Communication device and method for determining post-processing based on artificial intelligence / machine learning

Abstract

The invention discloses a communication device and method for determining post-processing based on AI / ML. A method performed by a communication device includes providing a communication device and method for determining post-processing based on AI / ML. The method comprises: the communication device determining post-processing of an AI / ML model according to at least one of the following: when the communication device is a UE, the AI / ML model is downloaded to the UE without post-processing, and the post-processing is performed on the UE, or the post-processing does not need to be performed before deployment, and when the communication device is the UE, the AI / ML model is downloaded to the UE without post-processing, and the post-processing does not need to be performed before deployment. The AI / ML model is downloaded to the UE after being subjected to the post-processing, and the post-processing is executed on a base station, or the UE is considered to be unable to execute the post-processing or not allowed to execute the post-processing. The post-processing-based autoencoder generates a CSI report and proposes one or more priority rules for determining a priority report level.

Description

Technical Field

[0001] The present application relates to the field of wireless communication systems, and more specifically to communication devices and methods for determining post-processing based on artificial intelligence (AI) / machine learning (ML). For example, the present application is related to the 18th version (Release 18) AI / ML new study item description (study item description, SID) on the new radio (NR) air interface established at the 94e plenary meeting of the 3rd generation partnership project (3GPP) radio access network (RAN) held in December 2022. Specifically, the present application relates to CSI discarding rules when multiple CSI reports overlap. CSI includes AI / ML model outputs generated by AI / ML model compression. Background Art

[0002] AI / ML has been applied to 3GPP RAN1 and has decided to study multiple use cases. They are CSI feedback enhancement, beam management and positioning. As pointed out in the 3GPP SID, although specific AI / ML algorithms and models may be studied for evaluation, AI / ML algorithms and models are implementation-specific and are not expected to be standardized.

[0003] Therefore, a communication device and method for determining post-processing based on artificial intelligence (AI) / machine learning (ML) is needed to solve the problems in the prior art, provide a method for allowing UE or gNB to perform post-processing, define a compression ratio of an automatic encoder, define a discarding rule for a CSI report, reduce system overhead, and provide good communication performance and / or high reliability. Summary of the invention

[0004] The purpose of this application is to propose a communication device and method for determining post-processing based on artificial intelligence (AI) / machine learning (ML) to solve the problems in the prior art, provide a method for allowing UE or gNB to perform post-processing, define a compression ratio of an automatic encoder, define a discarding rule for a CSI report, reduce system overhead, and provide good communication performance and / or high reliability.

[0005] In a first aspect of the present application, a method for determining post-processing based on artificial intelligence (AI) / machine learning (ML) performed by a communication device includes: the communication device determines the post-processing of the AI / ML model according to at least one of the following: when the communication device is a user equipment (UE), the AI / ML model is downloaded to the UE without post-processing, and the post-processing is performed on the UE, or the post-processing is not required before deployment; when the communication device is the UE, the AI / ML model is downloaded to the UE after the post-processing, and the post-processing is performed on a base station, or the UE is considered unable to perform the post-processing or is not allowed to perform the post-processing; when the communication device is the base station, the AI / ML model is uploaded to the base station without the post-processing, and the post-processing is performed on the base station, or no post-processing is required before deployment; when the communication device is the base station, the AI / ML model is uploaded to the base station after the post-processing, and the post-processing is performed on the base station, or the post-processing is not required before deployment; when the communication device is the base station, the AI / ML model is uploaded to the base station after the post-processing, and the post-processing is performed on the UE.

[0006] In a second aspect of the present application, a user equipment (UE) includes a memory, a transceiver, and a processor coupled to the memory and the transceiver. The processor is configured to execute the above method.

[0007] In a third aspect of the present application, a base station includes a memory, a transceiver, and a processor coupled to the memory and the transceiver. The processor is configured to execute the above method.

[0008] In a fourth aspect of the present application, a non-volatile machine-readable storage medium stores instructions which, when executed by a computer, cause the computer to execute the above method.

[0009] In a fifth aspect of the present application, a chip includes a processor, which is configured to call and run a computer program stored in a memory so that a device on which the chip is installed executes the above method.

[0010] In a sixth aspect of the present application, a computer-readable storage medium stores a computer program for causing a computer to execute the above method.

[0011] In a seventh aspect of the present application, a computer program product comprises a computer program, which enables a computer to execute the above method.

[0012] In an eighth aspect of the present application, a computer program enables a computer to execute the above method. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the embodiments of the present application or related technologies, the following briefly introduces the drawings to be described in the embodiments. Obviously, the drawings are only some embodiments of the present application, and ordinary technicians in this field can obtain other drawings based on these drawings without paying any price.

[0014] Figure 1A is a schematic diagram showing an example of the preprocessing, model training and post-processing processes according to an embodiment of the present application.

[0015] Figure 1B is a schematic diagram showing an example of a basic model of an automatic encoder according to an embodiment of the present application.

[0016] Figure 2 It is a block diagram showing communication between one or more user equipment (UE) and a base station (such as a gNB) in a communication network system according to an embodiment of the present application.

[0017] Figure 3 It is a flowchart showing wireless communication based on artificial intelligence (AI) / machine learning (ML) determination post-processing performed by a UE according to an embodiment of the present application.

[0018] Figure 4 It is a flowchart showing wireless communication based on artificial intelligence (AI) / machine learning (ML) determined post-processing performed by a first base station according to an embodiment of the present application.

[0019] Figure 5 2 is a schematic diagram showing an example in which post-processing according to an embodiment of the present application may be related to various factors.

[0020] Figure 6 It is a schematic diagram showing an example of post-processing adaptation of UE power level when the UE is capable of performing post-processing according to an embodiment of the present application.

[0021] Figure 7 It is a schematic diagram showing an example of post-processing adaptation of UE power level in a case where the gNB side is capable of performing post-processing according to an embodiment of the present application.

[0022] Figure 8 is a schematic diagram showing a quantization example for enhancing CSI feedback according to an embodiment of the present application.

[0023] Fig. 9is a schematic diagram illustrating an example of quantization performed to reconstruct a complex number according to an embodiment of the present application.

[0024] Fig.10 It is a schematic diagram showing an example of determining / defining a compression ratio for enhanced CSI feedback according to an embodiment of the present application.

[0025] Fig.11 is a block diagram of a system for wireless communication according to an embodiment of the present application. DETAILED DESCRIPTION

[0026] The technical content, structural features, objectives and effects of the present application are described in detail below in conjunction with the accompanying drawings. Specifically, the terms in the embodiments of the present application are only used to describe the purpose of specific embodiments, and are not intended to limit the present application.

[0027] The terms are used in the discussion of channel state information (CSI) reporting based on artificial intelligence (AI) / machine learning (ML). The description of the terms is shown in Table 1.

[0028] Table 1:

[0029]

[0030]

[0031] AI / ML has been applied to 3GPP RAN1, and it has been decided to study multiple use cases, namely CSI feedback enhancement, beam management, and positioning. In the prior art, some discussions are as follows: Space-frequency domain CSI compression using a dual-end AI model was selected as a representative sub-use case, but the possibility of studying other sub-use cases is not excluded. All pre-processing / post-processing, quantization / dequantization are included in the scope of this sub-use case. The conclusions of the discussion include: After the discussion of the evaluation method, the use of a dual-end model for time-space-frequency domain CSI compression is further explored as a possible sub-use case of CSI feedback enhancement. After the discussion of the evaluation method, the use of a single-end model to improve CSI accuracy based on traditional codebook design is further explored as a possible sub-use case of CSI feedback enhancement. After the discussion of the evaluation method, the use of a single-end model for CSI prediction is further explored as a possible sub-use case of CSI feedback enhancement. After the discussion of the evaluation method, CSI-RS configuration and overhead reduction are further explored as possible sub-use cases of CSI feedback enhancement. After the discussion of the evaluation method, resource allocation and scheduling are further explored as possible sub-use cases of CSI feedback enhancement. After the discussion of the evaluation methodology, CSI prediction combined with compression is further explored as a possible sub-use case for CSI feedback enhancement.

[0032] Figure 1A It is a schematic diagram illustrating an example of preprocessing, model training and post-processing process according to an embodiment of the present application. Figure 1A It is shown that, in some embodiments, for a machine learning model, there are usually three steps before deployment, namely preprocessing, model training, and post-processing. Preprocessing includes feature extraction and normalization to facilitate subsequent model training. Model training is used to adjust the parameters in the black box, which are initially randomized. Subsequently, post-processing processes the model through parameter quantization and neuron pruning to find a balance between complexity and performance. In the present application, some embodiments discuss various factors related to post-processing, such as UE capabilities, UE power, and processing delay requirements.

[0033] Figure 1B It is a schematic diagram illustrating an example of a basic model of an autoencoder according to an embodiment of the present application. Figure 1B It is shown that some embodiments study an autoencoder model for enhancing CSI feedback. An example of an autoencoder model is as follows: The input can be the original CSI (original CSI-RS value) or the (maximum) eigenvector. At the gNB end, the input may be decompressed and restored.

[0034] Figure 2 One or more user equipment (UE) 10 and a base station (such as a gNB) 20 in a communication network system 40 according to an embodiment of the present application are described. The communication network system 40 includes one or more UEs 10 and base stations 20 (such as a first base station or a second base station). One or more UEs 10 may include a memory 12, a transceiver 13, and a processor 11 coupled to the memory 12 and the transceiver 13. The base station 20 may include a memory 22, a transceiver 23, and a processor 21 coupled to the memory 22 and the transceiver 23. The processor 11 or 21 may be configured to implement the functions, processes, and / or methods described in this description. The layers of the wireless interface protocol may be implemented in the processor 11 or 21. The memory 12 or 22 is operatively coupled to the processor 11 or 21 and stores various information for operating the processor 11 or 21. The transceiver 13 or 23 is operatively coupled to the processor 11 or 21 for sending and / or receiving wireless signals.

[0035] The processor 11 or 21 may include an application-specific integrated circuit (ASIC), other chipsets, logic circuits and / or data processing devices. The memory 12 or 22 may include a read-only memory (ROM), a random access memory (RAM), flash memory, a memory card, a storage medium and / or other storage devices. The transceiver 13 or 23 may include a baseband circuit for processing radio frequency signals. When the embodiment is implemented in software, the techniques described herein may be implemented by modules (such as procedures, functions, etc.) that perform related functions. These modules may be stored in the memory 12 or 22 and executed by the processor 11 or 21. The memory 12 or 22 may be implemented as a memory built into the processor 11 or 21, or as a memory external to the processor 11 or 21, in which case it may be communicatively coupled to the processor 11 or 21 in various known ways.

[0036] In the following embodiments, please note that machine model update or ML model post-processing is a general operation applicable to machine learning based use cases such as beam management and positioning. This is not limited to the autoencoder embodiments described here. For other use cases, the machine learning model can be deployed as only one node, which can be a gNB or a UE. The model is updated based on specific criteria, such as performance indicators, bit error rate (BER), spectral efficiency, or mean square error (MSE). The machine learning model can be post-processed based on various criteria, such as latency requirements, power levels, and UE complexity levels.

[0037] In some embodiments, a method for determining post-processing based on artificial intelligence (AI) / machine learning (ML) performed by a communication device includes: the communication device determines the post-processing of the AI / ML model according to at least one of the following: when the communication device is a user equipment (UE), the AI / ML model is downloaded to the UE without post-processing, and post-processing is performed on the UE, or the AI / ML model does not need to be post-processed before deployment; when the communication device is a UE, the AI / ML model is downloaded to the UE after post-processing, and post-processing is performed on the base station, or the UE is considered unable to perform post-processing, or post-processing is not allowed; when the communication device is a base station, the AI / ML model is uploaded to the base station without post-processing, and post-processing is performed on the base station, or the AI / ML model does not need to be post-processed before deployment; when the communication device is a base station, the AI / ML model is uploaded to the base station after post-processing, and post-processing is performed on the UE. The communication device may be a UE 10 or a base station 20.

[0038] In some embodiments, the processor 11 is configured to determine the post-processing of the artificial intelligence (AI) / machine learning (ML) model according to at least one of the following: when the communication device is a user equipment (UE) 10, the AI / ML model is downloaded to the UE 10 without post-processing, and the post-processing is performed on the UE 10, or the AI / ML model does not need to be post-processed before deployment; when the communication device is a UE 10, the AI / ML model is downloaded to the UE 10 after post-processing, and the post-processing is performed on the base station 20, or the UE 10 is considered unable to perform post-processing, or the post-processing is not allowed. This can solve the problems in the prior art, provide a method for performing post-processing for the UE or gNB, define the compression ratio of the autoencoder, formulate CSI report discarding rules, reduce system overhead, and provide good communication performance and / or high reliability.

[0039] In some embodiments, the processor 21 is configured to determine the post-processing of the artificial intelligence (AI) / machine learning (ML) model according to at least one of the following: when the communication device is a base station 20, the AI / ML model is uploaded to the base station 20 without post-processing, and the post-processing is performed on the base station 20, or the AI / ML model does not need to be post-processed before deployment; when the communication device is a base station 20, the AI / ML model is uploaded to the base station 20 after post-processing, and the post-processing is performed on the user equipment (UE) 10. This can solve the problems in the prior art, provide a method for performing post-processing for the UE or gNB, define the compression ratio of the autoencoder, formulate CSI report discarding rules, reduce system overhead, and provide good communication performance and / or high reliability.

[0040] Figure 3 A wireless communication method 300 is shown, which is used for determining a method for performing post-processing by a user equipment (UE) according to artificial intelligence (AI) / machine learning (ML), and the method is part of an embodiment of the present application. In some embodiments, the method 300 includes: step 302, the communication device determines the post-processing of the artificial intelligence (AI) / machine learning (ML) model according to at least one of the following: when the communication device is a user equipment (UE), the AI / ML model is downloaded to the UE without post-processing, and post-processing is performed on the UE, or there is no need to perform post-processing on the AI / ML model before deployment; when the communication device is a UE, the AI / ML model is downloaded to the UE after post-processing, and post-processing is performed at the base station, or it is considered that the UE cannot perform post-processing, or is not allowed to perform post-processing. The method can solve the problems in the prior art, provide a method for performing post-processing for the UE or gNB, define the compression ratio of the autoencoder, formulate CSI report discarding rules, reduce system overhead, and provide good communication performance and / or high reliability.

[0041] Figure 4A wireless communication method 400 is shown, which is used for determining a method for performing post-processing by a base station according to artificial intelligence (AI) / machine learning (ML), and the method is part of an embodiment of the present application. In some embodiments, the method 400 includes: step 402, the communication device determines the post-processing of the artificial intelligence (AI) / machine learning (ML) model according to at least one of the following: when the communication device is a base station, the AI / ML model is uploaded to the base station without post-processing, and post-processing is performed at the base station, or there is no need to perform post-processing on the AI / ML model before deployment; when the communication device is a base station, the AI / ML model is uploaded to the base station after post-processing, and post-processing is performed on the user equipment (UE). The method can solve the problems in the prior art, provide a method for performing post-processing for the UE or gNB, define the compression ratio of the autoencoder, formulate CSI report discarding rules, reduce system overhead, and provide good communication performance and / or high reliability.

[0042] In some embodiments, the post-processing related parameters of the post-processing include a pruning ratio or an AI / ML parameter quantization level. In some embodiments, the ratio of randomly pruned neurons is at least one of the following: signaling sent by the base station to the UE, including radio resource control (RRC) parameters, media access control elements (MAC-CE) and downlink control information (DCI) fields, or reported by the UE to the base station if the UE is able to perform pruning as the post-processing. In some embodiments, if the RRC parameter, the MAC-CE or the DCI field is empty or not indicated, it is considered that the UE is unable to perform the post-processing or is not allowed to perform the post-processing. In some embodiments, if the base station will perform pruning as the post-processing, and the AI / ML model or a portion of the AI / ML model will be deployed on the UE, the pruning ratio is reported from the UE to the base station or a third node, or the pruning ratio is indicated to the UE by the base station.

[0043] In some embodiments, at least one of the following conditions is met: the pruning ratio is an integer between 0 and 100, indicating the percentage of neurons that need to be pruned or removed; the pruning ratio is a decimal between 0 and 1, indicating the proportion of the neurons that need to be pruned or removed; the pruning ratio is an integer between 0 and 100, indicating the percentage of the neurons that need to be retained; and the pruning ratio is a decimal between 0 and 1, indicating the proportion of the neurons that need to be retained. In some embodiments, the pruning ratio or the AI / ML parameter quantization level is tabulated. In some embodiments, the AI / ML parameter quantization level includes at least one or more of 4 bits, 8 bits, 16 bits and 32 bits. In some embodiments, the pruning ratio or the AI / ML parameter quantization level is reported by the UE as a UE capability, or is adapted to the UE capability. In some embodiments, the pruning ratio or the AI / ML parameter quantization level is configured to the UE by the base station using RRC signaling (e.g., MAC-CE or DCI).

[0044] In some embodiments, the post-processing includes quantization of the AI / ML model output. In some embodiments, the quantization of the AI / ML model output adopts uniform quantization, differential quantization, or non-differential quantization, with at least one level of 4 bits, 8 bits, 16 bits, and 32 bits, and / or the quantization of the AI / ML model output is performed by model training without post-processing, or the quantization of the AI / ML model output is differential quantization. In some embodiments, the AI / ML model output is channel state information (CSI) compressed by the AI / ML model. In some embodiments, when multiple CSI reports containing the AI / ML model output overlap, one or more parts / groups of a CSI report of the multiple CSI reports are sorted and deleted according to a rule, wherein the AI / ML model output is the output of an encoder of an autoencoder. In some embodiments, some or all elements of a CSI report or only one CSI report of the multiple CSI reports are sorted according to a priority rule, and some or all elements of the CSI report are divided into three groups, including Group 0, Group 1 and Group 2, wherein Group 0 includes at least one of the following: a compression ratio of the autoencoder, a quantization level of the encoder output of the autoencoder, a rank indicator, and a strongest coefficient indicator.

[0045] In some embodiments, the priority rule is used to determine the elements in the group 1 and the group 2, wherein at least one of the following conditions is satisfied: the priority rule is to arrange the elements in descending order according to the indices of the reconstructed complex vector or the compressed output of the encoder of the autoencoder; the priority rule is to arrange the elements in ascending order according to the indices of the reconstructed complex vector or the compressed output of the encoder of the autoencoder; the priority rule is to arrange the elements in descending order according to the magnitudes of the reconstructed complex vector or the encoder output of the autoencoder; the priority rule is to arrange the elements in ascending order according to the magnitudes of the reconstructed complex vector or the encoder output of the autoencoder; the priority rule is to arrange the elements in ascending order according to the indices of the non-zero elements in the reconstructed complex vector or the compressed output of the encoder of the autoencoder; the priority rule is to arrange the elements in descending order according to the magnitudes of the non-zero elements in the reconstructed complex vector or the output of the encoder of the autoencoder; and the priority rule is to arrange the elements in ascending order according to the magnitudes of the non-zero elements in the reconstructed complex vector or the output of the encoder of the autoencoder.

[0046] In some embodiments, priority rules are used to determine elements in group 1 and group 2, where at least one of the following conditions is satisfied: the priority rule is an ascending arrangement of the indices of non-zero elements of the reconstructed complex vector or the output of the autoencoder encoder; the priority rule is a descending arrangement of the amplitudes of non-zero elements of the reconstructed complex vector or the output of the autoencoder encoder; the priority rule is an ascending arrangement of the amplitudes of non-zero elements of the reconstructed complex vector or the output of the autoencoder encoder.

[0047] Some embodiments of the present application address the following identified issues: For post-processing, how will the UE or gNB perform post-processing? The ML model for deployment needs to meet both complexity and performance requirements. In addition, post-processing should adapt to various factors. As a hyperparameter of an ML model, the compression ratio of the autoencoder indicates its processing capability. The compression ratio can be a key factor in determining model selection or model adaptation and should be adaptable. However, there is currently no discussion of related issues. The third issue involves CSI report discarding rules. The CSI report is divided into Part 1 CSI and Part 2 CSI. When multiple CSI reports overlap, they may be discarded. Therefore, it is necessary to design a discarding rule applicable to CSI reports compressed by ML.

[0048] ML model adaptation and post-processing:

[0049] Figure 5 is a schematic diagram showing an example of post-processing related to various factors according to an embodiment of the present application. Figure 5Note that in some embodiments, the AI / ML model can be trained in an online or offline manner on the gNB, UE, or a third node. After the model training is completed, the model will be deployed. Before deployment, the AI / ML model can be post-processed. The main purpose of post-processing is to reduce the model complexity, such as by pruning and discarding some neurons, or quantizing the neural network parameters in the AI / ML model. In some examples, the AI / ML model is downloaded to the UE without post-processing. Post-processing can be performed on the UE side, or there is no need to post-process the AI / ML model before deployment. In some examples, the AI / ML model is downloaded to the UE after post-processing. Post-processing can be performed on the gNB side to save power consumption of the UE, or it is considered that the UE cannot or is not allowed to perform post-processing. In some examples, the AI / ML model is uploaded to the gNB without post-processing. Post-processing can be performed on the gNB side, or there is no need to post-process the AI / ML model before deployment. In some examples, the AI / ML model is uploaded to the gNB after post-processing. Post-processing can be performed on the UE side, one of the motivations is to reduce upload overhead.

[0050] In the above examples, parameters related to post-processing (such as pruning ratios or AI / ML parameter quantization levels) can be signaled in one of the following ways. In some examples, if the UE is capable of performing pruning as post-processing, the ratio of randomly pruning some neurons is signaled by the gNB to the UE. In some examples, if the UE is capable of performing pruning as post-processing, the ratio of randomly pruning some neurons is reported by the UE to the gNB. In some examples, whether the UE has the ability to perform post-processing is implicitly indicated. For example, when the RRC parameter / MAC-CE / DCI field is empty or not indicated, the UE is considered unable to perform post-processing or is not allowed to perform post-processing.

[0051] Alternatively, if the gNB plans to perform pruning as post-processing and the model or part of the model needs to be deployed on the UE, the pruning ratio is reported by the UE to the gNB or a third node (e.g., another UE, gNB, server, or unit capable of performing computational tasks). The pruning ratio is indicated by the gNB to the UE. In some examples, the pruning ratio is an integer between 0 and 100, indicating the percentage of neurons that need to be (randomly) pruned or discarded. In some examples, the pruning ratio is a decimal between 0 and 1, indicating the percentage of neurons that need to be (randomly) pruned or discarded. In some examples, the pruning ratio is an integer between 0 and 100, indicating the percentage of neurons that need to be (randomly) retained. In some examples, the pruning ratio is a decimal between 0 and 1, indicating the percentage of neurons that need to be (randomly) retained.

[0052] In some examples, the pruning ratio or quantization level can be expressed in a table. For example, the value of the pruning ratio can be expressed as follows, where "10" means pruning 15% of the neurons:

[0053] 00 01 10 11 5％ 10％ 15％ 20％

[0054] Similarly, the value of the quantization level can be expressed as follows, where "10" indicates 16-bit quantization:

[0055] 00 01 10 11 4-digit 8-bit 16-bit 32-bit

[0056] In some examples, the quantization level of the ML model parameters may include at least one of {4 bits, 8 bits, 16 bits, 32 bits}. The supported quantization levels are reported by the UE as part of its capabilities. In some examples, the supported quantization levels or pruning parameters are determined by the complexity (level) of the AI / ML model supported by the UE. There is a mapping relationship between these supported quantization levels or pruning parameters and complexity levels. For example, if the UE only supports low-complexity models, only the corresponding low quantization levels are supported, and only parameters with higher pruning ratios (i.e., settings for pruning more neurons) are supported.

[0057] In some examples, the pruning ratio or quantization level is adapted according to the capabilities of the UE. For example, when the UE cannot support a high-complexity AI / ML model, the AI / ML model should be post-processed with a low quantization level and / or a higher pruning ratio (indicating the proportion of pruned neurons). As an alternative, the quantization level can be configured by the gNB to the UE. In some examples, the configuration is based on parameters such as {CPU, memory, storage, bus bandwidth, FLOPs} of the UE. In addition, when the UE is able to perform post-processing, the pruning ratio or quantization level can also be adapted according to the UE's power consumption level, latency, etc.

[0058] Figure 6 It is a schematic diagram showing an example of adapting post-processing according to the UE power consumption level when the UE is capable of post-processing, which is consistent with an embodiment of the present application. Figure 6In some embodiments, the procedure is as follows: First, the AI / ML model is deployed and activated on the UE side. As time goes by, the UE power consumption drops below a certain threshold. The UE requests post-processing, such as specifying a pruning ratio. The UE intends to switch to a new ML model after post-processing. The request or post-processing can be completed through RRC / MAC-CE or UCI. The request for model switching can also be completed through RRC / MAC-CE or UCI. The gNB may confirm the UE's request and check whether the UE can perform post-processing. If it can, the gNB will indicate the post-processing parameters, such as the pruning ratio. If not, the gNB may indicate a new ML model or complete the post-processing itself. The confirmation can be completed through DCI / RRC / MAC-CE. The indication of the post-processing parameters can also be completed through DCI / RRC / MAC-CE. Subsequently, the UE performs post-processing according to the indicated parameters. After the post-processing is completed, the UE may report this result to the gNB, and the report can be completed through DCI / RRC / MAC-CE. The gNB may trigger the model switch to the new AI / ML model and activate the model. Model triggering and switching can be completed through one or two signaling. These signalings can be DCI / RRC / MAC-CE. The UE may perform model switching and model activation. As an alternative, the adaptation can also be targeted at latency. When the UE requires lower latency and the latency is below a certain threshold, a simplified model is preferred, which may be obtained by pruning more neurons or using post-processing at a lower quantization level.

[0059] Figure 7 is a schematic diagram showing an example of adapting post-processing according to UE power consumption level when post-processing is performed by the gNB side, which is consistent with an embodiment of the present application. Figure 7 Note that in some embodiments, the procedure is as follows: First, the AI / ML model or the UE-side model in the bilateral model is deployed and activated on the UE side. Over time, the UE power consumption drops below a certain threshold. The UE requests post-processing, such as specifying a pruning ratio. The UE intends to switch to a new AI / ML model after post-processing. The request or post-processing can be completed through RRC / MAC-CE or UCI. The request for model switching can also be completed through RRC / MAC-CE or UCI. The gNB may confirm the UE's request and check whether the UE can perform post-processing. If possible, the gNB will indicate the post-processing parameters, such as the pruning ratio. If not, the gNB may indicate a new AI / ML model or complete the post-processing on its own. Figure 7This is the case shown. Confirmation can be done via DCI / RRC / MAC-CE. Indication of post-processing parameters can also be done via DCI / RRC / MAC-CE. The AI / ML model or the UE-side model in the dual-sided model is downloaded to the UE via the air interface. The gNB may trigger a model switch to a new ML model and activate the model. Model triggering and switching can be done via one or two signalings. These signalings can be DCI / RRC / MAC-CE. The UE may perform model switching and model activation. Alternatively, the adaptation can also be for latency. When the UE requires lower latency and the latency is below a certain threshold, a simplified model is preferred, which may be obtained by pruning more neurons or using lower quantization level post-processing.

[0060] In some examples, whether the AI / ML model can be post-processed is a property of the model. The UE, gNB, or third node can check the property during model download, upload, retraining, or post-processing. In some examples, whether the AI / ML model can be pruned is a property. The UE, gNB, or third node can check the property during model download, upload, retraining, or post-processing. In some examples, whether the AI / ML model parameters can be quantized is a property. If the model parameters can be quantized, the quantization level is also a property of the model. The UE, gNB, or third node can check these properties during model download, upload, retraining, or post-processing. In some examples, the model attributes are packaged with the model image. The model attributes facilitate the deployment, training, and monitoring of the model. Alternatively, the model attributes are not packaged with the model image, but are managed separately by the gNB or third node. The model attributes and the model image are associated by the model ID. The model attributes facilitate the deployment, training, and monitoring of the model.

[0061] In some examples, the AI / ML model attributes include at least one of the following: {possible quantization level of ML model output, compression ratio, whether the AI / ML model can be post-processed}. The UE, gNB, or third node can check these model attributes during model download, upload, retraining, or post-processing. In some examples, if the model attributes match the request of the UE or gNB, the model will be accepted; otherwise, a new model will be requested. In some examples, the post-processing can be performed on the UE, gNB, or third node and performed during the fine-tuning phase. Note that this post-processing applies to all types of AI / ML models and usage scenarios, not just CSI feedback enhanced by AI / ML.

[0062] Quantization of model output:

[0063] Figure 8is a schematic diagram showing a quantization example for enhancing CSI feedback according to an embodiment of the present application. Figure 8 The quantization process for enhancing CSI feedback is shown. Fig. 9 is a schematic diagram showing an example of quantization performed to reconstruct a complex number according to an embodiment of the present application. Figure 8 and Fig. 9 Note that in some examples, quantization of the model output is a post-processing method. Quantization can be uniform quantization or non-differential quantization, and the supported quantization levels include at least one of {4-bit, 8-bit, 16-bit, 32-bit}. In some examples, the quantization of the model output is performed simultaneously with model training. In some examples, the quantization is differential quantization. The element with the strongest amplitude in the output vector is quantized to 1, and the remaining elements are quantized with reference to this value, for example according to Or complete the quantification through a table.

[0064]

[0065] If two elements have the same strongest amplitude, one of them will be randomly selected as the strongest coefficient, and its index is the strongest coefficient index. In some examples, quantization directly quantizes real numbers. In some examples, quantization quantizes complex numbers, including quantization of amplitude and phase. Amplitude can be quantized by uniform quantization (non-differential) or differential quantization; phase can be quantized using at least one of the following methods: {BPSK, QPSK, 8PSK, 16PSK, 32PSK, or 64PSK}.

[0066] In some examples, quantization is used to reconstruct complex numbers according to the following steps: Let a1, a2, ... a8 be real numbers, which are the outputs of the encoder. The reconstructed complex vector is C = [a1 + j·a5, a2 + j·a6, a3 + j·a7, a4 + j·a8]. Then, the amplitude and phase of C are quantized respectively. At the decoder end, the quantized vector will be dequantized accordingly. Let C' be the dequantized complex vector, and reconstruct the real numbers a1', a2', ..., a8' according to the mapping relationship from a1, a2, ... a8 to the complex vector C, that is, perform reverse mapping. C' = [a1' + j·a5', a2' + j·a6', a3' + j·a7', a4' + j·a8']. Where a1', a2', ..., a8' are the inputs of the decoder. Please note that the real number set a1, a2, ... a8 is only used as an example and is not necessarily limited to 8 elements.

[0067] In another example, as an alternative, the complex vector C is constructed as C = [a1 + j·a2, a3 + j·a4, a5 + j·a6, a7 + j·a8]. The corresponding dequantized complex vector is C' = [a1' + j·a2', a3' + j·a4', a5' + j·a6', a7' + j·a8']. In some examples, the quantization of the model output can adapt to various factors. For example, channel conditions, SINR, Doppler effect, latency, and UE power consumption level. For example, if the SINR drops below a certain threshold, the UE will request a higher quantization level from the gNB, and the gNB will select and confirm the quantization level to the UE through signaling.

[0068] In some examples, quantization does not require reconstruction of complex numbers, but directly quantizes real numbers a1, a2, ... a8, and can use a differential quantization method or a non-differential method. The method described in the above embodiment, such as Figure 8 , Fig. 9 And the methods in the related description can be used as examples.

[0069] Enhanced compression ratio adaptation of CSI feedback:

[0070] Fig.10 It is a schematic diagram showing an example of determining / defining a compression ratio for enhanced CSI feedback according to an embodiment of the present application. In some examples, the compression ratio is defined as follows: compression ratio = output size / input size. Or, compression ratio = input size / output size. In some examples, the input size is the size of the input vector, and the output size is the size of the encoder output vector. In some examples, the compression ratio can be indicated based on a table. For example, "10" can be represented by two bits, indicating that the compression ratio is 1 / 16, that is, the ratio of the output size to the input size is 1:16.

[0071] 00 01 10 11 1 / 4 1 / 8 1 / 16 1 / 32

[0072] In some examples, the compression ratio can be indicated based on a table. For example, "10" can be represented by two bits, indicating a compression ratio of 16, that is, a ratio of input size to output size of 16:1.

[0073] 00 01 10 11 4 8 16 32

[0074] In some examples, the compression ratio of the model output can be an adaptability factor, and the adaptation can be done explicitly or implicitly. In some examples, the explicit indication means that the compression ratio is configured by the gNB. In some examples, the compression ratio is reported by the UE and confirmed by the gNB. In some examples, the implicit indication means that the compression ratio is limited by at least one of the following factors: {SNR / SINR value, model complexity, complexity level supported by the UE, antenna port, antenna geometry, channel ranks}.

[0075] CSI discard rules:

[0076] When multiple CSI reports overlap, they are sorted according to the priority rules and partially discarded. Rep In some examples, the CSI report is generated by an autoencoder model based on the post-processing example above. In some examples, the CSI report is not generated by an autoencoder model with post-processing. In some examples, Figure 8 and Fig. 9 The quantization method described in can be applied to amplitude quantization and / or phase quantization in this section.

[0077] Table: New priority reporting levels envisioned for Part II CSI.

[0078]

[0079] A subset or all of the elements in the CSI report are divided into three groups, which are composed as described below. Typically, a subset or all of the elements in the CSI report are sorted according to some priority rule. The elements of Group 1 and Group 2 are determined based on this priority rule. The groups of CSI reports are then sorted according to the above table (new priority reporting levels designed for the second part of CSI). The user equipment (UE) will decide whether to discard or ignore the CSI report based on these reporting levels. Next, various grouping methods are provided.

[0080] In some examples, the priority rule is to sort in descending order according to the element index of the complex vector C reconstructed in the encoder compressed output of the autoencoder. Group 0 contains the compression ratio and / or quantization level (if reported) and the rank indication (RI) (if reported), and the strongest coefficient index (SCI). Group 1 includes the indication The index of the quantization amplitude of the highest priority elements and their index information in the complex vector C, as well as the index of the quantization phase of these highest priority elements (based on the quantization table / method). NZ Represents the total number of non-zero elements in C after encoder output quantization, calculated across layers. Group 2 includes instructions The index of the quantization amplitude of the lowest priority element and the index of its quantization phase (based on the quantization table / method). The values ​​of parameters a and b can be taken from at least the following set: {(2,2), (3 / 2,3), (3,3 / 2), (4,4 / 3), (4 / 3,4), (5 / 3,5 / 2), (5 / 2,5 / 3), (5,5 / 4), (5 / 4,5), ...}.. The parameter v represents the number of elements of the strongest coefficients across layers, where v = 1 is used for a single layer, v = 2 is used for two layers, and v = 3 is used for three layers. In some examples, v is fixed to 1 and does not change with the number of layers or rank.

[0081] In some examples, the priority rule is based on descending order of element index in the encoder compressed output of the autoencoder. Group 0 includes the compression ratio and / or quantization level (if reported) and the rank indication (RI) (if reported), and the strongest coefficient index (SCI). Group 1 includes the indication The index of the quantization amplitude of the highest priority elements and the index information of these elements in the encoder output vector. Group 2 includes the index indicating The index of the quantization amplitude of the lowest priority elements and the index information of these elements in the encoder output vector. The values ​​of parameters a and b can be taken from at least the following set: {(2,2), (3 / 2,3), (3,3 / 2), (4,4 / 3), (4 / 3,4), (5 / 3,5 / 2), (5 / 2,5 / 3), (5,5 / 4), (5 / 4,5), ...}.. The parameter v represents the number of elements of the strongest coefficients across layers, where v = 1 applies to a single layer, v = 2 applies to two layers, and v = 3 applies to three layers. In some examples, v is fixed to 1 and does not change with the number of layers or rank.

[0082] As an alternative to the above example, the priority may be an ascending order of the index of the elements in C or in the autoencoder output. As another alternative to the above example, the priority may be an ascending order of the (quantized) magnitudes of the elements in C or in the autoencoder output (vector). As another alternative to the above example, the priority may be a descending order of the (quantized) magnitudes of the elements in C or in the autoencoder output (vector).

[0083] In some examples, the elements of the second part CSI are not sorted and there is no priority rule. Based on the non-zero elements in the complex vector C output by the autoencoder encoder compression, the even-indexed elements are assigned to group 1 and the odd-indexed elements are assigned to group 2. Group 0 contains the compression ratio and / or quantization level (if reported) and the rank indication (RI) (if reported), as well as the strongest coefficient index (SCI). Group 1 includes the indication The index of the quantized amplitude of the even-indexed elements and their position index information in the complex vector C, as well as the index of the quantized phase of these even-indexed elements (based on the quantization table / method). NZ Indicates the total number of non-zero elements of the quantized complex vector C output by the encoder. Group 2 includes the indicator The index of the quantized amplitude of the odd-indexed elements and their position index information in the complex vector C, as well as the index of the quantized phase of these odd-indexed elements (based on the quantization table / method).

[0084] Alternatively, Group 0 contains the compression ratio and / or quantization level (if reported) and the rank indication (RI) (if reported), and the strongest coefficient index (SCI). Group 1 includes the indication The index of the quantized amplitude of the odd-indexed elements and their position index information in the complex vector C, as well as the index of the quantized phase of these odd-indexed elements (based on the quantization table / method). NZ Indicates the total number of non-zero elements in the quantized complex vector C output by the encoder. Group 2 includes the indicator The index of the quantized amplitude of the even-indexed elements and their position index information in the complex vector C, as well as the index of the quantized phase of these even-indexed elements (based on the quantization table / method).

[0085] In some examples, there is no priority rule. The total number of non-zero elements K in the autoencoder encoder compression output vector NZ The even-indexed elements of are assigned to group 1 and the odd-indexed elements are assigned to group 2. Group 0 contains the compression ratio and / or quantization level (if reported) and the rank indication (RI) (if reported), as well as the strongest coefficient index (SCI). Group 1 includes the following: The index of the quantized amplitude of even-indexed elements and their position information in the encoder output vector, and / or the index of the quantized phase of these even-indexed elements (based on the quantization table / method). NZ Indicates the total number of non-zero elements in the encoder output vector after quantization. Group 2 includes the following: Indicates The indices of the quantized amplitudes of the odd-indexed elements and their position information in the encoder output vector, and / or the indices of the quantized phases of these odd-indexed elements (based on the quantization table / method).

[0086] As an alternative, Group 0 contains the compression ratio and / or quantization level (if reported) and the rank indication (RI) (if reported), and the strongest coefficient index (SCI). Group 1 includes the following: The indices of the quantized amplitudes of the odd-indexed elements and their position information in the encoder output vector, as well as the indices of the quantized phases of these odd-indexed elements (based on the quantization table / method). NZIndicates the total number of non-zero elements in the encoder output vector after quantization. Group 2 includes the following: Indicates The indices of the quantized amplitudes of the even-indexed elements and their position information in the encoder output vector, and / or the indices of the quantized phases of these even-indexed elements (based on the quantization table / method).

[0087] In some examples, a scaling factor is quantized and reported with the compressed CSI. CSI can be quantized differentially. The element with the strongest amplitude is quantized to 1, and the scaling factor is the original amplitude of this element after quantization (i.e., the strongest amplitude). In some examples, the strongest amplitude is quantized using 4 bits, 8 bits, 16 bits, or 32 bits. This scaling factor can be included in the first part of the CSI or the second part of the CSI. If included in the second part of the CSI, it belongs to group 0. In some examples, the encoder output is non-differential quantization, so there is no strongest coefficient. The reporting of the scaling factor can be used in conjunction with any of the above examples.

[0088] Group 0 contains the compression ratio and / or quantization level (if reported) and / or compression ratio (if reported). Group 1 includes indices indicating the magnitudes of high priority even / odd non-zero elements and their corresponding position information (the positions of these even / odd elements). Group 2 includes indices indicating the magnitudes of low priority odd / even non-zero elements and their corresponding position information (the positions of these odd / even elements).

[0089] In summary, in certain embodiments of the present application, the machine learning model can be trained online or offline. After the AI / ML model training is completed, post-processing is required. Post-processing may include quantization and / or pruning of model parameters. This process determines the complexity of the model and also determines the performance of the model. A flexible model post-processing solution is provided for different factors (such as UE capabilities, UE power consumption, and processing delay requirements). Next, the CSI feedback problem is studied. The CSI is compressed by the AI / ML model, and the compressed CSI is reported to the gNB by the UE. The report is made after the compressed CSI is quantized. Detailed information on the quantization of the AI / ML model output is provided in the article. In addition, the compression ratio is used as a factor to determine the report size. Finally, when multiple such CSIs are reported, priority rules are required, and some CSIs are discarded according to the priority rules. The effects of some embodiments of the present invention include at least one of the following: a model adaptation mechanism based on post-processing is provided. When the UE has different power consumption levels or capabilities, the model can be post-processed accordingly. Post-processing supports model selection and provides greater flexibility for ML model customization. In the absence of an available ML model, an ML model can be generated by post-processing. Post-processing enables the ML model to adapt to different scenarios and settings, thereby improving the flexibility of model selection. The compression ratio can be used as a model attribute and designed to have an adaptive mechanism. The CSI feedback enhanced by ML can adapt to channel conditions, UE capabilities, or UE power consumption, etc. The CSI feedback enhanced by ML is divided into different groups and the groups are given priorities. When multiple CSI reports overlap, some or all groups of CSI reports can be discarded according to the discard order. The present invention solves the problem of CSI report overlap in CSI feedback enhanced by ML.

[0090] Fig.11 7 is a block diagram of an exemplary wireless communication system 700 according to an embodiment of the present application. The embodiments described herein may be implemented into a system using any appropriately configured hardware and / or software. Fig.11 A system 700 is shown, which includes a radio frequency (RF) circuit 710, a baseband circuit 720, an application circuit 730, a memory / storage device 740, a display 750, a camera 760, a sensor 770, and an input / output (I / O) interface 780, which are coupled to each other at least as shown. The application circuit 730 may include circuits, such as but not limited to one or more single-core or multi-core processors. The processor may include any combination of general-purpose processors and special-purpose processors, such as a graphics processor, an application processor. The processor may be coupled to a memory / storage device and configured to execute instructions stored in the memory / storage device to enable various applications and / or operating systems to run on the system.

[0091] While the present disclosure has been described in connection with what is considered to be the most practical and preferred embodiment, it is to be understood that the present disclosure is not limited to the disclosed embodiment, but is intended to cover various arrangements made without departing from the scope of the broadest interpretation of the appended claims.

Claims

1. A method for determining post-processing based on artificial intelligence AI / machine learning ML, performed by a communication device, characterized in that: include: The communication device determines post-processing of the AI / ML model according to at least one of the following: When the communication device is a user equipment UE, the AI / ML model is downloaded to the UE without post-processing, and the post-processing is performed on the UE, or the post-processing is not required before deployment; When the communication device is the UE, the AI / ML model is downloaded to the UE after the post-processing, and the post-processing is performed on the base station, or the UE is deemed unable to perform the post-processing or is not allowed to perform the post-processing; When the communication device is the base station, the AI / ML model is uploaded to the base station without the post-processing, and the post-processing is performed on the base station, or no post-processing is required before deployment; When the communication device is the base station, the AI / ML model is uploaded to the base station after the post-processing, and the post-processing is performed on the UE.

2. The method according to claim 1, characterized in that: The post-processing related parameters of the post-processing include a pruning ratio or an AI / ML parameter quantization level.

3. The method according to claim 2, characterized in that The rate of randomly pruned neurons is at least one of the following: signaling sent by the base station to the UE, including radio resource control RRC parameters, media access control elements MAC-CE and downlink control information DCI fields, or reported by the UE to the base station if the UE is capable of performing pruning as the post-processing.

4. The method according to claim 3, characterized in that If the RRC parameter, the MAC-CE or the DCI field is empty or not indicated, it is considered that the UE is unable to perform the post-processing or is not allowed to perform the post-processing.

5. The method according to claim 2, characterized in that: If the base station will perform pruning as the post-processing and the AI / ML model or part of the AI / ML model will be deployed on the UE, the pruning ratio is reported from the UE to the base station or a third node, or the pruning ratio is indicated by the base station to the UE.

6. The method according to any one of claims 2 to 5, characterized in that At least one of the following conditions is met: The pruning ratio is an integer between 0 and 100, indicating the percentage of neurons that need to be pruned or removed; The pruning ratio is a decimal between 0 and 1, indicating the proportion of the neurons that need to be pruned or removed; The pruning ratio is an integer between 0 and 100, indicating the percentage of the neurons to be retained; and The pruning ratio is a decimal between 0 and 1, indicating the proportion of the neurons that need to be retained.

7. The method according to any one of claims 2 to 6, characterized in that The pruning ratios or the AI / ML parameter quantization levels are tabulated.

8. The method according to any one of claims 2 to 7, characterized in that The AI / ML parameter quantization level includes at least one or more of 4 bits, 8 bits, 16 bits and 32 bits.

9. The method according to any one of claims 2 to 8, characterized in that The pruning ratio or the AI / ML parameter quantization level is reported by the UE as a UE capability, or is adapted as the UE capability.

10. The method according to any one of claims 2 to 8, characterized in that The pruning ratio or the AI / ML parameter quantization level is configured by the base station to the UE.

11. The method according to any one of claims 1 to 10, characterized in that The post-processing includes quantization of the AI / ML model output.

12. The method according to claim 11, characterized in that The quantization of the AI / ML model output adopts uniform quantization, differential quantization or non-differential quantization, with at least one level of 4 bits, 8 bits, 16 bits and 32 bits, and / or the quantization of the AI / ML model output is performed by model training without post-processing, or the quantization of the AI / ML model output is differential quantization.

13. The method according to claim 11 or 12, characterized in that: The AI / ML model output is channel state information CSI compressed by the AI / ML model.

14. The method according to claim 11 or 12, characterized in that: When multiple CSI reports containing the AI / ML model output overlap, one or more parts / groups of a CSI report of the multiple CSI reports are sorted and deleted according to a rule, wherein the AI / ML model output is the output of the encoder of the autoencoder.

15. The method according to claim 14, characterized in that Part or all of the elements in one CSI report or only one CSI report of the multiple CSI reports are sorted according to a priority rule, and part or all of the elements in the CSI report are divided into three groups, including group 0, group 1 and group 2, wherein group 0 includes at least one of the following: a compression ratio of the autoencoder, a quantization level of the encoder output of the autoencoder, a rank indicator, and a strongest coefficient indicator.

16. The method according to claim 15, characterized in that The priority rule is used to determine the elements in the group 1 and the group 2, where at least one of the following conditions is satisfied: The priority rule is to arrange in descending order according to the index of the elements in the compressed output of the encoder of the reconstructed complex vector or the autoencoder; The priority rule is to arrange in ascending order according to the index of the elements in the compressed output of the encoder of the reconstructed complex vector or the autoencoder; The priority rule is to arrange the elements in descending order according to the magnitude of the reconstructed complex vector or the output of the autoencoder; The priority rule is to arrange the elements in ascending order according to the magnitudes of the reconstructed complex vector or the autoencoder encoder output; The priority rule is to arrange in ascending order according to the index of the non-zero elements in the compressed output of the encoder of the reconstructed complex vector or the autoencoder; The priority rule is to arrange the non-zero elements in the reconstructed complex vector or the output of the encoder of the autoencoder in descending order according to the magnitudes thereof; and The priority rule is to arrange in ascending order according to the magnitudes of the non-zero elements in the output of the encoder of the reconstructed complex vector or the autoencoder.

17. A user equipment, characterized in that: include: Memory; Transceiver; as well as a processor coupled to the memory and the transceiver; The processor is configured to execute the method according to any one of claims 1 to 16.

18. A base station, characterized in that: include: Memory; Transceiver; as well as a processor coupled to the memory and the transceiver; The processor is configured to execute the method according to any one of claims 1 to 16.

19. A non-transitory machine-readable storage medium, characterized in that Instructions are stored thereon, and when a computer executes the instructions, the instructions cause the computer to perform the method of any one of claims 1 to 16.

20. A chip, characterized in that: include: A processor configured to call and run a computer program stored in a memory so that a device equipped with the chip executes a method according to any one of claims 1 to 16.

21. A computer-readable storage medium, characterized in that: A computer program is stored therein, wherein the computer program causes a computer to execute the method of any one of claims 1 to 16.

22. A computer program product, characterized in that Comprising a computer program, wherein the computer program causes a computer to perform the method of any one of claims 1 to 16.

23. A computer program, characterized in that The computer program causes a computer to execute the method according to any one of claims 1 to 16.

Images