Encoder training method, feedback method, terminal device, network device and system based on decoder-side assistance

Abstract

The application discloses an encoder training method based on decoder side assistance, a feedback method, a terminal device, a network device and a system. The method comprises the following steps: a terminal device acquires decoder side assistance training data; the terminal device encodes target channel state information in a local training data set to obtain feedback load, and reconstructs the feedback load by using the decoder side assistance training data to obtain reconstructed channel state information; finally, a loss function constructed by using the target channel state information and the reconstructed channel state information is used to train an encoder to obtain a trained encoder. The application can make the encoder more adaptive to the channel characteristics of the terminal device, better exert the active adaptation capability of the terminal device, and improve the reconstruction quality of the channel state information while ensuring the intercommunication of the encoder and the decoder, thereby improving the transmission performance of the whole communication system.

Description

Technical Field

[0001] This application relates to the field of wireless communication technology, and in particular to an encoder training method, feedback method, terminal device, network device and system based on decoder-side assistance. Background Technology

[0002] In channel state information compression feedback based on artificial intelligence / machine learning, terminal devices typically need to compress the target channel state information they measure or calculate into a feedback payload jointly defined by the encoder and decoder. Network devices then reconstruct the target channel state information based on this feedback payload. However, in practical applications, due to the incompatibility of encoder / decoder parameters between different vendors, the network device's decoder cannot correctly decode the feedback payload generated by the encoder, thus affecting the performance of channel state information compression feedback.

[0003] Therefore, in channel state information compression feedback based on artificial intelligence / machine learning, the network device typically trains the encoder and directly provides the trained encoder parameters to the terminal device. While this allows the terminal device to quickly obtain an encoder that interoperates with the decoder deployed on the network device, the local channel environment, receiving antenna capabilities, receiving link implementation, and current mobility status of the terminal device may differ from the training conditions of the network device. This means that the encoder parameters directly provided by the network device cannot be well adapted to the channel state information of the terminal device, failing to leverage the terminal device's local adaptation capabilities. Consequently, this affects the reconstruction quality of the channel state information feedback, ultimately leading to a decline in the transmission performance of the entire communication system.

[0004] Therefore, existing technologies still need to be improved and enhanced. Summary of the Invention

[0005] The technical problem to be solved by this application is to provide an encoder training method, feedback method, terminal device, network device and system based on decoder-side assistance, in order to address the shortcomings of the existing technology.

[0006] To address the aforementioned technical problems, the first aspect of this application provides a decoder-side assisted encoder training method, wherein the decoder-side assisted encoder training method specifically includes: The terminal device acquires decoder-side auxiliary training data, wherein the decoder-side auxiliary training data includes auxiliary decoder information and payload information. The auxiliary decoder information is used to describe the reconstruction behavior of the decoder deployed on the network device for reproducing, approximating, constraining or verifying the terminal device, and the payload information is used to describe the target object output by the encoder of the terminal device. The terminal device trains the encoder of the terminal device based on the decoder-side auxiliary training data and the local training dataset to obtain the trained encoder.

[0007] The encoder training method based on decoder-side assistance, wherein after the terminal device trains the encoder of the terminal device based on the decoder-side assistance training data and the local training dataset to obtain the trained encoder, the method further includes: The terminal device verifies the trained encoder to obtain the verification result; The terminal device generates ready feedback information based on the verification result, enabling the network device to perform subsequent operations based on the ready feedback information. The subsequent operations include one or more of the following: activating the payload configuration file, updating the decoder-side auxiliary training data, triggering a rollback process, or updating the auxiliary decoder.

[0008] The encoder training method based on decoder-side assistance, wherein the correspondence between the ready feedback information and the subsequent operation includes at least one of the following: The readiness feedback message indicates that training is ready and applicable; the subsequent operation is to activate all payload configuration files. The readiness feedback information is partially applicable; the subsequent operation is to activate the partially applicable load profile. The readiness feedback message is "not applicable," and the subsequent operation is to not activate any load profile. The readiness feedback information is to supplement the decoder-side auxiliary training data, and the subsequent operation is to update the decoder-side auxiliary training data; The readiness feedback message indicates verification failure, and the subsequent operation is to trigger a rollback process; The readiness feedback message is a request for the network device to update the auxiliary decoder, and the subsequent operation is to update the auxiliary decoder.

[0009] The encoder training method based on decoder-side assistance, wherein the ready feedback information includes a bitmap, list, or set of indexes generated according to at least one of the following: payload profile, rank, layer, subband, transmit antenna port, receive capability category, or configuration index.

[0010] The encoder training method based on decoder-side assistance includes at least one of the following: an auxiliary decoder, a decoder actually deployed on the network device, a reference decoder, decoder parameters, decoder parameter increments, a decoder header, a decoder adaptation module, a decoder identifier, or a base decoder identifier; and at least one of the following: a decoder behavior descriptor, a distillation target, a reconstruction output constraint, a consistency check information, or a decoder behavior version. Furthermore, the auxiliary decoder information is associated with an auxiliary training data identifier, a payload configuration file identifier, or a consistency version.

[0011] The encoder training method based on decoder-side assistance, wherein when training sample set, encoder parameters, reference encoder, decoder parameters or reference decoder are used as supplementary / backup information, the terminal device needs to perform local adaptation training, assisted reconstruction verification and readiness feedback information reporting on the encoder of the terminal device using the local training dataset.

[0012] The encoder training method based on decoder-side assistance, wherein the decoder-side assistance training data is sent from the network device to the terminal device, and the generation process of the decoder-side assistance training data specifically includes: The network device acquires a preset training dataset and constructs an auxiliary decoder based on the preset training dataset. When the auxiliary decoder is not a decoder deployed by the network device, the auxiliary decoder and the decoder deployed by the network device meet the consistency requirements. The network device uses the auxiliary decoder to generate payload information and verifies whether the auxiliary decoder and the payload information meet a preset reconstruction quality threshold. If the preset reconstruction quality threshold is met, the network device obtains the auxiliary decoder information corresponding to the auxiliary decoder, and generates decoder-side auxiliary training data based on the auxiliary decoder information and the payload information.

[0013] The encoder training method based on decoder-side assistance, wherein after constructing the auxiliary decoder based on the preset training dataset, the method further includes: If the auxiliary decoder is not a decoder deployed by the network device, the consistency between the auxiliary decoder and the decoder deployed by the network device is detected; When the consistency meets the preset consistency requirements, the network device executes the step of generating payload information using the auxiliary decoder; When the consistency does not meet the preset consistency requirements, the network device re-executes the steps of obtaining the preset training dataset and constructing an auxiliary decoder based on the preset training dataset.

[0014] The encoder training method based on decoder-side assistance includes, in which the decoder-side assistance training data further includes one or more of the following: assistance training data identifier, applicability conditions, performance evaluation configuration, and lifecycle. The performance evaluation configuration is bound to at least one of the following: decoder behavior version, consistency threshold, local closed-loop verification stage, channel state information report configuration, or configuration index.

[0015] The encoder training method based on decoder-side assistance, wherein the performance evaluation configuration includes at least one of the following: metric type, evaluation object, evaluation stage, aggregation range, quantile type or metric threshold, wherein the evaluation object includes at least one of coding payload, dequantization payload, auxiliary decoder output channel state information or network-side actual decoder output channel state information, and the evaluation stage includes local training, local verification and verification before reporting ready feedback information.

[0016] The encoder training method based on decoder-side assistance, wherein the terminal device acquires decoder-side assisted training data specifically including: The terminal device receives the first descriptive information of the decoder-side auxiliary training data generated by the network device, sent by the network device. The terminal device filters the first description information based on its own capabilities to obtain the second description information, and then feeds back the second description information to the network device; The terminal device receives decoder-side auxiliary training data sent by the network device, wherein the decoder-side auxiliary training data sent by the network device is obtained by the network device filtering the decoder-side auxiliary training data generated by the network device based on the second description information.

[0017] The encoder training method based on decoder-side assistance, wherein the first descriptive information includes one or more of the following: a list of payload profiles, decoder-side assistance training data identifiers, applicability conditions, size of decoder-side assistance training data, required training capabilities, performance thresholds, validity period, candidate decoder-side assistance training data, decoder behavior version, required local training budget, required storage budget, supported training modes, and validation granularity, wherein the list of payload profiles includes one or more payload profiles.

[0018] The encoder training method based on decoder-side assistance, wherein the second descriptive information includes one or more of the following: a list of accepted payload profiles, decoder-side assisted training data acceptable to the terminal device, supported training modes, available training budget, available storage budget, supported rank, layer, subband, transmit antenna port range, and ready feedback information reporting granularity; wherein some or all of the payload profiles in the payload profile list are included in the list of accepted payload profiles.

[0019] The encoder training method based on decoder-side assistance, wherein before the terminal device acquires decoder-side assisted training data, the method further includes: The terminal device sends its own capabilities to the network device based on the query request of the network device. The own capabilities include at least one of the following: the ability to parse decoder-side auxiliary training data, the ability to obtain local target channel state information, the ability to train or fine-tune the encoder according to the payload profile, the number of supported payload profiles, the supported rank / layer / subband / transmit antenna port range, the receive capability category, the local training budget, the storage budget, whether it supports low-rank adaptation parameter updates, whether it supports reporting applicability according to the configuration index, and the ability to update only the encoder adaptation layer or only the encoder head.

[0020] The encoder training method based on decoder-side assistance, wherein the target object output by the trained encoder satisfies a preset reconstruction quality threshold after being reconstructed by the auxiliary decoder corresponding to the auxiliary decoder information.

[0021] The encoder training method based on decoder-side assistance, wherein the terminal device trains the encoder of the terminal device based on the decoder-side assistance training data and the local training dataset to obtain the trained encoder specifically includes: The terminal device acquires the encoder; The terminal device inputs the target channel state information in the local training dataset into the encoder, and outputs the feedback payload through the encoder; The terminal device uses the auxiliary decoder information to reconstruct the feedback payload to obtain second reconstructed channel state information, and updates the preset network layer in the encoder based on the second reconstructed channel state information and the target channel state information to obtain the trained encoder.

[0022] The encoder training method based on decoder-side assistance, wherein the terminal device acquires the encoder specifically as follows: The terminal device reads the encoder that it has actually deployed.

[0023] The encoder training method based on decoder-side assistance, wherein the acquisition of the encoder by the terminal device specifically includes: The terminal device acquires payload data from the decoder-side auxiliary training data, wherein the payload data is one of intermediate feedback payload, payload distribution, or codeword statistics; The terminal device uses the auxiliary decoder information to reconstruct the first reconstructed channel state information corresponding to the payload data; The terminal device forms pre-training samples based on the payload data and the first reconstructed channel state information; The terminal device uses the pre-trained samples to pre-train the encoder corresponding to the terminal device, and uses the pre-trained encoder as the encoder of the terminal device.

[0024] The encoder training method based on decoder-side assistance, wherein when the decoder-side assisted training data supports multiple payload profiles, the terminal device trains an encoder head for each payload profile or trains a shared encoder head for all payload profiles, and the multiple payload profiles are all or part of the payload profiles corresponding to the network device.

[0025] The encoder training method based on decoder-side assistance, wherein the preset network layer is all the network layers of the encoder, or it is one or more of the following: encoder head, encoder adaptation module, last preset number layer, branch module corresponding to the payload configuration file, and rank or layer related module.

[0026] The encoder training method based on decoder-side assistance, wherein during the encoder training process on the terminal device, the method further includes: When the target channel state information used to train the encoder contains multiple layers, the terminal device selects the effective layer from the multiple layers; The terminal device trains the encoder only for the effective layer.

[0027] The encoder training method based on decoder-side assistance, wherein the method further includes: The terminal device verifies the trained encoder only for the effective layer and reports the readiness feedback information corresponding to the effective layer, payload profile, rank, or subband condition to the network device.

[0028] The encoder training method based on decoder-side assistance, wherein the list of payload configuration files corresponding to the decoder-side assisted training data consists of the minimum payload configuration files that can meet the coverage requirements and do not exceed the terminal device budget.

[0029] The encoder training method based on decoder-side assistance is described above. In this method, the network device sends out auxiliary decoder information in a one-time manner or sends out decoder-side auxiliary training data in a progressive manner. When the network device sends out decoder-side auxiliary training data in a progressive manner, the network device first sends out the basic payload configuration file, auxiliary decoder information or basic auxiliary package, and then sends out the supplementary payload configuration file, decoder header, decoder adaptation module, incremental package or fallback relationship according to the ready feedback information of the terminal device.

[0030] The encoder training method based on decoder-side assistance, wherein the triggering condition for progressively distributing decoder-side auxiliary training data is triggered by resending, rolling back, or partially activating the data based on the readiness feedback information of the terminal device, such as verification failure, partial applicability, request for supplementary decoder-side auxiliary training data, or inconsistency between the decoder behavior version or consistency version.

[0031] A second aspect of this application provides a decoder-side assisted encoder training method, wherein the decoder-side assisted encoder training method specifically includes: The network device constructs decoder-side auxiliary training data, wherein the decoder-side auxiliary training data includes auxiliary decoder information and payload information. The auxiliary decoder information is used to describe the reconstruction behavior of the decoder deployed on the network device for reproducing, approximating, constraining or verifying on the terminal device, and the payload information is used to describe the target object output by the encoder of the terminal device. The network device sends the decoder-side auxiliary training data so that the terminal device can train the encoder of the terminal device based on the decoder-side auxiliary training data and the local training dataset to obtain the trained encoder.

[0032] The decoder-side assisted encoder training method, wherein the network device constructs decoder-side assisted training data specifically including: The network device acquires a preset training dataset and constructs an auxiliary decoder based on the preset training dataset. When the auxiliary decoder is not a decoder deployed by the network device, the auxiliary decoder and the decoder deployed by the network device meet the consistency requirements. The network device uses the auxiliary decoder to generate payload information and verifies whether the auxiliary decoder and the payload information meet a preset reconstruction quality threshold. If the preset reconstruction quality threshold is met, the network device obtains the auxiliary decoder information corresponding to the auxiliary decoder, and generates decoder-side auxiliary training data based on the auxiliary decoder information and the payload information.

[0033] The decoder-side assisted encoder training method, wherein after constructing the auxiliary decoder based on the preset training dataset, the method further includes: If the auxiliary decoder is not a decoder deployed by the network device, the consistency between the auxiliary decoder and the decoder deployed by the network device is detected; When the consistency meets the preset consistency requirements, the network device executes the step of generating payload information using the auxiliary decoder; When the consistency does not meet the preset consistency requirements, the network device re-executes the steps of obtaining the preset training dataset and constructing an auxiliary decoder based on the preset training dataset.

[0034] The decoder-side assisted encoder training method further includes one or more of the following: auxiliary training data identifier, applicability conditions, performance evaluation configuration, and lifecycle. The performance evaluation configuration is bound to at least one of the following: decoder behavior version, consistency threshold, local closed-loop verification stage, channel state information report configuration, or configuration index.

[0035] The decoder-side assisted encoder training method, wherein the performance evaluation configuration includes at least one of the following: metric type, evaluation object, evaluation stage, aggregation range, quantile type or metric threshold, wherein the evaluation object includes at least one of coding payload, dequantization payload, auxiliary decoder output channel state information or network-side actual decoder output channel state information, and the evaluation stage includes local training, local verification and verification before reporting ready feedback information.

[0036] The encoder training method with decoder-side assistance, wherein the auxiliary decoder information includes at least one of the following: auxiliary decoder, decoder actually deployed on the network device, reference decoder, decoder parameters, decoder parameter increment, decoder header, decoder adaptation module, decoder identifier or base decoder identifier, and at least one of the following: decoder behavior descriptor, distillation target, reconstruction output constraint, consistency verification information or decoder behavior version, and the auxiliary decoder information is associated with auxiliary training data identifier, payload configuration file identifier or consistency version.

[0037] The decoder-side assisted encoder training method further includes a training sample set used by the decoder of the network device as supplementary or backup decoder-side assisted data. When the training sample set, encoder parameters, reference encoder, decoder parameters, or reference decoder are used as supplementary / backup information, the terminal device needs to perform local adaptation training, assisted reconstruction verification, and ready feedback information reporting on the encoder of the terminal device using the local training dataset.

[0038] The decoder-side assisted encoder training method, wherein the list of payload profiles corresponding to the decoder-side assisted training data consists of the minimum payload profiles that can meet the coverage requirements and do not exceed the terminal device budget, wherein the selection of the minimum payload profiles is based on the coverage rank / layer / subband / transmit antenna port conditions, terminal device budget, performance threshold and readiness feedback information.

[0039] The encoder training method with decoder-side assistance, wherein the network device sends out auxiliary decoder information in a one-time manner or sends out decoder-side auxiliary training data in a progressive manner. When the network device sends out decoder-side auxiliary training data in a progressive manner, the network device first sends out the basic payload configuration file, auxiliary decoder information or basic auxiliary package, and then sends out the supplementary payload configuration file, decoder header, decoder adaptation module, incremental package or fallback relationship according to the ready feedback information of the terminal device.

[0040] A third aspect of this application provides a decoder-side assisted encoder training system, wherein the decoder-side assisted encoder training system includes a network device and a terminal device, wherein the network device provides decoder-side assisted training data to the terminal device, wherein the decoder-side assisted training data includes auxiliary decoder information and payload information, wherein the auxiliary decoder information is used to describe the reconstruction behavior of the decoder deployed on the network device for reproducing, approximating, constraining or verifying the terminal device, and the payload information is used to describe the target object output by the encoder of the terminal device; the terminal device trains the encoder of the terminal device based on the decoder-side assisted training data and a local training dataset to obtain a trained encoder, wherein the target object output by the trained encoder, after being reconstructed by the auxiliary decoder corresponding to the auxiliary decoder information, satisfies a preset reconstruction quality threshold.

[0041] A fourth aspect of this application provides a terminal device, including: a processor and a memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps in the decoder-side-assisted encoder training method described above.

[0042] A fifth aspect of this application provides a network device, including: a processor and a memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps in the decoder-side-assisted encoder training method described above.

[0043] The sixth aspect of this application provides a channel state information feedback method based on artificial intelligence / machine learning, wherein the method includes: The terminal device uses an encoder to encode the target channel state information to obtain the encoded payload, and sends the encoded payload to the network device. The encoder is trained by the encoder training method based on decoder side assistance as described above. The network device uses the decoder corresponding to the encoder to reconstruct the encoded payload to obtain the target channel state information.

[0044] The aforementioned channel state information feedback method for artificial intelligence / machine learning includes a coding payload associated with a channel state information report configuration, a decoder behavior version, an auxiliary training data identifier, a pairing context identifier, or a payload configuration file identifier. The network device activates the corresponding channel state information feedback configuration based on the ready feedback information reported by the terminal device. When the ready feedback information indicates verification failure, partial applicability, or a request for supplementary auxiliary training data on the decoder side, the network device performs a rollback configuration, partial activation, or supplementary auxiliary training data on the decoder side.

[0045] Beneficial effects: 1. After obtaining the decoder-side auxiliary training data, this application uses the decoder-side auxiliary training data and the local training dataset to train the encoder of the terminal device, making the encoder more adaptable to the local channel characteristics of the terminal device, better leveraging the active adaptation capability of the terminal device, and improving the reconstruction quality of channel state information while ensuring the interoperability between the encoder and the decoder, thereby improving the transmission performance of the entire communication system.

[0046] 2. This application utilizes decoder-side auxiliary training data sent by the network device to train the encoder in the terminal device. The network device does not need to transmit a complete target channel state information sample set to the terminal device, which reduces the signaling overhead of the network device transmitting training data to the terminal device, reduces the amount of data transmission required for the terminal device to obtain the adapted encoder, and also avoids the risk of leakage of manufacturer parameters caused by the transmission of complete parameters, thus balancing data transmission efficiency and information security.

[0047] 3. After the terminal device completes encoder training, this application will send the verification results to the network device. The network device will then perform operations such as activation, update, or rollback based on the verification results. This can ensure the reliability of training while flexibly adapting to the deployment requirements of the network device and improving the adaptability of the entire training process.

[0048] 4. This application uses the reconstructed channel state information corresponding to the feedback payload output by the encoder to verify the compatibility between the encoder and the decoder. This achieves the interoperability of the encoder and decoder without explicitly exposing the decoder implementation method in the network device, thereby reducing the cost of interfacing technical solutions between different manufacturers.

[0049] 5. This application prompts the network device to activate the payload configuration file, update the decoder-side auxiliary training data, trigger the rollback process, or update the auxiliary decoder by sending the verification result to the network device. When the local training dataset of the terminal device exceeds the capability of the decoder-side auxiliary training data, the terminal device is not forced to continue training the encoder. Instead, the network device is prompted to update the decoder-side auxiliary training data, trigger the rollback process, or update the auxiliary decoder. This avoids the waste of computing power caused by invalid training and reduces the training resource consumption of the terminal device while ensuring the training effect. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 A flowchart illustrating an encoder training method based on decoder-side assistance provided in this application embodiment.

[0052] Figure 2 This is a flowchart of the handshake process between network devices and terminal devices.

[0053] Figure 3 This is a flowchart illustrating the process of acquiring auxiliary training data for the decoder side.

[0054] Figure 4 A flowchart illustrating the process of confirming auxiliary training data on the decoder side for terminal devices.

[0055] Figure 5 A flowchart illustrating the process of training the encoder locally on the terminal device.

[0056] Figure 6 A flowchart illustrating the process of verifying and reporting the trained encoder to the terminal device.

[0057] Figure 7 A flowchart of another encoder training method based on decoder-side assistance provided in this application embodiment.

[0058] Figure 8 This is a block diagram illustrating the principle of an encoder training system based on decoder-side assistance, as provided in an embodiment of this application. Detailed Implementation

[0059] This application provides an encoder training method, feedback method, terminal device, network device, and system based on decoder-side assistance. To make the objectives, technical solutions, and effects of this application clearer and more explicit, the following detailed description, with reference to the accompanying drawings and embodiments, further illustrates this application. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0060] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0061] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0062] It should be understood that the sequence number and size of each step in this embodiment do not imply the order of execution. The execution order of each process is determined by its function and internal logic, and should not constitute any limitation on the implementation process of this application embodiment.

[0063] The application content will be further explained below with reference to the accompanying drawings and the description of the embodiments.

[0064] Example 1 This embodiment provides an encoder training method based on decoder-side assistance, such as... Figure 1 and Figure 2 As shown, the encoder training method based on decoder-side assistance specifically includes: S10, The terminal device acquires auxiliary training data from the decoder side.

[0065] Specifically, the decoder-side auxiliary training data is sent from the network device to the terminal device. The terminal device contains an encoder, and the network device contains a decoder. The encoder in the terminal device compresses the initial object to form the target object, and the decoder decodes the target object to obtain the reconstructed object corresponding to the initial object. The network device can be a base station, a network device training entity, a network device model management entity, or other entities capable of generating and distributing training auxiliary information. The terminal device can be an entity with local encoder training or fine-tuning capabilities. For example, in channel state information compression feedback based on artificial intelligence / machine learning, the terminal device is a user terminal with an encoder deployed, and the network device is a base station with a decoder deployed.

[0066] The decoder-side auxiliary training data is obtained by distilling a preset training dataset used to construct the auxiliary decoder in the network device. This data may include auxiliary decoder information and payload information. The auxiliary decoder information describes the reconstruction behavior of the decoder deployed on the network device, used to reproduce, approximate, constrain, or verify the network device. The payload information describes the target object output by the encoder of the terminal device. Specifically, the auxiliary decoder information may include at least one of the following: auxiliary decoder, the decoder actually deployed on the network device, reference decoder, decoder parameters, decoder parameter increments, decoder header, decoder adaptation module, decoder identifier, or base decoder identifier; and at least one of the following: decoder behavior descriptor, distillation target, reconstruction output constraint, consistency verification information, or decoder behavior version. Furthermore, the auxiliary decoder information is associated with an auxiliary training data identifier, a payload configuration file identifier, or a consistency version. The auxiliary training data identifier is a label for the auxiliary decoder information and may include one or more of the following: auxiliary decoder identifier, pairing context identifier, decoder version number, decoder validity period, fallback relationship, and security verification information. The decoder behavior version constrains the decoder reconstruction behavior, and the consistency version verifies the consistency between the local auxiliary decoder on the terminal device and the decoder actually deployed on the network device.

[0067] Specifically, auxiliary decoder information includes auxiliary decoders, decoders actually deployed on network devices, reference decoders, decoder parameters, decoder parameter increments, decoder headers, decoder adaptation modules, decoder identifiers, or base decoder identifiers. The auxiliary decoder can be a nominal decoder, the actual decoder deployed on the network device, or a reference decoder, etc. Payload information can be payload domain information or intermediate feedback payloads. The payload domain information or intermediate feedback payloads are obtained by the encoder encoding the information source. The information source can be channel state information, channel matrix, precoding matrix, feature matrix, images, audio, video, text, sensor data, or multimodal data, etc. For example, the information source corresponding to the target object is channel state information. The payload information includes payload domain information, which includes one or more of the following: the available range, length, quantization method, codebook, distribution, priority, operating point, and performance requirements of the feedback payload output by the encoder in the terminal device.

[0068] It should be noted that, in this embodiment, the terminal device obtains decoder-side auxiliary training data generated by the network device, and then uses the decoder-side auxiliary training data to train the encoder in the terminal device. This reduces the signaling overhead of the network device transmitting training data to the terminal device, reduces the amount of data transmission required for the terminal device to obtain the adapted encoder, and avoids the risk of vendor parameter leakage caused by the transmission of complete parameters, thus balancing data transmission efficiency and information security. In practical applications, the training sample set used by the decoder of the network device can also be used. The training sample set is used as supplementary or backup decoder-side auxiliary data. When using the training sample set, encoder parameters, reference encoder, decoder parameters, or reference decoder as supplementary / backup information, the terminal device needs to perform local adaptation training, assisted reconstruction verification, and readiness feedback reporting using the local training dataset to train the encoder of the terminal device.

[0069] For example, the network device sends a complete training sample set to the terminal device, which then performs subsequent encoder training based on the local training dataset and performs local verification and applicability reporting according to the payload profile, rank, layer, or subband. Alternatively, the network device sends encoder parameters or reference model parameters to the terminal device. The terminal does not directly use these encoder parameters or reference model parameters, but instead performs subsequent encoder training based on the local training dataset and performs local verification and applicability reporting according to the payload profile, rank, layer, or subband. In other words, the embodiments of this application can form a closed loop through decoder-side auxiliary training data—local training—performance verification—applicability reporting, i.e., a closed-loop process.

[0070] In one embodiment, the decoder-side auxiliary training data may include, in addition to auxiliary decoder information and payload information, auxiliary training data identifier, applicability conditions, performance evaluation configuration, and one or more of the following: lifecycle. Specifically, since the number of receiving antennas, receiving links, receiving algorithms, and maximum supported rank of different terminal devices will affect the locally calculated target channel state information, the applicability conditions may include one or more of the following: port, subband, rank, layer, bandwidth portion, receiving capability category, maximum supported rank, whether local fine-tuning is required, whether multi-payload profile training is supported, and whether local encoder updates are supported. The lifecycle may include one or more of the following: training auxiliary package update rules, incremental rules, failure rules, fallback rules, and rules for requesting supplementary decoder-side auxiliary training data. The performance evaluation configuration includes at least one of the following: metric type, evaluation object, evaluation stage, aggregation range, quantile type, or metric threshold, and the performance evaluation configuration is bound to at least one of the following: decoder behavior version, consistency threshold, local closed-loop verification stage, channel state information reporting configuration, or configuration index.

[0071] The metric types may include one or more of the following: the average / percentile of SGCS (Spatial Gram Cosine Similarity) / NMSE (Normalized Mean Square Error) / PSNR (Peak Signal-to-Noise Ratio), the number of local validation samples, and configuration-level performance evaluation configurations. The evaluation objects include at least one of the following: coded payload, inverse quantization payload, auxiliary decoder output channel state information, or network-side actual decoder output channel state information. The evaluation phase includes local training, local validation, and validation before reporting ready feedback information.

[0072] Specifically, in a typical implementation, the performance evaluation configuration can be represented as: M_perf={metricType,evaluationObject,stage,scope,aggregation,percentileType,threshold}, Where M_perf represents the performance evaluation configuration, metricType represents the metric type, evaluationObject represents the evaluation object, stage represents the evaluation stage, scope represents the evaluation granularity, aggregation represents the aggregation method, percentileType represents the percentile type, and threshold represents the threshold.

[0073] Furthermore, metricType can include SGCS, NMSE, PSNR, SSIM or semantic similarity, evaluationObject can include payload vector, de-quantized vector or decoder output, stage can include before quantization, after dequantization or after decoder output, and scope can include by layer, by rank, by payload profile or by subband.

[0074] In one specific embodiment, the decoder-side auxiliary training data includes an auxiliary training data identifier, auxiliary decoder information, payload information, at least one applicability condition, and at least one performance evaluation configuration, wherein the performance evaluation configuration is used as a preset reconstruction quality threshold. For example, the decoder-side auxiliary training data includes an auxiliary training data identifier, an auxiliary decoder, payload domain information, receiver capability category, and SGCS. Of course, in practical applications, the data items included in the decoder-side auxiliary training data can be adjusted according to the device capabilities of the terminal device and the transmission path; this is only an example and not a specific limitation.

[0075] The decoder-side auxiliary training data is generated by the network device and sent to the terminal device. In other words, the terminal device obtains this decoder-side auxiliary training data by communicating with the network device. Figure 3 As shown, the process of generating auxiliary training data on the decoder side specifically includes: H10. The network device obtains a preset training dataset and constructs an auxiliary decoder based on the preset training dataset; H2O: The network device uses the auxiliary decoder to generate payload information and verifies whether the auxiliary decoder and the payload information meet the preset reconstruction quality threshold. H30. If the preset reconstruction quality threshold is met, the network device obtains the auxiliary decoder information corresponding to the auxiliary decoder, and generates decoder-side auxiliary training data based on the auxiliary decoder information and the payload information.

[0076] In step H10, the preset training dataset includes a preset number of target channel states, and the sources of these target channel states may include: Network devices autonomously collect data, such as target channel state information obtained by network devices based on historical feedback, simulation data, field measurement data, or network device training platforms; Terminal devices report at the L1 layer, for example, information provided to network devices by terminal devices through physical layer channel state information reports, periodic / semi-persistent / aperiodic reports, or other low-layer feedback processes; Terminal devices report at L3 or higher layers, for example, target channel state information or a sample set thereof reported by the terminal device through higher layer measurement reports, log measurements, network device data collection or performance monitoring processes; Reports carried by the uplink shared channel, such as target channel state information, paired target CSI and feedback, data acquisition samples or diagnostic samples sent by the terminal device to the base station via PUSCH or other uplink bearers; Small-scale reporting carried by the uplink control channel, for example, terminal devices can report indexes, statistics, quality status or small sample summaries via PUCCH or other lightweight control channels; TDD reciprocity or network device uplink measurement, for example, in available scenarios, network devices can derive channel or precoding related information that can be used for training based on SRS, uplink channel estimation, or reciprocity; Data obtained from the operator's network or model management system, such as data from the core network, operator model management platform, network equipment training entity, OAM system, offline data warehouse, or cross-cell historical data; Synthetic or simulated data, such as target channel state information generated based on standard channel models, ray tracing, digital twins, or laboratory measurements.

[0077] It should be noted that in practical applications, the preset training dataset can be obtained from one of the above sources, or from two or more of the above sources. For example, the preset training data includes simulation data and data reported by the L3 layer of the terminal device. The network device first uses the simulation data to train the initial auxiliary decoder, and then uses the data reported by the L3 layer of the terminal device or higher layers to calibrate the initial auxiliary decoder. Alternatively, the preset training data includes simulation data, data reported by the L3 layer of the terminal device, and data on TDD reciprocity or uplink measurements of the network device. The network device first uses the simulation data to train the initial auxiliary decoder, and then uses the data reported by the L3 layer of the terminal device and the data on TDD reciprocity or uplink measurements of the network device to fine-tune the initial auxiliary decoder.

[0078] Furthermore, after obtaining the preset training dataset, an auxiliary decoder is constructed using the preset training dataset. The construction method of the auxiliary decoder may include: 1) Joint training of temporary encoder and auxiliary decoder: The network device trains a temporary encoder and auxiliary decoder based on the target channel state data samples in the preset training dataset, but only uses the auxiliary decoder to construct the decoder-side auxiliary training data; 2) Direct optimization of payload codewords and auxiliary decoder: The network device does not explicitly retain the temporary encoder, but optimizes an intermediate payload for each target channel state data sample and trains the auxiliary decoder at the same time, so that the intermediate payload is close to the target channel state data sample after reconstruction by the auxiliary decoder. 3) The method of deriving from the actual decoder deployed on the network device: If the network device already has an actual deployed decoder, the actual deployed decoder, the compressed version of the decoder, the approximate version of the decoder, or a part of the decoder module can be used as an auxiliary decoder. 4) Obtained by adding increments from the reference decoder: When both the network device and the terminal device know the same reference decoder, the network device uses the reference decoder with parameter increments as the auxiliary decoder; 5) Header or adapter module approach: Use the decoder header / adapter module, which includes a shared decoder backbone and corresponding different payload levels, ranks, layers, subbands or scenarios, as an auxiliary decoder. 6) Codebook index-based approach: Select codebook identifiers or codebook indices as auxiliary decoders, such as different codebooks corresponding to different decoder branches or dequantization interpretation methods.

[0079] Furthermore, during the construction of the auxiliary decoder, the network device also trains the auxiliary decoder to meet the network device's reconstruction quality threshold. The training process for the auxiliary decoder can be as follows: Suppose that the target channel state information sample obtained by the network device is Its corresponding intermediate feedback load is The auxiliary decoder is The objective of a network device for training or determining an auxiliary decoder can be represented as: , in, Represents the load domain regularization term. Indicates the allowed load domains or load profiles. Lambda Indicates the weight of the regularization term. The reconstruction loss can be represented by at least one of the following: , , , in, = This represents a sample of reconstructed channel state information. Vectorization is represented. This represents the peak value or reference amplitude after normalization of the target channel state information sample. For complex matrices, the error can be calculated for the real and imaginary parts separately, or the NMSE and SGCS can be calculated directly for complex vectors.

[0080] Furthermore, the network device reconstruction quality threshold corresponding to the auxiliary decoder may include at least one of the following: Average SGCS not lower than the first threshold ; The specified percentile SGCS is not lower than the second threshold. For example, the 5th percentile or the 10th percentile; The average NMSE is not higher than the first error threshold. ; The specified percentile NMSE is not higher than the second error threshold. ; Average PSNR not lower than PSNR threshold ; Each load profile, rank, layer, or subband condition satisfies its corresponding preset threshold; The parameter size, inference complexity, and transmission size of the auxiliary decoder shall not exceed the preset upper limit.

[0081] The thresholds mentioned above can be configured by the network device or implemented internally within the network device. For example, the first threshold... The value can be set between 0.95 and 0.99. For example, 0.95 or 0.97 can be used for decoder-side auxiliary training data of general accuracy, 0.98 can be used for high-precision decoder-side auxiliary training data, and 0.99 can be used for decoder-side auxiliary training data in high-fidelity or rigorous verification scenarios. However, it is not advisable to fix the value at 0.99 in all deployments to avoid insufficient available samples or difficulty in generating training packages. This is the first error threshold for NMSE. The PSNR threshold can be set to -25dB, -30dB, or -35dB, depending on the system normalization method. It can be used as an optional indicator in scenarios where the target channel state information sample is normalized into an image-like matrix or a real-valued matrix.

[0082] It should be noted that in practical applications, in addition to verifying the auxiliary decoder, the intermediate feedback payload can also be verified. The verification process for the intermediate feedback payload can be as follows: For each target channel state information sample and intermediate feedback load The auxiliary decoder in the network device will process the intermediate feedback load. Reconstruction is performed to obtain a reconstructed channel state information sample. = If the target channel state information sample With reconstructed channel state information samples If the reconstruction index meets the preset threshold, then the intermediate feedback load It can be marked as a qualified load sample.

[0083] For each payload profile, qualification conditions can be set for that payload profile. These qualification conditions can include the network device requiring at least [a certain number of] samples from the target channel state information set corresponding to the payload profile. The proportion of target channel state information samples meets a preset threshold, or the average index and percentile index of the payload profile simultaneously meet their respective thresholds. The simultaneous satisfaction of the average index and percentile index of the payload profile by their respective thresholds can be expressed as follows: , in, Indicates the first A load configuration file This indicates the threshold corresponding to the average metric of the load profile. This indicates the threshold corresponding to the percentile index of the load profile.

[0084] Therefore, if a payload profile does not meet the eligibility criteria, it will not be included in the decoder-side auxiliary training data. Conversely, if a payload profile meets the eligibility criteria, it can be included in the decoder-side auxiliary training data. In other words, when a network device's decoder supports multiple payload profiles, the decoder-side auxiliary training data can include all payload profiles supported by the network device's decoder, or all payload profiles supported by the network device's decoder that meet the eligibility criteria, or a selection of payload profiles that meet the eligibility criteria from those supported by the network device's decoder based on preset selection criteria. These selection criteria can comprehensively consider factors such as reconstruction performance coverage, terminal capabilities, training budget, transmission overhead, and redundancy between configuration profiles.

[0085] For example, the preset selection criteria are: , in, Indicates the selected load profile. This indicates the payload profiles supported by the network device's decoder. Indicates refactoring performance coverage. This represents the refactoring performance coverage threshold. Indicates the number of training data. This represents the threshold for the amount of training data.

[0086] Furthermore, when selecting a payload profile according to preset selection criteria, if multiple payload profiles meet the preset selection criteria, a further selection can be made among these profiles based on the capabilities of the auxiliary decoder, terminal equipment, and transmission capacity to obtain a target payload profile. This target payload profile only needs to cover the main performance and complexity operation points required for the target scenario. For example, during the further selection, the network device can evaluate one or more of the following for each payload profile that meets the preset selection criteria: average SGCS, average NMSE, optional X-percentile index, complexity, and transmission size. Based on the evaluation results of the payload profiles that meet the preset selection criteria, the target payload profile that meets the coverage and budget constraints is selected. This reduces the scale of auxiliary training data on the decoder side and lowers the local training burden on the terminal side. Existing search algorithms, such as greedy search and least squares search, can be used when selecting the target payload profile that meets the coverage and budget constraints.

[0087] To ensure consistency between the auxiliary decoder and the decoder actually deployed on the network device, if the auxiliary decoder is not the decoder actually deployed on the network device during the construction process, such as if the auxiliary decoder is a nominal decoder or a reference decoder, then a consistency check needs to be performed between the auxiliary decoder and the decoder actually deployed on the network device to determine whether the auxiliary decoder and the decoder actually deployed on the network device meet the consistency requirements.

[0088] Based on this, after constructing the auxiliary decoder based on the preset training dataset, the method further includes: If the auxiliary decoder is not a decoder deployed by the network device, the consistency between the auxiliary decoder and the decoder deployed by the network device is detected; When the consistency meets the preset consistency requirements, the network device executes the step of generating payload information using the auxiliary decoder; When the consistency does not meet the preset consistency requirements, the network device re-executes the steps of obtaining the preset training dataset and constructing an auxiliary decoder based on the preset training dataset.

[0089] Specifically, consistency requirements can be: , , in, Indicates the nominal decoder or reference decoder. This refers to the decoder actually deployed on the network device. Indicates feedback load, Indicates similarity. Indicates distance, Indicates the similarity threshold. This represents the distance threshold.

[0090] Furthermore, in step H20, after constructing the auxiliary decoder, decoder-side auxiliary training data can be constructed based on the auxiliary decoder. The data information included in the decoder-side auxiliary training data is as described in the above embodiments, and will not be repeated here. For example, the auxiliary decoder information can be directly adopted from the auxiliary decoder. The decoder-side auxiliary training data includes auxiliary training data identifier, auxiliary decoder, payload information, applicability conditions, and performance evaluation configuration. Then, the preset training sample set can be distilled into decoder-side auxiliary training data. This distillation process can be represented as follows: , in, This represents the preset training sample set. Indicates an auxiliary decoder, Indicates load information, Indicates the performance evaluation configuration. Indicates the applicability conditions. This indicates the identifier for auxiliary training data.

[0091] In practical applications, the decoder-side auxiliary training data can adopt a hierarchical structure so that the terminal can gradually parse and determine whether to accept it. The hierarchical structure can be determined according to the data information included in the decoder-side auxiliary training data. For example, as shown in Table 1, the decoder-side auxiliary training data can include an auxiliary training data identification layer, a decoder behavior version layer, a consistency verification layer, an auxiliary decoder information layer, a payload information layer, an applicability condition layer, a performance evaluation configuration layer, and a lifecycle layer.

[0092] Table 1. Hierarchical structure of decoder-side auxiliary training data

[0093] Furthermore, it's important to note that, to avoid explicitly exposing the internal scalability methods of the decoder within the network device, the network device does not need to explicitly inform the terminal device about which payload scalability methods are used internally, such as truncation, masking, hierarchical latency, or accumulative latency. The terminal device only needs to acquire the decoder-side auxiliary training data and then use this data as a reconstruction prior, determining compatibility through local training and reconstruction performance verification. This allows the network device to maintain the flexibility of its internal model implementation, and the terminal device does not rely on specific internal model mechanisms, but rather uses actual reconstruction performance as the criterion for usability.

[0094] In one embodiment, to avoid sending large amounts of decoder-side auxiliary training data to terminal devices that lack local training capabilities, the network device can employ a phased handshake during the transmission of decoder-side auxiliary training data to the terminal device. Based on this, as... Figure 4 As shown, the terminal device acquires decoder-side auxiliary training data specifically including: S11. The terminal device receives the first description information of the decoder-side auxiliary training data generated by the network device and sent by the network device. S12. The terminal device filters the first description information based on its own capabilities to obtain the second description information, and feeds back the second description information to the network device. S13. The terminal device receives decoder-side auxiliary training data sent by the network device.

[0095] In step S11, the first description information is a description of the decoder-side auxiliary training data, which is not the complete decoder-side auxiliary training data. The first description information includes one or more of the following: a list of payload profiles, decoder-side auxiliary training data identifiers, applicability conditions, size of decoder-side auxiliary training data, required training capabilities, performance thresholds, validity period, candidate decoder-side auxiliary training data, decoder behavior version, required local training budget, required storage budget, supported training modes, and validation granularity. The list of payload profiles includes one or more payload profiles.

[0096] It should be noted that, in order to ensure that the terminal device can train the encoder, the network device can also perform capability and willingness verification on the terminal device before sending the first description information to the terminal device to confirm the terminal device's own capabilities. The own capabilities include at least one of the following: the ability to parse decoder-side auxiliary training data; the ability to obtain local target channel state information; the ability to train or fine-tune the encoder according to the payload profile; the number of supported payload profiles; the supported rank / layer / subband / transmit antenna port range; the receive capability category; the local training budget; the storage budget; whether it supports low-rank adaptation parameter updates; whether it supports reporting applicability according to the configuration index; and the ability to update only the encoder adaptation layer or only the encoder head.

[0097] For example, verifying the capabilities and intentions of terminal devices to confirm their own capabilities includes: The ability to receive and parse auxiliary decoder information; Local target channel state information acquisition capability; Local encoder training or fine-tuning capabilities; The ability to train and validate according to payload profiles; The ability to report applicability by rank / layer / subband / transmit antenna port or receive capability category; The ability to store training aids and local training samples; The ability to update only the encoder adapter layer or the encoder head.

[0098] Further, in step S12, after obtaining the first description information, the terminal device filters the first description information based on its own capabilities, that is, it selects the acceptable information from the first description information that can be accepted by the terminal device. For example, the terminal device determines the payload configuration file that can be accepted by the terminal device. Then, it generates second description information based on the acceptable information and feeds the second description information back to the network device. The second description information includes one or more of the following: an acceptable payload configuration file list, decoder-side auxiliary training data acceptable to the terminal device, supported training modes, available training budget, available storage budget, supported rank, layer, subband, transmit antenna port range, and ready feedback information reporting granularity. Some or all of the payload configuration files in the payload configuration file list are included in the acceptable payload configuration file list.

[0099] Further, in step S13, the network device determines the decoder-side auxiliary training data based on the second description information. If the terminal device can accept all payload configuration files in the first description information (i.e., the second description information includes all payload configuration files), the network device directly sends the decoder-side auxiliary training data to the terminal device. If the terminal device can accept some payload configuration files in the first description information (i.e., the second description information includes some payload configuration files), the decoder-side auxiliary training data sent by the network device includes the decoder header, adaptation module, and payload information of the second description information including some payload configuration files. Of course, when the decoder-side auxiliary training data includes one or more of auxiliary training data identifiers, applicability conditions, performance evaluation configurations, and lifecycles, the decoder-side auxiliary training data sent by the network device also includes the corresponding data information. In other words, the decoder-side auxiliary training data sent by the network device is obtained by the network device filtering the decoder-side auxiliary training data generated by the network device based on the second description information.

[0100] Furthermore, after the terminal device acquires the decoder-side auxiliary training data, it can send feedback confirmation information to the network device. This feedback confirmation information may include confirmation of completion of reception, confirmation of training start, reasons for training failure, or requests for supplementary decoder-side auxiliary training data. Moreover, if no feedback confirmation information is received from the terminal device after a preset time, the network device can release the decoder-side auxiliary training data or switch to other mechanisms.

[0101] S20. The terminal device trains the encoder of the terminal device based on the decoder-side auxiliary training data and the local training dataset to obtain the trained encoder.

[0102] Specifically, the local training dataset is a set of target channel state information obtained by the terminal device based on local measurements. This target channel state information can be a channel matrix, a precoding matrix, a feature matrix, an eigenvector, normalized target channel state information, historical measurement samples, or target channel state information obtained during monitoring. If the target channel state information is a precoding matrix, the precoding matrix itself typically does not explicitly include the receiving antenna dimension, but its calculation result is affected by the terminal device's receiving capability, receiving algorithm, and local channel environment. Therefore, the terminal device can further fine-tune the encoder based on its own receiving capability and the local target precoding matrix. In this embodiment, the local training dataset includes normalized target channel state information and / or target channel state information obtained from historical measurement samples or during monitoring.

[0103] Furthermore, the target object output by the trained encoder, after being reconstructed by the auxiliary decoder corresponding to the auxiliary decoder information, meets a preset reconstruction quality threshold. In other words, during the encoder training process, the auxiliary decoder corresponding to the auxiliary decoder information serves as prior reconstruction knowledge, constraining the encoder training process.

[0104] For example, such as Figure 5 As shown, the terminal device trains its encoder based on the decoder-side auxiliary training data and the local training dataset to obtain the trained encoder, specifically including: S21. The terminal device acquires the encoder; S22. The terminal device inputs the target channel state information in the local training dataset into the encoder and outputs the feedback payload through the encoder. S23. The terminal device uses the auxiliary decoder information to reconstruct the feedback payload to obtain the second reconstructed channel state information, and updates the preset network layer in the encoder based on the second reconstructed channel state information and the target channel state information to obtain the trained encoder.

[0105] In step S21, when the terminal device acquires the encoder, it can directly read the encoder actually deployed by the terminal device as the encoder to be trained, or it can first pre-train the encoder actually deployed by the terminal device, and then use the pre-trained encoder as the encoder to be trained. The pre-training can utilize the intermediate feedback payload included in the payload information and the first reconstructed channel state information generated based on the intermediate feedback payload and the auxiliary decoder information.

[0106] It is understandable that when training the encoder, one can directly use local training or a combined pre-training and local training approach. Local training refers to training the encoder using the target channel state information in the local training dataset. The combined pre-training and local training approach involves first pre-training the encoder using the payload information, including the intermediate feedback payload and the first reconstructed channel state information generated based on the intermediate feedback payload and auxiliary decoder information, and then training the encoder using the target channel state information in the local training dataset. This reduces the difficulty of cold starts, and the combined pre-training and local training approach can utilize the payload domain priors provided by the network device while adapting to the local channel distribution of the terminal device.

[0107] In one embodiment, the terminal device employs a local training method. Accordingly, the terminal device acquires the encoder by reading its own actually deployed encoder.

[0108] In another embodiment, the terminal device employs a combined pre-training and local training method. Accordingly, the terminal device acquiring the encoder specifically includes: The terminal device acquires payload data from the decoder-side auxiliary training data, wherein the payload data is one of intermediate feedback payload, payload distribution, or codeword statistics; The terminal device uses the auxiliary decoder information to reconstruct the first reconstructed channel state information corresponding to the payload data; The terminal device forms pre-training samples based on the payload data and the first reconstructed channel state information; The terminal device uses the pre-trained samples to pre-train the encoder corresponding to the terminal device, and uses the pre-trained encoder as the encoder of the terminal device.

[0109] Further, in step S22, after obtaining the encoder, the target channel state information in the local training dataset is input into the encoder. The encoder encodes the target channel state information to obtain the feedback payload, which can be expressed as: , in, Indicates feedback load, Indicates encoder, This indicates the target channel state information.

[0110] It should be noted that when the target channel state information includes multiple layers, the terminal device can train all or some of the layers in the target channel state information. In particular, when the actual effective feedback layers of the terminal device are fewer than the layers included in the target channel state information, only the effective layers can be encoded, trained, and validated. Different layers can share the encoder backbone but use different output headers or adapter modules. Accordingly, during the encoder training process of the terminal device, the method further includes: When the target channel state information used to train the encoder contains multiple layers, the terminal device selects the effective layer from the multiple layers; The terminal device trains the encoder only for the effective layer.

[0111] Furthermore, when the terminal device trains the encoder only for the effective layer, the feedback load can be expressed as: , in, Indicates the first The feedback load corresponding to the layer, Indicates the first The layer's output header, Indicates the first Layer load profile, Represents the set of effective layers. Indicates the encoder backbone, This indicates the target channel state information.

[0112] Furthermore, in step S23, after obtaining the feedback payload, the feedback payload is reconstructed using the auxiliary decoder corresponding to the auxiliary decoder information to obtain the second reconstructed channel state information, wherein the second reconstructed channel state information can be expressed as: , in, This indicates the second reconstructed channel state information. Indicates an auxiliary decoder, This indicates the feedback load.

[0113] Then, the second reconstructed channel state information and the target channel state information are used to construct a training objective function, and the encoder is trained using the training objective function (this training can be training or fine-tuning) to obtain the trained encoder. The training objective function includes three function terms: a reconstruction loss term, a payload domain constraint term, and a parameter offset regularization term. The reconstruction loss term is used to constrain the reconstruction loss, the payload domain constraint term is used to constrain the payload domain information, and the parameter offset regularization term is used to limit the encoder parameters from deviating too much from the original parameters, so as to reduce the instability caused by local fine-tuning.

[0114] In one embodiment, the training objective function can be expressed as: , in, Indicates encoder parameters, This indicates the number of samples in the local training dataset. Represents the training objective function. Represents the reconstruction loss term. Represents load domain constraint terms. This indicates the parameter offset regularization term. and Both represent loss weights. Indicates the initial encoder parameters. This indicates the load domain information.

[0115] Furthermore, during encoder training, the network parameters of all network layers in the encoder can be optimized, or the network parameters of a preset network layer in the encoder can be optimized. Therefore, the preset network layer can be all network layers of the encoder, or it can be one or more of the following: the encoder head, the encoder adaptation module, the final preset number layer, the branch module corresponding to the payload configuration file, and the rank or layer related module. For example, the encoder in the terminal device includes a basic encoder and one or more adaptation layers. The decoder-side auxiliary training data provided by the network device does not adjust the basic encoder, but only trains the adaptation layers to adapt to the local target channel state information through the encoder output of the adaptation layer. Therefore, the encoder can be represented as: , in, Indicates encoder, Indicates encoder parameters, Indicates the basic encoder. This represents the basic encoder parameters, which are frozen (i.e. remain unchanged) during training. Indicates the adapter layer. This represents the parameters of the adaptation layer, which are updated during training. This indicates channel state information.

[0116] Furthermore, when the decoder-side auxiliary training data supports multiple payload profiles, the terminal device trains an encoder head for each payload profile or trains a shared encoder head for all payload profiles. The multiple payload profiles are all or part of the payload profiles corresponding to the network device. For example, if an encoder head is trained for each payload profile, the feedback payload corresponding to each payload profile can be represented as: , in, Indicates the first The encoder header corresponding to each load profile. Indicates the encoder backbone, Indicates the number of payload configuration files.

[0117] In one embodiment, during the training of the encoder using a local training dataset, it can be monitored whether the local training dataset of the terminal device exceeds the capability range of the decoder-side auxiliary training data. If it does, the terminal device will not continue local training of the encoder, but will instead trigger the network device to update the decoder-side auxiliary training data. The criteria for determining the capability range of the decoder-side auxiliary training data include at least one of the following: The terminal device fails to reach the SGCS threshold or NMSE threshold within T consecutive verification windows; The difference between the local target channel state information distribution of the terminal device and the applicable scope of the auxiliary training data on the decoder side exceeds the distribution drift threshold. Terminal devices can only enable fewer than The proportional load profile meets the requirements; The number of local training sessions or energy consumption of the terminal device exceeds the limit; The terminal device detects that the rank, subband, or reception capability conditions of the target channel state information exceed the applicable range of the auxiliary training data on the decoder side. The terminal device indicated that the network device needs to update the auxiliary decoder.

[0118] After triggering the network device to update the decoder-side auxiliary training data, the terminal device can send sample data to the network device. The sample data may include complete target channel state information samples, subsets of target channel state information, compact summaries of target channel state information, low-dimensional statistics, outlier indexes, or representative samples. The network device uses the sample data to update the auxiliary decoder, decoder header, adapter module, payload domain information, or performance evaluation configuration, and then sends new decoder-side auxiliary training data to the terminal device. Furthermore, to avoid confusion between the embodiments of this application and the transmission path patent, the specific transmission path of the sample data is not limited to the embodiments of this application; the sample data can be transmitted through L1, L3, higher-layer data collection, PUSCH, operator networks, OAM / core networks, or other available mechanisms.

[0119] In one embodiment, the terminal device trains its encoder based on the decoder-side auxiliary training data and the local training dataset to obtain the trained encoder, such as... Figure 6 As shown, the method further includes: S31. The terminal device verifies the trained encoder to obtain the verification result; S32. The terminal device generates ready feedback information based on the verification result, enabling the network device to perform subsequent operations based on the ready feedback information.

[0120] In step S31, the trained encoder can be obtained by training the encoder deployed on the terminal device, or by fine-tuning the encoder deployed on the terminal device. After obtaining the trained encoder, in order to verify the model performance of the trained encoder, the terminal device will use local validation samples to perform performance validation on the trained encoder. The validation conditions can include one or more of the following: average SGCS, percentile SGCS, average NMSE, percentile NMSE, PSNR, performance per layer, performance per payload profile, and performance under each rank / layer / subband / transmit antenna port / receive capability category.

[0121] Furthermore, for each payload profile, SGCS threshold readiness, NMSE threshold readiness, and comprehensive applicability are pre-set. These metrics are used to verify the readiness of the trained encoder. The SGCS threshold readiness, NMSE threshold readiness, and comprehensive applicability can be expressed as follows: , , , in, This indicates the SGCS threshold readiness index. This indicates the NMSE threshold readiness indicator. The comprehensive applicability index is considered satisfied when all of its constituent indicators are met. Indicates the first Average SGCS of each load profile Indicates the first The average NMSE of each load profile This represents the low-tailed percentile of the SGCS. Represents the high tail percentile of NMSE, Indicates the SGCS threshold. Indicates the NMSE threshold. This indicates the effectiveness of the auxiliary training data on the decoder side.

[0122] It's understandable that since SGCS is a metric where a larger value is better, and NMSE is a metric where a smaller value is better, the lower tail quantile of SGCS and the upper tail quantile of NMSE are selected. The thresholds and limits for each metric can be configured using auxiliary training data on the decoder side, provided by other network device configurations, or configured by the terminal device itself using default thresholds and limits. No specific restrictions are imposed here; they can be determined based on the deployment scenario and the target reconstruction accuracy. Furthermore, the evaluation object for NMSE can be the floating-point payload before quantization, the floating-point payload after dequantization, the decoder input representation, or the decoder output representation.

[0123] Furthermore, it should be noted that in multi-layer or multi-rank scenarios, the terminal device can train only the actual effective layers. Therefore, when validating the encoder, it can also validate only the trained, effective layers. The actual effective layers can be determined based on network configuration, actual rank, or the capabilities of the terminal device. In other words, the terminal device can validate only the trained, effective layers. The layers in the algorithm are used to train or validate the encoder, and report the readiness of the corresponding layer, configuration file, or rank condition. Represents the actual effective set of layers. Presentation layer set.

[0124] Furthermore, in step S32, after obtaining the verification result, the terminal device reports the verification result to the network device. When reporting the verification result to the network device, the terminal device can generate ready feedback information based on the verification result and report this ready feedback information to the network device. The ready feedback information includes at least one of the following: Auxiliary training data labeling; Pairing context identifier; Training ready or not; Applicable, not applicable, or partially applicable; List of applicable load profiles; Applicable rank, layer, sub-band, transmit antenna port, or receive capability category; Has it undergone local fine-tuning? Does the local verification result meet the threshold? Do you request supplementary decoder-side auxiliary training data? Should we request the network device to update the auxiliary decoder? Should we revert to another feedback method?

[0125] For example, the readiness feedback information includes applicable, or the readiness feedback information includes partially applicable and applicable payload profile lists, or the readiness feedback information includes training completed but performance insufficient.

[0126] Furthermore, after receiving the readiness feedback information reported by the terminal device, the network device performs subsequent operations based on the readiness feedback information. These subsequent operations may be one or more of the following: activating the payload configuration file, updating the decoder-side auxiliary training data, triggering a rollback process, or updating the auxiliary decoder. The correspondence between the readiness feedback information and subsequent operations is as follows: If the ready feedback information indicates that training is ready and applicable, the subsequent operation is to activate the payload configuration file corresponding to the auxiliary training data on the decoder side. If the ready feedback information is partially applicable, the subsequent operation is to activate the applicable payload configuration file corresponding to the auxiliary training data on the decoder side. If the ready feedback information is not applicable, the subsequent operation is to not activate the payload configuration file corresponding to the auxiliary training data on the decoder side. If the ready feedback message is a request for additional decoder-side auxiliary training data, the subsequent operation is to update the decoder-side auxiliary training data. If the ready feedback message is verification failure, the subsequent operation will trigger a rollback process; If the ready feedback message is a request for the network device to update the auxiliary decoder, then the subsequent operation is to update the auxiliary decoder.

[0127] In one embodiment, when training the encoder in a terminal device using decoder-side auxiliary training data, the network device can either send auxiliary decoder information in a one-time manner or send decoder-side auxiliary training data in a progressive manner. When the network device sends decoder-side auxiliary training data progressively, it first sends the basic payload configuration file, auxiliary decoder information, or basic auxiliary packets, and then sends supplementary payload configuration files, decoder headers, decoder adaptation modules, incremental packets, or fallback relationships based on the terminal device's readiness feedback information. Specifically, the specific process of this progressive training method can be as follows: The terminal device acquires basic decoder-side auxiliary training data determined based on the basic payload configuration file set and decoder-side auxiliary training data, and trains the encoder using the basic decoder-side auxiliary training data. The basic decoder-side auxiliary training data includes decoder-side auxiliary information and payload information corresponding to the basic payload configuration file set. The terminal device verifies the trained encoder to determine the applicable and inapplicable payload profiles, and requests the payload information corresponding to the inapplicable payload profile from the network device. The terminal device uses supplementary payload information and decoder-side auxiliary information to train the encoder, which has been trained with basic decoder-side auxiliary training data.

[0128] This application utilizes a progressive training method to train the encoder, which allows for phased and gradual completion of encoder adaptation training, reducing the computational and storage requirements of each training session on the terminal device. It also addresses the issue of poor adaptation of some payload configuration files, thereby improving training efficiency and the final encoder adaptation performance.

[0129] In one embodiment, besides the encoder using decoder-side auxiliary training data for training and validation, the encoder can be internally trained within the terminal device, and the decoder can be internally trained within the network device. Here, the training process of the encoder by the terminal device without using decoder-side auxiliary training data and the training process of the decoder by the network device are denoted as internal training, while the training process of the encoder by the terminal device using decoder-side auxiliary training data is denoted as external training. The performance evaluation configurations used for internal and external training differ. The performance evaluation configuration used for external training may include average SGCS, average NMSE, and optional X-percentile SGCS or X-percentile NMSE. The performance metrics used for internal training may include weighted metrics based on layer, subband, CQI / SNR, service importance, or task importance. These weighted metrics are only used for internal training or payload profile selection and do not replace the performance evaluation configuration of average SGCS / average NMSE used in external training.

[0130] Furthermore, the performance evaluation configuration can also associate the source of verification samples, such as verification sample sets in network devices, local verification sample sets on terminal devices, field monitoring samples, representative sample subsets, or hybrid verification sets. When a terminal device reports readiness, it can simultaneously report the number of verification samples, the verification time window, the metric summary, the confidence level, or a list of applicable configuration files, without having to upload the complete verification sample set. In practical applications, the confidence level can be calculated based on the number of verification samples, the verification time window, sample coverage, and metric stability, and this confidence level can be used in conjunction with an internal weighted score for readiness reporting or configuration file priority ranking.

[0131] For example, confidence level can be expressed as: Score_conf=Score_weighted·C_val, C_val=f(N_val,W_val,Coverage_val,Stability_val), Wherein, Score_conf represents the confidence level, C_val is the score that can be determined by the number of validation samples, the validation time window, the sample coverage, and the stability of the metric, =core_weighted represents the weighting coefficient, N_val represents the number of validation samples, W_val represents the validation time window, Coverage_val represents the coverage of the validation samples to the configuration file / layer / subband, and Stability_val represents the stability of the metric.

[0132] Furthermore, in practical applications, for target channel state information based on the precoding matrix, SGCS is used to evaluate the directional similarity between the reconstructed precoding vector and the target precoding vector. To avoid ambiguity regarding rank, layer, and subband dimensions, SGCS can be standardized. This standardization process can be described as follows: The s-th subband, the m-th payload profile, the k-th target channel state information set, the rank v, and the SGCS at layer l can be represented as: , in, Represents SGCS, Represents the channel precoding vector label. Represents the predicted channel precoding vector. This represents the conjugate transpose of a vector. This represents the number of target channel state information items in the k-th target channel state information set.

[0133] Furthermore, for the m-th payload profile, the k-th target channel state information set, the rank v, and the average SGCS of the l-th layer, it can be expressed as: .

[0134] If X-percentile SGCS is required, you can Sort by size from smallest to largest, and use the low-tailed quantile to determine the threshold, because SGCS is an indicator where a larger value is better.

[0135] The aforementioned SGCS can be used for qualification verification before network devices send auxiliary decoders, or for local verification (i.e., readiness verification) of terminal devices, or for suitability reporting at the configuration file level or layer level, etc.

[0136] Example 2 This application provides an encoder training method based on decoder-side assistance, such as... Figure 2 and Figure 7 As shown, the encoder training method based on decoder-side assistance specifically includes: The network device constructs decoder-side auxiliary training data, wherein the decoder-side auxiliary training data includes auxiliary decoder information and payload information. The auxiliary decoder information is used to describe the reconstruction behavior of the decoder deployed on the network device for reproducing, approximating, constraining or verifying on the terminal device, and the payload information is used to describe the target object output by the encoder of the terminal device. The network device sends the decoder-side auxiliary training data so that the terminal device can train the encoder of the terminal device based on the decoder-side auxiliary training data and the local training dataset to obtain the trained encoder.

[0137] The decoder-side assisted encoder training method, wherein the network device constructs decoder-side assisted training data specifically including: The network device acquires a preset training dataset and constructs an auxiliary decoder based on the preset training dataset. When the auxiliary decoder is not a decoder deployed by the network device, the auxiliary decoder and the decoder deployed by the network device meet the consistency requirements. The network device uses the auxiliary decoder to generate payload information and verifies whether the auxiliary decoder and the payload information meet a preset reconstruction quality threshold. If the preset reconstruction quality threshold is met, the network device obtains the auxiliary decoder information corresponding to the auxiliary decoder, and generates decoder-side auxiliary training data based on the auxiliary decoder information and the payload information.

[0138] The decoder-side assisted encoder training method, wherein after constructing the auxiliary decoder based on the preset training dataset, the method further includes: If the auxiliary decoder is not a decoder deployed by the network device, the consistency between the auxiliary decoder and the decoder deployed by the network device is detected; When the consistency meets the preset consistency requirements, the network device executes the step of generating payload information using the auxiliary decoder; When the consistency does not meet the preset consistency requirements, the network device re-executes the steps of obtaining the preset training dataset and constructing an auxiliary decoder based on the preset training dataset.

[0139] The decoder-side assisted encoder training method further includes one or more of the following: auxiliary training data identifier, applicability conditions, performance evaluation configuration, and lifecycle. The performance evaluation configuration is bound to at least one of the following: decoder behavior version, consistency threshold, local closed-loop verification stage, channel state information report configuration, or configuration index.

[0140] The decoder-side assisted encoder training method, wherein the performance evaluation configuration includes at least one of the following: metric type, evaluation object, evaluation stage, aggregation range, quantile type or metric threshold, wherein the evaluation object includes at least one of coding payload, dequantization payload, auxiliary decoder output channel state information or network-side actual decoder output channel state information, and the evaluation stage includes local training, local verification and verification before reporting ready feedback information.

[0141] The encoder training method with decoder-side assistance, wherein the auxiliary decoder information includes at least one of the following: auxiliary decoder, decoder actually deployed on the network device, reference decoder, decoder parameters, decoder parameter increment, decoder header, decoder adaptation module, decoder identifier or base decoder identifier, and at least one of the following: decoder behavior descriptor, distillation target, reconstruction output constraint, consistency verification information or decoder behavior version, and the auxiliary decoder information is associated with auxiliary training data identifier, payload configuration file identifier or consistency version.

[0142] The decoder-side assisted encoder training method further includes a training sample set used by the decoder of the network device as supplementary or backup decoder-side assisted data. When the training sample set, encoder parameters, reference encoder, decoder parameters, or reference decoder are used as supplementary / backup information, the terminal device needs to perform local adaptation training, assisted reconstruction verification, and ready feedback information reporting on the encoder of the terminal device using the local training dataset.

[0143] The decoder-side assisted encoder training method, wherein the list of payload profiles corresponding to the decoder-side assisted training data consists of the minimum payload profiles that can meet the coverage requirements and do not exceed the terminal device budget, wherein the selection of the minimum payload profiles is based on the coverage rank / layer / subband / transmit antenna port conditions, terminal device budget, performance threshold and readiness feedback information.

[0144] The encoder training method with decoder-side assistance, wherein the network device sends out auxiliary decoder information in a one-time manner or sends out decoder-side auxiliary training data in a progressive manner. When the network device sends out decoder-side auxiliary training data in a progressive manner, the network device first sends out the basic payload configuration file, auxiliary decoder information or basic auxiliary package, and then sends out the supplementary payload configuration file, decoder header, decoder adaptation module, incremental package or fallback relationship according to the ready feedback information of the terminal device.

[0145] It should be noted that the execution and interaction processes of the network device and terminal device in this embodiment are the same as those in Embodiment 1 above, and will not be repeated here.

[0146] Example 3 Based on the above-described encoder training method with decoder-side assistance, this embodiment provides an encoder training system with decoder-side assistance, such as... Figure 8 As shown, the encoder training system based on decoder-side assistance includes a network device 110 and a terminal device 120. The network device 110 provides decoder-side assistance training data to the terminal device. This data includes auxiliary decoder information and payload information. The auxiliary decoder information describes the reconstruction behavior of the decoder deployed on the network device for reproducing, approximating, constraining, or verifying the data on the terminal device. The payload information describes the target object output by the encoder of the terminal device. The terminal device 120 trains its encoder based on the decoder-side assistance training data and a local training dataset to obtain a trained encoder. The target object output by the trained encoder, after reconstruction by the auxiliary decoder corresponding to the auxiliary decoder information, meets a preset reconstruction quality threshold.

[0147] Example 4 Based on the above-described encoder training method with decoder-side assistance, this embodiment provides a channel state information feedback method for artificial intelligence / machine learning, wherein the method includes: The terminal device uses an encoder to encode the target channel state information to obtain the encoded payload, and sends the encoded payload to the network device. The encoder is trained by the encoder training method based on decoder side assistance as described above. The network device uses the decoder corresponding to the encoder to reconstruct the encoded payload to obtain the target channel state information.

[0148] The aforementioned channel state information feedback method for artificial intelligence / machine learning includes a coding payload associated with a channel state information report configuration, a decoder behavior version, an auxiliary training data identifier, a pairing context identifier, or a payload configuration file identifier. The network device activates the corresponding channel state information feedback configuration based on the ready feedback information reported by the terminal device. When the ready feedback information indicates verification failure, partial applicability, or a request for supplementary auxiliary training data on the decoder side, the network device performs a rollback configuration, partial activation, or supplementary auxiliary training data on the decoder side.

[0149] Example 5 Based on the above encoder training method with decoder-side assistance, this application also provides a terminal device, including: a processor and a memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps in the decoder-side-assisted encoder training method described above.

[0150] Example 6 Based on the above encoder training method based on decoder-side assistance, this application also provides a network device, including: a processor and a memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps in the decoder-side-assisted encoder training method described above.

[0151] Furthermore, when the logical instructions in the memory of the terminal device in Embodiment 5 and the network device in Embodiment 6 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0152] As a computer-readable storage medium, the memory can be configured to store software programs, computer-executable programs, such as program instructions or modules corresponding to the methods in the embodiments of this disclosure. The terminal device in Embodiment 5 and the processor in Embodiment 6 execute functional applications and data processing by running the software programs, instructions, or modules stored in the memory, thereby implementing the methods in the above embodiments.

[0153] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory may include high-speed random access memory (RAM) and non-volatile memory. Examples include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks; these can also be transient storage media.

[0154] Furthermore, the specific process of loading and executing multiple instruction processors in the aforementioned storage medium and terminal device has been described in detail in the above method, and will not be repeated here.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A decoder-side assisted encoder training method, characterized in that, The encoder training method based on decoder-side assistance specifically includes: The terminal device acquires decoder-side auxiliary training data, wherein the decoder-side auxiliary training data includes auxiliary decoder information and payload information. The auxiliary decoder information is used to describe the reconstruction behavior of the decoder deployed on the network device for reproducing, approximating, constraining or verifying the terminal device, and the payload information is used to describe the target object output by the encoder of the terminal device. The terminal device trains the encoder of the terminal device based on the decoder-side auxiliary training data and the local training dataset to obtain the trained encoder.

2. The encoder training method based on decoder-side assistance according to claim 1, characterized in that, After the terminal device trains its encoder based on the decoder-side auxiliary training data and the local training dataset to obtain the trained encoder, the method further includes: The terminal device verifies the trained encoder to obtain the verification result; The terminal device generates ready feedback information based on the verification result, enabling the network device to perform subsequent operations based on the ready feedback information. The subsequent operations include one or more of the following: activating the payload configuration file, updating the decoder-side auxiliary training data, triggering a rollback process, or updating the auxiliary decoder.

3. The encoder training method based on decoder-side assistance according to claim 2, characterized in that, The correspondence between the readiness feedback information and the subsequent operation includes at least one of the following: The readiness feedback message indicates that training is ready and applicable; the subsequent operation is to activate all payload configuration files. The readiness feedback information is partially applicable; the subsequent operation is to activate the partially applicable load profile. The readiness feedback message is "not applicable," and the subsequent operation is to not activate any load profile. The readiness feedback information is to supplement the decoder-side auxiliary training data, and the subsequent operation is to update the decoder-side auxiliary training data; The readiness feedback message indicates verification failure, and the subsequent operation is to trigger a rollback process; The readiness feedback message is a request for the network device to update the auxiliary decoder, and the subsequent operation is to update the auxiliary decoder.

4. The encoder training method based on decoder-side assistance according to claim 2 or 3, characterized in that, The readiness feedback information includes a bitmap, list, or set of indexes generated according to at least one of the following: payload profile, rank, layer, subband, transmit antenna port, receive capability category, or configuration index.

5. The encoder training method based on decoder-side assistance according to claim 1 or 2, characterized in that, The auxiliary decoder information includes at least one of the following: auxiliary decoder, decoder actually deployed on the network device, reference decoder, decoder parameters, decoder parameter increment, decoder header, decoder adaptation module, decoder identifier or base decoder identifier, and at least one of the following: decoder behavior descriptor, distillation target, reconstruction output constraint, consistency verification information or decoder behavior version. The auxiliary decoder information is associated with auxiliary training data identifier, payload configuration file identifier or consistency version.

6. The encoder training method based on decoder-side assistance according to claim 1 or 2, characterized in that, The decoder-side auxiliary training data also includes the training sample set used by the decoder of the network device, which is used as supplementary or backup decoder-side auxiliary data. When the training sample set, encoder parameters, reference encoder, decoder parameters or reference decoder are used as supplementary / backup information, the terminal device needs to perform local adaptation training, assisted reconstruction verification and readiness feedback information reporting on the encoder of the terminal device using the local training dataset.

7. The encoder training method based on decoder-side assistance according to claim 1, characterized in that, The decoder-side auxiliary training data is sent from the network device to the terminal device, and the generation process of the decoder-side auxiliary training data specifically includes: The network device acquires a preset training dataset and constructs an auxiliary decoder based on the preset training dataset. When the auxiliary decoder is not a decoder deployed by the network device, the auxiliary decoder and the decoder deployed by the network device meet the consistency requirements. The network device uses the auxiliary decoder to generate payload information and verifies whether the auxiliary decoder and the payload information meet a preset reconstruction quality threshold. If the preset reconstruction quality threshold is met, the network device obtains the auxiliary decoder information corresponding to the auxiliary decoder, and generates decoder-side auxiliary training data based on the auxiliary decoder information and the payload information.

8. The encoder training method based on decoder-side assistance according to claim 7, characterized in that, After constructing the auxiliary decoder based on the preset training dataset, the method further includes: If the auxiliary decoder is not a decoder deployed by the network device, the consistency between the auxiliary decoder and the decoder deployed by the network device is detected; When the consistency meets the preset consistency requirements, the network device executes the step of generating payload information using the auxiliary decoder; When the consistency does not meet the preset consistency requirements, the network device re-executes the steps of obtaining the preset training dataset and constructing an auxiliary decoder based on the preset training dataset.

9. The encoder training method based on decoder-side assistance according to claim 1, characterized in that, The decoder-side auxiliary training data also includes one or more of the following: auxiliary training data identifier, applicability conditions, performance evaluation configuration, and lifecycle. The performance evaluation configuration is bound to at least one of the following: decoder behavior version, consistency threshold, local closed-loop verification stage, channel state information report configuration, or configuration index.

10. The encoder training method based on decoder-side assistance according to claim 9, characterized in that, The performance evaluation configuration includes at least one of the following: indicator type, evaluation object, evaluation stage, aggregation range, quantile type or indicator threshold. The evaluation object includes at least one of the following: coding payload, inverse quantization payload, auxiliary decoder output channel state information or network-side actual decoder output channel state information. The evaluation stage includes local training, local verification and verification before reporting ready feedback information.

11. The encoder training method based on decoder-side assistance according to claim 1, characterized in that, The acquisition of decoder-side auxiliary training data by the terminal device specifically includes: The terminal device receives the first descriptive information of the decoder-side auxiliary training data generated by the network device, sent by the network device. The terminal device filters the first description information based on its own capabilities to obtain the second description information, and then feeds back the second description information to the network device; The terminal device receives decoder-side auxiliary training data sent by the network device, wherein the decoder-side auxiliary training data sent by the network device is obtained by the network device filtering the decoder-side auxiliary training data generated by the network device based on the second description information.

12. The encoder training method based on decoder-side assistance according to claim 11, characterized in that, The first description information includes: The payload configuration file list includes one or more of the following: a list of payload configuration files, decoder-side auxiliary training data identifiers, applicability conditions, decoder-side auxiliary training data size, required training capabilities, performance thresholds, validity period, candidate decoder-side auxiliary training data, decoder behavior version, required local training budget, required storage budget, supported training modes, and validation granularity. The payload configuration file list includes one or more payload configuration files.

13. The encoder training method based on decoder-side assistance according to claim 12, characterized in that, The second description information includes one or more of the following: a list of accepted payload profiles, decoder-side auxiliary training data acceptable to the terminal device, supported training modes, available training budget, available storage budget, supported rank, layer, subband, transmit antenna port range, and ready feedback information reporting granularity. Among these, some or all of the payload profiles in the payload profile list are included in the list of accepted payload profiles.

14. The encoder training method based on decoder-side assistance according to claim 1 or 11, characterized in that, Before the terminal device acquires the decoder-side auxiliary training data, the method further includes: The terminal device sends its own capabilities to the network device based on the query request of the network device. The own capabilities include at least one of the following: the ability to parse decoder-side auxiliary training data, the ability to obtain local target channel state information, the ability to train or fine-tune the encoder according to the payload profile, the number of supported payload profiles, the supported rank / layer / subband / transmit antenna port range, the receive capability category, the local training budget, the storage budget, whether it supports low-rank adaptation parameter updates, whether it supports reporting applicability according to the configuration index, and the ability to update only the encoder adaptation layer or only the encoder head.

15. The encoder training method based on decoder-side assistance according to claim 1, characterized in that, The target object output by the trained encoder satisfies the preset reconstruction quality threshold after being reconstructed by the auxiliary decoder corresponding to the auxiliary decoder information.

16. The encoder training method based on decoder-side assistance according to claim 1, characterized in that, The terminal device trains its encoder based on the decoder-side auxiliary training data and the local training dataset to obtain the trained encoder, specifically including: The terminal device acquires the encoder; The terminal device inputs the target channel state information in the local training dataset into the encoder, and outputs the feedback payload through the encoder; The terminal device uses the auxiliary decoder information to reconstruct the feedback payload to obtain second reconstructed channel state information, and updates the preset network layer in the encoder based on the second reconstructed channel state information and the target channel state information to obtain the trained encoder.

17. The encoder training method based on decoder-side assistance according to claim 16, characterized in that, The terminal device acquires the encoder specifically as follows: The terminal device reads the encoder that it has actually deployed.

18. The encoder training method based on decoder-side assistance according to claim 16, characterized in that, The acquisition of the encoder by the terminal device specifically includes: The terminal device acquires payload data from the decoder-side auxiliary training data, wherein the payload data is one of intermediate feedback payload, payload distribution, or codeword statistics; The terminal device uses the auxiliary decoder information to reconstruct the first reconstructed channel state information corresponding to the payload data; The terminal device forms pre-training samples based on the payload data and the first reconstructed channel state information; The terminal device uses the pre-trained samples to pre-train the encoder corresponding to the terminal device, and uses the pre-trained encoder as the encoder of the terminal device.

19. The encoder training method based on decoder-side assistance according to claim 16, characterized in that, When the decoder-side auxiliary training data supports multiple payload profiles, the terminal device trains an encoder head for each payload profile or trains a shared encoder head for all payload profiles. The multiple payload profiles are all or part of the payload profiles corresponding to the network device.

20. The encoder training method based on decoder-side assistance according to any one of claims 16-19, characterized in that, The preset network layer can be any of the encoder's network layers, or it can be one or more of the following: encoder head, encoder adapter module, last preset quantity layer, branch module corresponding to the load configuration file, and rank or layer related modules.

21. The encoder training method based on decoder-side assistance according to claim 16, characterized in that, During the encoder training process on the terminal device, the method further includes: When the target channel state information used to train the encoder contains multiple layers, the terminal device selects the effective layer from the multiple layers; The terminal device trains the encoder only for the effective layer.

22. The encoder training method based on decoder-side assistance according to claim 21, characterized in that, The method further includes: The terminal device verifies the trained encoder only for the effective layer and reports the readiness feedback information corresponding to the effective layer, payload profile, rank, or subband condition to the network device.

23. The encoder training method based on decoder-side assistance according to claim 1, characterized in that, The list of payload configuration files corresponding to the decoder-side auxiliary training data consists of the minimum payload configuration files that can meet the coverage requirements and do not exceed the terminal device's budget.

24. The encoder training method based on decoder-side assistance according to claim 1, characterized in that, The network device may send auxiliary decoder information in a one-time manner or send decoder-side auxiliary training data in a progressive manner. When the network device sends decoder-side auxiliary training data in a progressive manner, the network device first sends the basic payload configuration file, auxiliary decoder information or basic auxiliary package, and then sends the supplementary payload configuration file, decoder header, decoder adaptation module, incremental package or fallback relationship according to the ready feedback information of the terminal device.

25. The encoder training method based on decoder-side assistance according to claim 24, characterized in that, The triggering conditions for the progressive distribution of decoder-side auxiliary training data are based on the verification failure, partial applicability, request for supplementary decoder-side auxiliary training data, or inconsistency in decoder behavior version or consistency version in the ready feedback information of the terminal device, which triggers resending, rollback, or partial activation.

26. A decoder-side assisted encoder training method, characterized in that, The encoder training method based on decoder-side assistance specifically includes: The network device constructs decoder-side auxiliary training data, wherein the decoder-side auxiliary training data includes auxiliary decoder information and payload information. The auxiliary decoder information is used to describe the reconstruction behavior of the decoder deployed on the network device for reproducing, approximating, constraining or verifying on the terminal device, and the payload information is used to describe the target object output by the encoder of the terminal device. The network device sends the decoder-side auxiliary training data so that the terminal device can train the encoder of the terminal device based on the decoder-side auxiliary training data and the local training dataset to obtain the trained encoder.

27. The encoder training method based on decoder-side assistance according to claim 26, characterized in that, The network device constructs decoder-side auxiliary training data, specifically including: The network device acquires a preset training dataset and constructs an auxiliary decoder based on the preset training dataset. When the auxiliary decoder is not a decoder deployed by the network device, the auxiliary decoder and the decoder deployed by the network device meet the consistency requirements. The network device uses the auxiliary decoder to generate payload information and verifies whether the auxiliary decoder and the payload information meet a preset reconstruction quality threshold. If the preset reconstruction quality threshold is met, the network device obtains the auxiliary decoder information corresponding to the auxiliary decoder, and generates decoder-side auxiliary training data based on the auxiliary decoder information and the payload information.

28. The encoder training method based on decoder-side assistance according to claim 27, characterized in that, After constructing the auxiliary decoder based on the preset training dataset, the method further includes: If the auxiliary decoder is not a decoder deployed by the network device, the consistency between the auxiliary decoder and the decoder deployed by the network device is detected; When the consistency meets the preset consistency requirements, the network device executes the step of generating payload information using the auxiliary decoder; When the consistency does not meet the preset consistency requirements, the network device re-executes the steps of obtaining the preset training dataset and constructing an auxiliary decoder based on the preset training dataset.

29. The encoder training method based on decoder-side assistance according to claim 26 or 28, characterized in that, The decoder-side auxiliary training data also includes one or more of the following: auxiliary training data identifier, applicability conditions, performance evaluation configuration, and lifecycle. The performance evaluation configuration is bound to at least one of the following: decoder behavior version, consistency threshold, local closed-loop verification stage, channel state information report configuration, or configuration index.

30. The encoder training method based on decoder-side assistance according to claim 29, characterized in that, The performance evaluation configuration includes at least one of the following: indicator type, evaluation object, evaluation stage, aggregation range, quantile type or indicator threshold. The evaluation object includes at least one of the following: coding payload, inverse quantization payload, auxiliary decoder output channel state information or network-side actual decoder output channel state information. The evaluation stage includes local training, local verification and verification before reporting ready feedback information.

31. The encoder training method based on decoder-side assistance according to claim 26, characterized in that, The auxiliary decoder information includes at least one of the following: auxiliary decoder, decoder actually deployed on the network device, reference decoder, decoder parameters, decoder parameter increment, decoder header, decoder adaptation module, decoder identifier or base decoder identifier, and at least one of the following: decoder behavior descriptor, distillation target, reconstruction output constraint, consistency verification information or decoder behavior version. The auxiliary decoder information is associated with auxiliary training data identifier, payload configuration file identifier or consistency version.

32. The encoder training method based on decoder-side assistance according to claim 26, characterized in that, The decoder-side auxiliary training data also includes the training sample set used by the decoder of the network device, which is used as supplementary or backup decoder-side auxiliary data. When the training sample set, encoder parameters, reference encoder, decoder parameters or reference decoder are used as supplementary / backup information, the terminal device needs to perform local adaptation training, assisted reconstruction verification and readiness feedback information reporting on the encoder of the terminal device using the local training dataset.

33. The encoder training method based on decoder-side assistance according to claim 26, characterized in that, The list of payload profiles corresponding to the decoder-side auxiliary training data consists of the minimum payload profiles that can meet the coverage requirements and do not exceed the terminal device budget. The selection of the minimum payload profiles is based on the coverage rank / layer / subband / transmit antenna port conditions, terminal device budget, performance threshold, and readiness feedback information.

34. The encoder training method based on decoder-side assistance according to claim 26, characterized in that, The network device may send auxiliary decoder information in a one-time manner or send decoder-side auxiliary training data in a progressive manner. When the network device sends decoder-side auxiliary training data in a progressive manner, the network device first sends the basic payload configuration file, auxiliary decoder information or basic auxiliary package, and then sends the supplementary payload configuration file, decoder header, decoder adaptation module, incremental package or fallback relationship according to the ready feedback information of the terminal device.

35. An encoder training system based on decoder-side assistance, characterized in that, The encoder training system based on decoder-side assistance includes network devices and terminal devices, wherein, A network device is used to provide decoder-side auxiliary training data to a terminal device. The decoder-side auxiliary training data includes auxiliary decoder information and payload information. The auxiliary decoder information is used to describe the reconstruction behavior of the decoder deployed on the network device in reproducing, approximating, constraining, or verifying the terminal device. The payload information is used to describe the target object output by the encoder of the terminal device. The terminal device trains the encoder of the terminal device based on the auxiliary training data on the decoder side and the local training dataset to obtain the trained encoder. The target object output by the trained encoder satisfies the preset reconstruction quality threshold after being reconstructed by the auxiliary decoder corresponding to the auxiliary decoder information.

36. A terminal device, characterized in that, include: Processor and memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps of the encoder training method based on decoder-side assistance as described in any one of claims 1-25.

37. A network device, characterized in that, include: Processor and memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps of the encoder training method based on decoder-side assistance as described in any one of claims 26-34.

38. A channel state information feedback method for artificial intelligence / machine learning, characterized in that, The method includes: The terminal device uses an encoder to encode the target channel state information to obtain the encoded payload, and sends the encoded payload to the network device. The encoder is trained by the encoder training method based on decoder side assistance as described in any one of claims 1-25, or by the encoder training method based on decoder side assistance as described in any one of claims 26-24. The network device uses the decoder corresponding to the encoder to reconstruct the encoded payload to obtain the target channel state information.

39. The channel state information feedback method for artificial intelligence / machine learning according to claim 38, characterized in that, The encoded payload is associated with the channel state information report configuration, decoder behavior version, auxiliary training data identifier, pairing context identifier, or payload configuration file identifier. The network device activates the corresponding channel state information feedback configuration based on the ready feedback information reported by the terminal device. When the ready feedback information indicates verification failure, partial applicability, or a request for supplementary auxiliary training data on the decoder side, the network device performs a rollback configuration, partial activation, or supplementary auxiliary training data on the decoder side.

Images