signaling configuration methods. Furthermore, the various processing time drawbacks related to the timeline need to be considered in conjunction with the specific parameters of the AI / ML processing.

[0061] The following describes the process using the model monitoring phase as an example. The method includes at least one of the following steps (it is worth noting that only some of the steps in Figure 3 are shown in the following example):

[0062] Step S100: The user equipment reports the first capability information supporting AI / ML to the base station.

[0063] Specifically, user equipment or base stations are just one example; in practice, they can also be transformed into other nodes. The first capability information may include at least one of the following: whether it supports CSI processing based on AI / ML models, supported AI / ML-related capabilities, whether it supports serial / parallel processing of CSI determination based on non-AI / ML and CSI determination based on AI / ML, whether it supports parallel operation of multiple model tasks, total CPU usage, currently occupied CPU, or the total number of IPUs used to determine AI / ML computing resources. The IPUs are described below.

[0064] Specifically, whether the first capability information supports CSI processing based on AI / ML models includes at least one of the following:

[0065] 1. AI / ML CSI reports to a dedicated resource pool for computation; or

[0066] 2. Does it support the use of independent resource pools for AI / ML CSI reporting calculations across different use cases / functions?

[0067] Among them, the supported AI / ML capabilities in the first capability information include at least one of the following: supported AI use cases, supported AI functions, supported AI features, supported AI feature groups, and / or supported AI models.

[0068] Among them, the total CPU in the first capability information can be N, which is used to determine the CSI processing resources for CSI reporting. CPU .

[0069] Among them, the currently occupied CPU in the first capability information can be the currently occupied CSI processing resource L. CPU .

[0070] Specifically, the total number of IPUs used to determine AI / ML computing resources in the first capability information can be N used to determine AI computing resources. IPU .

[0071] Optionally, the first capability information may also include the currently occupied CSI processing resources L. IPU .

[0072] User equipment reports first capability information to the base station so that the base station can configure CSI reporting appropriately based on the first capability information.

[0073] Step S200: Receive the first reporting configuration of Channel State Information (CSI), wherein the first reporting configuration is determined based on the first capability information.

[0074] Specifically, the first reporting configuration instructs the user equipment to report based on at least one of the following: CSI based on non-AI / ML and / or CSI based on AI / ML.

[0075] Once the user equipment receives the first report configuration of the Channel State Information (CSI), it can determine the resource and other information reported by the CSI.

[0076] Step S300: Receive a CSI request or activation model message; wherein the message includes at least one of the following: Downlink Control Information (DCI), Media Access Control Element (MACCE), and / or Radio Resource Control (RRC).

[0077] Specifically, the message can be a downlink message. The user equipment executes the CSI processing flow according to the instructions or configuration of the downlink message, as shown in Figure 3.

[0078] Step S400: Determine the CPU required for CSI reporting. Based on the total number of CPUs, the CPUs already occupied, and the CPUs required for CSI reporting, determine the CSIs that need to be reported.

[0079] In some embodiments of this disclosure, when CSI reporting based on non-AI / ML and CSI reporting based on AI / ML share the CPU, the total CPU amount reported by non-AI / ML and the total CPU amount reported by AI / ML are determined by at least one of the following methods: the total CPU amount reported by non-AI / ML and the total CPU amount reported by AI / ML each account for a certain proportion of the total CPU supported by the UE, and / or the proportion of each total CPU amount related to AI / ML CSI reporting in the total CPU supported by the user equipment is determined according to the Lifecycle Management (LCM) link and / or AI attributes.

[0080] It is worth noting that there is no restriction on the sharing of CPU between CSI reporting based on non-AI / ML and CSI reporting based on AI / ML. In some cases, this restriction may not be necessary.

[0081] Specifically, if non-AI / ML-based CSI reporting and AI / ML-based CSI reporting share the same CSI Processing Unit (CPU), the total number of CPUs supported by the UE for non-AI / ML-based CSI reporting and the total number of CPUs supported by AI / ML-based CSI reporting shall be allocated in at least one of the following ways:

[0082] 1. The total CPU used for non-AI / ML-based CSI reporting and the total CPU used for AI / ML-based CSI reporting each account for a certain proportion of the total CPU supported by the UE. Assuming the proportion of CPU used for non-AI / ML-based CSI reporting is r1, and the proportion of CPU used for AI / ML-based CSI reporting is r2, then r1 + r2 ≤ 1. The above proportions are determined through at least one of the following methods: predefined, reported by the UE to the base station, and / or configured by the base station to the UE; or...

[0083] 2. Allocate the proportion of total CPU related to AI / ML-based CSI reporting to the total CPU supported by the UE based on the Life Cycle Management (LCM) stage and / or AI attributes. The LCM stage includes model training, model inference, and model monitoring. The AI ​​attributes are defined using at least one of the following methods:

[0084] 1) AI attributes are defined as the model inference method. At this time, AI attributes include one-sided models and two-sided models;

[0085] 2) AI attributes are defined as functions, including but not limited to AI / ML-based beam management, AI / ML-based CSI prediction, and / or AI / ML-based CSI compression;

[0086] or

[0087] 3) AI attributes are defined as AI use cases, where AI attributes include, but are not limited to, AI / ML-based spatial beam prediction (BM-Case1), AI / ML-based temporal beam prediction (BM-Case2), AI / ML-based CSI prediction, AI / ML-based spatial-frequency CSI compression, and / or AI / ML-based spatiotemporal-frequency CSI compression.

[0088] For example, the proportion of the total CPU related to each AI / ML-based CSI reporting to the total CPU supported by the UE is allocated according to the LCM stage and AI attributes. This proportion can be indicated by Table 1, where r p,q This represents the proportion of the total CPU corresponding to the p-th LCM stage and the q-th AI attribute to the total CPU supported by the UE, satisfying... The p-th type of LCM step can be one of the following: model training, model monitoring, and / or model inference.

[0089] Table 1 shows the percentage of total CPUs related to AI / ML-based CSI reporting out of the total CPUs supported by the UE.

[0090] It should be noted that Table 1 is an example of an indicator table showing the proportion of total CPU usage related to AI / ML-based CSI reporting to the total CPU usage supported by the UE. Each indicator table may include more or fewer LCM stages, AI attributes, and the corresponding relationships between these stages and the total CPU usage related to AI / ML-based CSI reporting. Furthermore, there are no restrictions on the relationships between LCM stages, AI attributes, and the total CPU usage related to AI / ML-based CSI reporting in Table 1. That is, adding, deleting, modifying, or altering the relationships between LCM stages, AI attributes, and the total CPU usage related to AI / ML-based CSI reporting to the total CPU usage supported by the UE can result in the indicator table for the proportion of total CPU usage related to AI / ML-based CSI reporting to the total CPU usage supported by the UE that is protected by this disclosure.

[0091] Alternatively, assume that the total number of CPUs supported by the UE is N. CPUIf L is the number of CPUs used for CSI reporting on the current symbol, then N is the number of CPUs remaining available. CPU -L, if there are N CSI reports on the current symbol that require CPU usage, the UE will execute M of the CSI reports according to the priority rules, where 0 ≤ M ≤ N. The maximum value of M, The CPU required for the nth CSI report is determined by the order of the CSI report. The earlier the CSI report number, the higher its priority. That is, the CSI report with n=0 has the highest priority, and the CSI report with n=N-1 has the lowest priority.

[0092] Optionally, assume that the total number of CPUs supported by the UE for non-AI / ML-based CSI reporting and the total number of CPUs supported for AI / ML-based CSI reporting are N respectively. CPU,non-AI and N CPU,AI The number of CPUs used for non-AI / ML-based CSI reporting and the number of CPUs used for AI / ML-based CSI reporting on the current symbol are L respectively. non-AI and L AI Then, the remaining available CPUs for non-AI / ML-based CSI reporting and the remaining available CPUs for AI / ML-based CSI reporting are N and N, respectively. CPU,non-AI -L non-AI and N CPU,AI -L AI If there are N CSI reports on the current symbol that require CPU usage, of which N1 are non-AI / ML based CSI reports and N2 are AI / ML based CSI reports, then the UE shall take at least one of the following actions:

[0093] 1. Execute M1+M2 CSI reports according to the priority rules, where 0≤M1≤N1 is satisfied. The maximum value of M1, The CPU usage required for the n1th non-AI / ML based CSI report is determined by the order of the CSI report; the earlier the CSI report number, the higher its priority, where 0 ≤ M² ≤ N². The maximum value of M2, The CPU usage required for the n2nd AI / ML-based CSI report is determined by the order of the CSI report; the earlier the CSI report number, the higher its priority.

[0094] 2. Execute M1+M2 CSI reports according to the priority rules, where 0≤M1≤N1 is considered satisfactory. The maximum value of M1, The CPU usage required for the n1th non-AI / ML-based CSI report is determined by the order of the CSI report; the earlier the CSI report number, the higher its priority, provided that 0 ≤ M2 ≤ N1 + N2. The maximum value of M2, To determine the required CPU usage for the n2nd CSI report in both AI / ML-based CSI reporting and non-AI / ML-based CSI reporting (where CPU is not used for non-AI / ML-based CSI reporting), the earlier the CSI report number, the higher its priority; or

[0095] 3. Execute M1+M2 CSI reports according to the priority rules, where 0≤M2≤N2 is considered satisfactory. The maximum value of M2, The CPU usage required for the n2nd AI / ML-based CSI report is determined by the order of the CSI report; the earlier the CSI report number, the higher its priority, where 0 ≤ M1 ≤ N1 + N2. The maximum value of M1, The CPU required for the n1st CSI report in the CSI report based on AI / ML (excluding AI / ML-based CSI reports) is determined by the order of the CSI report. The earlier the CSI report number, the higher its priority.

[0096] It is worth noting that the priority of the above CSI reporting depends on at least one of the following parameters: reporting configuration ID, serving cell index, reporting content, reporting method, CSI solution method, LCM stage, and / or model inference method. Among them, the CSI solution method includes AI / ML-based CSI solution and non-AI / ML-based CSI solution, which are represented by indexes. Each index indicates a CSI solution method, and the index value range is {0,1}. The LCM stage includes model training, model monitoring, and / or model inference, which are represented by indexes. Each index indicates an LCM stage, and the index value range is {0,1,2}. The model inference method includes single-sided model and two-sided model, which are represented by indexes. Each index indicates an inference method, and the index value range is {0,1}.

[0097] Optionally, if AI / ML-based CSI reporting has a higher priority than non-AI / ML-based CSI reporting, the priority value of CSI reporting can be expressed as follows: Pri CSI (f,y,k,c,s)=8·N cells ·M s ·f+2·N cells ·M s ·y+N cells ·Ms ·k+M s ·c+s

[0098] In the above formula, f=0 represents CSI solution based on AI / ML, f=1 represents CSI solution based on non-AI / ML, y=0 represents aperiodic CSI reporting based on PUSCH, y=1 represents semi-persistent CSI reporting based on PUSCH, y=2 represents semi-persistent CSI reporting based on PUCCH, y=3 represents periodic CSI reporting based on PUCCH, k=0 represents CSI reporting carrying L1-RSRP or L1-SINR, k=1 represents CSI reporting without carrying L1-RSRP or L1-SINR, c represents serving cell index, s represents reporting configuration ID, and N... cells M represents the maximum number of serving cells. s This indicates the maximum number of CSI reporting configuration IDs. The lower the priority value for a CSI report, the higher its priority.

[0099] Optionally, if the priority of AI / ML-based CSI reporting is higher than that of non-AI / ML-based CSI reporting, and the priority of model inference is higher than that of model monitoring and the priority of model monitoring is higher than that of model training, then in the above formula, f=0 represents model inference, f=1 represents model monitoring, f=2 represents model training, and f=3 represents non-AI / ML-based CSI solving.

[0100] Optionally, if the priority of AI / ML-based CSI reporting is higher than that of non-AI / ML-based CSI reporting, and the priority of a one-sided model is higher than that of a two-sided model, then f=0 in the above formula represents a one-sided model, f=1 represents a two-sided model, and f=2 represents a non-AI / ML-based CSI solution.

[0101] In some embodiments of this disclosure, when AI / ML-based CSI reporting is based on a CPU dedicated to AI / ML, and several functions of CSI reporting correspond to a CPU dedicated to AI / ML, the CPU occupied by the AI / ML-based CSI reporting is determined in at least one of the following ways: offset relative to the CPU occupied by non-AI / ML-based CSI reporting, and / or CPU associated with LCM links and / or AI attributes.

[0102] It's worth noting that the AI / ML-based CSI reporting mentioned above is based on a dedicated AI / ML CPU, and the condition that several CSI reporting functions correspond to one dedicated AI / ML CPU only applies to certain situations. In other cases, this restriction may not apply. Furthermore, the functions can be the same or different functions.

[0103] Specifically, if AI / ML-based CSI reporting uses a dedicated CPU for AI / ML, that is, a CPU independent of non-AI / ML-based CSI reporting, while CSI reporting for different functions shares the dedicated AI / ML CPU, then the CPU used by AI / ML-based CSI reporting can be represented in at least one of the following forms:

[0104] 1. An offset relative to the CPU usage of non-AI / ML-based CSI reporting is used to reduce indication overhead. This offset value is related to the LCM (Local Management Module) and / or AI attributes. A description of the LCM and AI attributes is provided in step S400. This offset value can be determined through at least one of the following methods: predefined, reported by the UE to the base station, and / or configured by the base station to the UE. An example of determining the offset value is when there are few AI attribute candidates, such as when the AI ​​attribute is defined as a model inference method; in this case, a predefined method can be used. Another example is when the UE reports applicable functionality or activated functionality, or when the base station indicates the required function; in this case, the UE can report an offset value associated with the AI ​​attribute to the base station to further reduce indication overhead. An example of the above offset value is that the offset value is related to the LCM and AI attributes; in this case, the offset value can be indicated by Table 2, where Δ... p,q This represents the offset value corresponding to the p-th LCM step and the q-th AI attribute. The p-th LCM step can be one of the following: model training, model monitoring, or model inference; or

[0105] Table 2 shows the CPU offset relative to non-AI / ML-based CSI reporting.

[0106] It should be noted that Table 2 is an example of an offset value indication table relative to the CPU usage of CSI reports based on non-AI / ML. This table may include more or fewer LCM steps, AI attributes, and their corresponding offset values. Furthermore, there are no restrictions on the correspondence between LCM steps, AI attributes, and their corresponding offset values ​​in Table 2. In other words, adding, deleting, modifying, or altering the correspondence between LCM steps, AI attributes, and their corresponding offset values ​​in Table 2 can yield the offset value indication table for CPU usage of CSI reports based on non-AI / ML protected by this disclosure.

[0107] 2. CPU associated with LCM and / or AI attributes. The description of LCM and AI attributes is provided in step S400. This value can be determined through at least one of the following methods: predefined, reported by the UE to the base station, and / or configured by the base station to the UE. One implementation for determining the CPU is when there are few AI attribute candidates, such as when the AI ​​attribute is defined as model inference mode; in this case, a predefined method can be used. Another implementation is that when the UE reports available or activated functions, or when the base station indicates the required function, the UE can report the CPU associated with the AI ​​attribute to the base station to reduce indication overhead. One implementation of the CPU occupied by the above CSI reporting is based on the fact that the CPU occupied by AI / ML CSI reporting is related to LCM and / or AI attributes. In this case, the CPU occupied by CSI reporting can be indicated by Table 3, where... This represents the CPU usage corresponding to the p-th LCM step and the q-th AI attribute. The p-th LCM step can be one of the following: model training, model monitoring, and / or model inference.

[0108] Table 3 CPU usage of AI / ML-based CSI reporting

[0109] It should be noted that Table 3 is an example of a CPU usage indicator table for AI / ML-based CSI reporting. Such a table may include more or fewer LCM steps, and the correspondence between AI attributes and CPU usage in AI / ML-based CSI reporting. Furthermore, there are no restrictions on the correspondence between LCM steps, AI attributes, and CPU usage in AI / ML-based CSI reporting in Table 3. In other words, the CPU usage indicator table for AI / ML-based CSI reporting that is protected in this disclosure can be obtained by adding, deleting, modifying, or altering the correspondence between LCM steps, AI attributes, and CPU usage in AI / ML-based CSI reporting based on Table 3.

[0110] Optionally, when the LCM step is model monitoring, considering that the monitoring data may simultaneously include CSI obtained from non-AI / ML measurements and CSI obtained from AI / ML inference, the CPU occupied by AI / ML-based CSI reporting can be composed of the CPU used for non-AI / ML-based CSI reporting and / or the CPU used for AI / ML-based CSI reporting, i.e., O CPU,AI =f(O CPU,AI,a +O CPU,AI,b )

[0111] In the above formula, O CPU,AI,a This indicates the portion of CSI reporting to the CPU based on non-AI / ML methods, O CPU,AI,brepresents the part of AI / ML-based CSI reporting CPU.

[0112] For an implementation of the above formula, the monitoring data includes both CSI obtained based on non-AI / ML measurement and CSI obtained based on AI / ML inference, that is, the CPU occupied by AI / ML-based CSI reporting is composed of the CPU for non-AI / ML-based CSI reporting and the CPU for AI / ML-based CSI reporting. Meanwhile, the non-AI / ML-based CSI solving and the AI / ML-based CSI solving operate independently, and the above formula can be expressed as O CPU,AI =O CPU,AI,a +O CPU,AI,b .

[0113] For another implementation of the above formula, the monitoring data includes both CSI obtained based on non-AI / ML measurement and CSI obtained based on AI / ML inference, that is, the CPU occupied by AI / ML-based CSI reporting is composed of the CPU for non-AI / ML-based CSI reporting and the CPU for AI / ML-based CSI reporting. Meanwhile, the non-AI / ML-based CSI solving and the AI / ML-based CSI solving share the measurement of reference signals, and the above formula can be expressed as O CPU,AI =a·O CPU,AI,a +b·O CPU,AI,b , wherein coefficients 0<a≤1 and 0<b≤1 are determined by at least one of the following manners: predefinition, reporting by a UE to a base station, and / or configuration by a base station to a UE.

[0114] In some embodiments of the present disclosure, when the AI / ML-based CSI reporting is based on a CPU dedicated to AI / ML, and the CSI reporting of each function corresponds to an exclusive CPU respectively, the CPU occupied by the AI / ML-based CSI reporting is determined in at least one of the following manners: an offset relative to the CPU occupied by non-AI / ML-based CSI reporting, an offset relative to a CPU associated with an AI attribute, and / or a CPU associated with an LCM link and / or model complexity.

[0115] It is worth noting that the above condition that AI / ML-based CSI reporting is based on a CPU dedicated to AI / ML and CSI reporting of multiple functions corresponds to one exclusive CPU is only for certain cases, and in other cases, the above condition restriction may also be omitted. In addition, the functions may be the same functions or different functions.

[0116] Specifically, if AI / ML-based CSI reporting uses a dedicated CPU for AI / ML, i.e., independent of the CPU used for non-AI / ML-based CSI reporting, and different functions of CSI reporting use their own dedicated CPUs, then the total number of dedicated CPUs for each function can be expressed in at least one of the following forms:

[0117] 1. An offset relative to the total CPU usage reported for non-AI / ML-based CSI, to reduce indication overhead, wherein the offset value is related to AI attributes and / or UE performance. A description of AI attributes is provided in the corresponding content of step S400. UE performance includes at least one of the following parameters: number of cores, number of threads, core operating frequency, core operating voltage, core power, and / or battery capacity. This offset value can be determined in at least one of the following ways: predefined, reported by the UE to the base station, and / or configured by the base station to the UE. In one implementation, the offset value is determined when there are few AI attribute candidates, such as when the AI ​​attribute is defined as model inference mode, in which case a predefined method can be used. In another implementation, the offset value is determined when the UE reports available or activated functions or when the base station indicates the required function; the UE can report an offset value associated with the AI ​​attribute to the base station to further reduce indication overhead. In one implementation, the offset value is related to AI attributes, which can be indicated by Table 4, where Δ... q This represents the offset value corresponding to the q-th AI attribute; or

[0118] Table 4 shows the offset values ​​relative to the total CPU usage for non-AI / ML-based CSI reporting.

[0119] It should be noted that Table 4 is an example of an offset value indication table relative to the total CPU volume reported by CSI based on non-AI / ML. Such an offset value indication table may include more or fewer offset values ​​corresponding to AI attributes. Furthermore, there are no restrictions on the correspondence between offset values ​​and AI attributes in Table 4. That is, adding, deleting, modifying, or altering the correspondence between offset values ​​and AI attributes based on Table 4 can yield the offset value indication table relative to the total CPU volume reported by CSI based on non-AI / ML that is protected by this disclosure.

[0120] 2. Total CPU count associated with AI attributes and / or UE performance. A description of AI attributes is provided in step S400. UE performance includes at least one of the following parameters: number of cores, number of threads, core operating frequency, core operating voltage, core power, and / or battery capacity. The total CPU count can be determined in at least one of the following ways: predefined, reported by the UE to the base station, and / or configured by the base station to the UE. In one implementation, the total CPU count is determined when there are few AI attribute candidates, such as when the AI ​​attribute is defined as model inference mode, a predefined method can be used. In another implementation, when the UE reports available or activated functions, or when the base station indicates the required function, the UE can report the total CPU count associated with the AI ​​attributes to the base station to reduce indication overhead. In one implementation, the total CPU count for each function is related to the AI ​​attribute, which can be indicated by Table 5. This represents the total amount of dedicated CPU corresponding to the q-th AI attribute.

[0121] Table 5 shows the total dedicated CPU for each function.

[0122] It should be noted that Table 5 is an example of a dedicated CPU total quantity indication table for each function. Each function's dedicated CPU total quantity indication table may include more or fewer CPU total quantity and AI attribute correspondences. Furthermore, there are no restrictions on the correspondence between CPU total quantity and AI attributes in Table 5. That is, by adding, deleting, modifying, or altering the correspondence between CPU total quantity and AI attributes based on Table 5, a dedicated CPU total quantity indication table for each function protected by this disclosure can be obtained.

[0123] The CPU usage of AI / ML-based CSI reporting can be represented in at least one of the following forms:

[0124] 1. The CPU usage is offset relative to non-AI / ML-based CSI reporting to reduce indication overhead. Assume that non-AI / ML-based CSI reporting uses O CPU. CPU The CPU usage of AI / ML-based CSI reporting is represented as O. CPU,AI =O CPU +Δ AI,1 +Δ AI,2 The first offset value Δ AI,1Regarding AI attributes, a description of the AI ​​attributes is provided in step S400. The first offset value can be determined through at least one of the following methods: predefined, reported by the UE to the base station, and / or configured by the base station to the UE. In one implementation, the first offset value is determined when there are few AI attribute candidates, such as when the AI ​​attribute is defined as a model inference method; in this case, a predefined method can be used. In another implementation, when the UE reports available or activated functions, or when the base station indicates the required function, the UE can report the first offset value associated with the AI ​​attribute to the base station to further reduce indication overhead. In one implementation, the first offset value can be indexed through Table 6, where... This represents the first offset value corresponding to the q-th AI attribute.

[0125] Table 6 shows the first offset of CPU usage relative to non-AI / ML-based CSI reporting.

[0126] It should be noted that Table 6 is an example of a first offset value indication table for CPU usage relative to non-AI / ML-based CSI reports. Such a table may include more or fewer correspondences between first offset values ​​and AI attributes. Furthermore, there are no restrictions on the correspondences between first offset values ​​and AI attributes in Table 6. That is, the table of first offset values ​​for CPU usage relative to non-AI / ML-based CSI reports, which is protected by this disclosure, can be obtained by adding, deleting, modifying, or altering the correspondences between first offset values ​​and AI attributes in Table 6.

[0127] Second offset value Δ AI,1 Related to the LCM stage and / or model complexity, the description of the LCM stage can be found in the corresponding content in step S400. The model complexity is defined using at least one of the following methods:

[0128] 1) Model complexity is defined as the model backbone, where model complexity includes, but is not limited to, Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and / or Transformer; or

[0129] 2) Model complexity is defined as computational cost. The unit of model complexity includes, but is not limited to, floating-point operations per second (FLOPS), MFLOPS, GFLOPS, TFLOPS, multiply-accumulate operations (MACs), MMACs, GMACs, and / or TMACs. Model complexity can be expressed in at least one of the following forms:

[0130] ① Directly indicates the value of model complexity;

[0131] ② Use index indicators, where each index corresponds to at least one of the following values: model complexity range, minimum model complexity, and / or maximum model complexity;

[0132] ③ Model complexity is defined as model structure. In this case, model complexity is represented by at least one of the following parameters: number of layers, number of neurons, number of modules, and / or number of self-attention heads. The definition of a module depends on the model type; for example, the modules of a Long Short-Term Memory (LSTM) network are cells, and the modules of a Transformer are blocks. These parameters can be directly indicated or indicated by indices. When using indices, each index corresponds to at least one of the following values: parameter range, minimum parameter value, and / or maximum parameter value; or

[0133] 4. Model complexity is defined as the size of the model parameters. The unit of model complexity includes, but is not limited to, thousands, millions, billions, and / or trillions. Model complexity can be indicated directly or by index. When using index, each index corresponds to at least one of the following values: model complexity range, minimum model complexity, and / or maximum model complexity.

[0134] The second offset value can be determined in at least one of the following ways: predefined, reported by the UE to the base station, and / or configured by the base station to the UE. In one implementation, the second offset value is determined when the UE reports at least one of the following information: available function, activated function, applicable model, and / or activated model, or when the base station indicates the function or model to be used. The UE may report a second offset value associated with the model complexity corresponding to the interacted information to the base station to further reduce indication overhead. In one implementation, the second offset value is related to the LCM link and model complexity; in this case, the offset value can be indicated by Table 7, where... This represents the second offset value corresponding to the p-th LCM stage and the s-th model complexity. The p-th LCM stage can be one of the following: model training, model monitoring, or model inference.

[0135] Table 7 shows the second offset of CPU usage relative to non-AI / ML-based CSI reporting.

[0136] It should be noted that Table 7 is an example of a second offset value indication table relative to the CPU usage of CSI reports based on non-AI / ML. This table may include more or fewer LCM stages, and the correspondence between model complexity and second offset values. Furthermore, there are no restrictions on the correspondence between LCM stages, model complexity, and second offset values ​​in Table 7. That is, the table of second offset values ​​relative to the CPU usage of CSI reports based on non-AI / ML can be obtained by adding, deleting, modifying, or altering the correspondence between LCM stages, model complexity, and second offset values ​​based on Table 7.

[0137] 2. Offset relative to the CPU associated with the AI ​​attribute to reduce indication overhead. Assume the CPU associated with the AI ​​attribute is O'. CPU,AI The CPU usage of AI / ML-based CSI reporting is represented as O. CPU,AI =O' CPU,AI +Δ AI,3 , where O' CPU,AIRegarding AI attributes, a description of the AI ​​attributes is provided in step S400. The CPU can be determined through at least one of the following methods: predefined, reported by the UE to the base station, and / or configured by the base station to the UE. In one implementation, the CPU is determined when there are few AI attribute candidates, such as when the AI ​​attribute is defined as a model inference method; in this case, a predefined method can be used. In another implementation, when the UE reports available or activated functions, or when the base station indicates the required function, the UE can report the CPU associated with the AI ​​attribute to the base station to reduce indication overhead. In one implementation, the CPU can be indicated through Table 8, where... This represents the CPU corresponding to the q-th AI attribute.

[0138] Table 8 CPU associated with AI attributes

[0139] It should be noted that Table 8 is an example of a CPU indicator table associated with AI attributes. A CPU indicator table associated with AI attributes may include more or fewer correspondences between CPUs and AI attributes. Furthermore, there are no restrictions on the correspondences between CPUs and AI attributes in Table 8. That is, by adding, deleting, modifying, or altering the correspondences between CPUs and AI attributes based on Table 8, a CPU indicator table associated with AI attributes, as protected by this disclosure, can be obtained.

[0140] Offset value Δ AI,3 The offset value is related to the LCM process and / or model complexity. Details regarding the LCM process and model complexity are provided in step S400. This offset value can be determined through at least one of the following methods: predefined, reported by the UE to the base station, and / or configured by the base station to the UE. In one implementation, the offset value is determined when the UE reports at least one of the following information: available functions, activated functions, available models, and / or activated models, or when the base station indicates the required function or model. The UE can then report an offset value associated with the model complexity corresponding to the interacted information to the base station to further reduce indication overhead. In one implementation, the offset value is related to the LCM process and model complexity. In this case, the offset value can be indicated using Table 9, where... This represents the offset between the p-th LCM stage and the s-th model complexity. The p-th LCM stage can be one of the following: model training, model monitoring, and / or model inference; or

[0141] Table 9 shows the CPU offset values ​​relative to AI attributes.

[0142] It should be noted that Table 9 is an example of an offset value indication table relative to the CPU associated with the AI ​​attribute. This table may include more or fewer LCM stages, and the correspondence between model complexity and offset values. Furthermore, there are no restrictions on the correspondence between LCM stages, model complexity, and offset values ​​in Table 9. That is, the offset value indication table relative to the CPU associated with the AI ​​attribute, which is protected by this disclosure, can be obtained by adding, deleting, modifying, or altering the correspondence between LCM stages, model complexity, and offset values ​​in Table 9.

[0143] 3. CPU associated with LCM stage and / or model complexity. The descriptions of LCM stage and model complexity are detailed in step S400. This CPU can be determined through at least one of the following methods: predefined, reported by the UE to the base station, and / or configured by the base station to the UE. In one implementation, the CPU is determined when the UE reports at least one of the following information: available functions, activated functions, available models, and / or activated models, or when the base station indicates the required function or model. The UE can then report the CPU associated with the model complexity corresponding to the interacted information to the base station to reduce indication overhead. In one implementation, the CPU occupied by AI / ML-based CSI reporting is related to LCM stage and model complexity. In this case, the CPU occupied by CSI reporting can be indicated by Table 10, where... This represents the CPU usage corresponding to the p-th LCM stage and the s-th model complexity. The p-th LCM stage can be one of the following: model training, model monitoring, and / or model inference.

[0144] Table 10 CPU usage of AI / ML-based CSI reporting

[0145] It should be noted that Table 10 is an example of a CPU indicator table for AI / ML-based CSI reporting. A CPU indicator table for AI / ML-based CSI reporting may include more or fewer LCM stages and a mapping between model complexity and CPU usage. Furthermore, there are no restrictions on the mapping between LCM stages, model complexity, and CPU usage in Table 10. That is, the CPU indicator table for AI / ML-based CSI reporting protected in this disclosure can be obtained by adding, deleting, modifying, or altering the mapping between LCM stages, model complexity, and CPU usage in Table 10.

[0146] Step S500: Determine the AI / ML-based CSI processing time related to the CSI to be reported, and report the CSI to be reported based on the AI / ML-based CSI processing time related to the CSI to be reported and / or the first reporting configuration.

[0147] In the existing technology, when the CSI Request is transmitted as DCI, only the DCI decoding time is consumed, so the DCI decoding time in the processing time is consistent with the traditional method.

[0148] In some embodiments of this disclosure, the AI / ML-based CSI processing time includes the processing time of the CSI request. When the message is transmitted through the MAC CE, the processing time of the CSI request includes at least one of the following: the time for processing the MAC CE, the DCI decoding time, and / or the data decoding time.

[0149] Specifically, the UE determines that the processing time for a CSI request includes DCI decoding time, while Data Decoding time is optional. When the CSI request is transmitted via MAC-CE, the processing time includes the time for processing MAC-CE (i.e., the time for the MAC message to be parsed and transmitted at the upper layer), DCI decoding time, and / or data decoding time.

[0150] In some embodiments of this disclosure, the AI / ML-based CSI processing time includes the processing time of the CSI request. When the message is transmitted through the RRC, the processing time of the CSI request includes at least one of the following: the time for processing the RRC, the DCI decoding time, and / or the data decoding time.

[0151] Specifically, the UE determines that the processing time for a CSI request includes DCI decoding time, while data decoding time is optional. When the CSI request is configured to use RRC, the processing time includes RRC processing time (the time it takes for the RRC message to be parsed and transmitted at the upper layer), DCI decoding time, and / or data decoding time.

[0152] In some embodiments of this disclosure, the AI / ML-based CSI processing time is determined based on a first value, and the AI / ML-based CSI processing time includes at least one of the following: an offset of the non-AI / ML-based CSI processing time and / or the product of the non-AI / ML-based CSI processing time and an offset factor.

[0153] Specifically, assuming the processing time for traditional CSI reporting (i.e., non-AI / ML-based CSI reporting) is Z, the processing time for AI / ML-based CSI is Z0. AI It may be in at least one of the following forms:

[0154] 1. The processing time for AI / ML-based CSI is offset from the processing time for non-AI / ML-based CSI to reduce indication overhead, i.e., Z. AI =Z+Δ Z , where Δ Z The offset value is related to the AI ​​attribute and / or model complexity. Descriptions of the AI ​​attribute and model complexity are detailed in step S400. This offset value can be determined through pre-definition and / or by the UE reporting to the base station. In one implementation, the offset value is determined when the UE reports at least one of the following information: available function, activated function, available model, and / or activated model, or when the base station indicates the required function or model. The UE can report an offset value associated with the AI ​​attribute and / or model complexity corresponding to the interacted information to the base station to further reduce indication overhead. In one implementation, the offset value is determined by the model complexity, and in this case, the offset value can be indicated through Table 11, where Δ... Z,s This represents the offset value corresponding to the complexity of the s-th model; or

[0155] Table 11 CSI processing time offset values ​​associated with model complexity

[0156] It should be noted that Table 11 is an example of a CSI processing time offset value indicator table associated with model complexity. A CSI processing time offset value indicator table associated with model complexity may include more or fewer correspondences between model complexity and offset values. Furthermore, there are no restrictions on the correspondence between model complexity and offset values ​​in Table 11. That is, the CSI processing time offset value indicator table associated with model complexity, which is protected by this disclosure, can be obtained by adding, deleting, modifying, or altering the correspondence between model complexity and offset values ​​in Table 11.

[0157] 2. The processing time for AI / ML-based CSI is the product of the processing time for non-AI / ML-based CSI and the offset factor, i.e. To reduce indication overhead, where k is an offset factor related to AI attributes and / or model complexity, the specific details of which are described in step S400, can be found. This offset factor can be determined through predefinition and / or by the UE reporting to the base station. In one implementation, the offset factor is determined when the UE reports at least one of the following information: available functions, activated functions, available models, and / or activated models, or when the base station indicates the required function or model. The UE can then report an offset factor associated with the AI ​​attributes and / or model complexity corresponding to the interacted information to the base station, further reducing indication overhead. In another implementation, the offset factor is determined by the model complexity, and in this case, the offset factor can be indicated through Table 12, where k... s This represents the offset factor corresponding to the complexity of the s-th model.

[0158] Table 12 CSI processing time offset factor associated with model complexity

[0159] It should be noted that Table 12 is an example of a CSI processing time offset factor indicator table associated with model complexity. A CSI processing time offset factor indicator table associated with model complexity may include more or fewer correspondences between model complexity and offset factors. Furthermore, there are no restrictions on the correspondence between model complexity and offset factors in Table 12. That is, the CSI processing time offset factor indicator table associated with model complexity, which is protected by this disclosure, can be obtained by adding, deleting, modifying, or altering the correspondence between model complexity and offset factors in Table 12.

[0160] In some embodiments of this disclosure, the AI / ML-based CSI processing time is a predefined value.

[0161] Specifically, the CSI processing time Z based on AI / ML AI Predefined values ​​are used, and the optional predefined values ​​include, but are not limited to, Z2, 2Z2, 3Z2, 4Z2, 5Z2, and / or 6Z2.

[0162] In addition, the device needs to perform the following two types of measurements in both the AI-based and non-AI-based sections:

[0163] Measurement Behavior 1: In order to perform model inference, data is collected to obtain the model's input.

[0164] At this point, the time occupied only includes traditional beam switching.

[0165] Measurement Action 2: To verify the performance of the model inference results, measure the ground truth of the corresponding model inference results.

[0166] At this time, the time consumed includes Beam switching and CSI Computation (GT).

[0167] Step S501, Model Inference.

[0168] Specifically, the model needs to perform model inference, which includes at least one of the following steps:

[0169] 1. Model Transmission. Model transmission is necessary when there is no available model on the UE side, or when the model needs to be updated. It's worth noting that model transmission is optional.

[0170] 1) Additional Model Transfer time is introduced when the model transfer is not completed before the relevant CSI-RS measurements are performed;

[0171] 2) The Model Transfer time can be obtained from NW or from OTT Server depending on the model transmission method, and there are two corresponding configurations.

[0172] 2. Loading the model.

[0173] 1. After receiving the model activation command from DCI / MAC-CE / RRC, the user needs to perform preparatory work for model activation, such as the time required to load all model parameters into memory. In some embodiments of this disclosure, the time including a series of model preparation tasks is uniformly referred to as the model activation time.

[0174] 3. Reasoning of the model.

[0175] Model Inference is the time required for users to perform model inference calculations and output results while completing model loading and user data collection.

[0176] In some embodiments of this disclosure, the processing time related to model inference is determined based on at least one of the following parameters: model migration time, model activation time, and / or model inference time.

[0177] Specifically, for model inference, the processing time Z related to model inference is... AI Determined by at least one of the following parameters: model transfer time Z transfer Model activation time Z activation And / or model inference time Z inference , can be represented as: Z AI =f(Z) transferZ activation Z inference )

[0178] The definitions of the above parameters are as follows:

[0179] 1. If the model required for model inference is not present within the UE, the UE needs to obtain the model from other devices. The model migration time is Z. transfer Use at least one of the following definition methods:

[0180] 1) The model is transmitted to the UE from the network side or other UEs via the air interface. The model is carried on the PDSCH. In this case, the model migration time is the time interval between the DCI carrying the CSI request and the completion of decoding the PDSCH carrying the model; or

[0181] 2) When the model is transmitted to the UE from an external 3GPP system device, such as an OTT server, the model migration time can be determined in at least one of the following ways: predefined, reported by the UE to the base station, and / or configured by the base station to the UE. In one implementation, the model migration time is the actual model migration time reported by the UE to the base station when a cell handover or OTT server change occurs, or the estimated model migration time indicated by the base station to the UE. When the base station indicates the model migration time to the UE, this time can take at least one of the following forms: maximum migration time, minimum migration time, and / or a migration time range;

[0182] 2. If the model required for inference within the UE is not yet enabled, the UE needs to activate the model. The model activation time is Z. activation The deactivation model unloading time and / or activation model loading time are determined by the deactivation model unloading time and / or activation model loading time. The deactivation model unloading time exists when memory is insufficient, requiring the deactivated model to be erased. This time can be determined through predefinition and / or by the UE reporting to the base station. The activation model loading time represents the time required to write the desired model into memory, and this time can also be determined through predefinition and / or by the UE reporting to the base station. In one implementation, the deactivation model unloading time and / or activation model loading time are determined when the UE performs model switching or when the base station indicates the desired model. The UE can report the deactivation model unloading time and / or activation model loading time to the base station.

[0183] 3. Model inference time Z inferenceThe inference time is determined by AI attributes and / or model complexity, with descriptions of AI attributes and model complexity provided in step S400. This time can be determined through predefinition and / or by the UE reporting to the base station. In one implementation, the inference time is determined when the UE reports at least one of the following information: available functions, activated functions, available models, and / or activated models, or when the base station indicates the required function or model. The UE may then report the model inference time associated with the AI ​​attributes and / or model complexity corresponding to the interacted information to the base station, thereby reducing indication overhead.

[0184] For example, one implementation of the above-mentioned model inference time is that the model inference time is determined by AI attributes, and in this case, the model inference time can be indicated by Table 13, where... This represents the model inference time corresponding to the q-th AI attribute.

[0185] Table 13 Model inference time associated with AI attributes

[0186] It should be noted that Table 13 is an example of a model inference time indicator table associated with AI attributes. A model inference time indicator table associated with AI attributes may include more or fewer correspondences between model inference time and AI attributes. Furthermore, there are no restrictions on the correspondence between model inference time and AI attributes in Table 13. That is, by adding, deleting, modifying, or altering the correspondence between model inference time and AI attributes based on Table 13, a model inference time indicator table associated with AI attributes, as protected by this disclosure, can be obtained.

[0187] Another implementation is that the model inference time is determined by the model complexity. In this case, the model inference time can be indicated by Table 14, where... This represents the model inference time corresponding to the complexity of the s-th model.

[0188] Table 14 Model inference time related to model complexity

[0189] It should be noted that Table 14 is an example of a model inference time indicator table related to model complexity. A model inference time indicator table related to model complexity may include more or fewer correspondences between model inference times for different model complexities. Furthermore, there are no restrictions on the correspondences between model inference times for different model complexities in Table 14. That is, by adding, deleting, modifying, or altering the correspondences between model inference times for different model complexities based on Table 14, a model inference time indicator table related to model complexity, as protected by this disclosure, can be obtained.

[0190] One implementation of the above formula is when the UE does not contain the model required for model inference. In this case, the processing time related to model inference is determined by the model migration time, model activation time, and model inference time, i.e., Z. AI =Z transfer +Z activation +Z inference +Δ offset

[0191] Another implementation of the above formula is when the UE contains a model required for model inference, but the model is not activated. In this case, the processing time related to model inference is determined by the model activation time and the model inference time, i.e., Z. AI =Z activation +Z inference +Δ offset

[0192] Another implementation of the above formula is that the UE contains a model required for model inference, and this model is already enabled. In this case, the processing time related to model inference is determined by the model inference time, i.e., Z. AI =Z inference +Δ offset

[0193] The above Δ offset This represents the offset of the model inference time, assuming the time interval between DCI reception and the reception of the last reference signal corresponding to model inference is Z. base And the sum of the CSI request processing time and beam switching time (optional) is Z. preparation When Z preparation +Z transfer +Z activation ≥Z base At that time, Δ offset =0, otherwise Δ offset =Z base -Z preparation -Z transfer -Z activation .

[0194] Step S502: Obtain the model's performance metrics; when the model is monitored by a user device, the AI / ML-based CSI processing time includes the performance metric calculation time.

[0195] Specifically, after the device obtains the model inference results and the ground truth, it needs to use the two to calculate the model performance metrics according to the calculation principles specified in the protocol, in order to measure whether the model's output meets the requirements.

[0196] In one implementation, when the model is monitored by the user equipment (i.e., UE-side model monitoring): the UE needs to perform performance metric calculations, which can be divided into two paths: monitoring using statistical values ​​and monitoring using instantaneous values. Both paths involve computation of large-scale data; therefore, the performance metric calculation time (Metrics Computation) also needs to be considered.

[0197] In another implementation, when the model is monitored by the access network equipment (i.e., the base station's model monitoring): the UE does not need to perform performance metric calculations, and the performance metric calculation time (Metrics Computation) does not need to be considered.

[0198] It is worth noting that the preparation time for Tx is the same as the traditional method.

[0199] Step S503: Report the required CSI.

[0200] It is worth noting that, in practice, there are no restrictions on the execution order of steps S100 to S500. That is, any number of the described steps can be skipped or combined in any order to implement the method or an alternative method. Similarly, for step S500, there are no restrictions on the execution order of steps S501 to S503. That is, any number of the described steps can be skipped or combined in any order to implement the method or an alternative method.

[0201] The following describes the process using the model inference stage as an example. The method includes at least one of the following steps:

[0202] Step H100: The user equipment reports its first AI / ML capability information to the base station. Step H100 is identical in detail to step S100 in the model monitoring phase.

[0203] Step H200: Receive the first reporting configuration of Channel State Information (CSI), wherein the first reporting configuration is determined based on the first capability information. Step H200 is identical in detail to step S200 in the model monitoring phase.

[0204] Step H300: Receive a CSI request or a model activation message; wherein the message includes at least one of the following: Downlink Control Information (DCI), Media Access Control Element (MACCE), and / or Radio Resource Control (RRC). Step H300 is identical in detail to step S300 in the model monitoring phase.

[0205] Step H400: Determine the CPU required for CSI reporting. Based on the total number of CPUs, the CPUs already occupied, and the CPUs required for CSI reporting, determine the CSIs to be reported. Step H400 is identical in detail to step S400 in the model monitoring phase.

[0206] Step H500: Determine the required AI / ML-based CSI processing time for CSI. The required AI / ML-based CSI processing time for CSI in step H500 is the same as the required CSI-based processing time in step S500 of the model monitoring phase.

[0207] It is worth noting that in practice, there is no restriction on the execution order of steps H100 to H500. That is to say, any number of the described steps can be skipped or combined in any order to implement the method or an alternative method.

[0208] Regarding the model inference phase, some aspects differ from the model monitoring phase, so this part will be described separately below:

[0209] The model inference stage does not require the computation of obtaining the Ground Truth, i.e., CSI Computation (GT).

[0210] Model inference includes at least one of the following steps:

[0211] 1. Model Transfer. When there is no available model on the UE side, or when the model needs to be updated, model transfer is required, which introduces additional Model Transfer time. It is worth noting that model transfer is optional.

[0212] 2. Loading the model.

[0213] Specifically, after receiving the model activation command from DCI / MAC-CE / RRC, the user needs to perform preparatory work for model activation, such as the time required to load all model parameters into memory. We collectively refer to the time for this series of model preparation tasks as the Model Activation time.

[0214] 3. Reasoning of the model.

[0215] Specifically, this refers to the time required for the user to perform model inference calculations and output results while completing model loading and user data collection (Model Inference).

[0216] The performance metric computation time does not need to be considered during the model inference phase.

[0217] It is worth noting that the preparation time for Tx is the same as the traditional method.

[0218] The following describes the process using the model training phase as an example. The method includes at least one of the following steps:

[0219] Step L100: The user equipment reports its first AI / ML capability information to the base station. Step L100 is identical in detail to step S100 in the model monitoring phase.

[0220] Step L200: Receive the first reporting configuration of Channel State Information (CSI), wherein the first reporting configuration is determined based on the first capability information.

[0221] Specifically, the first reporting configuration instructs the UE to perform CSI reporting. Since the data collection for model training involves traditional measurement and calculation results, it can be reported as non-AI / ML-based CSI. However, since data collection falls within the scope of AI / ML, it can also be reported as AI-specific CSI. (Optional) The base station may need to configure the reporting method.

[0222] Step L300: Receive a CSI request or a message to activate the model; wherein the message includes at least one of the following: Downlink Control Information (DCI), Media Access Control Element (MAC CE), and / or Radio Resource Control (RRC). In other words, the CSI transmission for AI / ML air interface model training data collection can be achieved using L1-signaling or higher-layer message transmission.

[0223] Step L400: Determine the CPU required for CSI reporting. Based on the total number of CPUs, the CPUs already occupied, and the CPUs required for CSI reporting, determine the CSIs to be reported. Step L400 is the same as step S400 in the model monitoring stage.

[0224] Step L500: Determine the required CSI-related processing time. The required CSI-related processing time in step L500 is the same as the required CSI-related processing time in step S500 of the model monitoring phase.

[0225] It is worth noting that in practice, there is no restriction on the execution order of steps L100 to L500. That is to say, any number of the described steps can be skipped or combined in any order to implement the method or an alternative method.

[0226] The model training phase does not involve model inference, so this part will be discussed separately below:

[0227] The model training phase does not require consideration of obtaining the ground truth, i.e., it does not require consideration of CSI Computation (GT).

[0228] The model training phase does not require the user experience (UE) to perform model inference, therefore there is no need to consider model transfer, model activation, or model inference.

[0229] The model training phase does not require calculating model performance metrics, nor does it require considering the computation time for performance metrics.

[0230] It is worth noting that the transmitter preparation time (Tx preparation) is the same as in the traditional method.

[0231] In summary, the unified CSI processing flowchart in Figure 3 covers the determination of all possible time-related parameters at different stages of LCM. By associating different LCM stages with different time parameters, this processing flow can be universally applied.

[0232] Figure 3-4 illustrates one of the flowcharts of the wireless communication method based on artificial intelligence or machine learning (AI / ML) provided in this disclosure. As shown in Figure 3-4, for model monitoring, the embodiments of this disclosure design two processing methods: serial and parallel. When the UE can use two sets of computing units based on non-AI and AI to process the model monitoring task in parallel, the processing latency can be effectively reduced.

[0233] In parallel processing, while performing model inference, the UE utilizes the idle air interface after model data collection to perform CSI measurements and uses a separate traditional computing unit to process CSI Computation (GT), as shown in Figure 4. At this point, the impact of CSI Computation (GT) does not need to be considered when configuring the reference time.

[0234] As can be seen from Figures 3 and 4, the difference between them lies in the model monitoring phase: during serial processing, model inference and CSI calculation are processed sequentially; during parallel processing, model inference and CSI calculation are processed simultaneously. Other steps in Figures 3 and 4 remain unchanged. Figures 3 and 4 use a common flow, with differences described in a differentiated manner. More detailed flow steps can be found in Figures 3 and 4. The common flow includes at least one of the following steps:

[0235] M100 and user equipment report information on AI / ML-related secondary capabilities to the access network equipment.

[0236] Specifically, no restrictions are placed on user equipment and access network equipment; in practice, other equipment can also be included. Support for AI / ML-related secondary capability information may include at least one of the following: serial / parallel processing capability information and / or processing time related to model monitoring.

[0237] In some embodiments of this disclosure, the serial / parallel processing capability information may include at least one of the following: supporting serial processing, supporting parallel processing based on non-AI / ML CSI determination and AI / ML CSI determination, supporting parallel processing based on non-AI / ML CSI determination, AI / ML CSI determination, and multiple AI / ML CSI determinations, and / or supporting serial processing based on non-AI / ML CSI determination and multiple AI / ML CSI determinations in parallel processing.

[0238] Specifically, UE capability reporting needs to be based on functionality, feature, feature group, and model, including N. cpu O cpu And / or reference time, etc. For specific LCM scenarios, such as model performance monitoring or multi-model parallel processing, it is also necessary to report parallel processing capabilities. Combined with traditional serial reporting, serial / parallel processing may occur. Serial / parallel processing capability information includes at least one of the following:

[0239] 1. Field 00 represents parallel processing capability; 1: only serial processing is supported.

[0240] 2. Field 01 represents parallel processing capability 2: Supports parallel processing of CSI determination based on non-AI / ML and CSI determination based on AI / ML;

[0241] 3. Field 02 represents parallel processing capability 3: supports CSI determination based on non-AI / ML and CSI determination based on AI / ML, as well as parallel processing of multiple AI / ML CSI determinations; or

[0242] 4. Field 03 represents parallel processing capability 4: Supports serial processing based on non-AI / ML CSI determination, and multiple parallel processing based on AI / ML CSI determination.

[0243] The AI / ML-based CSI processing time during model performance monitoring includes at least one of the following:

[0244] 1. Model performance monitoring based on non-AI / ML CSI measurement and computation processing time

[0245] In both the AI ​​and non-AI components, the device needs to perform at least one of the following measurements:

[0246] Measurement Behavior 1: Data collection to obtain model input for model inference.

[0247] The time it occupies only includes traditional Beamswitching.

[0248] Measurement Action 2: To verify the performance of the model inference results, the ground truth value of the corresponding model inference results is measured.

[0249] The time occupied includes beam switching and / or CSI Computation (GT).

[0250] 2. AI inference-related processing time for model performance monitoring

[0251] At the same time, the model needs to perform model inference, which includes at least one of the following steps:

[0252] 1) Loading the model.

[0253] Specifically, after receiving the model activation command issued by DCI / MAC-CE / RRC, the time required for the user to load all model parameters into memory is used for model loading.

[0254] 2) Reasoning of the model.

[0255] Specifically, model inference is the time required for the user to perform model inference calculations and output results after completing model loading and user data collection.

[0256] 3. Performance metrics calculation and processing time for model performance monitoring

[0257] After the device obtains the model inference results and Ground Truth, it needs to use the two to calculate the model performance metrics according to the preset calculation principles, in order to measure whether the model's output meets the requirements.

[0258] The calculation of performance metrics for model monitoring can be divided into two paths: monitoring using statistical values ​​and monitoring using instantaneous values. Both paths involve computation on large-scale data, therefore, the metric computation time also needs to be considered.

[0259] 4. Other processing time for model performance monitoring, including at least one of the following:

[0260] 1. For model performance monitoring, if the reporting trigger mechanism still uses DCI triggering, the DCI decoding time within the processing time is consistent with that of CSI based on non-AI / ML.

[0261] 2. Optionally, since model monitoring reports are not sensitive to latency requirements, model performance monitoring reports can also be triggered using MAC-CE. In this case, the processing time of MAC-CE, i.e., DCI+Data decoding, and the time for MAC messages to be parsed and transmitted at the upper layer need to be considered.

[0262] 3. Similarly, model performance monitoring reports can also use RRC configuration. In this case, the processing time of RRC, i.e., DCI+Data decoding, and the time for RRC messages to be parsed and transmitted at the upper layer need to be considered.

[0263] It is worth noting that the transmitter preparation (Tx preparation) time is the same as the traditional method.

[0264] In some embodiments of this disclosure, the computation time related to model monitoring includes the computation time for the actual target value and / or the computation time for the metric.

[0265] Specifically, for model performance monitoring, the computation time related to model monitoring includes the computation time for the ground truth and / or the computation time for metrics, which means that an additional amount of time needs to be introduced that varies depending on the model type or function type:

[0266] In the above formula, X AI,comp The value can vary depending on different use cases or functions, different LCM stages, and different serial / parallel processing methods.

[0267] In one implementation, when monitoring the performance of a model undergoing serial processing, since the calculation of metrics must be performed after obtaining the ground truth, X... AI,comp =X CSI,comp +X metr,comp .

[0268] In another implementation, when monitoring the performance of the parallel processing model, since obtaining the ground truth can be done simultaneously during model inference, there is no need to consider X separately. CSI,comp Therefore, X AI,comp =X metr,comp .

[0269] In another implementation, when network-side model monitoring is performed and serial processing is executed, and the UE side does not perform performance metric calculations, there is no need to consider X additionally. metr,comp Therefore, X AI,comp =X CSI,comp .

[0270] In another implementation, when network-side model monitoring is performed and parallel processing is executed, and the UE side does not perform performance metric calculations, there is no need to consider X additionally. metr,comp Furthermore, obtaining the ground truth can be done simultaneously during model inference without needing to consider X separately. CSI,comp Therefore, X AI,comp =0.

[0271] Additionally, in other situations, such as model inference or training scenarios, X AI,comp =0.

[0272] In some embodiments of this disclosure, the processing time related to model monitoring is associated with AI / ML functions, AI / ML features, AI / ML feature groups, and / or AI / ML models.

[0273] Specifically, X metr,comp This can be determined by looking up a table. For example, Table 15 shows the relationship between the processing time related to model monitoring and the AI / ML functionality and the model.

[0274] Table 15 shows an example of processing time related to model monitoring.

[0275] It should be noted that Table 15 is an example of a table indicating processing time related to model monitoring. This table may include more or fewer correspondences between model monitoring-related processing times and AI / ML functions and models. Furthermore, there are no restrictions on the correspondences between model monitoring-related processing times and AI / ML functions and models in Table 15. That is, the table indicating processing time related to model monitoring, as protected by this disclosure, can be obtained by adding, deleting, modifying, or altering the correspondences between model monitoring-related processing times and AI / ML functions and models in Table 15.

[0276] Table 16 shows the correlation between computation time related to model monitoring and AI / ML features and AI / ML feature groups.

[0277] Table 16 shows an example of processing time related to model monitoring.

[0278] It should be noted that Table 16 is an example of a table indicating the processing time related to model monitoring. This table may include more or fewer correspondences between model monitoring-related processing times and AI / ML features and feature groups. Furthermore, there are no restrictions on the correspondences between model monitoring-related processing times and AI / ML features and feature groups in Table 16. That is, adding, deleting, modifying, or altering the correspondences between model monitoring-related processing times and AI / ML features and feature groups based on Table 16 yields the table that constitutes the model monitoring-related processing time indication protected in this disclosure.

[0279] In some embodiments of this disclosure, the computation time associated with model monitoring is correlated with performance monitoring metrics.

[0280] Specifically, as shown in Table 17, Table 17 illustrates the correlation between computation time and performance metrics related to model monitoring.

[0281] Table 17 shows an example of processing time related to model monitoring.

[0282] It should be noted that Table 17 is an example of a table for determining processing time indicators related to model monitoring. Such a table may include more or fewer correspondences between processing times related to model monitoring and performance metrics. Furthermore, there are no restrictions on the correspondences between processing times related to model monitoring and performance metrics in Table 17. That is, the table for determining processing time indicators related to model monitoring, as protected by this disclosure, can be obtained by adding, deleting, modifying, or altering the correspondences between processing times related to model monitoring and performance metrics in Table 17.

[0283] In Table 17, model monitoring performance metrics can be categorized into those based on single measurements and those based on statistical measurements. Furthermore, different types can correspond to different performance monitoring metrics, as illustrated in the following examples:

[0284] Example 1: Type A corresponds to NMSE (Normalized Mean Square Error)

[0285] Example 2: Type B corresponds to SGCS (Squared generalized cosine similarity)

[0286] Example 3: Type C corresponds to the average NMSE of multiple measurements.

[0287] Example 4: Type D corresponds to the average SGCS of multiple measurements.

[0288] In event-based triggering mechanisms, different types can also correspond to different performance monitoring events, as shown in the following examples:

[0289] Example 1: Corresponding to Type A,

[0290] 1) Event 1: Within a certain time period, the number of NMSE values ​​less than a certain threshold is greater than N; or,

[0291] 2) Event 2: The NMSE of N consecutive measurements is less than a certain threshold;

[0292] Example 2: Corresponding to Type B,

[0293] 1) Event 1: Within a certain time period, the number of SGCS values ​​less than a certain threshold is greater than N; or,

[0294] 2) Event 2: The SGCS measured N times consecutively is less than a certain threshold;

[0295] Example 3: Type C corresponds to:

[0296] 1) Event 1: Within a certain time period, the number of instances where the average NMSE based on k statistical analyses is less than a certain threshold is greater than N; or,

[0297] 2) Event 2: The average NMSE based on k statistical measurements of N consecutive measurements is less than a certain threshold;

[0298] Example 4: Type D corresponds to:

[0299] 1) Event 1: Within a certain time period, the number of instances where the average SGCS based on k statistical analyses is less than a certain threshold is greater than N; or,

[0300] 2) Event 2: The average SGCS based on k statistical measurements of N consecutive measurements is less than a certain threshold;

[0301] In some embodiments of this disclosure, the computation time related to model monitoring can be determined by multiple parameters, which are associated with at least one of the following parameters: AI / ML functionality, use case, model, feature, feature group, performance metrics, and / or other related parameters. Other related parameters may be QCM-ID, associated ID.

[0302] In addition, the present disclosure also defines a reference time.

[0303] In UE capability reporting, the reference time needs to be defined based on functionality, feature, feature group, and model, and the reference time and calculation time must meet at least one of the following constraints:

[0304] 1. Reference time Z ref,AI ;

[0305] 1) The time offset between the completion of CSI processing and the receipt of the CSI Request;

[0306] 2)Z ref,AI =X load,AI +X inf,AI +X AI,comp ;

[0307] 2. Reference time Z' ref,AI ;

[0308] 1) The time offset from the completion of CSI processing to the receipt of the LastCSI-RS;

[0309] 2)Z' ref,AI =X inf,AI +X AI,comp ;

[0310] For specific steps M200, please refer to S200.

[0311] For specific steps M300, please refer to S300.

[0312] For specific steps M400, please refer to S400.

[0313] For specific steps M500, please refer to S500.

[0314] It is worth noting that at least one Channel State Information Reference Signal (CSI-RS) is two CSI-RS. It is important to note that the two CSI-RS mentioned here is just an example and is not intended as a limitation.

[0315] In one implementation, after the UE receives the first CSI-RS, the UE performs model inference. After the UE receives the second CSI-RS, the UE performs CSI calculation, that is, the UE performs serial processing, as shown in Figure 3.

[0316] In another implementation, when the UE receives two CSI-RS at different times, the UE performs model inference and CSI calculation simultaneously, that is, the UE performs parallel processing, as shown in Figure 4.

[0317] Specifically, in parallel processing, while the UE performs model inference, it utilizes the idle air interface after completing model data collection to perform CSI measurements and uses a separate traditional computing unit to process CSI Computation (GT), as shown in Figure 4. At this point, the impact of CSI Computation (GT) does not need to be considered when configuring the reference time.

[0318] It is worth noting that in practice, there is no restriction on the execution order of steps M100 to M600. That is to say, any number of the described steps can be skipped or combined in any order to implement the method or an alternative method.

[0319] Figure 5 illustrates one of the flowcharts of a wireless communication method based on artificial intelligence or machine learning (AI / ML) provided in this disclosure. As shown in Figure 5, the wireless communication method includes at least one of the following steps:

[0320] G100, Determine a first processing unit (e.g., IPU), wherein the first processing unit is used to calculate the computing resources available to the user equipment;

[0321] Specifically, in traditional designs, the CPU is primarily used to calculate the available reporting resources of the UE. However, in AI-driven 6G scenarios, a new processing unit needs to be defined to calculate the available computing resources of the UE. The first processing unit, the Intelligent Processing Unit (IPU), serves as the unit for measuring the AI-related computing resources of the UE and is used as the new processing unit in AI-driven 6G scenarios. The upper limit N of the UE's IPU is... IPU Determined by hardware capabilities, the computational resource consumption of different tasks can be defined as O based on the model, function, and / or feature group, etc. IPU The answer can be determined by looking up the table in Table 18.

[0322] Table 18 Example of IPU Definition

[0323] It should be noted that Table 18 is an example of an IPU definition instruction table, which may include more or fewer correspondences between IPUs and models, functions, and / or feature groups. Furthermore, there are no restrictions on the correspondences between IPUs and models, functions, and / or feature groups in Table 18. That is, the IPU definition instruction table protected in this disclosure can be obtained by adding, deleting, modifying, or altering the correspondences between IPUs and models, functions, and / or feature groups based on Table 18.

[0324] In some embodiments of this disclosure, the first processing unit determines based on at least one of the following parameters: AI / ML functionality, AI / ML use case, AI / ML model, AI / ML feature, AI / ML feature group, AI / ML performance metrics, and / or other related parameters (e.g., QCM-ID, associated ID).

[0325] G200, based on third capability information and / or predefined principles, determines the first processing unit;

[0326] In some embodiments of this disclosure, the first processing unit is determined based on third capability information; the third capability information includes at least one of the following: computing power, storage and transmission capability, and / or energy efficiency.

[0327] Specifically, the third capability information is shown in Tables 19-23. The computing capability includes at least one of the following: computing unit type, instruction set type, and / or computing speed metric. The computing unit type includes the number of cores in the central processing unit (CPU), the number of cores in the graphics processing unit (GPU), and / or the matrix operation units of the tensor processing unit (TPU), all of which affect computing capability. The instruction set type includes SIMD vectorized instruction sets (e.g., Intel AVX-512 VNNI, AMD VOPD), matrix extension instruction sets (e.g., Intel AMX), dedicated AI accelerator instruction sets (e.g., NVIDIA Tensor Core, Google TPU), memory-to-machine instruction sets (Samsung HBM-PIM), and / or sparse operation instruction sets (e.g., Tesla Dojo), etc. Under the same hardware, the choice of different instruction sets affects computing capability. The computing speed metric can be a general computing speed metric: TeraFLOPS. Storage and transmission capabilities can be categorized by storage cell type and capacity: for example, HBM, HBM2, GDDR6, GDDR6X, etc.; and whether on-chip cache optimization is included. Energy efficiency can be categorized by the general energy efficiency metric: TOPS / W.

[0328] Table 19 Example of IPU Definition - Based on Computing Capacity 1

[0329] It should be noted that Table 19 is an example of an IPU definition instruction table, which may include more or fewer IPUs and their corresponding computing capabilities. Furthermore, there are no restrictions on the correspondence between IPUs and computing capabilities in Table 19. That is, the IPU definition instruction table protected in this disclosure can be obtained by adding, deleting, modifying, or altering the correspondence between IPUs and computing capabilities based on Table 19.

[0330] Table 20 IPU Definition Example - Based on Computing Capacity 1

[0331] It should be noted that Table 20 is an example of an IPU definition instruction table, which may include more or fewer IPUs and computing power correspondences. Furthermore, there are no restrictions on the correspondences between IPUs and computing power in Table 20. That is, the IPU definition instruction table protected in this disclosure can be obtained by adding, deleting, modifying, or altering the correspondences between IPUs and computing power based on Table 20.

[0332] Table 21 IPU Definition Example - Based on Storage Capacity

[0333] It should be noted that Table 21 is an example of an IPU definition instruction table, which may include more or fewer IPUs and storage capacity correspondences. Furthermore, there are no restrictions on the correspondences between IPUs and storage capacity in Table 21. That is, the IPU definition instruction table protected in this disclosure can be obtained by adding, deleting, modifying, or altering the correspondences between IPUs and storage capacity based on Table 21.

[0334] Table 22 Example of IPU Definition - Based on Energy Efficiency Ratio

[0335] It should be noted that Table 22 is an example of an IPU definition indication table, and the IPU definition indication table may include more or fewer correspondences between IPUs and energy efficiency. Furthermore, there are no restrictions on the correspondences between IPUs and energy efficiency in Table 22. That is to say, the IPU definition indication table protected by this disclosure can be obtained by adding, deleting, modifying, or altering the correspondences between IPUs and energy efficiency based on Table 22.

[0336] Table 23 IPU Definition Example - General

[0337] It should be noted that Table 23 is an example of an IPU definition instruction table, and the IPU definition instruction table may include more or fewer correspondences between IPUs and synthesis. Furthermore, there are no restrictions on the correspondences between IPUs and synthesis in Table 23. That is, the IPU definition instruction table protected in this disclosure can be obtained by adding, deleting, modifying, or altering the correspondences between IPUs and synthesis based on Table 23.

[0338] G300, report mapping information, third capability information, and / or fourth capability information;

[0339] Specifically, when a UE reports mapping information to a base station, it needs to involve the configuration information of computing resources.

[0340] In some embodiments of this disclosure, the configuration information of the computing resources includes at least one of the following: the proportion of total computing resources, the proportion of the number of cores or clusters of GPU pipelined processing units, the proportion of the number of threads, and / or the proportion of total storage resources.

[0341] In one implementation, the configuration information of the computing resources includes at least one of the following:

[0342] 1. Overall computing resource allocation percentage;

[0343] 1) Assume the number of cores is N, and the computing resources of a single core i are C. i pk If the proportion of computing resources allocated to model k is determined, then the maximum computing resources for model k are...

[0344] 2) At this time, N IPU =NC i ,

[0345] 3) Or more generally, N IPU Related to the total computing resources of the hardware, O IPU The percentage of computing resources required to associate with the model.

[0346] 2. Percentage of cores / clusters of GPU pipelined units;

[0347] 1) Assume that model k uses N cores. k Assuming each core has the same capabilities, the maximum computational resource for model k is...

[0348] 2) At this time, N IPU =NC i ,

[0349] 3) Or more generally, N IPU Related to the total number of cores in the hardware, O IPU The number of cores required to associate with the model.

[0350] 3. Thread count percentage;

[0351] 1) Assume that the maximum number of threads supported by a single core i is M. i The number of cores used in model k is N. k The number of threads used by model k in each core is M. k Then the maximum computational resource for model k is

[0352] 2) At this time, N IPU =NC i ,

[0353] 3) Or more generally, N IPU Related to the total number of threads in the hardware, O IPU The number of threads required to associate with the model.

[0354] 4. Overall storage resource ratio;

[0355] 1) Assume the number (capacity) of storage units is N, p k If the storage resource allocation ratio for model k is given, then the maximum storage resource for model k is p. k N.

[0356] 2) At this time, N IPU =N, Need to meet

[0357] Optional, N IPU Related to the total number of storage units in the hardware, O IPU The number of storage units required to associate with the model.

[0358] The fourth capability information includes at least one of the following: not supporting AI / ML models, supporting single AI / ML models, and / or supporting parallel processing of multiple AI / ML models.

[0359] Specifically, in UE capability reporting, such as in the RRC message UECapabilityInformation, the fourth capability information (i.e., multi-model parallel processing capability) is reported based on the UE hardware capabilities.

[0360] The fourth capability information includes at least one of the following:

[0361] 1. Field 00 represents parallel processing capability 1: AI / ML models are not supported;

[0362] 2. Field 01 represents parallel processing capability 1: supports single AI / ML model;

[0363] 3. Field 02 represents parallel processing capability 2: Supports parallel processing of multiple AI / ML models.

[0364] Optionally, when supporting multi-model parallelism, capability reporting needs to indicate N. IPU .

[0365] G400, send configuration information.

[0366] Specifically, the base station configures users to simultaneously perform a certain number and / or types of AI inference tasks; analogous to a non-AI / ML-based UE reporting its CPU, the base station then distributes the CSI report configuration. In reality, the base station understands the CPU to ensure more reasonable CSI report configuration and avoids unnecessary configuration beyond the UE's capabilities.

[0367] G500, based on the first processing unit, the user equipment determines the tasks to be processed in parallel.

[0368] Specifically, when multiple models process tasks concurrently, the UE determines the tasks to be processed in parallel based on the number of IPUs. Assume the total number of IPUs in the UE is N. IPU The number of parallel model computation tasks it receives is N, and the computing resources it has already occupied or reserved are L. At this point, the maximum number of parallel tasks it can achieve, M, satisfies:

[0369] When N<M, the UE can use an AI model to process all tasks without requiring special operations.

[0370] When N>M, the UE can use an AI model to process part of the tasks, and the following methods are available for processing the remaining parallelly processed tasks.

[0371] In some embodiments of the present disclosure, the parallelly processed tasks are determined based on the behavior rules of user equipment. The behavior rules include: determining the number of parallelly executed AI models according to IPUs and / or processing methods for AI models that cannot be executed immediately. Where the behavior rule is a processing method for an AI model that cannot be executed immediately, the processing method for an AI model that cannot be executed immediately includes at least one of the following: model fallback processing method, model deferral processing method, indication method for model fallback or deferral, and determining a reporting time after completing fallback and deferral processing.

[0372] Specifically, the manner in which the remaining parallelly processed tasks are determined based on the behavior rules of user equipment includes at least one of the following:

[0373] Method 1: When the delay requirement is strict, the number of AI tasks processed by the UE using the AI model is M, and the remaining tasks fall back to be processed by a non-AI / ML-based CSI algorithm. In this case, the UE needs to provide an indication in the CSI report. The processing method includes at least one of the following:

[0374] 1. A 1~2-bit task processing identifier, which identifies that the result in the CSI report is processed based on a non-AI / ML CSI algorithm.

[0375] 2. Whether N-M tasks fall back to a non-AI / ML-based algorithm can be determined by the following method:

[0376] 1) Determined according to the allowFallback field in the CSI-ReportConfig corresponding to the task.

[0377] Wherein, 0 indicates that fallback is not allowed, and 1 indicates that fallback is allowed.

[0378] 2) According to the O of the task IPU the size determines (in ascending or descending order).

[0379] Specifically, the determination is performed according to the principle of maximizing the utilization of AI computing resources, that is, for any two task sets each containing M tasks and

[0380] 3. The UE indicates tasks processed in parallel by AI, or tasks that fall back to non-AI / ML-based algorithms.

[0381] 1) The task number or task name can be used as an indicator in the task configuration (e.g., in the relevant CSI-ReportConfig).

[0382] 2) The task instructions can be followed using a bitmap, where 0 indicates a rollback task and 1 indicates a task processed by AI. For example, if the CSI-ResourceConfigIds corresponding to the task configurations are 12345 in chronological order, and the returned bitmap is 10101, then the tasks with CSI-ResourceConfigIds 2 and 4 have been rolled back to traditional algorithm processing.

[0383] Method 2: When latency requirements are not strict, the UE first uses the AI ​​model to process M AI tasks in parallel, and the remaining NM are delayed.

[0384] No special handling is required when the delayed parallel processing task is no later than the configured CSI report time.

[0385] When a delayed parallel processing task is later than the CSI report time, it can be handled using at least one of the following methods:

[0386] 1. NW is configured with a set of optional times, and the UE sends the CSI report at the time most recently after the task is completed.

[0387] 2. NW is configured with a set of CSI report time offsets. The UE is scheduled to send the CSI report at the original time plus the offset time.

[0388] 3. The UE sends an uplink short format UCI, MAC-CE, or RRC (associated with delayed CSI processing) to trigger the NW reconfigure the corresponding CSI report time.

[0389] 4. The UE sends a CSI report according to the configuration, using specific fields, such as reporting quantity as none, to indicate that the task is delayed and trigger NW reconfiguration.

[0390] Whether NM tasks are delayed can be determined by at least one of the following methods:

[0391] 1) According to the priority of the task.

[0392] This priority can be calculated or indicated by the functions / models / FGs associated with the task, according to the rules predefined by the base station.

[0393] 2) According to the O of this task IPU Size determines (ascending or descending order).

[0394] 3) Determined according to the principle of maximizing the utilization of AI computing resources. That is, for any two task sets containing M tasks... and

[0395] 6. The UE indicates a task with delayed parallel processing, including at least one of the following:

[0396] 1) The task number or task name can be used as an indicator in the task configuration (e.g., in the relevant CSI-ReportConfig).

[0397] 2) The task can be indicated by Bitmap in the order of task instructions, where 0 indicates a delayed task and 1 indicates a task to be processed by AI.

[0398] The choice between Method 1 and Method 2 can be indicated by the NW in the CSI reporting or related configuration of CSI resources, such as CSI-ReportConfig or CSI-ResourceConfig, or it can be determined by the UE and informed by the NW (or access network device) in UCI or RRC.

[0399] It is worth noting that in practice, there is no restriction on the execution order of steps G100 to G500. That is to say, any number of the described steps can be skipped or combined in any order to implement the method or an alternative method.

[0400] Through the above embodiments, this disclosure achieves at least one of the following technical effects through the above methods: The embodiments of this disclosure provide AI / ML-based CPU working mechanisms in different scenarios to support AI / ML-based CSI-related use cases working on independent dedicated CPU resources. Traditional CPU working mechanisms and AI / ML-based CPU working mechanisms can run in parallel, improving computational efficiency and system performance, and reducing processing latency. The unified CSI processing flow design of the embodiments of this disclosure is suitable for obtaining the computation time corresponding to AI / ML CSI reports. In addition, through simple parameter configuration, it can be applied to different LCM stages. The IPU mechanism of the embodiments of this disclosure can support the allocation of computing resources in parallel models for multi-model parallel scenarios.

[0401] This document describes a wireless communication method based on artificial intelligence or machine learning (AI / ML), applicable to communication such as between a user equipment (UE) and a base station. However, these inventive concepts, methods, apparatuses, devices, computer-readable storage media, chips, and computer program products are not limited to AI / ML-based wireless communication and can be extended to other communication scenarios to achieve the same technical benefits and effects.

[0402] In these scalable communication scenarios, entities such as base stations (e.g., gNB, eNodeB, Transmitter Receiving Point (TRP), NodeB for next-generation communication, or Wi-Fi access points) or network elements are involved. User equipment (UE) refers to devices used for communication at the user end, such as mobile phones; it can also be called a terminal, mobile station, or mobile terminal. UEs can be various devices, including but not limited to mobile phones, tablets, virtual reality (VR) devices, augmented reality (AR) devices, wireless terminals for industrial control, wireless terminals for autonomous driving, wireless terminals for remote medical surgery, wireless terminals for smart grids, wireless terminals for environmental monitoring, wireless terminals for smart cities, and wireless terminals for smart homes, etc.

[0403] Furthermore, UEs and base stations can be deployed in various environments, including but not limited to indoor, outdoor, handheld devices, vehicle-mounted devices, or even on water, in the air, on airplanes, drones, or on satellites.

[0404] Therefore, although this paper describes wireless communication methods and devices based on artificial intelligence or machine learning (AI / ML), the inventive concepts and techniques contained herein can be extended to other communication scenarios and are expected to achieve the same technical benefits and effects. It is readily apparent that these inventive concepts have broad applicability and scalability, whether for communication between different types of base stations and user equipment, or for communication in different deployment environments.

[0405] It should be noted that the above steps are merely examples and do not limit the scope of the invention. Various modifications and variations can be made to the steps without departing from the spirit and scope of the invention.

[0406] The order of the described steps (signaling / boxes) is not intended to be construed as a limitation, and any number of the described steps (signaling / boxes) can be skipped or combined in any order to implement the method or an alternative method.

[0407] This disclosure describes examples of communication between terminals and network element components in the network architecture described in the above embodiments, which are primarily for illustrative purposes and not for limitation.

[0408] The order of the described steps (signaling / blocks) is not intended to be construed as limiting, and any number of the described steps (signaling / blocks) can be skipped or combined in any order to implement the method or alternative methods. Generally, any of the components, modules, methods, and operations described herein can be implemented using software, firmware, hardware (e.g., fixed logic circuitry), manual processing, or any combination thereof. Some operations of the example methods can be described in the general context of executable instructions stored on computer-readable storage located locally and / or remotely on a computer processing system, and implementations can include software applications, programs, functions, etc. Alternatively or additionally, any functionality described herein can be performed at least in part by one or more hardware logic components, such as, but not limited to, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), etc.

[0409] Furthermore, the signaling transmission described in the embodiments of this disclosure can be implemented in any manner known in the art. For example, signaling transmission can be explicit and / or implicit. Moreover, the illustrated steps (signaling / blocks) are for illustrative purposes only and are not intended to limit this application.

[0410] Figure 6 is a schematic structural diagram of a wireless communication device 900 provided in this disclosure. The wireless communication device includes a processor and a memory, the memory for storing computer programs, and the processor for calling and running the computer programs stored in the memory, executing instructions for at least one of the operations described above.

[0411] The wireless communication device can be a user equipment, a base station, or a network element. The wireless communication device 900 shown in Figure 6 includes a processor 910, which can call and run computer programs from memory to implement the methods in the embodiments of this application.

[0412] Optionally, as shown in FIG6, the wireless communication device 900 may further include a memory 920. The processor 910 can retrieve and run computer programs from the memory 920 to implement the methods in the embodiments of this application. The memory 920 may be a separate device independent of the processor 910, or it may be integrated into the processor 910.

[0413] Optionally, as shown in Figure 6, the wireless communication device 900 may further include a transceiver 930. The processor 910 can control the transceiver 930 to communicate with other devices. Specifically, it can send information or data to other devices or receive information or data sent by other devices. The transceiver 930 may include a transmitter and a receiver. The transceiver 930 may further include an antenna, and the number of antennas may be one or more.

[0414] Optionally, the wireless communication device 900 may specifically be a base station in the embodiments of this application, and the wireless communication device 900 may implement the corresponding processes implemented by the base station in the various methods of the embodiments of this application. For the sake of brevity, it will not be described in detail here.

[0415] Optionally, the wireless communication device 900 may specifically be a mobile user equipment / user equipment in the embodiments of this application, and the wireless communication device 900 may implement the corresponding processes implemented by the mobile user equipment / user equipment in the various methods of the embodiments of this application. For the sake of brevity, it will not be described in detail here.

[0416] Optionally, the wireless communication device 900 may specifically be a network element in the embodiments of this application, and the wireless communication device 900 may implement the corresponding processes implemented by the network element in the various methods of the embodiments of this application. For the sake of brevity, it will not be described in detail here.

[0417] According to an example embodiment, a chip is provided, the chip including: a processor for calling and running a computer program from a memory, causing a device on which the chip is installed to perform the method according to any one of the above embodiments, examples, or example embodiments.

[0418] According to an example embodiment, a computer-readable storage medium is provided for storing a computer program that causes a computer to perform a method according to any one of the above embodiments, examples, or example embodiments.

[0419] According to an example embodiment, a computer program product is provided, including a computer program / instructions that, when executed by a processor (e.g., by the processor or an apparatus, device, computer, or machine including the processor), implement the method according to any one of the above embodiments, examples, or example embodiments.

[0420] Embodiments of this disclosure are combinations of technologies / processes that can be employed in 3GPP specifications to create a final product.

[0421] While this disclosure has been described in conjunction with what are considered to be the most practical and preferred embodiments, it should be understood that this disclosure is not limited to the disclosed embodiments, but is intended to cover various arrangements made without departing from the broadest interpretation of the appended claims.

Claims

1. A wireless communication method based on artificial intelligence or machine learning (AI / ML), executed on a user equipment, the method comprising: Report information on the first capabilities that support AI / ML.

2. The method according to claim 1, wherein, The first capability information includes at least one of the following: Whether it supports CSI processing based on AI / ML models, the supported AI / ML capabilities, whether it supports serial / parallel processing of CSI determination based on non-AI / ML and CSI determination based on AI / ML, whether it supports parallel operation of multiple model tasks, total CPU usage, or currently occupied CPU.

3. The method according to claim 1, wherein, The method further includes: receiving a first reporting configuration of Channel State Information (CSI), wherein the first reporting configuration is determined based on the first capability information.

4. The method of claim 3, wherein, The first reporting configuration instructs the user equipment to report based on at least one of the following: CSI based on non-AI / ML and / or CSI based on AI / ML.

5. A wireless communication method based on artificial intelligence or machine learning (AI / ML), executed on a user equipment, the method comprising: Receive a CSI request or activation model message; wherein the message includes at least one of the following: Downlink Control Information (DCI), Media Access Control Element (MAC CE), and / or Radio Resource Control (RRC).

6. The method of claim 5, wherein, The method further includes: Determine the AI / ML-based CSI processing time, and report the CSI based on the AI / ML-based CSI processing time.

7. The method according to claim 6, wherein, The AI / ML-based CSI processing time includes determining the processing time for the CSI request. When the message is transmitted through the MAC CE, the processing time includes at least one of the following: MAC CE processing time, DCI decoding time, and / or data decoding time.

8. The method according to claim 6, wherein, The AI / ML-based CSI processing time includes determining the processing time for the CSI request. When the message is transmitted through the RRC, the processing time includes at least one of the following: the time for processing the RRC, the DCI decoding time, and / or the data decoding time.

9. The method according to claim 6, wherein, For the model inference, the AI / ML-based CSI processing time includes model inference-related processing time, which is determined based on at least one of the following parameters: model migration time, model activation time, and / or model inference time.

10. The method according to claim 6, wherein, The AI / ML-based CSI processing time is determined based on a first value, and the AI / ML-based CSI processing time includes at least one of the following: an offset of the non-AI / ML-based CSI processing time and / or the product of the non-AI / ML-based CSI processing time and an offset factor.

11. The method according to claim 6, wherein, The AI / ML-based CSI processing time is a predefined value.

12. The method according to claim 6, wherein, The method further includes: Obtain the model's performance metrics; When the model is monitored by a user device, the AI / ML-based CSI processing time includes the time for calculating performance metrics.

13. A wireless communication method based on artificial intelligence or machine learning (AI / ML), executed on a user equipment, wherein, The method includes: Determine the CPU required for Channel State Information (CSI) reporting. Based on the total number of CPUs, the CPUs already occupied, and the CPUs required for CSI reporting, determine the CSIs that need to be reported. Report the required CSI information.

14. The method according to claim 13, wherein, The total number of CPUs reported based on non-AI / ML CSI and the total number of CPUs reported based on AI / ML are determined by at least one of the following methods: the total number of CPUs reported based on non-AI / ML CSI and the total number of CPUs reported based on AI / ML each account for a certain proportion of the total number of CPUs supported by the UE, and / or the proportion of each total number of CPUs related to AI / ML CSI reporting in the total number of CPUs supported by the user equipment is determined according to the Lifecycle Management (LCM) link and / or AI attributes.

15. The method according to claim 13, wherein, The CPU used by AI / ML-based CSI reporting is determined using at least one of the following methods: the offset relative to the CPU used by non-AI / ML-based CSI reporting, and / or the CPU associated with the LCM stage and / or AI attributes.

16. The method according to claim 13, wherein, The total CPU count corresponding to the CSI reports of each function is determined in at least one of the following ways: offset relative to the total CPU count of non-AI / ML based CSI reports, and / or the total CPU count associated with AI attributes and / or UE performance.

17. The method according to claim 13, wherein, The CPU usage of AI / ML-based CSI reporting is determined in at least one of the following ways: offset relative to CPU usage of non-AI / ML-based CSI reporting, offset relative to CPU associated with AI attributes, and / or CPU associated with LCM stages and / or model complexity.

18. A wireless communication method based on artificial intelligence or machine learning (AI / ML), executed on a user equipment, wherein, The method includes: Report information supporting AI / ML-related secondary capabilities.

19. The method according to claim 18, wherein, The second capability information includes at least one of the following: serial / parallel processing capability information and / or processing time related to model monitoring.

20. The method according to claim 19, wherein, The serial / parallel processing capability information includes at least one of the following: supporting serial processing, supporting parallel processing based on non-AI / ML CSI determination and AI / ML CSI determination, supporting parallel processing based on non-AI / ML CSI determination, AI / ML CSI determination, and multiple AI / ML CSI determinations, and / or supporting serial processing based on non-AI / ML CSI determination and multiple AI / ML CSI determinations in parallel processing.

21. The method according to claim 19, wherein, The processing time related to model monitoring includes the calculation time for the actual target value and / or the calculation time for the indicator.

22. The method according to claim 21, wherein, The processing time related to model monitoring is associated with the functions of AI / ML, the features of AI / ML, the feature groups of AI / ML, and / or the AI / ML model.

23. The method according to claim 21, wherein, The processing time related to the model monitoring is correlated with the performance monitoring indicators.

24. The method according to claim 19, wherein, The processing time related to model monitoring is associated with at least one of the following parameters: AI / ML function, use case, model, feature, feature group, performance monitoring metric, and / or other relevant parameters.

25. A wireless communication method based on artificial intelligence or machine learning (AI / ML), executed on a user equipment, wherein, The method includes: A first processing unit is determined, wherein the first processing unit is used to calculate the computing resources available to the user equipment; Based on the first processing unit, the user equipment determines the tasks to be processed in parallel.

26. The method of claim 25, wherein, The first processing unit determines based on at least one of the following parameters: AI / ML function, AI / ML use case, AI / ML model, AI / ML feature, AI / ML feature group, AI / ML performance monitoring metric, and / or other relevant parameters.

27. The method according to claim 25, wherein, The method further includes: The first processing unit is determined based on third capability information and / or predefined principles; Report mapping information and / or third-party capability information.

28. The method according to claim 27, wherein, Based on third capability information and / or predefined principles, the first processing unit is determined to include the first processing unit based on third capability information; the third capability information includes at least one of the following: computing power, storage and transmission capability, and / or energy efficiency.

29. The method according to claim 28, wherein, The computing capability includes at least one of the following: computing unit type, instruction set type, and / or computing speed index.

30. The method according to claim 28, wherein, The storage includes storage unit type and / or capacity.

31. The method according to claim 28, wherein, The energy efficiency mentioned includes energy efficiency indicators.

32. The method according to claim 29, wherein, The types of processing units include the number of cores in the central processing unit (CPU), the number of cores in the graphics processing unit (GPU), and the matrix operation units in the tensor processing unit (TPU).

33. The method according to claim 25, wherein, The configuration information of the computing resources includes at least one of the following: The percentage of total computing resources, the percentage of cores or clusters of GPU pipelined processing units, the percentage of threads, and / or the percentage of total storage resources.

34. The method according to claim 25, wherein, The tasks for parallel processing are determined based on the behavioral criteria of the user equipment.

35. The method according to claim 25, wherein, The code of conduct includes: determining the number of AI models to be executed in parallel and / or how to handle AI models that cannot be executed immediately based on IPU.

36. The method according to claim 25, wherein, The behavioral guidelines are the handling methods for AI models that cannot be executed immediately. The handling methods for AI models that cannot be executed immediately include at least one of the following: model rollback handling method, model postponement handling method, model rollback or delay indication method, and reporting time after determining rollback and postponement handling.

37. The method according to claim 25, wherein, The method further includes reporting fourth capability information, wherein the fourth capability information includes at least one of the following: It does not support AI / ML models, supports single AI / ML models and / or supports parallel processing of multiple AI / ML models.

38. A wireless communication method based on artificial intelligence or machine learning (AI / ML), executed in an access network device, the method comprising: Receive first capability information supporting AI / ML; A first reporting configuration for transmitting Channel State Information (CSI), wherein the first reporting configuration is determined based on the first capability information.

39. The method according to claim 38, wherein, The first capability information includes at least one of the following: Whether it supports CSI processing based on AI / ML models, the supported AI / ML capabilities, whether it supports serial / parallel processing of CSI determination based on non-AI / ML and CSI determination based on AI / ML, whether it supports parallel operation of multiple model tasks, total CPU usage, and currently occupied CPU.

40. The method of claim 38, wherein, The method further includes: sending a first reporting configuration of Channel State Information (CSI), wherein the first reporting configuration is determined based on the first capability information.

41. The method according to claim 40, wherein, The first reporting configuration instructs the user equipment to report based on at least one of the following: CSI based on non-AI / ML and / or CSI based on AI / ML.

42. A wireless communication method based on artificial intelligence or machine learning (AI / ML), executed in an access network device, the method comprising: Send a message requesting a CSI or activating a model; wherein the message includes at least one of the following: Downlink Control Information (DCI), Media Access Control Element (MAC CE), and / or Radio Resource Control (RRC).

43. The method according to claim 42, wherein, The method further includes: Receive CSI.

44. A wireless communication method based on artificial intelligence or machine learning (AI / ML), wherein, Executed on an access network device, the method includes: Receive information on secondary capabilities related to AI / ML.

45. The method according to claim 44, wherein, The second capability information includes at least one of the following: serial / parallel processing capability information and / or processing time related to model monitoring.

46. ​​The method according to claim 45, wherein, The serial / parallel processing capability information includes at least one of the following: supporting serial processing, supporting parallel processing based on non-AI / ML CSI determination and AI / ML CSI determination, supporting parallel processing based on non-AI / ML CSI determination, AI / ML CSI determination, and multiple AI / ML CSI determinations, and / or supporting serial processing based on non-AI / ML CSI determination and multiple AI / ML CSI determinations in parallel processing.

47. The method according to claim 45, wherein, The processing time related to model monitoring includes the calculation time for the actual target value and / or the calculation time for the indicator.

48. The method according to claim 47, wherein, The processing time related to model monitoring is associated with the functions of AI / ML, the characteristics of AI / ML, and / or the AI / ML model.

49. The method according to claim 47, wherein, The correlation between the processing time and performance monitoring metrics related to the model monitoring is described.

50. The method of claim 45, wherein, The processing time related to model monitoring is associated with at least one of the following parameters: AI / ML function, use case, model, feature, feature group, performance monitoring metric, and / or other relevant parameters.

51. A wireless communication method based on artificial intelligence or machine learning (AI / ML), executed in an access network device, wherein, The method includes: Receive mapping information, third capability information, and / or fourth capability information.

52. The method according to claim 51, wherein, Receive instructions on model rollback or delay.

53. A wireless communication device, wherein, The wireless communication device includes a processor and a memory for storing computer programs, the processor for calling and running the computer programs stored in the memory to perform the method as described in any one of claims 1 to 52.

54. A readable storage medium for storing a computer program that is invoked and executed by a processor to perform the method as described in any one of 1-52.

Images